Using nationwide 'big data' from linked electronic health records to help improve outcomes in cardiovascular diseases: 33 studies using methods from epidemiology, informatics, economics and social science in the ClinicAl disease research using LInked Bespoke studies and Electronic health Records (CALIBER) programme

The global political importance of data

People increasingly leave digital ‘traces’ relevant to health in multiple different information systems as they interact with health and other services and go about their everyday lives. The potential for such data to inform, and potentially transform, our understanding of health and disease has been recognised by world leaders. For example, George Osborne, the UK Chancellor of the Exchequer from 2010 to 2016, said in a Science Speech to the Royal Society in November 2013 that the UK has ‘some of the world’s best and most complete data-sets in healthcare’ (© Crown copyright; contains public sector information licensed under the Open Government Licence v3.0). 1 Following the 2015 State of the Union address, the White House signalled an initiative that would ‘catalyse a new era of data-based and more precise medical treatment’ (reproduced under a Creative Commons Attribution 3.0 License). 2

Data tapestry

The electronic record relevant to health is diverse. This ‘data tapestry’ of information (Figure 1) spans records in primary care and in hospitals, structured data (codes, values) and unstructured data (text). It includes information on genomics, through to social media, the sensed self and wider social determinants of disease. Deriving information and knowledge from one or more sources of these rapidly developing sources of health-relevant data might be of great benefit to inform improvements in care and outcomes of disease.

FIGURE 1.

The tapestry of potentially high-value information sources that may be linked to an individual for use in health care. CPT, current procedural terminology; ECG, electrocardiogram; EPA, US Environmental Protection Agency; GIS, geographic information system; GPS, global positioning system; ICD-9, International Classification of Diseases, Ninth Edition; LOINC, Logical Observation Identifiers Names and Codes; NDC, National Drug Code; OTC, over the counter; SNP, single nucleotide polymorphism. Reproduced with permission from Weber et al. Finding the missing link for big biomedical data. JAMA 2014;311:2479–80. 3 Copyright © 2014 American Medical Association. All rights reserved.

Translational gap

One model that links the process of science and improvements in health care and that has attracted considerable attention in the UK is the translational model, as laid out, for example, in the Cooksey Review. 4 This identified two gaps in the translation of biomedical science to health care, construed as an essentially linear process. The first translational gap is the generation of new therapies to improve health care based on biomedical science. Epidemiological research using electronic health records (EHRs) can help to close the first translational gap by enabling researchers to quantify the need for new therapies, identify risk factors and (with genomic data) suggest potential new drug targets. Randomised clinical trials with follow-up through EHRs can be cheaper and simpler than conventional investigator-led trials and may be particularly helpful for increasing the evidence base for non-drug interventions.

However, it is in the second translational gap, that is, the introduction of new therapies into clinical practice, on which EHR-based research programmes may have the greatest impact.

For example, new prognostic risk scores may help interventions to be targeted most appropriately, decision support tools embedded in clinical systems may help to improve clinical decision-making and health economic analyses can ensure that treatments are managed in a cost-effective way.

Learning health systems

A second, more recent, model linking data, information and knowledge to improvements in health care comes from the concept of a learning health system5 in which science and informatics, involving the real-time access to knowledge and digital capture of the care and outcomes, are a central component.

Electronic health records: what does the UK contribute?

Although there may be a strong scientific and health system case for making much better use of the diversity of the ‘data tapestry’, in reality, existing research efforts have been largely focused on structured data collected as part of an interaction with one or more components of the health system.

The UK is one of the few countries in the world in which a picture of the patient journey can be traced through EHRs spanning primary care, hospitals and, ultimately, death registries. The UK has the potential to address research questions that would currently not be possible in Denmark and Sweden (Table 1), countries that do have outstanding national registries but that have not brought national primary care records to research in the same way as is possible in the UK.

A combination of features in the UK underpin the potential of EHRs:

Free at point of use

+

unique health identifier

+

≈ every citizen registered with a general practitioner (GP)

+

≈ every general practice uses EHRs

+

wide range of linkage possibilities, including to the national heart attack registry [Myocardial Ischaemia National Audit Project (MINAP)].

Linkage of EHR and registry data, which are commonly held separately across multiple sources, with bespoke phenotypic and genetic information can generate a unique platform to explore why diseases occur and progress, to investigate quality of care and to identify opportunities to improve health outcomes. This new data revolution has attracted substantial investment and promotion with initiatives to improve access to and use of linked EHR data, which will serve to decrease the burden of obtaining data in a research-ready format and encourage research collaboration.

In what senses are linked electronic health records ‘big data’?

One widely used definition of big data concerns the 4 Vs: volume, variety, veracity and velocity. EHR data are potentially ‘big data’ in every sense; not only do they cover a large number of people (entire populations) with a large amount of information [e.g. 2 million patients, 5 billion rows of data in our ClinicAl disease research using LInked Bespoke studies and Electronic health Records (CALIBER) data platform, which forms the foundation of this programme] per patient, but the data are highly varied and complex, with different ways of coding information (e.g. by means of International Classification of Diseases and Read codes). The data validity (‘veracity’) is simultaneously a concern for clinical care and for research. Clearly, having data in real time (‘velocity’) can be crucial for clinical decision support. Harnessing EHR data for research requires a deep understanding of the health-care system from which the data originated, as well as the statistical and computing skills to analyse large data sets.

In 2011 the UK government published a Strategy for UK Life Sciences,16 which placed EHR research as a central part of the strategy to accelerate health research in the UK. The UK is unique in being the only country with national cardiovascular disease (CVD) registries and primary care data (including important information on cardiovascular risk factors and their management in the community) available at a scale for research. The recent report on personalised health care by the National Information Board and Genomics England17 urged the NHS to transform its use of information in order to provide better care as well as to enable better research. This will require a substantial increase in the maturity of EHRs in the UK as well as a framework to make these data available for research.

Current state of electronic health records in UK hospitals relevant to quality and outcomes research

Although UK general practices have, via the Clinical Practice Research Datalink (CPRD)6 and other initiatives, contributed EHR data that have led to > 1000 peer-reviewed publications, the use of data within hospitals beyond the simple Hospital Episode Statistics (HES) returns to generate useful knowledge has led to very few publications. One of the reasons for this is that UK hospitals, like those in Europe, have been at a low level of maturity, according, for example, to the Healthcare Information Management Systems Society classification (Table 2). This is changing rapidly in the UK in light of the emerging strategy from the National Information Board and NHS Transformation programmes such as the Genomics England-led 100,000 Genomes Project. 19

However, part of the government’s plan for making health-care information available for research, the care.data programme,20 aroused national anxieties about patient confidentiality and had to be suspended pending further public consultation. This illustrates the sensitivities that must be respected in EHR research. Technical solutions such as data pseudonymisation and secure computer workspaces (‘safe havens’) can enable sensitive patient data to be safely used for research, but it is vital to involve patients and the public in balancing privacy concerns against the benefits of research.

Cardiovascular diseases: current public health impact

Cardiovascular diseases remain the leading cause of premature death in the UK and the developed world. 21 Between 1990 and 2010 there were only minor improvements (3% decrease) in years lived with disability from coronary heart disease (CHD),22 and CVD is estimated to cost the NHS about £15.7B a year – 21% of the total NHS budget. 23 Preventing the onset or progression of CVDs remains an important national priority. Interventions can be targeted at different time points along the patient journey: before disease manifests (by risk factor modification), during initial presentation of disease (with prompt and accurate diagnosis, delivery of effective care on referral and admission to hospital) and after discharge from hospital and beyond (optimising secondary prevention). Temporal changes in CVD management have been driven by the availability of new treatment targets [e.g. National Institute for Health and Care Excellence (NICE) guidance and Joint British Societies recommendations on the prevention of CVD guidance] and availability and use of effective treatments for CVDs. 24 Increases in the number of prescriptions of lipid-lowering drugs and number of percutaneous coronary interventions (PCIs) carried out in the UK, together with changes in CVD risk assessment and prevalence, have contributed towards decreases in CVD mortality over the past decade, although the extent to which improved mortality can be attributed to changes in major risk factors, such as smoking, blood pressure (BP) and cholesterol, or specific CVD treatments, remains unclear.

Concerns of a ‘broken pipeline’ in cardiovascular research

There is widespread concern that current models of discovering new interventions, evaluation in trials and implementation in clinical practice take too long (an average of 17 years), are too costly (US$5–11B for each new licensed drug) and are too risky for investors. There is a major need for new research models in general and trial models in particular in cardiovascular research, for several reasons. First, because major unmet needs remain [e.g. acute myocardial infarction (AMI) 1-year mortality is still 20% despite five proven secondary preventative medications]. Second, because late drug failures occurring within Phase III trials (each costing US$50M–100M) are particularly problematic and have been recently seen for drugs that raise high-density lipoprotein (HDL) cholesterol and for heart rate-lowering agents [ivabradine (Procoralan^®, Servier, Neuilly-sur-Seine, France)]. Third, non-drug interventions, based on clinical algorithms and decision support, are of growing importance in CVD but are rarely trialled at all. Fourth, growing use of antiplatelet and anticoagulant drugs is associated with major side effects (bleeding is the second most common cause of iatrogenic hospital admission). Fifth, rare CVDs are undertrialled, partly owing to difficulties in recruitment. Finally, the observational evidence that leads to trials often comes from primary prevention settings; however, the populations in which drugs are first trialled are often in secondary preventative and high-risk settings. Planning a trial can be complex because of the lack of suitable populations from which to estimate accurate event rates and select the most appropriate end points. It is too slow and costly to evaluate feasibility, and opportunities are frequently missed to collect valuable data that will adequately inform late-phase trials. It is inefficient to collect clinical data separately for trials and for clinical care, and the process of randomisation and outcome ascertainment is not integrated into the pathway of clinical care. Trial participants are highly selected, which may limit the generalisability of the trial results, and clinical outcomes in clinical care may not be as promising as suggested by the trials.

Electronic health records data opportunities in coronary disease

It is precisely because hospital information systems have historically been unfit for any purpose relating to understanding the quality and outcomes of care that professional societies have organised themselves and established post hoc collections of data that populate disease registries.

There are very few countries (possibly only the UK and Sweden) that have national mandatory ongoing heart attack registries in which every hospital in the country returns data on care and outcome. MINAP is one of the only national registries recording continuous data from every hospital that manages acute coronary syndrome (ACS) patients. The information it provides about clinical phenotypes is unavailable from any other national source and has shown that the reduced rates of heart attack in the early part of this century have been largely confined to ST-segment elevation acute myocardial infarction (STEMI), whereas rates of non-ST-segment elevation acute myocardial infarction (NSTEMI) have remained static. 25 This detailed epidemiology is less easily available for other manifestations of CVD, such as abdominal aortic aneurysm and ischaemic stroke, because disease-specific registries either have not been developed or have become nationwide (all hospitals) only in the past 5 years. 26 This represents an important gap in our knowledge, and it is not clear whether or not the risk factors and changing epidemiology for CHD can be generalised to other manifestations of CVD.

In the UK, rates of hospitalised AMI (heart attack) and coronary death can be extracted from national data sets such as HES and the Office for National Statistics (ONS) death registry. Richer data on CVDs and procedures are available from national cardiovascular registries. The national registry of percutaneous intervention was originally managed by the British Cardiovascular Interventional Society, and the British Cardiovascular Society and the Royal College of Physicians originally managed the MINAP heart attack registry. In 2010, the Healthcare Quality Improvement Partnership handed responsibility for the six cardiovascular registries to the newly established National Institute for Cardiovascular Outcomes Research (NICOR), which now not only has responsibility for the annual audit reports but also makes the registry data available to research groups. 27

The overarching aim of CALIBER is to harness diverse data sources to undertake novel research in CVDs, including opportunities improving care. Our research programme adopted the ‘patient journey’ concept to explore missed opportunities in primary prevention, diagnosis, acute care and secondary prevention (Figure 2). By harnessing the potential of linked EHRs, we seek to provide valuable information to guide clinical decision-making and health policy-making, with the potential to improve quality of care and outcomes for patients.

FIGURE 2.

Overview of the 33 studies in our programme: the patient journey. SCAD, stable coronary artery disease.

Owing to the evolution of coronary atherosclerosis and symptomatic manifestations (angina, heart attack), which can be viewed as chronic, long-term conditions, our research programme adopted the ‘patient journey’ concept to explore missed opportunities in primary prevention, diagnosis, acute care and secondary prevention. In particular, we have attempted to exploit the potential offered by linked EHRs spanning the patient journey through primary care, the hospital system and ultimately to death. By identifying the nature and scale of such missed opportunities over time, we seek to provide valuable information to guide clinical decision-making and health policy-making, with the potential to improve quality of care and outcomes for patients. A summary of our main findings and research recommendations for 33 studies included in our programme can be found in Table 3.

Public and user group involvement

Electronic health records are increasingly considered to belong to individual patients. The National Information Board recommended online access to the full primary care record by April 2015. 17 In CALIBER we have had an active public and user involvement group which has had biannual meetings since 2006. Following the INVOLVE guidelines,66 this group represents a high level of user and lay participation in research, which includes contributing to the design and conduct of research. The group includes people relatively new to the ‘patient journey’ and others with many years’ experience of being cared for within the UK NHS. Our programme has benefited greatly from the group’s engagement, which has informed the development of studies from inception through to analysis and interpretation of findings. Central to the vision of the Farr Institute of Health Informatics Research is the need to transform the relationship between the patient, their health record and its uses for clinical care and translational research. The Farr Patient and Public Working Group meets biannually and is actively involved in national health informatics research network projects, research and events which are also open to lay members of the public.

Chapter 2 (studies 1–3): the ClinicAl disease research using LInked Bespoke studies and Electronic health Records platform

We have established a framework for studying CVD across the patient journey, introducing novel approaches and scrutinising underlying methodology.

• Linked EHRs: to map the patient journey, we exploited the previously untapped resource of the EHR for research purposes and, for the first time, linked data across primary care (CPRD), hospital admission (HES), a disease-specific registry (MINAP) and the national death registry (ONS). The resulting CALIBER data set comprises > 2 million patient records. The governance and analytic challenges were considerable, requiring a series of permissions, a contractual relationship with the Medicines and Healthcare products Regulatory Agency, engagement of a trusted third party to perform the linkage, safe storage of the linked data sets, development of detailed algorithms to define clinical phenotypes of interest and, finally, application of complex analytic strategies to take account of the volume of data, missing data and hidden confounders.

• Promoting high-quality prognostic research leading to patient benefit: in order to establish a high-quality framework for prognosis research, we created the international PROGnosis RESearch Strategy (PROGRESS) partnership, with representatives from academia and medical journal editors, to evaluate the field of prognosis research and make specific recommendations on improving the quality and impact of prognosis research.

Clinical context of diagnosis, risk prediction and treatment of coronary diseases: synopses of studies

The studies in the chapters that follow are a combination of synopses of previously published (n = 26) and currently unpublished (n = 15) studies and are summarised in Table 3 and Figure 2. We have included more detail for the unpublished studies. A summary of the published papers relating to individual studies is presented in Appendix 1.

Background

Electronic health records are generated at an unprecedented scale during patients’ interactions with the NHS. The UK is unique in being able to trace a complete patient journey through coded EHRs spanning primary care, hospitals, disease registries and, ultimately, death registries. Large-scale linkage of EHRs in CVD has major challenges, and the quality of prognosis research, which provides crucial evidence for translating research findings into clinical practice, is poor.

Objectives

(1) To establish an EHR-based research platform and data resource in CVD (CALIBER) and (2) to develop a framework for exemplary prognosis research.

Methods

(1) We linked coded EHR data from four sources spanning primary care (CPRD), secondary care (HES) and national registry databases (MINAP and the ONS death registry). (2) We established the international PROGRESS partnership.

Results

(1) Through CALIBER, we linked and curated data covering almost 2 million people, and curated ≈ 600 EHR disease and risk factor phenotyping algorithms based on linked EHR data. (2) We made 24 recommendations to address challenges and opportunities across four prognosis research themes.

Conclusions

The CALIBER platform promotes the transparent and scalable use of EHR data, enabling the wealth of information in UK EHRs to be exploited for research purposes. Our CALIBER data resource and recommendations for improving the quality of prognosis research provide a framework for high-quality CVD research.

Linked data sources

The four linked data sets that we used are shown in Figure 3, which illustrates the longitudinal, population-based nature of CALIBER data sets. Each source captures a different aspect of the patient journey:

• CPRD contains longitudinal primary care coded data on symptoms, diagnosis, prescriptions, investigations, referrals and vaccinations. 85 Data are recorded prospectively as part of clinical care and are not subject to recall bias or differential bias related to outcome. Diagnostic data are coded using Read codes, which map to Systematic Nomenclature of Medicine – Clinical Terms. Of the 684 primary care practices in CPRD (13.6 million patients), 226 consented to data linkage across the CALIBER data sources. These practices contained 3.9% of the population of England in 2006.

• HES is a national data set of all admissions to NHS hospitals in the UK (see www.hscic.gov.uk/hes). Diagnoses are recorded using the International Statistical Classification of Diseases and Health-Related Problems, Tenth Edition (ICD-10) and procedures using the Office of Population Censuses and Surveys’ Classification of Interventions and Procedures version 4.

• MINAP is a national registry of patients admitted to all acute hospitals in England and Wales with ACS events (n = 230). For each admission, MINAP records patient demographic characteristics, limited medical history, clinical features and investigations, drug treatment (pre, during and post admission) and final diagnosis, including differentiation into STEMI, NSTEMI and unstable angina. Data undergo annual quality assessments. A descriptive review of the MINAP registry has been published. 86

• The ONS provides data on cause-specific mortality for all residents of the UK. 87 Cause-specific mortality data are extracted from death certificates and recorded using ICD-10. Small-area measures of social deprivation are recorded using the 2007 Index of Multiple Deprivation (IMD)88 and Townsend scores.

FIGURE 3.

Longitudinal nature of four linked primary care and disease registry data sources in the CALIBER data resource. AAA, abdominal aortic aneurysm; ACEI, angiotensin-converting enzyme inhibitor; ECG, electrocardiogram; MI, myocardial infarction. Reproduced from Denaxas et al. 15 Published by Oxford University Press on behalf of the International Epidemiological Association. © The Author 2012. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by-nc/3.0/), which permits non-commercial reuse, distribution, and reproduction in any medium, provided the original work is properly cited.

Converting raw data to research-ready data: curated common data model

Important challenges remain in relation to converting raw EHR data into research-ready variables. The manner and reason for the generation, capture and recording of EHR data vary substantially between health-care settings. In addition, different medical classification systems are often used for different data sources and, consequently, clinical information may be recorded in multiple sources but at different levels of clinical detail. Integrating data from these different sources is therefore complex, but doing so is advantageous for research and consequently for clinical purposes. For example, our research illustrated how linkage of EHR data sources substantially improves the identification of AMI cases.

The complexities of EHR data mean that a structured method for combining diverse data sources and creating EHR ‘phenotypes’ is needed, but no internationally recognised framework currently exists. An EHR phenotyping algorithm translates these clinical requirements into queries that leverage multiple linked EHR sources, identify relevant patients and extract disease onset, severity and subphenotype information. Defining phenotyping algorithms for a study (disease end points, disease start points, exclusions, covariates and risk factors) consists of identifying relevant codes for diagnosis (e.g. diagnosis of cancer) or risk factors (e.g. recording of smoking status) and defining how they should be combined. However, these definitions, on which research findings are based, are poorly described in published literature, with the exception of cases in which the curation of phenotyping algorithms occurs upstream by the data provider [e.g. electronic medical records and genomics network (eMERGE)].

We created and evaluated a framework (Figure 4) for transforming raw EHR data into research-ready phenotypes in a collaborative and transparent manner by involving domain experts across disciplines (epidemiology, computer science, public health, health informatics) and using a centralised repository for their distribution and curation. Here, we illustrate the steps involved in creating a phenotype using the example of atrial fibrillation (Figure 5). The framework consists of multiple iterative steps from initial phenotype definition with domain experts to algorithm implementation, validation and distribution to the scientific community. All CALIBER studies are registered in the public domain on Clinicaltrials.gov, and their analytic protocol is made available for download.

FIGURE 4.

Development of a phenotype algorithm using the CALIBER programme. Reproduced from Morley et al. 89 © 2014 Morley et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

FIGURE 5.

Flow diagram illustrating the CALIBER phenotype for atrial fibrillation. AF, atrial fibrillation; DVT, deep-vein thrombosis; PE, pulmonary embolism. Reproduced from Morley et al. 89 © 2014 Morley et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

To aid the process of creating, validating and disseminating EHR phenotyping algorithms for epidemiological research, we created the CALIBER data portal (www.caliberresearch.org/portal), an interactive online repository of phenotyping algorithms. All algorithms are displayed on the portal, where registered external researchers can view and comment on them. A number of metadata fields is recorded for each algorithm and is made available. The data portal contains phenotyping algorithms that leverage multiple EHR data sources in order to define > 90 diseases and related risk factors. The data portal also facilitates the transportability of the developed algorithms on national EHR data from different countries where subtle differences might exist in the manner clinical information is recorded. Algorithms available through the data portal can be utilised for performing international comparisons between health systems (e.g. England and the USA) or act as the base on which phenotyping algorithms reflecting the local health system are created.

The CALIBER portal operates a collaboration-driven data-sharing model and prioritises collaborations based on scientific-added value, capacity development and sustainable funding. Details on CALIBER and how to access CALIBER resources are available at www.ucl.ac.uk/health-informatics/caliber.

Strengths and limitations of the data platform

The CALIBER portal spans ambulatory and hospitalised care in large population samples, permitting a ‘higher resolution’ approach to cardiovascular epidemiology in three respects:

1. resolution of broad CVD aggregates (e.g. CVD, CHD) into distinct disease phenotypes in different vascular beds:

   1. cerebral circulation – transient ischaemic attack, ischaemic stroke, subarachnoid haemorrhage, intracerebral haemorrhage

   2. peripheral circulation – PAD, abdominal aortic aneurysm

   3. coronary circulation – stable angina, unstable angina, AMI (STEMI and NSTEMI), unheralded CHD death, heart failure, cardiac arrest/sudden coronary death

2. sufficient temporal resolution to distinguish whether a first event (e.g. heart attack or stroke) was the first manifestation of CVD or if it was preceded by any of a range of non-fatal events

3. sufficient number of events and patients to generate accurate estimates of risk.

The main limitations of any EHR research platform relate to data quality:

1. Missing data. Data completeness is variable [e.g. 82.6% of people have ≥ 1 body mass index (BMI) measurement but only 44.9% have ≥ 1 cholesterol measurement]. Developments in imputation within longitudinal cohorts may offer a partial solution. 90

2. CALIBER makes use of coded data from multiple national EHR sources. A plethora of unstructured data exist in secondary care that are recorded within hospital information systems but remain inaccessible for research. This is mainly due to the lack of a centralised manner in which these data can be extracted, linked and utilised at scale. Significant national initiatives such as the National Institute for Health Research (NIHR) Health Informatics Collaborative (HIC) aim to create informatics infrastructure that will provide means for accessing these phenotypically rich data sources.

3. An important aspect of EHR phenotype development is validation, preferably against a ‘gold standard’ (such as a manual review of case notes). We could not validate the developed phenotypes in this manner as, for the CALIBER programme, the initiation and funding of a separate study is required for recontacting participants or clinicians to confirm diagnoses or review records. Previous studies, however, have shown the high degree of reliability of the data. 91

4. Currently, restrictive confidentiality legislation and information governance constraints in England prevent the use of anonymised clinical text on a large scale. Electronic free text in secondary care does not get captured or extracted in any form by HES and is thus not available for research. In primary care, free text from GPs contributing to CPRD is downloaded and held on government servers at CPRD, and is not readily available at scale to researchers based in universities (researchers can get access to very small corpora of text which has to be manually anonymised, a process that requires substantial resources and does not scale). Consequently, our approach to developing phenotyping algorithms has limited ability to fully utilise data elements extracted from clinical text as others have done.

Introduction

We assessed the concordance, completeness and diagnostic validity of AMI recording across CPRD, HES, MINAP and ONS in order to ensure the robustness of the data linkage. Our specific aims in this study were to compare the incidence, recording, agreement of dates and codes, risk factors and all-cause mortality of AMI.

Methods

Results

Some 21,482 patients had a record of fatal or non-fatal AMI recorded in any of the four data sources. Single-source crude estimates for incident AMI per 100,000 patient-years are shown in Figure 6. The crude incidence of AMI was 25% lower using only CPRD and 50% lower using only MINAP compared with using all three sources.

FIGURE 6.

Crude incidence of fatal and non-fatal AMI estimated using different combinations of data from primary care (CPRD), hospital admissions (HES), disease registry (MINAP) and death registry (ONS). MI, myocardial infarction. Reproduced from Herrett et al. 28 Copyright © 2013, British Medical Journal Publishing Group. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by-nc/3.0/), which permits non-commercial reuse, distribution, and reproduction in any medium, provided the original work is properly cited.

Death after acute myocardial infarction

Crude 30-day mortality was lower in MINAP [hazard ratio (HR) 10.8%, 95% confidence interval (CI) 10.2% to 11.4%] than in HES (HR 13.9%, 95% CI 13.3% to 14.4%) or CPRD (HR 14.9%, 95% CI 14.4% to 15.5%) (Figure 7a). At 1 year, however, mortality was similar in all three groups, at around 20% (Figure 7b).

FIGURE 7.

Kaplan–Meier curves. (a) Kaplan–Meier curves showing 30-day all-cause mortality, stratified by record source (CPRD, HES and MINAP) in 20,819 patients. CPRD, n = 15,819; HES, n = 13,831; and MINAP, n = 10,351. MIs recorded by the ONS are not shown, as they are, by definition, fatal on the date of AMI; and (b) Kaplan–Meier curves showing 12-month all-cause mortality, stratified by record source (CPRD, HES and MINAP) in 20,819 patients. MI, myocardial infarction. Reproduced from Herrett et al. 28 Copyright © 2013, British Medical Journal Publishing Group. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by-nc/3.0/), which permits non-commercial reuse, distribution, and reproduction in any medium, provided the original work is properly cited.

In the linked data, patients with AMI recorded in only one source had higher mortality than those recorded in more than one source (age- and sex-adjusted HR 2.29, 95% CI 2.17 to 2.42; p < 0.001); those recorded only in CPRD had the highest mortality on the first day but the lowest mortality thereafter, compared with HES and MINAP (adjusted HR for coronary mortality 0.49, 95% CI 0.40 to 0.60; p < 0.001; and for other causes 0.57, 95% CI 0.49 to 0.67; p < 0.001). Of the 3518 patients with AMI recorded in any of the four sources who died of any cause within 7 days, 54.4% (n = 1914) had an AMI code recorded in CPRD within 30 days. The underlying cause of death recorded in ONS was AMI in 2924 patients (83.0%). Of fatal AMIs identified by ONS, only 36.7% (n = 1072) were recorded in HES and 17.1% (n = 498) in MINAP within 30 days, but 55.9% (n = 1631) were recorded in CPRD.

Non-fatal acute myocardial infarction

Among the 17,964 patients with at least one record of non-fatal AMI, 13,380 (74.5%) were recorded in CPRD, 12,189 (67.9%) in HES and 9438 (52.5%) in MINAP. Overall, 5561 (31.0%) patients had the event recorded in all three sources and 11,482 (63.9%) patients had it recorded in at least two sources (Figure 8). When we extended the recording window from 30 to 90 days, the proportion recorded in all three sources increased only slightly, to 32.0% (n = 5747). Only 25.1% (3364/13,380) of the non-fatal AMIs recorded in CPRD stated the type (STEMI or NSTEMI), compared with 100% in MINAP.

FIGURE 8.

Number and percentage of records recorded in primary care (CPRD), hospital care (HES) and disease registry (MINAP) for non-fatal AMI across the three sources (n = 17,964 patients). Reproduced from Herrett et al. 28 Copyright © 2013, British Medical Journal Publishing Group. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by-nc/3.0/), which permits non-commercial reuse, distribution, and reproduction in any medium, provided the original work is properly cited.

Discussion

In comparing records of one major disease event – AMI – across CALIBER’s four linked sources, we found that each source misses a substantial proportion of events. However, our data confirm the validity of AMI recorded in these sources, in terms of risk factor profiles and mortality at 1 year. Taken together, these findings support the wider use of linkage of multiple record sources by clinicians, policy-makers and researchers in order to maximise diagnostic capture.

Conclusions

Our findings show that by linking EHRs we not only provide a longitudinal component with which to map the CVD journey from first diagnosis to death but also enrich the data sources and maximise capture of AMI with a high level of diagnostic validity. Assuming this extends to other manifestations of CVD, our findings provide reassurance that the CALIBER programme has the scale and diagnostic validity necessary for national EHR research.

Background

Understanding how CVDs first present is vital to organise primary and secondary prevention efforts, but little is known about CVDs other than AMI, stroke and coronary death. UK South Asian populations are particularly affected by CHD but it is unclear if this reflects a higher incidence of disease, worse prognosis or both.

Objectives

(1) To identify how the development of 12 CVDs is affected by age, sex, ethnicity and socioeconomic deprivation; and (2) to examine the complex relationships between incidence and prognosis of CHD among South Asian populations.

Methods

(1) An analysis of CALIBER linked data; and (2) an analysis of cohort data and a meta-analysis of published data.

Results

(1) Acute myocardial infarction, stroke and coronary death collectively account for just 34% of initial cardiovascular events. The strength and shape of the association of CVDs differed with age, and the risk of different CVDs differed by sex. Initial presentation of CVDs showed considerable heterogeneity with deprivation. Those of South Asian ethnicity were at substantially increased risk of all coronary presentations compared with white patients. (2) CHD incidence among those of South Asian ethnicity was markedly higher than among white patients, but their prognosis was significantly better.

Conclusions

These are the only contemporary, large-scale incidence estimates of a wide range of fatal and non-fatal cerebrovascular, cardiac and peripheral CVDs. Our findings are crucial for understanding the global burden of CVDs. Our findings argue against the use of composite measures of cardiovascular health, which may mask important variability across conditions.

The studies in this chapter relate to people without established disease (healthy population) (Figure 10).

FIGURE 10.

Overview of the 33 studies in our programme: population without established disease.

Introduction

An important principle in CVD management is that the first lifetime diagnosis indicates the failure of primary prevention and the need for measures to prevent subsequent events. The decades-long emphasis given to the prevention of AMI and stroke is reflected in the significant declines in their incidence in developed countries. 119 Incidence rates for angina, heart failure and other chronic CVD presentations do not appear to have similarly declined. 119–121 Consequently, the spectrum of initial lifetime presentations of CVD in contemporary practice is likely to have changed, compared with the end of the last century. Cohort studies that report only fatal end points (‘final presentations’)122 are less informative for primary preventative efforts than those that investigate initial lifetime presentations, and studies that include non-fatal events have generally focused on AMI and stroke while ignoring other CVDs. 123–126

Large-scale studies that evaluate the first lifetime diagnosis in different ages, sexes, socioeconomic groups and ethnicities across a wide range of acute and chronic CVDs including both fatal and non-fatal presentations are overdue.

Fundamental unanswered questions about initial lifetime CVD presentation arise. First, what is the relative frequency of different CVDs as they affect different ages, sexes, ethnicities and socioeconomic groups in contemporary practice? Second, what is the lifetime risk of the most common CVDs? Lifetime risks provide an intuitive measure of the burden of each disease in the population, relevant to public health policy and risk communication. Lifetime risks have recently been reported for CHD and CVD aggregates127–129 and heart failure,130 but rarely for a wide range of initial presentations of CVDs. 38 Third, is male sex an equally strong risk factor common to all CVDs, or does the association differ across a range of diseases?

The lack of large, contemporary, population-based cohorts with detailed clinical follow-up spanning hospital and ambulatory care has hindered the study of the initial presentation of a wide range of acute and chronic CVDs. It has been suggested that EHR data might be ‘meaningfully reused’131 to create ‘mega cohorts’ for such research. 132 We studied a contemporary, population-based cohort based on linked EHRs across primary, secondary, disease registry and death records to address these questions.

Methods

The CALIBER study population was 1,937,360 patients for age, sex and deprivation analyses, and 1,068,318 for ethnicity analyses. For ethnicity, patients were categorised as white, black, South Asian or other/mixed. Deprivation was measured by the IMD (2007 revision),88 derived from postcode of residence. Patients were categorised into quintiles, where the first quintile indicated the least deprived and the fifth quintile indicated the most deprived group. Primary end points for the analysis were the initial presentation of fatal and non-fatal CVDs across the CALIBER data sources. CVD presentations studied were stable angina, unstable angina, AMI, heart failure, a composite of ventricular arrhythmia, cardioversion, cardiac arrest or sudden cardiac death (SCD), transient ischaemic attack, subarachnoid haemorrhage, intracerebral haemorrhage, ischaemic stroke, abdominal aortic aneurysm and PAD. Data analysis was based on Cox proportional hazards models.

Results

Follow-up was for a median of 5.5 years, during which 114,859 subjects experienced one of the 12 CVDs (52.3% in men). AMI, stroke and coronary death accounted for less than one-third of the total CVD presentations recorded in CALIBER.

Age

The strength and shape of the association of CVDs with age (Figure 11) varied from predominantly linear (in angina and non-fatal AMI) to strongly quadratic (unheralded coronary death, stroke, abdominal aortic aneurysm), and from weak (subarachnoid haemorrhage, unstable angina and cardiac arrest-SCD) to very strong (heart failure and abdominal aortic aneurysm). 35

FIGURE 11.

Age-adjusted HRs of men vs. women for initial presentation of 12 different CVDs among a population of 1.93 million adults. HRs are derived from Cox proportional hazards models stratified by general practice and adjusted for baseline age (linear and quadratic terms) and interaction between sex and baseline age. MI, myocardial infarction. Reproduced from George et al. 35 Copyright © 2015, The Authors. Circulation is published on behalf of the American Heart Association, Inc., by Wolters Kluwer. This is an open access article under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/4.0), which permits use, distribution, and reproduction in any medium, provided that the original work is properly cited.

Sex

The most frequent initial CVD presentation was non-fatal AMI for men (17% of total CVD) and stable angina for women (12% of total CVD) (Figure 12). The cumulative lifetime incidence of CVD was 51.3% for men and 41.2% for women. Men were more likely than women to experience non-fatal AMI (7.1% vs. 3.9%) and abdominal aortic aneurysm (2.4% vs. 0.9%). Men experienced conditions affecting the myocardium, PAD and abdominal aortic aneurysm around 10 years earlier than women. However, the lifetime incidence of heart failure (8%), transient ischaemic attack (5%), ischaemic stroke (3%) and intracerebral haemorrhage (1%) was similar in men and women.

FIGURE 12.

Proportion of total initial lifetime presentations of CVDs by disease type in men and women. Columns in descending order by proportion of events in men aged 30–39 years. Number of events displayed at base of column. MI, myocardial infarction; NOS, not otherwise specified. Reproduced from George et al. 35 Copyright © 2015 The Authors. Circulation is published on behalf of the American Heart Association, Inc., by Wolters Kluwer. This is an open access article under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/4.0), which permits use, distribution, and reproduction in any medium, provided that the original work is properly cited.

Apart from subarachnoid haemorrhage (HR men vs. women 0.69; 95% CI 0.59 to 0.79), all presentations showed a positive association with male sex (Figure 13). However, the strength of this association varied markedly: for transient ischaemic attack, intracerebral haemorrhage and unstable angina, age-adjusted HRs for male sex were < 1.5; for stable angina, ischaemic stroke, PAD, heart failure and cardiac arrest/SCD they were between 1.5 and 2.0; and for abdominal aortic aneurysm, AMI and unheralded coronary death they were between 3.6 and 5.0.

FIGURE 13.

Hazard ratios of men vs. women for initial lifetime presentation of 12 different CVDs among a population of 1.93 million adults. The vertical grey dotted line corresponds to the HR of the composite CVD end point. Reproduced from George et al. 35 Copyright © 2015 The Authors. Circulation is published on behalf of the American Heart Association, Inc., by Wolters Kluwer. This is an open access article under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/4.0), which permits use, distribution, and reproduction in any medium, provided that the original work is properly cited.

Deprivation

Associations of deprivation with initial presentations of CVDs showed considerable heterogeneity in both women and men. In women (Figure 14), socioeconomic deprivation showed no association with abdominal aortic aneurysm. In men too there was no association with abdominal aortic aneurysm and no association with stable angina, cardiac arrest-SCD, subarachnoid haemorrhage or transient ischaemic attack. For other CVDs, hazard increased linearly with deprivation quintile, the increase being comparable between the sexes for unstable angina, AMI and heart failure, but steeper in men for PAD, ischaemic stroke and intracerebral haemorrhage.

FIGURE 14.

Hazard ratios for the association between the initial presentation of 12 CVDs and socioeconomic deprivation adjusted for cardiovascular risk factors measured at baseline in women. Reference: least deprived quintile. MI, myocardial infarction; Q, quintile. Reproduced from Pujades-Rodriguez et al. 34 © 2014 Pujades-Rodriguez et al. 34 This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Differences in lifetime risk between the least and most deprived quintiles were greatest for PAD in both women (4.3% vs. 5.8%) and men (4.6% vs. 7.8%). However, in both sexes, differences were small or non-existent for cardiac arrest-SCD, transient ischaemic attack, abdominal aortic aneurysm, and ischaemic and intracerebral haemorrhage.

Discussion

By linking EHRs, this study has analysed > 100,000 initial CVD presentations arising in a cohort of nearly 2 million people, providing sufficient scale to distinguish multiple types of CVD and assess inequalities in the risk of initial presentation according to age, sex, ethnicity and levels of deprivation. Through the use of linked EHRs, we were able to capture diseases first presenting in primary care and were not confined to hospitalised cases. Our cohort is population based, with > 99% of the English population estimated to be registered with a family physician,133 unlike other recent large recent cohort studies, such as the UK Biobank,107 which had response rates of < 10%. 134 The size of cohort possible with this EHR platform – nearly 2 million people – allows us to include serious but less common events such as abdominal aortic aneurysm and subarachnoid haemorrhage, as well as to have sufficient events in women to study sex differences reliably. We were able to follow up actual events in a clinically meaningful 5-year time frame, similar to the time horizon of randomised controlled trials (RCTs). More broadly, we demonstrate the potential of linked EHR cohorts to complement bespoke, investigator-led cohorts. 73 The UK Biobank,107,112 the Research Program for Genes Environment and Health in Kaiser12 and the ‘precision’ medicine initiatives135 all place a major emphasis on specific disease types and follow-up through health records. Furthermore, such large EHR cohorts hold out the promise of a lower cost to research funders for data collection, the intrinsic clinical relevance of ‘real-world’ data, the opportunities to study diseases with higher specificity given the cohort sizes possible, and the prospect for researchers and clinicians to work across the boundaries that currently impede the translation of new discoveries into public health benefit. 73

Although AMI and ischaemic stroke are the most commonly used end points in trials, they collectively account for just 34% of initial cardiovascular events. The majority of cardiovascular presentations are heart failure, angina, transient ischaemic attack and PAD, which are rarely evaluated individually in primary prevention studies.

Despite recent progress in prevention, there is still a 45% lifetime risk of any CVD between the ages of 30 and 95 years, with men at greater risk than women (51.3% vs. 41.2%). However, the effect of male sex varies significantly between the 12 different CVD presentations. Men are actually less likely to experience subarachnoid haemorrhage and show similar incidence levels to women for several conditions. Male sex showed a moderate association with five presentations and a strong association with abdominal aortic aneurysm, AMI and unheralded coronary death. However, because men typically experience some CVDs at a younger age, the lifetime incidence of several presentations (heart failure, transient ischaemic attack, ischaemic stroke and intracerebral haemorrhage) was similar in men and women.

Differences in the risk of CVDs associated with South Asian and black ethnicity were striking but often affected different vascular beds. Thus, the heightened risk of presenting with coronary disease among South Asian than white populations contrasted with the reduced risk for people of black ethnicity who were more subject to haemorrhagic stroke. However, those of both South Asian and black ethnicity had a reduced risk of presenting with PAD than people of white ethnicity, which was consistent with the findings of a recent meta-analysis. 136

Similarly, we found that simple associations between local socioeconomic deprivation and broad categories of CVD do not hold true for more granular CVD classifications, particularly in men. Age was an important influence, with associations between socioeconomic deprivation and CVDs falling with increasing age. Hence, the cardiovascular health risk linked to local socioeconomic deprivation falls most heavily on younger people, and on young women in particular.

Our findings suggest that future research on the primary prevention of CVDs should take account of current patterns of disease presentation and redress the imbalance of previous literature, which has focused overwhelmingly on heart attack and stroke. Estimates of national and global burden of CVDs need to take account of the modern epidemic in which heart failure, PAD and other chronic diseases are becoming more common than AMI and stroke. Male sex, older age, South Asian ethnicity and low levels of social deprivation, although associated with some CVDs, are not associated with all, and this heterogeneity has implications for research into risk prediction, aetiology and prognosis.

Our findings have implications for the design and interpretation of observational studies, randomised trials and meta-analyses investigating the primary prevention of CVDs. Because the fundamental risk factors of age and sex have such heterogeneous associations with different CVDs, and most studies are sufficiently powered to examine only CVD aggregates, it is important to account for the relative proportion of each disease in the composite end point in meta-analysis. These proportions have been shown to differ substantially between studies. 137 Despite an extensive literature on the underlying biological and behavioural pathways by which sex may influence aggregates of CVD and CHD, there is a lack of mechanism studies that investigate why sex has such heterogeneous associations with different CVDs.

Conclusions

These inequalities in CVD development support the use of separate risk models for women and men and for different ethnic groups. They argue against the use of composite measures of cardiovascular health, which may mask important variability across conditions. The findings suggest that policies and interventions targeting local socioeconomic deprivation are likely to have the greatest impact on specific forms of CVD, such as heart failure and PAD, and on women. Whether or not the complex inequalities in CVD development are associated with inequalities in CVD outcomes is an important question that we begin to address in the following study, which focuses on ethnicity and its association with incidence and prognosis of CHD.

Introduction

Data presented in study 4 point strongly to greater disease incidence as a key contributor to the higher coronary death rate in South Asian people than in white people living in the UK. Whether or not a worse prognosis among people with established coronary disease also contributes to this higher death rate remains unresolved. Understanding the relative importance of incidence and prognosis to health outcomes would provide important information to guide the most appropriate public health interventions. To generate a clearer picture, this study adopted a two-pronged strategy. A systematic review and meta-analysis was carried out to synthesise the evidence on coronary disease incidence and prognosis in South Asian populations. In addition, MINAP data were used to generate a contemporary cohort of South Asian patients from England and Wales, to examine possible factors affecting prognosis in this group.

Methods

The systematic review was based on an expert search strategy of the published literature in MEDLINE (PubMed) from inception until January 2012 to identify papers on incidence or prognosis in people of South Asian ancestry specifically related to coronary disease. Three reviewers independently extracted information according to a predetermined protocol. The new ACS cohort was based on consecutively collected patients from 230 hospitals in England and Wales recorded in MINAP between 2004 and 2008. The primary outcome measures were age-adjusted mortality at 1 year and at 30 days.

Results

Incidence

The meta-analysis was based on data for 111,555 people of South Asian ethnicity and 4,197,923 white people included in 11 eligible studies. It showed that South Asian participants had a higher incidence of coronary disease than white people (HR 1.35, 95% CI 1.30 to 1.40). This was consistent for the two cohorts examining incidence of fatal coronary disease and the studies examining incidence of non-fatal (alone or with fatal) coronary disease (Figure 15).

FIGURE 15.

Incidence of coronary disease in South Asian compared with white individuals: meta-analysis of 11 incidence populations including 111,555 South Asians. Estimates are HRs unless otherwise indicated. A, age; D, diabetes; F, female; F/NF MI, fatal/non-fatal myocardial infarction; G, sex; H, hypertension; L, dyslipidaemia; M, male; S, smoking; UKPDS, UK Prospective Diabetes Study. a, Odds ratio; b, incidence ratio. Reproduced from Zaman et al. 36 Copyright © 2013 BMJ Publishing Group Ltd and the British Cardiovascular Society. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by-nc/3.0/) which permits non-commercial reuse, distribution, and reproduction in any medium, provided the original work is properly cited.

Prognosis

By contrast, an analysis of 12 prognosis populations including 14,531 South Asian and 274,977 white patients showed that South Asian individuals had a better prognosis than white people (HR 0.78, 95% CI 0.74 to 0.82). This more favourable prognosis was particularly apparent in bigger, more recent studies on acute clinical presentations (Figure 16). In the ACS cohort, diabetes was markedly more common in South Asian than in white populations (42.4% vs. 16.9%) and had greater impact on mortality. Even so, prognosis was superior for South Asian people overall and in most subgroups analysed, including those with diabetes (HR 0.78, 95% CI 0.71 to 0.86), those of older age (≥ 80 years of age; HR 0.72, 95% CI 0.63 to 0.83) and those living in areas of highest social deprivation (HR 0.73, 95% CI 0.65 to 0.82). This superior prognosis was seen for patients of Indian, Pakistani and Bangladeshi origin. It was also seen over the 30-day time frame (HR 0.85, 95% CI 0.77 to 0.94, after adjustments). An analysis of the MINAP cohort showed that South Asian patients with ACS were more likely than white patients to receive revascularisation and to be discharged on four secondary prevention medications.

FIGURE 16.

Prognosis of coronary disease in South Asian individuals compared with white people: meta-analysis of 12 prognosis populations including 14,531 people of South Asian ethnicity. Estimates are HRs unless otherwise indicated. A, age; ACRE, Appropriateness of Coronary REvascularisation; D, diabetes; G, sex; H, hypertension; I, IMD; L, dyslipidaemia; P, phenotype of ACS; S, smoking. a, Odds ratio. Reproduced from Zaman et al. 36 Copyright © 2013 BMJ Publishing Group Ltd and the British Cardiovascular Society. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by-nc/3.0/) which permits non-commercial reuse, distribution, and reproduction in any medium, provided the original work is properly cited.

Discussion

The evidence synthesis suggests that South Asian ethnicity has different effects on the incidence and prognosis of coronary disease. Although South Asian patients show a markedly higher incidence than white patients, after adjustment for age, risk factors and disease severity, their prognosis is significantly better.

This study confirms ethnic inequality in the risk of developing coronary disease, consistent with our CALIBER analyses in study 4. The present meta-analysis confirmed that South Asian patients with SCAD were at no prognostic disadvantage. Indeed, among South Asian patients with ACS, prognosis was significantly better than for white patients. There was no evidence of undertreatment, with South Asian patients being more likely to be prescribed secondary prevention medications at discharge and more likely to have undergone coronary angiography. This may reflect awareness on the part of physicians of the great coronary disease risk in South Asian groups, or a more general awareness in the health system of factors relevant to the health of South Asian individuals.

Large cohort studies identified through linkage of primary care, ACS registry, hospitalisation and cause-specific death data are warranted to better understand how South Asian ethnicity influences the onset and progression of coronary disease and other CVDs. Such studies will allow longitudinal patterns of risk, and missed opportunities to lower risk before admission and after discharge, to be evaluated.

Conclusions

Although South Asian populations experience health inequalities in coronary disease, this study suggests that this reflects a higher incidence of disease rather than deficiencies in service provision affecting prognosis. The findings therefore suggest that public health measures to tackle ethnic inequalities in coronary mortality should focus on primary prevention and reducing the anomalous incidence of disease in South Asian populations.

Background

Reported failures to meet national guideline targets and treatment opportunities to optimise risk reduction among people without established disease represent an important missed opportunity for protecting patients against CVD.

Objectives

(1) To investigate and compare associations between type 2 diabetes and the risk of 12 initial cardiovascular presentations in men and women; (2) to investigate the association between BP and 12 presentations of CVDs; and (3) to identify the proportion of hypertensive patients attaining BP control according to national targets and factors associated with attainment of targets.

Methods

An analysis of CALIBER linked data.

Results

(1) Type 2 diabetes had the greatest impact on PAD and heart failure but was protective against abdominal aortic aneurysm. (2) There was substantial heterogeneity in associations with BP across different CVDs and ages, and the lifetime burden of hypertension was considerable. (3) Twenty per cent of newly diagnosed hypertensive patients did not achieve guideline-recommended reductions in BP, partly reflecting underuse of antihypertensive drugs and longer time periods between diagnosis and initiation of treatment. Consultation rates were significantly lower among those whose BP was not reduced to target (6.9 vs. 34 consultations per year).

Conclusions

This study confirms that there are high levels of undertreatment among hypertensive patients. The lifetime burden of hypertension remains substantial, despite modern therapy. More vigorous clinical management could increase the numbers achieving BP targets. In patients with type 2 diabetes, more rigorous glycaemic control could have benefits for cardiovascular health.

The studies in this chapter relate to people without established disease (healthy population) (see Figure 10)

Introduction

In the UK and globally, there is an epidemic of type 2 diabetes and growing concern over the contemporary impact of type 2 diabetes on a wide range of macrovascular consequences. The association of type 2 diabetes with CHD and stroke is well established156 but its wider associations across the range of CVDs affecting the myocardium and cerebral, abdominal and peripheral vasculature are less well documented. No previous studies have been sufficiently large to compare these associations and to estimate less common outcomes reliably. It is not clear, therefore, to what extent patients with diabetes are at risk of different CVDs, or whether or not risks are affected by glycaemic control as reflected by glycated haemoglobin (HbA_1c) levels. The possibility that these treatment opportunities are being missed prompted this cohort study, which used the CALIBER database of linked EHRs to investigate and compare associations between type 2 diabetes and future risk of 12 of the most common initial cardiovascular presentations in men and women.

Methods

The study period was January 1997 to March 2010, and patients were eligible for inclusion if they were at least 30 years of age. We excluded patients with a prior history of CVD, and patients with type 1 diabetes or diabetes of uncertain type. We defined patients as having type 2 diabetes based on coded diagnoses recorded in CPRD or HES on or before study entry. We assessed glycaemic control by the mean HbA_1c from 3 years before to 3 years after study entry. The primary end point was the first record of one of the following 12 cardiovascular presentations in any of the data sources: stable angina, unstable angina, AMI, unheralded coronary death, heart failure, transient ischaemic attack, ischaemic stroke, subarachnoid haemorrhage, intracerebral haemorrhage, PAD, abdominal aortic aneurysm and a composite ‘arrhythmia or sudden cardiac death’ consisting of cardioversion, ventricular arrhythmia, implantable cardioverter defibrillator, cardiac arrest or SCD.

Results

The study cohort included 1,887,062 patients without diabetes and 34,198 patient with type 2 diabetes. The use of statins and antihypertensive medication was higher in patients with type 2 diabetes and increased over time. Patients with type 2 diabetes were approximately twice as likely to have black or South Asian ethnicity as those without diabetes (Table 5).

We analysed 113,638 non-fatal and fatal cardiovascular events over 11.6 million patient-years of follow-up (median follow-up 5.5 years). PAD and heart failure were the most common end points among patients with type 2 diabetes (16.2% and 14.1%, respectively) (Figure 17).

FIGURE 17.

Distribution of initial presentations of 12 CVDs in people with and without type 2 diabetes and no history of CVD. MI, myocardial infarction. Reproduced from Shah et al. 37 Copyright © 2015 Shah et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Published by Elsevier Ltd.

In multiply adjusted Cox models (Figure 18), there were strong positive associations between type 2 diabetes and PAD (HR 2.98, 95% CI 2.76 to 3.22), ischaemic stroke (HR 1.72, 95% CI 1.52 to 1.95), stable angina (HR 1.62, 95% CI 1.49 to 1.77), heart failure (HR 1.56, 95% CI 1.45 to 1.69) and non-fatal AMI (HR 1.54, 95% CI 1.42 to 1.67). In contrast, type 2 diabetes was inversely associated with abdominal aortic aneurysm (HR 0.46, 95% CI 0.35 to 0.59) and subarachnoid haemorrhage (HR 0.48, 95% CI 0.26 to 0.89). There was no statistically significant association with arrhythmia/SCD (HR 0.95, 95% CI 0.76 to 1.19).

FIGURE 18.

Association of type 2 diabetes with 12 CVDs in patients aged ≥ 30 years. MI, myocardial infarction. Reproduced from Shah et al. 37 Copyright © 2015 Shah et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Published by Elsevier Ltd.

The risk of CVDs associated with the presence of type 2 diabetes was highest for patients with a HbA_1c of > 58 mmol/mol or those in whom the HbA_1c had not been recorded (Figure 19). Patients with type 2 diabetes and a HbA_1c of < 48 mmol/mol had an increased risk of PAD (HR 1.82, 95% CI 1.52 to 2.18) and ischaemic stroke (HR 1.53, 95% CI 1.18 to 1.98), but no greater risk of any of the other CVDs.

FIGURE 19.

Association of type 2 diabetes with initial presentation of 12 CVDs by level of glycaemic control. MI, myocardial infarction. Reproduced from Shah et al. 37 Copyright © 2015 Shah et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Published by Elsevier Ltd.

Adjusted HRs for different initial presentations of CVDs associated with type 2 diabetes, adjusted for age, sex, BMI, level of deprivation, HDL cholesterol, total cholesterol, systolic blood pressure (SBP), smoking status and statin and antihypertensive drug prescriptions. Two non-specific components of the primary end point are not shown in the figure for simplicity (they are imprecisely recorded versions of one of the 12 initial presentations): coronary disease not further specified (HR 1.58, 95% CI 1.45 to 1.73; p < 0.0001); stroke not further specified (HR 1.64, 95% CI 1.48 to 1.81; p < 0.0001).

Discussion

This population-based, contemporary study of nearly 2 million individuals with a median follow-up of 5.5 years provides strong evidence that associations between type 2 diabetes and 12 specific CVD outcomes differ in terms of direction and magnitude of effects. The most frequent cardiovascular presentations were PAD and heart failure. For those CVD presentations associated with type 2 diabetes, the risk is greatest when glycaemic control is poor, as reflected by a HbA_1c of > 58 mmol/mol. Our findings support the use of clinically collected real-world data from EHRs in studying risk of different CVDs among people with type 2 diabetes, and to complement findings from smaller, less contemporary, investigator-led cohort studies, and suggest that intensifying glycaemic control in type 2 diabetes provides an important opportunity for reducing the risk of major cardiovascular complications.

A key finding of our study was that PAD was among the most common first presentations of CVD, which is consistent with the increase in prevalence between 2000 and 2010 found in a recent systematic review. 157 Furthermore, of all the 12 diseases studied, PAD showed the strongest association with type 2 diabetes, with an adjusted HR of 3, which is consistent with previous studies. 157 Heart failure was also among the most common first presentations, with possible mechanisms including prolonged hypertension, chronic hyperglycaemia, microvascular disease, glycosylation of myocardial proteins, diabetic nephropathy and autonomic neuropathy. 156

Conversely, type 2 diabetes appeared protective against abdominal aortic aneurysm and subarachnoid haemorrhage, consistent with previous reports from cross-sectional screening studies158–160 and case–control studies. 161,162 Potential mechanisms for diabetic protection against aneurysmal disease, which might have therapeutic implications, include diabetes-mediated changes in the vascular extracellular matrix, such as glycation of the extracellular matrix and covalent cross-linking between elastin and collagen in the aortic wall. 163

Our finding that levels of HbA_1c were associated with the risk of many CVDs strengthens the case for the causal relevance of blood glucose. It also provides support for management strategies directed at intensifying glycaemic control in type 2 diabetes, although targets should probably fall short of ‘normalising’ HbA_1c levels in light of the negative findings with respect to cardiovascular death and non-fatal stroke and AMI in the Action to Control Cardiovascular Risk in Diabetes (ACCORD) trial. 164

The differences between relative risks of different CVDs in patients with type 2 diabetes have implications for clinical risk assessment and trial design. Our findings point to the importance of considering chronic disease end points such as heart failure, PAD and stable angina in the primary end point because they are common, have a high morbidity burden and might have different treatment effects from AMI or stroke.

Conclusions

This study has shown heterogeneous associations between type 2 diabetes and 12 different CVDs, the most frequent presentations with PAD, heart failure and angina, reinforcing guideline recommendations for treating type 2 diabetes patients with statins. Guidelines also recommend angiotensin-converting enzyme inhibitors (ACEIs) to protect against microvascular complications but our data suggest that these drugs may also have a role in protecting against left ventricular remodelling and progressive heart failure. Our finding that the risk of many CVDs increases substantially with increasing HbA_1c concentrations identifies enhanced glycaemic control as a potentially important opportunity for improving outcomes in these patients.

Introduction

Hypertension is a major risk factor for all forms of major CVD events and the leading risk factor for overall recent global burden of disease. 165 A meta-analysis of 61 cohorts containing individual data for 1 million adults aged between 40 and 89 years recruited between 1950 and 1990 reported log-linear associations between SBP and diastolic blood pressure (DBP) and ischaemic heart disease and stroke mortality. 122 The study found no apparent threshold below which any further reduction in risk is observed, down to a BP of 115/75 mmHg. However, no previous study has compared the associations of BP across a wide range of incident CVDs, across a broad age range (among people aged < 40 years of age) and a wider range of BP values (including < 115/75 mmHg). In this novel contemporary study using the CALIBER linked data set, we set out to investigate the association between BP and 12 different presentations of CVD. Study findings could help to refine strategies for primary prevention and inform the design of future clinical trials.

Methods

Using CALIBER linked EHR data, we selected 1.25 million patients from 225 general practices in England registered between January 1997 and March 2010. Patients were eligible for inclusion if they were aged ≥ 30 years at study entry, had been registered with their practice for at least 1 year before the index date, had not been diagnosed with CVD and had at least one baseline BP measurement. Baseline BP, recorded during consultations in primary care, was based on mean SBP or DBP readings taken within 2 years of the index date. Where multiple BP measurements were available, an average reading was calculated. In sensitivity analyses, we used the BP measurement recorded closest to the index date. Hypertension was defined as baseline BP of ≥ 140/90 mmHg, a recorded diagnosis of hypertension, or evidence of repeat prescriptions (at least two monthly packs) for BP-lowering drugs, including thiazide diuretics, beta-blockers, ACEIs, angiotensin receptor blockers (ARBs) or calcium channel blockers. The most recent BP measurement recorded in the year before the index date was used for covariate adjustments. Missing covariate values were imputed using multiple imputation using the ‘mice’ algorithm in the statistical package R version 15.2 (The R Foundation for Statistical Computing, Vienna, Austria). Patients were censored on the date of first CVD presentation, death from other causes, deregistration from the practice or final practice data collection, whichever occurred first. Initial presentation of CVD was defined as 12 CVDs diagnosed in primary or secondary care, or death, and total CVD (all 12 CVDs combined). HRs were based on disease-specific Cox models with time since study entry as the timescale, adjusted for baseline age (linear and quadratic term) and stratified by sex and primary care practice.

Results

Among 1,258,006 patients included in the cohort, a total of 83,098 initial CVD presentations were recorded over a median of 5.2 years’ follow-up. SBP and DBP showed a range of associations with different CVD outcomes (Figure 20). A 20-mmHg rise in SBP showed marked associations with stable angina (HR 1.41, 95% CI 1.36 to 1.46), subarachnoid haemorrhage (HR 1.43, 95% CI 1.25 to 1.63) and intracerebral haemorrhage (HR 1.44, 95% CI 1.32 to 1.58). DBP showed substantially weaker associations with stable angina, PAD, AMI and with total CVD. The strongest associations of all CVDs were observed for abdominal aortic aneurysm: with DBP (HR 1.45, 95% CI 1.34 to 1.56) and mean arterial pressure (HR per 10 mmHg 1.61, 95% CI 1.48 to 1.75). It also showed the weakest association with SBP (HR 1.08, 95% CI 1.00 to 1.17) per 20 mmHg, and was the only CVD presentation where the association with higher pulse pressure was reversed (HR per 10 mmHg 0.91, 95% CI 0.86 to 0.98). AMI had a stronger association with SBP in women than in men (p < 0.0001), but associations with other CVD presentations were similar in both sexes.

FIGURE 20.

Forest plot of HRs and 95% CIs for 12 CVDs per 20/10-mmHg increase in SBP or DBP, adjusted for age and sex. SBP is represented by a black box, DBP is represented by a green box. Vertical dashed lines correspond to the associations of SBP (black) or DBP (green) with total CVD. Adjustments include age, quadratic age and stratification by sex and primary care practice. CIs are Bonferroni corrected (13 end points × 2 variables = 26 tests). MI, myocardial infarction. Reproduced from Rapsomaniki et al. 38 Copyright © 2014 Rapsomaniki et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Associations with both SBP and DBP decreased with age for all outcomes at varying rates for different CVD presentations (Figure 21).

FIGURE 21.

Forest plots of HRs and 95% CIs for 12 CVDs for 20/10-mmHg changes in BP in different age groups, adjusted for age and sex. Models included continuous age, age group, interaction between BP and age group (which is the association reported in the forest plot), and stratification by sex and primary care practice. CIs are Bonferroni corrected (3 age groups × 12 end points = 36 tests). MI, myocardial infarction. Reproduced from Rapsomaniki et al. 38 Copyright © 2014 Rapsomaniki et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

The shape of associations for SBP and DBP showed substantial variation across different CVDs and ages. Between the ages of 30 and 79 years, we observed linear associations between SBP and DBP with stable angina, AMI, unstable angina, unheralded coronary death and subarachnoid haemorrhage. Risk was lowest among those with a SBP of 90–114 mmHg and was significantly lower than among those with a higher SBP of 115–129 mmHg. Associations with other CVDs were mostly log-linear, with little or no reduction in risk for BP of < 130/85 mmHg. However, modest increases in either SBP or DBP of > 130/85 mmHg were associated with large increases in risk, especially for unheralded coronary death, heart failure, ischaemic stroke, intracerebral haemorrhage and PAD.

The observed association between SBP and subarachnoid haemorrhage or abdominal aortic aneurysm disappeared in people aged ≥ 60 years of age. Among people aged ≥ 80 years of age (men and women), SBP was strongly associated with stable angina, AMI, intracerebral haemorrhage and PAD.

Receipt of BP-lowering medication at baseline did not decrease the associations with SBP and DBP. Hypertensive patients (diagnosed as having a BP of ≥ 140/90 mmHg or being in receipt of BP-lowering drugs) developed CVD 5 years earlier than patients without hypertension (95% CI 4.8 to 5.2) and had a lifetime risk of overall CVD at 30 years of age of 63.3% (95% CI 62.9% to 63.8%) compared with non-hypertensive people (HR 46.1%, 95% CI 45.5% to 46.8%).

The mean number of CVD-free life-years lost associated with hypertension was 5.0 years (95% CI 4.8 to 5.4 years) from 30 years of age, 3.4 years (95% 3.3 to 3.6 years) from 60 years and 1.6 years (95% CI 1.5 to 1.7 years) from 80 years (Figure 22). The largest proportions of CVD-free years of life lost associated with hypertension at index 30 years of age were attributable to stable angina (22%), unstable angina (21%) and AMI (15%). However, from an index age of 80 years, stable angina and heart failure accounted for most (19% each) of all CVD-free years of life lost, followed by unstable angina (15%), AMI (12%) and ischaemic stroke (10%).

FIGURE 22.

Years of life lost to CVD associated with hypertension at index ages of 30, 60, and 80 years, adjusted for sex, smoking, diabetes, and total and HDL cholesterol. A total of 83,098 CVD events occurred. MI, myocardial infarction. Reproduced from Rapsomaniki et al. 38 Copyright © 2014 Rapsomaniki et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Discussion

In this contemporary study of 1.25 million adults, we show that, despite availability of modern preventative medication, the lifetime burden of hypertension remains considerable. Our findings challenge previous assumptions that BP has strong associations with incidence of all CVDs and across a wide age range, and that diastolic and systolic associations are concordant.

Our study has important advantages over previous research. First, we report associations of BP across 12 CVDs in a validated EHR linkage setting, including younger adults and a lower range of BP values than in other studies. Second, our findings are generalisable to the general population (99% of the UK population are registered with a primary care practice) and focus on measurements obtained as part of clinical practice, and, as such, are clinically relevant.

Our findings have important clinical implications. In this study, we show the high burden of CVDs such as stable and unstable angina, and heart failure in older age (≥ 80 years). The lifetime burden of BP in relation to different CVDs that we observed in this study can be used to inform decision-making and extend counselling treatments of patients, as emphasised in recent published national and international guidance. 24,166 Our findings also emphasise an unmet need, in the form of improved BP management and better management of other CVD risk factors as part of global risk estimation, to reduce risks of CVD.

Our findings can be used to help inform the design of new clinical trials and to aid interpretation of existing trials to assess new BP-lowering strategies. Although heart failure and PAD are observed to be common initial presentations of CVD, they are included less frequently in the primary outcome of BP-lowering trials. 167

Conclusions

Our novel, contemporary, high-resolution study showed that SBP and DBP show heterogeneous associations across a wide range of CVDs and at different ages. Our findings have implications for the design of new trials and preventative strategies to address the substantial contemporary unmitigated lifetime burden of hypertension.

Introduction

The value of risk factor modification for protecting against CVD has long been recognised and is written into all international guidelines. In the landmark INTERHEART study, just nine potentially modifiable risk factors were found to account for > 90% of the population-attributable risk of AMI in all regions of the world. 68 Among these risk factors, hypertension is particularly important because it is common, easily diagnosed and readily treatable.

In patients with diagnosed hypertension, guidelines set treatment targets to optimise risk reduction, with the recent report from the Eighth Joint National Committee (JNC 8) recommending < 140 mmHg for SBP (< 150 mmHg for patients aged ≥ 60 years without diabetes and chronic kidney disease) and < 90 mmHg for DBP. 168 Historically, however, there is evidence that guideline targets for BP reduction are not consistently achieved in clinical practice,169–171 although contemporary data for primary care populations are unavailable. A continuing failure to meet guideline targets would represent an important missed opportunity for protecting patients against CVD. We have used the CALIBER data set to analyse the scale of this missed opportunity in contemporary UK practice, with the primary objectives of describing (1) the proportion of hypertensive patients attaining BP control according to JNC 8 targets168 and (2) factors associated with attainment of BP targets.

Methods

We analysed a cohort of all individuals prospectively and newly identified with high BP among individuals aged ≥ 18 years, registered in the general practices contributing data to the CALIBER data set between January 1997 and March 2010 (the administrative censoring date of the CPRD database). Individuals had a minimum of 1 year of follow-up after practice registration, no prior history of CVD and a minimum of 6 months of follow-up since their high BP diagnosis. Individuals with drug-induced hypertension and those with pregnancy-induced hypertension were excluded (Figure 23).

FIGURE 23.

CALIBER patient inclusion flow diagram.

High BP was defined as the record of three or more systolic or diastolic high BP readings within a 1-year period, or two or more systolic or diastolic high BP readings within a 6-month period in CPRD. High BP measurements were defined as in the 2014 JNC 8 guidelines168 for patient management, that is, ≥ 140/90 mmHg for individuals aged ≤ 60 years and ≥ 150/90 mmHg for individuals aged ≥ 60 years. The date of diagnosis of high BP was the earliest date of the last BP reading that identified the individual as having high BP (e.g. the date of the third reading when defined as three or more systolic or diastolic high BP readings within 1 year). In sensitivity analyses, high BP was defined as the record of two or more systolic or diastolic high BP readings within a 1-year period.

Results

Of the 196,212 patients with newly diagnosed high BP, 156,587 (79.8%) attained BP control during follow-up (Table 6). The median yearly consultation rate was higher in patients who reached BP control than in those who did not (6.9 vs. 34 per year). BP-lowering medication was prescribed to 95,972 (48.9%) individuals during a median of 8.0 months after diagnosis, with the proportion receiving treatment being higher in those who did achieve BP control than in those who did not (56.3% vs. 19.9%). Lifestyle interventions were more frequently recorded for patients who achieved BP control.

Factors significantly associated with BP control (Table 7) in the fully adjusted model included female sex [odds ratio (OR) 2.34, 95% CI 2.25 to 2.43], BP-lowering treatment (OR 4.90, 95% CI 4.48 to 5.35), statin medication (OR 1.78, 95% CI 1.58 to 2.003) and diabetes (OR 2.53, 95% CI 2.32 to 2.77). Patients diagnosed earlier during the study period were also more likely to achieve BP control than those diagnosed later. BP control was less likely to be achieved in patients with more severe grades of hypertension (grade 2: OR 0.79, 95% CI 0.75 to 0.82; grade 3: OR 0.59, 95% CI 0.54 to 0.63, compared with grade 1), those with abnormal weight (lower or higher than normal) and those who did not receive thiazides as BP-lowering medication.

Discussion

In this population-based CALIBER cohort of patients presenting with newly diagnosed hypertension, BP control according to JNC 8168 targets was achieved in about 20% of patients. Patients in whom BP control was achieved were more likely to have grade 2 and 3 than grade 1 hypertension, to visit their doctor more often and to receive BP-lowering treatment and statins. The longer the delay between diagnosis of hypertension and the prescription of treatment, the less likely patients were to achieve control.

Importantly, despite the availability of a range of effective antihypertensive drug classes, the BP targets set by the JNC 8168 were not achieved in one in five hypertensive patients. The failure to optimally reduce cardiovascular risk in such a large proportion of patients must contribute significantly to the CVD burden, given its direct association with BP across a wide spectrum of measurements. 122,172 The relative underuse of antihypertensive drugs and delays between diagnosis and initiation of treatment are making a major contribution to this missed opportunity to reduce cardiovascular risk. Nevertheless, our data show that other factors, beyond utilisation of antihypertensive drugs, are also important, including higher rates of obesity and smoking, and lower rates of prescription of secondary prevention drugs, particularly statins, among patients who did not achieve BP control.

Conclusions

This study confirms high levels of undertreatment among hypertensive patients and shows that the missed opportunity to reduce cardiovascular risk includes a range of lifestyle factors that are all potentially amenable to modification. However, successful treatment of hypertension requires vigilant clinical supervision, and it was notable that consultation rates were significantly lower among patients whose BP was not treated to target (although we cannot be sure that frequent consultation to monitor BP and other cardiovascular risk factors would lead to more effective risk reduction). If the 20% of hypertensive patients who fail to achieve JNC 8168 targets were treated more vigorously in terms of BP reduction and lifestyle modification, the dividend for cardiovascular morbidity and mortality would be substantial.

Background

Delayed diagnosis in patients with symptoms suggestive of ischaemic heart disease is a common missed opportunity for treatment.

Objectives

(1) To determine the incidence of chest pain according to diagnostic type and the extent and nature of AMI heralding; and (2) to develop automated methods for identifying relevant texts and extracting coded information from unstructured free text in EHRs.

Methods

An analysis of CALIBER linked data. In a subset of CALIBER patients we developed and tested two free-text data-mining tools: the freetext matching algorithm (FMA) and a novel machine-learning algorithm [the ‘Semi-supervised Set Covering Machine’ (S³CM)].

Results

(1) Among 172,181 patients with a first recorded label of chest pain, we identified high levels (72%) of failure to attribute a cause for chest pain at first presentation, and many members of this group went on to experience cardiovascular events. (2) Among patients experiencing a first AMI, only 14% of STEMI events and 9% of NSTEMI events were unheralded. (3) FMA achieved high precision and recall. (4) The S³CM algorithm outperformed comparator programs on data-mining tasks.

Conclusions

(1) Failure to attribute a cause to chest pain is leading to missed opportunities for primary or secondary prevention of disease. Most AMIs are heralded by prior atherosclerotic disease or cardiovascular risk factors. (2) Further development of natural language processing algorithms such as FMA and S³CM may improve the usability of EHR for research on CVDs and symptoms.

The CALIBER programme provides a unique opportunity to explore the events leading up to an AMI, as it contains data from primary care (risk factors and consultations prior to the event) and secondary care (hospital admission and disease registry data). The primary care data in CALIBER includes access to limited samples of unstructured free text entered by GPs, and we have developed tools to make the free text a useful data source for research. However, detailed information about patients’ hospital admissions are not currently collected at scale; future work on hospital informatics should aim to address this.

The studies in this chapter relate to patients presenting for the first time with stable chest pain (Figure 24).

FIGURE 24.

Overview of the 33 studies in our programme: initial presentation with stable chest pain.

Introduction

Cardiac diagnosis that is delayed after first symptomatic presentation or not made until the advent of a major complication represents a clear missed opportunity for treatment and occurs not uncommonly in the patient presenting with cardiac chest pain. Thus, in a cohort analysis of patients attending a RACPC we have previously reported that about one-third of major events (non-fatal AMI or coronary death) during follow-up occurred in patients reassured with a non-cardiac diagnosis. 74 We have now undertaken a more detailed analysis using the CALIBER data set, with the aim of determining:

• the annual distribution of chest pain presentations in primary and secondary care according to diagnostic type (unattributed to a cause, attributed to non-coronary cause, recorded as angina)

• the common short-term (6-month) trajectories of presented chest pain in primary and secondary care

• the annual incidence of cardiovascular events in patients with a first chest pain presentation.

Methods

We included only patients with a first recorded label of chest pain (cause unattributed or attributed to non-coronary cause) or angina code or at least two anti-anginal drugs in primary or secondary care between 2002 and 2009. This first recorded label was defined as the index presentation. No patient who had a prior record of chest pain or of any prior cardiovascular event was included. Follow-up was for a median of 3.3 years [interquartile range (IQR) 1.6–5.5 years].

Results

Clinical characteristics of the 172,181 patients meeting inclusion criteria are shown in Table 8. There was a preponderance of women, who constituted 54% of the group, with no significant change across index years. Only 5% of the group received a diagnosis of angina, with the majority of chest pain presentations being ‘unattributed’. There was a significant difference in the distribution of chest pain categories across index years, rates of unattributed and non-coronary chest pain increasing from 2002 to 2009, while rates of diagnosed angina diminished (p < 0.001) (Figure 25). Deprivation quintile had no apparent effect on index chest pain presentations in contrast to age, the proportion of unattributed and non-coronary presentations tending to diminish as angina presentations increased with advancing age.

FIGURE 25.

Type of chest pain diagnosis by year among CALIBER patients with a first recorded label of chest pain.

Figure 26a and b shows follow-up consultations for chest pain by month after index diagnoses of unattributed and non-cardiac chest pain, respectively. Less than 25% of patients in either diagnostic group attended for follow-up consultations during the first 6 months, most commonly in the first month. For the majority of these patients, the consultation resulted in no modification of the index diagnosis, with < 1% per month of all patients in either diagnostic group converting to a diagnosis of angina.

FIGURE 26.

Number of follow-up consultations for chest pain by diagnosis and by month after index diagnosis of (a) unattributed chest pain and (b) non-coronary chest pain. Data are proportions (%) of patients attending for consultation stratified by diagnosis.

The incidence of cardiovascular events during follow-up for patients with index diagnoses of non-coronary chest pain and unattributed chest pain was < 10% and < 20%, respectively. Patients with an index diagnosis of angina were considerably older and 62% experienced cardiovascular events during follow-up (Table 9).

Discussion

The annual incidence of recorded chest pain from 2002 to 2009 has shown progressive reductions in rates of diagnosed angina as rates of unattributed and non-coronary chest pain have increased. Our data emphasise how trajectories of care and cardiovascular outcomes are largely predetermined by the initial diagnostic categorisation of chest pain. This initial categorisation was generally stable and in only a very small proportion of patients was it modified during follow-up consultations in the first 6 months. Although the data were generally reassuring with regard to the prognostic validity of the first recorded chest pain category, substantial numbers of patients with unattributed chest pain went on to experience cardiovascular events, including AMI and coronary death. In a significant proportion of these cases it must be assumed that failure to diagnose angina at first presentation resulted in an inappropriate trajectory of care, representing an important missed opportunity to provide treatment and lifestyle advice to protect against the progression of CVD.

Failure to attribute a cause for chest pain at first presentation occurred in > 70% of cases and it was in this group that prognostic consequences were particularly telling, with an 8% incidence of cardiovascular events during follow-up, accounting for more than four times the absolute number of events compared with the angina group. This has important clinical implications, indicating a need for contemporary prognostic models based on the characteristics of patients presenting with unspecified chest pain so that those at risk of cardiovascular events can be more accurately identified.

Conclusions

Only 5% of patients assessed for chest pain receive a diagnosis of angina and this proportion has been falling over the past decade. One in 12 of the > 70% of patients diagnosed with unattributed chest pain went to experience a cardiovascular event; indeed, more events occurred in this group than in the group diagnosed with angina. A failure to identify patients with chest pain of cardiac origin at first presentation is an important missed opportunity for initiating preventative treatment and reducing the incidence of CVD.

Introduction

In the weeks or months prior to AMI, heralding by major risk factors, established atherosclerotic disease or prodromal symptoms provide an important opportunity to intervene with preventative strategies. Studies that have retrospectively evaluated medical history suggest that risk factors and prior atherosclerotic disease are common in people experiencing AMI, particularly NSTEMI, but they may underestimate the burden of disease and may poorly reflect the timing of prodromal symptoms. 175–179 CALIBER’s linked registries permit, for the first time, an evaluation of the extent and nature of STEMI and NSTEMI heralding using prospectively collected information on the onset of atherosclerotic disease (in coronary, cerebral and peripheral circulations), cardiovascular risk factors and prodromal chest pain.

Methods

We included patients aged ≥ 18 years experiencing a first AMI between 1 January 2003 and 31 December 2008, with at least 1 year of prior observation in CPRD and at least one consultation during that period. Atherosclerotic disease, risk factors or chest pain recorded in MINAP or CPRD at any time prior to AMI were defined as ‘heralding’ the AMI. The earliest record of atherosclerotic disease in CPRD was taken as the date of onset (dates were not recorded in MINAP). Risk factors investigated were smoking, hypertension, dyslipidaemia and diabetes (diagnosed diabetes or insulin prescription) and were defined by codes in CPRD or MINAP. We also determined whether or not patients had been prescribed BP-lowering, lipid-lowering or antiplatelet medications in the 6 months before AMI. The frequency of primary care consultations for chest pain was assessed in patients without diagnosed atherosclerotic disease.

Results

There were 8174 eligible patients with a first AMI. Figure 27 shows that only 14% (95% CI 13% to 16%) of STEMIs and 9% (95% CI 9% to 10%) of NSTEMIs were unheralded by prior atherosclerotic disease, one or more cardiovascular risk factors or chest pain, leaving 14% (95% CI 13% to 16%) unheralded by these factors.

FIGURE 27.

Percentage of patients with previous atherosclerotic disease and risk factors in patients experiencing first STEMI (n = 3780) and NSTEMI (n = 4394). Reproduced from Herrett et al. 41 Copyright © The European Society of Cardiology 2013. Published by SAGE. This article is distributed under the terms of the Creative Commons Attribution License (www.creativecommons.org/licenses/by-nc/3.0/), which permits non-commercial reuse, distribution, and reproduction in any medium, provided the original work is attributed as specified on the SAGE and Open Access page (www.uk.sagepub.com/aboutus/openaccess.htm).

Some 41% of patients had previously diagnosed atherosclerotic disease (Table 10), more of whom had NSTEMI than STEMI (50% vs. 29%). Coronary disease was the most common presentation in the year before AMI, recorded in 21% of STEMI and 41% of NSTEMI patients, with rates rising rapidly in the 3 months before AMI (Figure 28a). A similar pattern was observed for chest pain consultations in patients without diagnosed atherosclerotic disease (Figure 28b). Among patients with previously diagnosed atherosclerotic disease, 87% were being prescribed one or more of aspirin, statins and BP-lowering treatment in the 6 months before AMI, but only 34% were receiving all three forms of medication. Some 59% of AMIs were unheralded by previously diagnosed atherosclerotic disease; 79% of these patients had one or more risk factors (a record of diabetes, hypertension, dyslipidaemia, current smoking, or a prescription for statins, BP-lowering medication or antiplatelet medication in the 6 months before AMI). This was the same in STEMI (79%) and NSTEMI (80%) patients. AMI was unheralded by all the factors discussed in 546 (14%) STEMI patients and 413 (9%) NSTEMI patients. These patients were more likely to be younger [median age 67 years (IQR 58–77 years ) vs. 71 years (IQR 60–80 years)], male (72% vs. 63%) and to have a lower rate of primary care consultation prior to AMI (median four vs. seven consultations per year) than those with heralded infarction.

FIGURE 28.

Rates of (a) coronary diagnosis and (b) chest pain consultations in the months leading to first STEMI and NSTEMI, with 95% CIs among CALIBER patients. Each time point covers a 1-month time band (0–1 months, 1–2 months, etc.). Part (b) represents chest pain consultations just in patients without previously diagnosed atherosclerotic disease; although these patients have not received a coronary disease diagnosis, they may well be heralded by possible coronary symptoms. Reproduced from Herrett et al. 41 Copyright © The European Society of Cardiology 2013. Published by SAGE. This article is distributed under the terms of the Creative Commons Attribution License (www.creativecommons.org/licenses/by-nc/3.0/), which permits non-commercial reuse, distribution, and reproduction in any medium, provided the original work is attributed as specified on the SAGE and Open Access page (www.uk.sagepub.com/aboutus/openaccess.htm).

Discussion

This study shows how important opportunities for the primary or secondary prevention of arterial disease are either being missed or failing to protect against AMI. In patients with previously diagnosed atherosclerotic disease or risk factors, AMI represents the unmet potential of secondary or primary prevention, respectively. In the majority of cases, prior atherosclerotic disease was diagnosed long in advance of both STEMI and NSTEMI, providing an extended period during which secondary prevention could be implemented. Our data describing the use of secondary prevention measures in the 6 months before AMI showed that most patients with diagnosed atherosclerotic disease were receiving one of either statins, aspirin or BP-lowering drugs, but only one-third were receiving all three, indicating that there are likely to be missed opportunities for secondary prevention in this group.

Coronary disease was the most common pre-AMI presentation, but 10% of both STEMIs and NSTEMIs were heralded by PAD and/or cerebrovascular disease alone. This emphasises the importance of further efforts to improve secondary prevention following diagnoses in the cerebral and peripheral arteries in order to prevent an important proportion of AMI. The high prevalence of risk factors in both STEMI and NSTEMI highlights the importance of tackling the widely reported missed opportunities for the implementation of existing interventions that are known to be effective.

In addition, the categorisation of continuous measures in this analysis may have been an oversimplification of cardiovascular risk. Although a binary indicator is simple to interpret in studies and a useful basis on which to prescribe treatment, it does not reflect the continuum of risk over the full range of measurements. A more detailed investigation of these risk factors might reveal borderline raised risk in many patients, and lowering BP and lipids in those not diagnosed as hypertensive or dyslipidaemic may also prevent AMI. However, the implications of our analysis are limited by a lack of comparison to AMI-free controls. Such a comparison may allow further conclusions to be drawn from these data.

Clinical experience and retrospective studies have long suggested that AMI might be preceded by premonitory symptoms of chest pain presenting to a GP or ambulatory care. 180,181 Our study extends knowledge in several respects. First, we confirmed this association with prospective data. Second, we found that there were increases in coronary disease diagnoses and chest pain consultations in both STEMI and NSTEMI. This is in contrast to the widely held view that STEMI is usually of sudden onset. Third, we showed that the increases were specific to coronary diagnoses and chest pain, rather than to disease in cerebral or peripheral circulations, suggesting a local rather than systemic pro-thrombotic state.

Further research is warranted to better characterise the phenotypes, causes and prognoses of unheralded STEMI and NSTEMIs to improve the usability of EHRs for research. Analyses comparing AMI patients with a control group without AMI, including a comparison of missed opportunities for care in measuring and controlling elevated risk, are needed.

Conclusions

New-onset chest pain, even in the absence of major risk factors or prior atherosclerotic disease, may identify patients at risk of imminent AMI. Interestingly, a further CALIBER analysis has shown that chest pain and other ischaemic presentations in the 7 days preceding AMI reduces the risk of death in the first week by more than one-third, perhaps owing to a natural preconditioning effect. 182 Nevertheless, the opportunity for careful triage of these patients, in accordance with the recommendations of recent guidelines,183 should not be missed to ensure that high-risk individuals can be identified and treated to prevent progression to AMI.

Introduction

Electronic health records research to date has predominantly used coded data, which are readily available for analysis, but free-text fields may contain important additional information. 43,184–186 Manual review of free-text records is time-consuming, so there has been interest in developing software algorithms to extract diagnoses and other clinical information from free text. Development of such software algorithms is challenging, however, as clinical text contains complex language structures and specialist terminology, as well as context-specific abbreviations. Some algorithms have been developed to extract information about specific subjects, such as smoking status, heart failure or family history. 187–191 Other programs have wider functionality but have tended not to be widely used outside the laboratories that developed them. 190 The linked CPRD record in CALIBER includes coded information but also free-text fields containing information recorded by GPs or derived from hospital letters. One potentially valuable use of free-text information is in investigations of cause of death, which are often recorded in free-text fields. This study describes the development of a software tool for extracting information about cause of death, as well as a selection of other clinical information, from free-text fields in CPRD.

Methods

The FMA was developed to extract diagnoses, dates, durations, laboratory results and heart rate and blood measures from free-text fields. Using lookup tables and synonyms, the program aims to match samples of free text to the standard Read codes used by GPs (Figure 29) (www.connectingforhealth.nhs.uk/systemsandservices/data/uktc/readcodes). The EHRs of a random sample of 3310 patients who died in 2001 were manually analysed to identify causes of death. The initial version of the program was developed to extract causes of death from these data, and iterations of the program were tested on random samples of several hundred free-text entries. The prototype program was developed in Visual Basic with a Microsoft Access^® 2000 (Microsoft Corporation, Redmond, WA, USA) user interface to facilitate reviewing. Following this development phase, FMA was tested on two samples of free text, from the CPRD (1000 texts associated with death in 2001) and 1000 general texts from a CAD study. Results were compared with those generated by the US Library of Medicine’s MetaMap program and manual review (the gold standard). 192

FIGURE 29.

Flow chart showing the stages of matching samples of text used by GPs in primary care using the FMA program and analysis of text. OXMIS, Oxford Medical Information System. Reproduced from Shah et al. 42 © 2012 Shah et al. ; licensee BioMed Central Ltd. This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Results

When tested on the 1000 texts associated with death, FMA successfully matched 683 out of 735 positive diagnoses, achieving a precision (PPV) of 98.4% (95% CI 97.2% to 99.2%) and a recall (sensitivity) of 92.9% (95% CI 90.8% to 94.7%). On the general sample, FMA successfully identified 346 of the 447 positive diagnoses, achieving a precision of 91.5% (95% CI 88.3% to 94.1%) and a recall of 77.4% (95% CI 73.2% to 81.2%). Its performance was similar to that of MetaMap (Table 11).

Discussion

Through the analysis of free text in CPRD, FMA was able to correctly assign Read terms of death in 93% of cases, achieving a precision of > 90% on various types of free text. FMA also proved capable of extracting information about test results such as pulse rate. FMA may provide a valuable tool for extracting information from EHRs for research, reducing the need for manual review and overcoming confidentiality concerns over access to free text for researchers. This could potentially generate important additional information; in the 2001 CPRD data, for example, almost 20% of causes of death are recorded in free-text rather than coded form. FMA can also extract useful supporting information such as timing and suddenness of death.

The code of FMA is available in an open-access repository, under the GNU General Public Licence (see https://github.com/anoopshah/freetext-matching-algorithm) to enable use and modification of the algorithm by other researchers and to facilitate further collaborative development.

Machine learning and natural language processing approaches to examining the large corpus of unstructured (free-text) data that exist in primary care and hospital records should be developed in order to identify opportunities for earlier diagnosis and to add phenotypic information and safety signals.

Finally, as well as research on EHRs, real-time natural language processing systems could support enhanced coding in clinical practice. Such systems are recommended in the Connecting for Health NHS Common User Interface guidance,193 to enable clinicians to enter information in a natural form that is automatically converted into the appropriate codes.

Conclusions

The FMA is an open source, free-text data-mining tool, which showed high precision and recall on a test data set from the CPRD and can be used to extract uncoded diagnosis or symptom information for EHR research.

Introduction

In developing our FMA we were able to match samples of free text to the standard Read codes used by GPs, thereby coding the samples to make them available for research. The analysis of clinical text is challenging, however, because of its specialist terminology, potentially complex language structures and use of context-specific abbreviations. In this study we adopted a different approach termed ‘machine learning’, in which we do encode detailed language rules by hand, but allow a learning computer algorithm to derive its own rules based on some training data. 187,194,195 In some studies, tools that incorporate machine learning components have been shown to perform better than purely rule-based or pattern-matching approaches. 196–198

Here, we describe the development of a machine learning algorithm (S³CM) related to previous models developed by Rosales et al. 199 to determine whether or not clinical free text contains specific information of interest. Its ability to detect coronary angiogram test results and a diagnosis of ovarian cancer was compared with a fully supervised predecessor system,200 the transductive support vector machine (TSVM),201 and FMA. 42 A potential advantage of S³CM was that it did not require a perfectly labelled training set, which means that the existing Read codes associated with each free text could be used as the label, rather than requiring time-consuming manual labelling of the training data set.

Methods

We tested S³CM on free-text records in CPRD that might contain coronary angiogram results. Just 4.2% of CPRD records have angiogram results coded, but a much larger number have codes indicating that an angiogram was performed. The results of these ‘missing’ angiograms may be available in free-text fields. A second task was to detect a suspected or definite diagnosis of ovarian cancer. This cancer is difficult to diagnose, and clinicians may note down suspicions before a formal diagnosis is recorded. An analysis of these earlier thoughts would provide useful insight into clinical decision-making processes.

The algorithms were trained using texts with Read codes relevant to the two tasks. The performance of S³CM and the other tools was compared with the gold standard of manual review by a specialist clinician.

Results

In the training set for the coronary angiogram task, only 60% of records with an angiogram result Read code actually had angiogram results in the free-text field. Nevertheless, presented with unlabelled texts, S³CM achieved 87% recall with 64% precision. It outperformed both the fully supervised SCM (60% precision, 78% recall) and the TSVM (3% precision, 2% recall). Similarly, in the ovarian cancer diagnosis task, S³CM outperformed both SCM and TSVM, detecting 303 out of 353 diagnoses (86% recall, 74% precision). FMA achieved greater precision (85%) but lower recall (62%). In a test set, S³CM identified 99% of patients with ovarian cancer, even though only 82% had a formally coded diagnosis.

Discussion

The S³CM algorithm was developed as a semisupervised learning algorithm for the classification of clinical test records. It has shown promising performance in two unrelated pilot tasks, even though the training data set for the first of these tests was far from perfect. One advantage of S³CM is that training examples can be obtained simply through the use of diagnostic codes rather than by manual review. Moreover, because it does not depend on a subject-specific knowledge base or set of linguistic rules, it is, in theory, applicable to other subject areas (and languages). Its main disadvantage is that it will not detect particularly complex or rare forms of language. Misspellings and abbreviations may also present issues.

Further testing of our semisupervised machine-learning tool should be conducted using other EHR data sets, such as discharge letters and electronic hospital notes.

Conclusions

The tool is likely to be of most value in applications in which diagnostic or other information is typically recorded in free-text form. It is semi-automatic and easy to run, and may have most use as a way of winnowing large data sets for more detailed review. It would also be suitable for use on other sources of data, such as electronic hospital notes, and is sufficiently flexible to be used across multiple disease areas, requiring only a small number of labelled data for training.

Background

Evidence from EHR- and investigator-led studies suggests that angina is not being adequately diagnosed in RACPCs and potential cardiac-related chest pain may be being inappropriately investigated. CDSSs provide an approach for guiding appropriate investigation and initial treatment decisions, but their integration into hospital information systems and usual practice remains challenging.

Objectives

To assess the feasibility, use and outcomes of a web-based CDSS [Optimising the Management of Angina (OMA)] in RACPCs.

Methods

Patients from three UK NHS RACPCs were recruited before and after the introduction of OMA. We carried out an ethnographic study of OMA use.

Results

Integration of the CDSS with hospital information systems and the clinical workflow was hindered by multiple technical and procedural problems. OMA was used in 84% of consultations. There were no differences in medication prescribing or investigation within 6 months between groups using OMA and delivering usual care. Qualitative data suggested that clinicians found OMA useful in confirming rather than guiding their decision-making, but it had no impact on the assessment of the most difficult-to-interpret cases. Difficulties implementing the CDSS and a strong belief in the primacy of clinicians’ judgements were associated with OMA’s lack of influence.

Conclusions

The OMA programme had a negligible effect on the decision-making behaviour of clinicians in RACPCs, and a trial powered on cardiac events was deemed unwarranted. Insight into the ways in which clinicians interacted with OMA could inform development of new tools to aid clinical decision-making with a greater likelihood of being adopted in practice.

The studies in this chapter relate to patients presenting for the first time with stable chest pain (see Figure 24).

Introduction

Rapid access chest pain clinics may incorrectly diagnose some patients, and not all of those patients diagnosed with angina receive appropriate secondary prevention measures. For example, in a cohort of 8762 patients followed for 3 years after attending a chest pain clinic, 33.1% of all cardiovascular events occurred in patients with a diagnosis of non-cardiac chest pain. 74 This study also reported underprescription of evidence-based treatments for angina: of those patients receiving a diagnosis of angina, only 28% were taking a statin. The need for better decision-making for patients presenting with new-onset stable chest pain was highlighted in the 2010 NICE chest pain guidelines,202 which recommended investigation decisions based on the pre-test probability of coronary disease.

The OMA CDSS is a web-based computerised version of the NICE Clinical Guideline (CG)95,202 with additional guidance provided on investigation and prescribing decisions derived from a RAND (RAND Corporation, Santa Monica, CA, USA) Delphi panel of 11 experts in the fields of general cardiology, imaging, radiology, general practice, chest pain clinics, clinical pharmacology and biomarkers/inflammation. The panel was convened in December 2009, and panel members were asked to rate the appropriateness of investigation and risk-lowering medication in patients with suspected stable angina pectoris attending RACPCs. Members rated the appropriateness (inappropriate, uncertain or appropriate) of 2400 hypothetical patient-specific indications for exercise ECG, coronary computerised tomography (CT), functional imaging and invasive coronary angiography and initiation of the risk-lowering medications aspirin, beta-blocker, ACEI and statin. This involves eliciting decision thresholds using pre-test probability (diagnostic) and 10-year risk (prognostic) scores and categorical clinical information.

Optimising the Management of Angina clinicians entered the following data into the CDSS: age, sex, typicality of chest pain, resting ECG result and cardiovascular risk factors (smoking status, diabetes and hyperlipidaemia). These data are used to calculate the individual’s pre-test probability for coronary heart disease and to give investigation recommendations in line with the NICE guideline.

The OMA programme (Figure 30) is a complex intervention delivered at the clinic level. The program has three stages: preparation, training and clinic tools. The aim of the preparation and training stages is to facilitate the use of the clinic tools and, in particular, the CDSS. The components of the intervention were developed with attention to the behavioural change domains identified by Michie et al. 203

FIGURE 30.

Summary of the OMA intervention. Source: reproduced from Johnson et al. 45 © Johnson et al. 2015. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Clinical decision support systems can be effective at modifying clinician behaviour,204 with four factors predicting improved clinical practice. We planned the design of OMA to incorporate all of these components, that is, provision of decision support at the time and place of decision-making, provision of recommendations rather than just assessments, computer-based decision support and real-time provision of decision support to clinicians during the consultation. However, less is known about the effect of CDSSs on patient outcomes, and what evidence does exist is conflicting. A systematic review205 of studies evaluating the effectiveness of computer-based guideline implementation systems providing real-time interaction during consultations in ambulatory care settings identified a lack of evidence of an effect of these systems on patient outcomes and conflicting evidence of effect on process outcomes.

A meta-regression of 162 randomised trials of CDSSs206 identified several factors that could explain why some systems succeed in terms of clinician behaviour change and others fail. Systems were more likely to be successful if they required practitioners to provide reasons when over-riding advice and if they provided advice concurrently to patients and practitioners. A linked editorial207 argued for an approach including qualitative contextual evaluation and observation to further our understanding of the factors that make CDSSs successful.

Implementation of CDSSs is a complex intervention. The Medical Research Council (MRC) framework208 for the development of complex interventions promotes a focus on feasibility and pilot work, using a mixture of qualitative and quantitative methods where appropriate, so that full evaluations, preferably in RCTs, are conducted only on interventions that are feasible and acceptable, and that have a likelihood of improving outcomes.

Methods

In this study we used a mixed-methods approach to answer the following research questions:

1. Is delivery and use of the OMA CDSS feasible in chest pain clinics?

2. In what proportion of consultations is the OMA CDSS used?

3. Are the OMA CDSS recommendations for investigation and medication followed?

4. Does the investigation and medication treatment of participants differ before and after delivery of the OMA programme, and is this likely to impact on outcomes?

Quantitative methods were used to answer questions 2–4. Qualitative ethnographic methods were used to answer question 1, and to help explain the quantitative results.

Results

Quantitative study

We recruited 294 patients to the quantitative study between November 2009 and May 2011 (Figure 31), 291 of whom were followed up for 6 months. The web-based CDSS was used in the chest pain clinic in two of the pilot sites, clinics A and C.

FIGURE 31.

Flow diagram showing recruitment of patients attending RACPCs to the OMA programme. a, Trust 1, Clinic A Before OMA and Trust 2, Clinic B; b, Trust 1, Clinic A After OMA and Trust 2, Clinic C. Reproduced from Johnson et al. 45 © Johnson et al. 2015. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Discussion

Clinical decision support systems are often considered one of the most important rationales for developing flexible and extendable hospital EHR systems. The long-term viability and sustainability of the CDSS requires modular access to the underlying EHR system and ongoing, bidirectional data flows between research and care components. In this study, we attempted to develop, implement and evaluate a CDSS in a clinical setting lacking a modular EHR system that can foster CDSS integration – a situation that is still widespread within both primary care and secondary physician information systems in the NHS. Our initial technical scoping exercise revealed significant variation both within and across hospitals, which rendered the option of creating an integrated CDSS unfeasible. As a result, we built and deployed a system that was accessible through the internet and required minimal interaction with the underlying hospital information systems.

Through our mixed-methods study, we have developed an in-depth understanding of the feasibility and plausibility of the proposed OMA CDSS, and compared process outcomes in patients managed in clinics before and after delivery of the OMA programme. The OMA CDSS, delivered within the OMA programme, has been shown not to be feasible in its present form and, in line with MRC guidance, these findings do not support a trial of the OMA programme.

Although the CDSS was used in 84% of consultations for which it was available, it had little impact on medication prescribing at 6 months and first investigation within 6 months, and we have, therefore, not carried out an expected value of information analysis. Of those patients recommended an investigation by the CDSS, 50% of patients before OMA and 59% of patients after OMA did not receive any investigation, and it is possible that this potential underinvestigation could lead to incorrect diagnoses of non-cardiac chest pain. The range of investigations requested by clinicians was more limited than recommended by the CDSS, with invasive angiography being the most commonly undertaken investigation.

The qualitative analysis describes how the CDSS was used in practice, highlighting structural and practical difficulties, issues with data input and how the data are processed by the CDSS problems with matching of all clinically available information to the CDSS algorithm, and difficulties with implementation of the CDSS recommendations.

The strengths of this study lie in its mixed-methods approach. The quantitative study used both hospital and primary care records to allow follow-up of patients at 6 months, and high rates of follow-up were achieved (99% at 6 months). The qualitative study was extensive and was conducted by experienced ethnographers who were familiar with the chest pain clinic setting. Although the quantitative data suggested that the CDSS was used, the qualitative data were essential in developing a more complete understanding of how it was used and provided multiple examples of difficulties in practice. Linking qualitative data to clinical and diagnostic information helped to make sense of the ways in which the CDSS was used and recommendations followed. The ethnographic method uses several sources of data (interviews, observations, focus groups), allowing comparisons to be made between what clinicians say they do and what they actually do.

The qualitative data collection was limited by the inclusion of observations after the delivery of the OMA programme in only one site, although we were able to interview one clinician from the second site at which the OMA programme was delivered. Each of the researchers involved in collection of qualitative data was involved in the development of the OMA programme.

For patients for whom the CDSS was not available or, when available, was not used during the clinical consultation, the assessment of cardiovascular risk factors was retrospectively made using routinely collected data for which there was more than one source (GP referral and hospital record). Data were collected from the GP referral only when it was not available in the hospital clinic record. We found disagreement between the sources for the presence of hyperlipidaemia and for smoking status, which may affect the comparison of the groups at baseline.

For the purposes of populating the CDSS for this group of patients, the clinic data were used to determine the typicality of chest pain. The complexity of categorising chest pain based on a patient’s description has been described before. 211 We found labelling chest pain typicality to be difficult, which resonated with the difficulties the chest pain clinicians described. Given that typicality is central to the pre-test probability of coronary disease and, hence, to CDSS recommendations, the difficulties experienced by both the research team and chest pain clinicians are important.

The OMA CDSS is a component of the OMA programme, which includes preparation and education stages. This study focuses on the CDSS, but clearly the other components of the OMA programme may have an important impact on the way in which the CDSS is received and used.

We have measured medication prescription in general practice 6 months after consultation, as a proxy for the use of secondary prevention medication. Clearly, there are many factors affecting whether or not medication is continued at 6 months. For example, medications may be prescribed for new indications or may be stopped following investigations that rule out coronary disease.

Tools to aid decision-making are introduced into the clinical workplace, a dynamic and complex system in which technologies, people and organisational routines dynamically interact. The OMA CDSS was introduced at a time of transition, NICE CG95202 having been published 4 months earlier. At this time, the ETT had a central role in chest pain clinic routines, and our main qualitative study site did not routinely use CT or CT coronary angiography for the diagnosis of angina. Although our original plan had been to develop a CDSS with both medication and investigation guidance taken from our expert panel, and allowing the incorporation of exercise test findings, given the primacy of NICE as a source of national guidance, it seemed prudent to incorporate NICE guidance. However, the absence of a test that was an entrenched part of chest pain clinic practice added to the difficulties of the clinicians in using the CDSS. Chest pain clinics are unlikely to be able to incorporate guidance immediately, and it is possible that the CDSS would have had more impact on clinic processes when clinics were more able to follow guidance.

Although largely based on national guidance, the CDSS did not ‘fit’ with established clinical practice in several ways. Gaining entry to the guideline requires clear categorisation of a patient, in this case through entering clinical data, which results in an estimate of a patient’s pre-test probability of significant CAD. This categorisation is dependent on multiple inputs, including typicality of chest pain and risk factors. The CDSS requires data entry in a dichotomised way, in contrast to the contextualised and nuanced information available to the clinician. Use of the CDSS could be viewed as a series of interactions during which the clinician first decides how to fit clinical inputs to the CDSS and subsequently considers how to act on the CDSS outputs, given the clinical context and structural factors, such as the availability of particular investigations. Each of these interactions requires the clinician to balance the requirements of the tool and the breadth of information available within the clinical context. Actions contrary to CDSS recommendations could be seen not as ‘non-compliance’ but as rational decision-making under conditions of uncertainty.

Conclusions

Our pilot study suggests that the OMA intervention has no significant impact on the decision-making behaviour of clinicians in chest pain clinics, and significant challenges were encountered when attempting to integrate it with standard clinical workflows and hospital EHR systems. It is not feasible in its current form and should not be evaluated further. The mixed-methods approach that we adopted, including qualitative ethnographic methods, has provided an in-depth understanding of the ways in which clinicians interacted with a CDSS. This understanding could support the development of new tools to aid clinical decision-making with a greater likelihood of being adopted in practice and with higher integration with hospital EHR systems.

Background

Cohorts recruited in clinical settings make important contributions to addressing translational research questions in prognosis research. Few large, well-characterised cohorts are available to support research into CVD.

Objectives

To establish a contemporary, prospective clinical cohort of patients with linked phenotypic, genetic and biomarker data as a resource.

Methods

Consecutive patients undergoing investigation for new-onset, non-acute chest pain were recruited between 2009 and 2014 from four UK NHS hospitals. Participants provided blood samples for biomarker and genetic analysis, and completed a health questionnaire at baseline. Detailed clinical information was manually extracted from hospital electronic databases. Participants are being followed up for CVD events, and health-related quality of life was assessed using a postal questionnaire (mean follow-up of 469 days).

Results

We enrolled 3345 consecutive patients {63.1% men; mean age 58.0 years [standard deviation (SD) ± 12.5 years]}. Blood samples were collected for 84% of patients, baseline questionnaire data for 89% of patients and clinical data for 99% of patients.

Conclusions

The Clinical Cohorts in Coronary disease Collaboration is a contemporary consented genetic bioresource linked to hospital EHRs, questionnaire data and health outcomes. Significant challenges remain for automatically extracting the entire patient EHR from health information systems and, as a result, a wealth of data remain inaccessible for research. The design, development and provision of common infrastructure, data standards and services that will allow researchers to access and exploit such data are crucial for expanding 4C in the future. Related initiatives such as the NIHR HIC will provide best practices and guidelines for doing so.

The studies in this chapter relate to patients presenting for the first time with stable chest pain (see Figure 24).

Introduction

Consented cohorts recruited in clinical settings make important contributions to addressing translational research questions in prognosis research. 29–32 However, despite their potential advantages, large prospective clinical cohorts are lacking across all disease areas, and only a handful exist for CAD. 214–217

There is a stark contrast between the significant scientific interest in and the lack of clinically useful novel biomarkers. New technologies such as mass spectrometry and nuclear magnetic resonance spectroscopy for measuring metabolites (metabolomics) or proteins (proteomics) have the potential to provide early insight into the aetiology of CAD and, thus, better predictive performance. However, the value of existing circulating biomarkers, or panels of biomarkers, for predicting or mediating disease prognosis has yet to be proven.

Although in its early stages, the adoption of information technology in health care and biosciences – ‘clinical informatics’ – is gaining rapid momentum and is predicted to revolutionise the way -omics data linked to phenotypes recorded within EHRs will drive translation of scientific discovery into clinical practice and patient benefit. In the USA, consented biorepositories linked to EHRs are increasingly successfully being used for translational research at scale. 218 For example, BioVU (Vanderbilt University Medical Centre, Nashville, TN, USA)219 is a biorepository containing over 200,000 DNA samples extracted from discarded blood linked to the full hospital EHR of consented patients attending the Vanderbilt University Medical Centre.

To advance research through collaboration, 4C was established in the UK in 2009 as a resource in which DNA and biomarker samples were obtained at the time of presentation with chest pain and linked to detailed phenotypic data obtained from EHRs and participant self-completed questionnaires. We sought to explore and assess the current potential of setting up a comparable consented research platform by collecting DNA samples and to quantify the extent to which diverse NHS hospital information systems are accessible for extracting secondary care data (structured and unstructured) for research purposes at scale.

Methods

Study cohort

Eligibility

Consecutive patients undergoing evaluation for chest pain at RACPCs or elective invasive coronary angiography at participating hospitals were approached for inclusion (Figure 32). Patients were not eligible for inclusion if they were diagnosed with structural heart disease, cardiomyopathy or arrhythmia, or were referred for ablation or placement of devices (e.g. pacemakers). No patients were excluded on the basis of age or coexisting conditions. Trained researchers identified eligible patients by checking clinic appointment databases and angiogram procedure lists on a daily basis.

FIGURE 32.

Clinical Cohorts in Coronary disease Collaboration study flow diagram.

Participant consent

Patients were approached directly by a researcher at the hospital at the time of their clinic appointment. Informed consent was sought for (1) completion of a baseline health questionnaire; (2) provision of a blood sample for long-term storage and analysis of DNA and biomarkers; (3) extraction of patient hospital data; and (4) linkage of patient data with national EHR data sources (e.g. coded hospitalisations and procedures, emergency admissions and cause-specific mortality) (see Appendix 4).

Results

Single-nucleotide polymorphism genotyping

Minor allele frequencies were replicated from published data (Figure 33).

FIGURE 33.

Genotyping of single-nucleotide polymorphisms in the 4C cohort and replication of allele frequencies from published data (CARDIoGRAM PlusC4D). A total of 51 single-nucleotide polymorphisms were successfully genotyped. Minor allele frequencies as expected from published/reported data.

Metabolic profiling

¹H-NMR spectroscopy was used to resource 225 cardiovascular-related metabolites in human serum. These included (1) lipoprotein lipids consisting of lipids, serum albumin and albumin-bound fatty acid resonances; (2) low-molecular-weight molecules, pyruvic acid, amino acids, ketone bodies and creatinine; and (3) serum constituents with detailed molecular information on serum lipid extracts including free and esterified cholesterol, sphingomyelin and omega-3 fatty acids. A total of 2323 samples were successfully analysed, and 2025 patients underwent angiography (1816 invasive coronary angiograms and 209 coronary CT angiograms).

Discussion

Although researchers carrying out large-scale studies were previously limited to utilising EHR data from single sources, the potential to access data spanning the entire translational cycle now exists, with new opportunities to link clinical health-care data with other sources of electronic information and bespoke data on a large scale. To exploit these new opportunities, we established the 4C cohort, a contemporary, prospective cohort of 3345 patients undergoing investigation for SCAD, with blood and questionnaire data linked to the hospital EHRs, who are being followed up for a range of disease end points recorded in national registry databases. The cohort was established as a resource to encourage collaborative research based on individual patient data, in order to strengthen prognostic model development and external validation.

We faced two significant challenges during the conduct of our study. First, despite attempts to co-ordinate permissions from NHS organisations to undertake clinical research with the introduction of the NIHR Co-ordinated System for Permissions in 2008, we experienced significant delays in obtaining the required approvals [mean NHS Trust site-specific approval time was 85 days (range 50–115 days) across four sites], delaying recruitment of sites and influencing participant accrual targets. Health Research Authority plans to unify the approvals process for research in the UK came into effect at the end of 2015 (see www.hra.nhs.uk/), and it remains to be seen whether or not efforts to reduce the time it takes to approve research are successful, in the same way that regulatory approval of clinical trials was expedited following a highly critical Academy of Medical Sciences report on the effect of delays on approvals on scientific discovery. 227 Contractual negotiations, human resources checks and governance relating to cross-site working within NHS hospital trusts further impacted on recruitment and must also be reviewed. 228

A second challenge was an inability to download data electronically from the hospital EHR, resulting in time-consuming manual collection and recording of clinical data. The lack of efficient means to access or transfer EHR data, including discharge summaries and biochemistry results stored on separate clinical systems, greatly increased the time and resulting financial cost required to collect complete data.

The design, development and provision of common infrastructure, data standards and services are required to allow researchers to seamlessly interact and extract clinical data for research purposes across hospitals. An example of such an initiative is the NIHR-funded HIC health informatics programme. Across five participating Biomedical Research Centres (University College London Hospitals, Imperial College Healthcare, Guy’s and St Thomas’ NHS Foundation Trust, Oxford University Hospitals and Cambridge University Hospitals), the HIC aims to enable the collaborative sharing of NHS data collected and captured as part of usual clinical care to facilitate more effective clinical research. HIC investigators have developed a common data model that is used to extract clinical data (retrospectively and prospectively) from hospital information systems across five scientific themes (viral hepatology, ACS, ovarian cancer, renal transplantation and clinical care). The extracted data are catalogued through a common metadata catalogue and shared via the use of secure safe havens that are accredited to international information security standards such as International Organization for Standardization (ISO) 27001:2013. The HIC approach has demonstrated how robust e-infrastructure and health information modelling approaches are required in order to enable the extraction of clinical data for research from hospitals and to tackle the challenges associated with diverse vendors and information systems being used within and across trusts. A streamlined technical approach also enables researchers to effectively address generic challenges related to data sharing, information governance and patient confidentiality more effectively through the use of standards.

In his State of the Union address in 2015, US President Barack Obama emphasised the power of big data in his personalised medicine initiative. 2 The new era of ‘big data’ brings unprecedented opportunities for prognosis research. Proposals to link clinical data systems locally and embed patient consent for research into clinical care would mitigate the complexities of obtaining complete data and recruiting patients in busy clinical settings, and thereby increase the utility of such cohorts. High-throughput, high-quality research will also be expedited by ambitious discovery projects that are linking EHR data to genomic data, such as the 100,000 Genomes Project,19 and large international EHR research platforms such as the eMERGE network,229 the Danish National Biobank Registry230 and the UK Biobank. 107 However, complex issues, such as how to repurpose EHR data for research (acquisition of ‘research-ready’ data) and extract accurate phenotypes from EHRs, will require careful consideration.

Conclusions

The 4C cohort is a novel contemporary DNA and plasma resource linked to detailed information collected from the hospital EHR and questionnaire data and health outcomes. Significant challenges remain for automatically extracting the entire patient EHR from health information systems and, as a result, a wealth of data remains inaccessible for research. The design, development and provision of common infrastructure, data standards and services that will allow researchers to access and exploit such data are crucial for expanding 4C in the future, driven by best-practice initiatives such as the NIHR HIC.

Background

The scope and impact of prognosis research and EHR research (in primary and secondary care, and in disease and procedure registries) are intimately related. Guidelines recommend measuring a wide range of clinical factors, but there are uncertainties in the reliability of individual biomarkers and their clinical usefulness in predicting clinical outcomes.

Objectives

To examine evidence supporting the use of five established and novel prognostic factors for predicting the progression of disease and outcomes among SCAD patients.

Methods

An analysis of CALIBER data and literature-based systematic reviews and meta-analyses.

Results

(1) Haemoglobin below 13.5 g/dl for men and 12.8 g/dl for women was associated with increased mortality. (2) Heart rate of > 87 b.p.m. was associated with increased hazard of heart failure and mortality. (3) Published evidence of associations between C-reactive protein (CRP) levels and risk of cardiovascular events was of variable quality, and the degree of association may have been overstated. (4) Pooled data showed that Ch9p21 was convincingly associated with increased risk of first but not subsequent cardiovascular events. (5) SBP in type 2 diabetes patients diagnosed with stable angina declined over time. Lower values of SBP significantly increased the risk of all-cause mortality, non-fatal AMI or non-fatal stroke, as did a greater decline in the rate of change of SBP.

Conclusions

Some biomarkers appear to be useful prognostic factors in SCAD, although poor methodology impairs the interpretation of many study findings.

The studies in this chapter relate to patients presenting for the first time with SCAD (Figure 34).

FIGURE 34.

Overview of the 33 studies in our programme: initial presentation with SCAD.

Introduction

Given the known links between myocardial hypoxia and later acute coronary events, low haemoglobin levels may be a useful prognostic biomarker in SCAD. Furthermore, haemoglobin levels are routinely monitored, and well-established interventions exist to boost haemoglobin levels.

Previous studies have identified an association between low haemoglobin levels and mortality in patients with CAD. 234,235 However, these studies have not been able to characterise the shape of the relationship between haemoglobin levels and mortality, or to determine whether or not threshold levels exist. In addition, they did not examine sex differences and did not distinguish patients with a history of either angina or AMI.

Methods

A retrospective cohort study using CPRD data was carried out. Among CPRD patients aged 35–90 years, we defined two study populations on the basis of coded diagnoses between 2001 and 2006: first 20,131 patients with new-onset stable angina and no prior ACS and second, 14,171 patients with an AMI with survival beyond 7 days. The date of the earliest stable angina or AMI diagnosis for each patient was designated the ‘index date’. For the stable angina population, we included patients with a code for angioplasty or coronary artery bypass but not angina only if they had evidence of previous angina symptoms (e.g. a code for chest pain or a nitrate prescription; the date of the earliest such entry was taken as the index date). We excluded patients with ACS within 7 days of the index date because their initial presentation may have been unstable rather than stable angina. From both populations we excluded patients who died within 7 days of the index date, patients with other cardiac conditions prior to the index date (myocarditis, endocarditis, pericarditis, valve disease or congenital cardiac abnormalities) and patients who were pregnant less than 3 months before the index date. We estimated the glomerular filtration rate from serum creatinine using the four-variable Modification of Diet in Renal Disease formula. Follow-up data were available for a mean of 3.2 years. We analysed data using semiparametric Cox regression with multiple adjustment.

We carried out a systematic review, which involved searching MEDLINE and EMBASE databases between 1966 and November 2008 for prospective studies in which blood biomarkers were measured in patients with stable angina or at least 2 weeks after ACS, and a relative risk for mortality or coronary events by haemoglobin concentration was reported with a p-value or CI. Two reviewers reviewed article titles and abstracts for eligibility and independently abstracted data from eligible articles using a pre-defined coding protocol. Studies with less than 1-year follow-up were excluded. Two reviewers independently abstracted data from eligible articles, with disagreements resolved by consensus, or adjudication by a third reviewer if required.

Results

Patients with lower haemoglobin tended to be older and were more likely to have diabetes, renal impairment, cancer, lower mean corpuscular volume and to be taking iron supplements and nitrates (p < 0.001 for all) (Figures 35 and 36).

FIGURE 35.

Adjusted HRs for death for patients with stable angina or AMI by haemoglobin concentration. (a) Schematic diagram of haemoglobin threshold model for adjusted HRs for mortality; (b) women with stable angina; (c) men with stable angina; (d) women with AMI; and (e) men with AMI. Reproduced from Shah et al. 47 © 2011 Shah et al. 47 This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

In both men and women, thresholds were identified below which mortality showed a linear inverse relationship with haemoglobin levels (Figures 36 and 37).

FIGURE 36.

Multiply adjusted Kaplan–Meier curves showing the probability in (a) women and (b) men with new-onset stable angina according to haemoglobin level. Adjustment was done by stratification on linear predictors from a Cox model, which included age, eGFR, SBP, total cholesterol, family history of diabetes, stroke and comorbidity (Charlson Index), but not sex or haemoglobin. Reproduced from Shah et al. 47 © 2011 Shah et al. 47 This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

For stable angina patients, the lower threshold was 11.9 g/dl (95% CI 11.5 to 13.5 g/dl) for women and 13.1 g/dl (95% CI 13.0 to 13.4 g/dl) for men. For men with AMI, this threshold was 13.5 g/dl (95% CI 13.2 to 13.9 g/dl). For the AMI population, the threshold for women with AMI was slightly lower, at 12.8 g/dl (95% CI 12.1 to 13.5 g/dl) compared with 13.5 g/dl (95% CI 13.2 to 13.9 g/dl) for men. The shape and strength of the association was similar in both sexes, and in both AMI and angina patients.

A total of 29.5% of male AMI patients with subthreshold haemoglobin levels were at statistically significantly greater risk of death (HR 2.00, 95% CI 1.76 to 2.29).

The systematic review and meta-analysis identified 10 published studies reporting just 1127 end points in total, not one of which examined risk thresholds.

Discussion

This study has reported more data than all previous studies combined (3174 deaths vs. 1127 events), providing the first estimates of haemoglobin level thresholds associated with a markedly enhanced risk of death. As well as identifying risk thresholds, the study indicates that, contrary to previous assumptions, the relationship between risk and haemoglobin levels is not linear.

Thresholds are higher than might have been anticipated, placing large numbers of patients (15–30%) in a high-risk category. Elevated risk associated with subthreshold haemoglobin levels was seen in both women and men, and in both AMI and angina patients, with risk being more pronounced in AMI patients. Attributable risk calculations suggest that the link between haemoglobin levels and mortality could be responsible for up to one-quarter of deaths after AMI. This level of risk is similar to that identified for LDL cholesterol and fibrinogen. There was some evidence of an association between high haemoglobin levels and poor outcomes, but the number of patients affected was small.

New clinical trials to assess whether or not haemoglobin levels are causal and whether or not clinicians should intervene to increase haemoglobin levels are warranted.

Conclusions

Our findings provide support for clinical trials to evaluate strategies to increase haemoglobin levels in patients with SCAD, through either iron supplementation or more aggressive treatments, such as intravenous iron infusions or erythropoiesis-stimulating agents. Further studies are needed to assess the potential prognostic value of haemoglobin in stratifying SCAD patients.

Introduction

The relationship between resting heart rate and CVD and mortality has been extensively investigated and acknowledged in patients with SCAD236 and in the general population. However, findings remain controversial. In 2008, two large trials of ivabradine, a pure heart rate-lowering agent without other known cardiovascular effects, showed controversial effects related to risk reduction for both stable and unstable coronary disease in patients with CAD and chronic heart failure,237 and in patients without reduced left ventricular function with SCAD;237 these results indicated conflicting scenarios for the mechanisms through which an increased heart rate could be involved in an increased risk of CVD. 238

Contemporary data regarding elevated heart rate and prognosis of CAD patients are derived from clinical trials or angiographic registries that are collected solely for research purposes and are confined to high-risk patients238–240 or those admitted for ACS,240,241 and not among the general population. 236,238 These data are biased by the selective nature of the populations studied and might compromise external validity. 233 In addition, much of the evidence draws on the effect of heart rate-lowering medication and not the actual heart rate marker on cardiovascular end points. 237,242,243 The majority of studies have also failed to examine potential differences in heart rate effects on cardiovascular events between men and women, probably owing to limited sample sizes and subsequent restricted statistical power. 240,244–246 The preponderance of composite cardiovascular end points in the published literature indicates the need for a study that will investigate the association of heart rate and specific cardiovascular events in a SCAD population.

Using linked EHRs we investigated the prognosis of SCAD in patients with low or elevated heart rate, using real-world data in a general population. The aim of this study was to examine the relationship between resting heart rate and a range of previously unexplored future cardiovascular phenotypes and non-cardiac complications, as well as cerebrovascular events, in a large population of patients with suspected or proven SCAD. We also aimed to determine the shape of any associations and to explore potential differences in the associations between men and women.

Methods

Results

Heart rate and presentation of cardiovascular diseases

Age- and sex-adjusted analyses showed a significantly increased hazard of heart failure, cardiovascular and all-cause death and atrial fibrillation in quintile 5 compared with quintile 1 (Figure 37). An increased risk of stroke and bleeding events was also identified. These associations remained after adjusting for sex, age, smoking, SBP, diabetes, cholesterol, HDL, BMI and IMD. Strong associations were found between high heart rate and hazard of heart failure independently of the type of previous coronary event (for overall coronary event types HR 2.42, 95% CI 2.19 to 2.68). Association with the risk of stroke was also found for participants in quintile 5 (HR 1.25, 95% CI 1.10 to 1.43) and for people who experienced a previous stable angina event (HR 1.28, 95% CI 1.11 to 1.48), but not in people who experienced an AMI or unstable angina, probably owing to the smaller number of cases and the subsequent loss of statistical power. Similarly, bleeding and atrial fibrillation events were higher in people in quintile 5 (HR 1.21, 95% CI 1.10 to 1.33) than in people in quintile 1 (HR 1.58, 95% CI 1.49 to 1.68). The association with atrial fibrillation was present when taking into account the different subtypes of prior stable coronary events. No association was found between heart rate and AMI, PAD, stable angina or SCD, the last of which perhaps as a result of the limited number of events.

FIGURE 37.

Multivariable-adjusted HRs for the association between heart rate (top vs. bottom quintile) and the risk of eight CVDs, atrial fibrillation and bleeding events in CALIBER patients. MI, myocardial infarction.

Increased risk of heart failure, AMI and stable angina were found among participants with increased heart rate quintile changes, irrespective of their direction (upwards or downwards). However, the small sample size resulted in low power and such effects were significant only at the 10% level (results not shown).

Discussion

In this prospective study, we report an association between elevated heart rate and risk of cardiovascular end points in patients with SCAD. Our study confirms previous reports that elevated heart rate is a predictor of mortality and heart failure in patients with SCAD. 236,238,239

However, our results also extend these observations to cerebrovascular events, atrial fibrillation and complications such as bleeding events. We found no association between increased heart rate and AMI, stable angina, SCD or PAD, whereas a strong association was identified with heart failure, cardiovascular death and all-cause mortality. An increased risk of stroke, atrial fibrillation and bleeding events was also found in both univariate and multivariate analyses in people with heart rates of > 87 b.p.m. in quintile 5 compared with people in quintile 1 (< 65 b.p.m.).

The effects of heart rate-lowering medication on mortality in patients with CAD have been extensively demonstrated. 247–249 However, the association between resting heart rate values (without the mediation of heart rate medication in analyses) and specific cardiovascular end points and atrial fibrillation has been underinvestigated. The lack of association between heart rate and CHD found in this study contrasts with results from previous post hoc analyses of trials with heart rate-lowering agents (e.g. ivabradine), suggesting that it would improve outcomes in patients with SCAD. 250

In patients with heart failure, the reduction in heart rate with ivabradine has been shown to improve clinical outcomes, beyond the improvements observed with beta-blockers. 251 The benefits of lowering heart rate in patients with heart failure but not in those with SCAD may reflect the fact that an elevated heart rate is attributable to different pathophysiological mechanisms in these two conditions,252 such as neurohormonal activation in patients with heart failure and further ventricular remodelling. Sympathetic activation and progressive disruption of autonomic balance may also be important in the pathogenesis of atrial fibrillation after coronary artery bypass grafting. 253,254 White et al. 255 found a significant increase in the incidence of atrial fibrillation among patients in whom the administration of beta-adrenoceptor antagonists was ceased compared with those whose drug treatment was continued after surgery. Further evidence of a mechanistic link of heart rate with atrial fibrillation comes from genome-wide association studies that have identified heart rate-increasing alleles at five loci associated with both increased and decreased risk of atrial fibrillation. 256

Published studies have suggested a threshold effect of heart rate on mortality outcomes, above which risk increases steeply. Data from the Coronary Artery Surgery Study (CASS) registry suggested an increased risk of all-cause mortality for a heart rate of > 77 b.p.m. and of CVD death at a heart rate of > 83 b.p.m.,236 whereas other studies have supported a slightly lower threshold of > 70 b.p.m. 246 Data from the ivabradine for patients with stable coronary artery disease and left-ventricular systolic dysfunction (BEAUTIFUL) study have suggested that the increase in death and heart failure outcomes rose continuously at ≥ 70 b.p.m., whereas the relationship was less pronounced for AMI and revascularisation. 238 Our study provides evidence for a similar shape in risk as the BEAUTIFUL study, with a steep increase in risk for heart failure and cardiovascular mortality at > 70 b.p.m., a threshold significantly lower than the majority of the available studies. A similar steep increase was also found for atrial fibrillation events above a heart rate of 70 b.p.m., whereas the risk of SCD shows an increase at higher heart rate values (> 90 b.p.m.).

Heart rate is one of the most easily accessible biomarkers; however, rates of adequate heart rate control appear to be low in current clinical practice, potentially representing a missed therapeutic opportunity. In designing randomised trials, the target population, the choice of composition of the primary end point and the likely rate at which the end point accrues can all be informed by the examination of linked EHRs. Future clinical trials on heart rate-specific medications, such as ivabradine, should refine our understanding of the potential role for specific heart rate lowering in the treatment and prevention of people at early stages of heart failure and resting tachycardia.

Conclusions

Resting heart rate is an easily accessible biomarker in primary and secondary care patients with heterogeneous associations with CVDs. Higher resting heart rates increase the hazard of heart failure events, stroke in women, atrial fibrillation and bleeding events, without substantially affecting the hazard of vascular events in the coronary arterial beds. Our findings show how disaggregation of CVD into specific phenotypes can help to disentangle the underlying disease mechanisms, with implications for clinical practice and the design and interpretation of clinical trials.

Introduction

Prognostic biomarkers have great potential in translational medicine research and could support clinical decision-making. However, based mainly on experience in cancer, concerns have been raised about the quality of prognostic biomarker research. 31 In particular, evidence syntheses have highlighted issues with publication bias and the selective reporting of results. 257,258 In addition, few evaluations have assessed the quality of research evidence related to individual biomarkers.

It is unclear whether or not similar biases affect the evidence base for biomarkers in CVD. To address this question, we systematically analysed the evidence related to use of CRP as a prognostic biomarker in SCAD. CRP has been widely studied in this context, and its measurement is recommended in US259 and European260 clinical guidelines (although actual clinical practice varies widely).

We undertook a systematic review, meta-analysis and metaregression of published evidence on CRP. Our key aims were to assess the quality of published evidence, the extent to which reported associations are independent of other prognostic factors, the scale of any bias attributable to small study size, any possible sources of heterogeneity and whether or not papers report clinically useful figures on the ability of CRP measurements to distinguish patients who experience a coronary event from those who do not.

Methods

As no agreed criteria exist for assessing reporting quality, we developed a 17-item list based on previous work in CVD and cancer (including the REMARK reporting guidelines). Notably, this list also included whether or not studies reported a reference to a study protocol or statistical protocol specified in advance of data collection.

MEDLINE (1966–2009) and EMBASE (1980–2009) were searched and selected prospective studies of patients with stable coronary disease, reporting a relative risk for the association of CRP with death and non-fatal cardiovascular events. We identified relevant papers and assessed individual papers according to the 17-item checklist. Two reviewers independently abstracted data from eligible articles (n = 116) using a predefined coding protocol. Individual item disagreement between the two reviewers was resolved by consensus or, rarely, adjudication by a third reviewer. Reported relative risk, ORs or HRs were extracted from each eligible paper.

Results

Of 1566 articles, 83 studies met inclusion criteria, covering 61,684 patients and 6485 events.

The median number of study-quality items reported was seven out of a possible 17, with no increase between 1997 and 2009. Just two studies reported a study protocol; none reported a statistical analysis protocol (Figure 38). The meta-analysis generated a relative risk for the top versus bottom third of the CRP distribution of 1.97 (95% CI 1.78 to 2.17), although the size of effect varied widely (I² = 79.5) (Figure 39). Just 13 studies adjusted for standard risk factors; adjustment lowered relative risk to 1.65 (95% CI 1.39 to 1.96). Ten different methods were used to compare CRP levels, with use of continuous measures giving lower relative risks.

FIGURE 38.

Quality of individual reports (n = 17 items, n = 83 studies) included in our systematic review of CRP as a prognostic biomarker in SCAD based on REMARK guidelines. Reproduced from Hemingway et al. 49 © 2010 Hemingway et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

FIGURE 39.

Forest plot of the effect of CRP on prognosis among patients with SCAD based on 83 studies included in our systematic review and meta-analysis. For example, Jánoskuti et al. 264 adjusted for age, sex, smoking and obesity and three other non-conventional risk factors, but not for diabetes. a, Four studies did not report the number of events; b, five studies made no adjustments at all; one (Minoretti 2006313) adjusted for age and sex. A, age; D, diabetes; G, sex; L, lipids; O, obesity; S, smoking; Wt, study weight percentage from random-effects meta-analysis. Reproduced from Hemingway et al. 49 © 2010 Hemingway et al. 49 This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Funnel plots and quantitative analyses provided strong evidence for publication bias (Figure 40). Further adjustment for publication bias reduced the relative risk to 1.19 (95% CI 1.13 to 1.25). Just two studies reported a measure of discrimination (c-statistic). When data from 20 suitable studies were pooled, the detection rate was calculated as 31% for a 10% false-positive rate, giving a c-statistic value of 0.61 (95% CI 0.57 to 0.66).

FIGURE 40.

Funnel plot showing extent of publication bias for studies included in our systematic review and meta-analysis of CRP as a prognostic biomarker in SCAD, with contours showing different levels of study significance. MA, meta-analysis. Reproduced from Hemingway et al. 49 © 2010 Hemingway et al. 49 This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Discussion

This large-scale systematic review and meta-analysis identified significant shortcomings in the evidence base on CRP as a prognostic biomarker, particularly the absence of a pre-specified statistical analysis protocol and strong evidence of publication bias having a sizeable impact on association.

There are no agreed comprehensive criteria for measuring the quality of study reports. The REMARK reporting guidelines for tumour prognostic markers provide a useful start, but are not currently in a form that lend themselves to measurement and omit reference to statistical analytic protocols.

Our review suggests that there is a need for new clinical cohorts with published pre-specified statistical analytic protocols and prospective registration of studies, which can be pooled to evaluate the true usefulness of the role of CRP in prognosis in SCAD. Better reporting is required and existing guidelines require development across disease areas and formalisation into data extraction tools.

Along with a widespread failure to adjust for potential confounders, the findings call into question the reported association between CRP and prognosis of patents with SCAD. Furthermore, the discriminatory potential of CRP was so low as to offer little clinically useful information.

Conclusions

The findings of this evaluation suggest that current CG recommendations on the measurement of CRP need to be revised. More generally, they highlight major flaws in the design of studies on prognostic biomarkers. There is a need for enhanced reporting and, in particular, for the registration of studies and statistical analysis protocols.

Introduction

Chromosomal locus 9p21 (Ch9p21) is a well-established genetic risk factor for CHD. Several lines of evidence have implicated Ch9p21 in risk of first coronary events,314 but it is less clear whether or not it also influences susceptibility to further coronary events, with conflicting results from mostly small-scale studies.

The distinction is important, as the predictive value of genetic variants at Ch9p21 may differ for first and subsequent coronary events. In addition, differences would also imply that the mechanisms of disease may not be the same in first and subsequent coronary events, which would have important implications for secondary prevention and drug development.

To address this question, we carried out a systematic review and meta-analysis to synthesise the evidence relating to Ch9p21 and the risk of first and subsequent coronary events, and to identify any sources of bias that could influence associations.

Methods

We searched PubMed and EMBASE from inception to 30 June 2013 for studies reporting associations between genetic variants at Ch9p21 and CHD up to 30 June 2013. Two reviewers extracted information, and discrepancies were resolved by consensus. We also searched studies reporting associations of genetic risk scores with CHD, as they may have incorporated data on the Ch9p21 locus [thus, our search strategy encompassed terms capturing the genetic locus (9p21 or 9p21.3), genetic risk score and CHD].

Prospective cohorts and nested case–control or cohort studies were included in the review. We excluded case–control or cross-sectional studies of Ch9p21 and CHD, editorials and reviews.

Two authors independently extracted data and any discrepancies were resolved by consensus.

Results

Of 327 papers identified, 25 satisfied inclusion criteria. As well as these data, from 33 cohorts, we added data from two unpublished studies. After the elimination of data from four cohorts reporting genetic results in a recessive format, we analysed 31 cohorts in total, covering 193,372 individuals and 20,100 coronary events.

In the 16 cohorts reporting first events, 15,664 events occurred among a population of 168,209. The pooled HR for Ch9p21 was 1.19 (95% CI 1.17 to 1.22) per risk allele. In cohorts of individuals who had already experienced CHD, 4436 events occurred among a population of 25,163. The pooled HR for Ch9p21 for subsequent events was 1.01 (95% CI 0.97 to 1.06) per risk allele. Funnel plots revealed no evidence of small-study bias (Figure 41). Risk allele frequencies were similar in population and clinical cohorts, arguing against any survival bias.

FIGURE 41.

Small-study bias illustrated by funnel plots of the association of Ch9p21 (per risk allele) with (a) first and (b) subsequent coronary heart disease events in 31 studies included in our systematic review and meta-analysis. Reproduced from Patel et al. 50 Copyright © 2014 American College of Cardiology Foundation. Published by Elsevier Inc. This article is available under the terms of the Creative Commons Attribution License (CC BY). You may copy and distribute the article, create extracts, abstracts and new works from the article, alter and revise the article, text or data mine the article and otherwise reuse the article commercially (including reuse and/or resale of the article) without permission from Elsevier.

Discussion

This systematic review and meta-analysis of data from more than 190,000 individuals, experiencing 20,000 events, has confirmed that genetic variants at Ch9p21 are associated with initial CHD events. However, it found no evidence to suggest that Ch9p21 contributes to the risk of further events in patients with established CHD. There was no evidence to suggest that this differential association reflected a lack of power or publication or survival bias.

Our study has important research and clinical implications. First, our findings suggest that future genetic association studies should be cautious in assuming that genetic variants have similar effects for both first and subsequent CHD events. Second, further research into mechanistic differences that may account for the differential genetic effect is warranted to enhance understanding of the underlying reasons for this. Our findings point to the need for a consortium of studies among individuals with established CHD to better understand the genomic susceptibility for subsequent CHD events and to carry out analysis, including assessment for selection biases, which is not possible with literature-based meta-analyses. More research is warranted to determine the effect of mechanistic differences for the differential genetic effect, to enhance understanding of the underlying reasons.

Conclusions

Our findings suggest that disease processes may differ between first and subsequent coronary events, with Ch9p21 playing a part only in the former. The precise role of Ch9p21 in disease is difficult to assess, as the molecular mechanisms affected by Ch9p21 genetic variants are unclear. A better understanding of Ch9p2’s role could help to clarify how pathogenic processes differ between first and subsequent experiences of CHD.

The findings call for caution in extrapolating the results of genetic association studies from first presentation of CHD to subsequent events. In addition, they raise questions about the use of genetic testing to assess risk – direct-to-consumer and physician-ordered testing options exist, but are unlikely to provide useful information in patients with existing disease.

Introduction

Biomarker levels, such as BP, are recorded in clinical practice at more than one time point, often as a monitoring tool to indicate patient status, as well as clinical event data such as time to AMI, stroke or death. Extensive studies in CVD have been conducted to assess the role of one or more biomarkers measured at a single time point in predicting clinical events, such as death from cardiovascular causes. 49,315–317 Repeated measurements have only recently received attention, demonstrating that variation in BP during follow-up is independently (of baseline BP) associated with risk of adverse events, such as stroke and AMI. 318 Many biomarkers, such as SBP and DBP, are now collected as part of usual care over time, and this vast resource of information, existing in registry sources such as CPRD, is yet to be fully explored. 319 Furthermore, as data linkage becomes more common,320 there are great opportunities for the development of real-time dynamic risk prediction models. 321

Through a joint modelling approach, we can both directly model a biomarker profile over time, accounting for measurement error and capturing between-patient variability, and directly link aspects of the biomarker profile to the time to event. This methodological approach is now well established, minimising bias and improving efficiency in estimates of association between longitudinal and clinical event data. 322–324 Alternatively, a joint model can be thought of modelling a biomarker profile over time, adjusting for informative drop-out, often owing to death.

In this study, our aim was to determine the extent to which current values of BP, trend over time or within-person variability adds information beyond a single baseline value in the prediction of death among people with stable angina and type 2 diabetes, AMI and stroke. We studied people with angina and type 2 diabetes as a high-risk population in whom clinical guidelines recommend close BP monitoring. Despite this, little is known about patterns of BP change (e.g. whether or not it increases with age as in the general population).

Methods

Results

Exploratory analysis

Figure 42 shows the observed SBP values for a patient with 336 repeated measures, illustrating the trajectory of SBP over time. Figure 43 shows the overall trend in observed SBP profiles for all patients.

FIGURE 42.

Observed SBP for a CALIBER patient with type 2 diabetes and stable angina with 336 repeated measurements of BP.

FIGURE 43.

Non-parametric smoothed estimates of observed SBP among CALIBER patients with type 2 diabetes and stable angina. (a) Time from diagnosis as the timescale; and (b) adjusted timescale, with each patient’s censoring/event time taken away from the measurement times.

Figure 43a shows a non-parametric smoother of SBP from time since diagnosis of stable angina, separately for patients who were censored at the end of follow-up, and those who experienced the composite end point. In Figure 43b we use the same non-parametric smoother of SBP, but adjust the timescale by taking away each patient’s observed censoring or event time. This gives an indication of any differences in the trend of SBP over time in the period leading up to an event, compared with patients who are censored. Figure 43b indicates that patients who experience the composite end point have a pronounced decline in SBP in the 2–3 years prior to an event, which could imply that the rate of change of SBP over time could play an important part in the risk of composite end point.

Modelling the time to death, acute myocardial infarction or stroke

For a female as opposed to a male of the same age at baseline and at the same value of SBP, the risk of an event is reduced by 17.2% (95% CI 8.9% to 24.8%) (Table 21).

TABLE 21 - Results from fitting a Weibull survival model to CALIBER patients with type 2 diabetes and stable angina, with SBP as a time-varying covariate

SBP = (SBP – 140)/10.

Age = Age – 68.

Parameter	Log-HR	p-value	95% CI
SBP	–0.0704	< 0.001	–0.1020 to –0.0389
SBP2	0.0345	< 0.001	0.0287 to 0.0403
SBP3	–0.0021	0.001	–0.0034 to –0.0009
Age	0.0554	< 0.001	0.0503 to 0.0605
Age2	0.0004	0.032	0.0000 to 0.0605
Female	–0.1888	< 0.001	–0.2848 to –0.0928
Intercept	–2.5811	< 0.001	–2.6798 to –2.4824
log-(gamma)	–0.0780	< 0.001	–0.1179 to –0.0380

Interpretation of covariate effects is complicated with the introduction of polynomial terms of SBP and age. Figure 44 shows the non-linear relationship between SBP and the risk of composite end point.

FIGURE 44.

Estimated HR for the effect of SBP on the risk of composite end point, relative to a reference value of SBP of 140 mmHg in CALIBER patients with type 2 diabetes and stable angina.

Joint modelling of the profile of systolic blood pressure and time to death, acute myocardial infarction or stroke

The results from the full joint model, linking repeatedly measured SBP and the time to composite end point are presented in Table 22.

TABLE 22 - Joint model of repeatedly measured SBP and time to composite end point of death, AMI and stroke among CALIBER patients with type 2 diabetes and stable angina

Longitudinal outcome is SBP/10.

Parameter	Coefficient	p-value	95% CI
Longitudinal
Time	–0.0951	< 0.001	–0.1061 to –0.0842
Age	0.0046	0.070	–0.0004 to 0.0097
Female	0.2710	< 0.001	0.1927 to 0.3494
Age × female	0.0138	< 0.001	0.0066 to 0.0211
Intercept	14.1143	< 0.001	14.0570 to 14.1717
Survival
Association (current value)	–0.1382	< 0.001	–0.1869 to –0.0894
Association (slope)	–1.0552	< 0.001	–1.4820 to –0.6285
Age	0.0560	< 0.001	0.0509 to 0.0612
Age2	0.0005	0.005	0.0001 to 0.0008
Female	–0.1574	0.002	–0.2549 to –0.0600
Intercept	–0.6142	0.070	–1.2786 to 0.0501
log-(gamma)	–0.0971	< 0.001	–0.1386 to –0.0556
Random effects
sd(time)	0.2048		0.1942 to 0.2161
sd(intercept)	1.4805		1.4448 to 1.5173
corr(time,intercept)	–0.5717		–0.6075 to –0.5336
sd(residual)	1.5423		1.5331 to 1.5516

We found highly statistically significant associations (p < 0.001) between the risk of composite end point and both the current value of SBP and the rate of change. Both association coefficients are negative, indicating that lower values of SBP increase the risk of an event, and a greater decline in the rate of change of SBP increases a patient’s risk.

This is illustrated in Figure 45, which shows predicted event probabilities for two patients of the same sex and similar age at baseline, conditional on patient-specific SBP profiles. Patient A shows a slightly increasing trend in SBP over time compared with patient B, who has a decreasing trend in SBP. Subsequently, patient A has an improved event probability prediction compared with patient B.

FIGURE 45.

Predicted event probabilities conditional on patient-specific SBP profiles among CALIBER patients with type 2 diabetes and stable angina for (a) patient A and (b) patient B.

Discussion

This linked EHR-based study of type 2 diabetes patients diagnosed with stable angina provides strong evidence of a declining trend in SBP following diagnosis. This is in contrast to a generally increasing trend among a healthy population. The current value and rate of change of SBP over time are strongly associated with the risk of death, AMI or stroke, indicating that lower values of SBP and a declining trend in SBP may have an important role in identifying patients at increased risk.

This is the first study to use repeated measures of SBP, both to model its trend over time adjusting for dropout and to investigate how time-dependent aspects such as the current value and rate of change are associated with risk of death, AMI or stroke. The joint model results provide a platform for dynamic risk prediction, whereby individual patients’ risk profiles can be updated when new measurements become available. To date, all current prediction models for cardiovascular risk use single baseline measurements for all modelled covariates (i.e. measurements taken at the start of follow-up). Methods for dynamic risk prediction in which the patterns of change of clinically measured biomarkers can be incorporated into clinical decision support models in clinical practice are required.

Conclusions

Patients with type 2 diabetes and diagnoses of stable angina have an overall decreasing trend in SBP over time. The current value and rate of change of SBP over time are both strongly associated with the risk of death, AMI or stroke, indicating that lower values of SBP and a declining trend in SBP may have an important role in identifying patients at increased risk.

Joint modelling provides the ability to investigate trends in a biomarker over time, accounting for measurement error and dropout (often as a result of death) or, alternatively, to assess the association between a biomarker profile (or aspects of the trajectory such as rate of change) and the time to event of interest.

Background

Real-world data based on linked EHRs can contribute important prognostic information in unselected patients. Improved survival after ACS and in an ageing population have resulted in a large population of patients with SCAD who require high health-care resource use. Validated prognostic models could assist in clinical management of SCAD patients and, in combination with information on resource use and costs, provide valuable tools for policy-makers.

Objectives

To develop prognostic models, examine health-care use and costs, and to provide a framework to evaluate the cost-effectiveness of treatments in SCAD patients.

Methods

Prognostic models and 5-year risk predictions were developed, and lifetime health-care costs were calculated using CALIBER data.

Results

Prognostic models achieved a C-index of 0.811 for all-cause mortality and 0.778 for non-fatal AMI or coronary death. The strongest predictors were initial diagnosis, CVD comorbidities and biomarker levels. Five-year and lifetime costs were £9181 (57% CVD related) and £107,052 (60% CVD related), respectively, for lowest-risk patients, versus £21,872 (72% CVD related) and £38,606 (72% CVD related) for highest-risk patients. A new treatment with 20% hazard reduction for AMI, ischaemic stroke or CVD death would be cost-effective if priced at < £61 per year for the lowest-risk group and < £657 per year for the highest-risk group.

Conclusions

Validated prognostic models could be used to support the clinical management of SCAD patients. Cost modelling can inform resource allocation decisions and be used to assess the cost-effectiveness of new treatments.

The studies in this chapter relate to patients presenting for the first time with SCAD (see Figure 34).

Introduction

Stable coronary artery disease is becoming increasingly common, as a result of an ageing population and enhanced survival after ACS. In the USA alone, it affects approximately 7% of the population. 332 Guidelines produced by the American College of Cardiology Foundation (ACCF)329 and American Heart Association (AHA)329 recommend that patients should be stratified into high-, intermediate- and low-risk groups on the basis of their annual mortality risk, a classification that should then guide further investigation and choice of treatment. However, prognostic tools developed to date have typically been based on data collected in clinical trials rather than ‘real-world’ data gathered as part of clinical practice. 260,329,333 Furthermore, despite recommendations in guidelines, models have not assessed both all-cause mortality and coronary events, and no model has been validated using an independent data set. None has yet been recommended for clinical use.

To address these gaps, we used the CALIBER platform to develop prognostic models consistent with the recommendations made in clinical guidelines. Using CALIBER linked data, we sought to identify associations between outcomes and a wide range of clinical and other measures. Prognostically useful factors were incorporated into prognostic models for all-cause mortality and non-fatal AMI or coronary death, and models were assessed to quantify the accuracy of their predictions and their clinical usefulness – in terms of life-years saved had they been implemented. The performance of prognostic models was validated using an independent data set.

Methods

Candidate prognostic factors were identified from recent guidelines on the management of SCAD and included demographic factors, disease subtype, nitrate use, previous experience of heart surgery, history of other conditions, various physiological measures, psychosocial factors and standard clinically assessed biomarkers.

Prognostic models were developed according to the checklist in the Transparent Reporting of a Model for Individual Prognosis or Diagnosis (TRIPOD) guidelines. Five-year risk predictions were developed, extending annual mortality risks to < 5% (low), 5–14% (intermediate) and > 15% (high), and a net reclassification improvement was calculated for each factor added to the model. The potential clinical impact of each model was estimated by calculating the number of additional years of life saved had the model been implemented. Models were validated using data from 4020 patients in the Appropriateness of Coronary REvascularisation (ACRE) study, which followed patients from 1996/97 to 2004. 334

Results

Of 102,023 patients identified, 44.7% had stable angina, 21.5% had another CHD, 13.1% had unstable angina, 4.7% had STEMI, 6.7% had NSTEMI and 9.4% had unclassified AMI. The 5-year mortality for the group as a whole was 20.6% (95% CI 20.3% to 20.9%) but varied markedly among disease subtypes. The 5-year risk for non-fatal AMI or coronary death was 9.7% (95% CI 9.4% to 9.9%). Factors with prognostic value included age, sex, diagnosis, socioeconomic deprivation, smoking, hypertension, diabetes, lipid levels, heart failure, PAD, atrial fibrillation, stroke, chronic kidney disease, chronic pulmonary disease, liver disease, cancer, depression, anxiety, heart rate, creatinine levels, white cell count and haemoglobin levels (Table 23).

Models incorporating these factors achieved a C-index of 0.811 for all-cause mortality (95% CI 0.806 to 0.816) and 0.778 for non-fatal AMI or coronary death (95% CI 0.770 to 0.785) (i.e. when comparing two patients, the model correctly predicted who would have an earlier event 81.1% of the time). The strongest predictors were initial diagnosis, CVD comorbidities and biomarker levels (Figure 46).

FIGURE 46.

Potential clinical impact of using CALIBER risk models for clinical evaluation of patients with SCAD: evaluation of prediction performance on incremental addition of different sets of predictors, using sociodemographic characteristics as the reference model. (a) All-cause mortality model; (b) non-fatal MI or coronary death model. a, Based on guideline definition of low (< 1%), intermediate (1–3%) and high (> 3%) annual risk; b, assuming that management is associated with HR 0.8 and is cost-effective at high risk. MI, myocardial infarction; NRI, net reclassification index. Reproduced from Rapsomaniki et al. 52 Copyright © 2014, © The author 2013. Published by Oxford University Press on behalf of the European Society of Cardiology. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by-nc/3.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited.

If the models had been used to identify patients at high risk (3% annual mortality), and a management programme with a HR of 0.8 had been implemented, an additional 13–16 life-years or 15–18 coronary-free years could have been achieved per 1000 patients screened, in comparison with models based only on sex, age and deprivation (Figure 47).

FIGURE 47.

Contribution of each variable included in the CALIBER model for predicting all-cause mortality in SCAD, assessed by the decrease in the C-index on its removal from the complete model. MI, myocardial infarction; TCHOL, total cholesterol. Reproduced from Rapsomaniki et al. 52 Copyright © 2014, © The author 2013. Published by Oxford University Press on behalf of the European Society of Cardiology. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by-nc/3.0/), which permits non-commercial reuse, distribution, and reproduction in any medium, provided the original work is properly cited.

In the ACRE validation, 872 deaths and 1238 non-fatal AMI or coronary deaths were seen during 7.5 years of follow-up. In this data set, the model’s C-index for discrimination was 0.735 for all-cause mortality (95% CI 0.715 to 0.755) and 0.718 for non-fatal AMI or coronary death (95% CI 0.700 to 0.736).

Discussion

In this population-based study of > 100,000 SCAD patients, prognostic models based on widely available clinical data could identify patients at high risk of long-term mortality and coronary events with a high degree of accuracy.

The models showed good calibration and discrimination, when applied to both original source data and an independent data set. Their relevance to clinical decision-making was confirmed through calculations demonstrating significant benefits had they been applied in practice to stratify patients. As there is no cost involved in acquiring new data, use of the models is likely to be highly cost-effective.

The models have been developed to be consistent with ACCF/AHA definitions and guidelines. 329 The findings support use of the models in early stages of stepwise risk assessment, to identify patients for intensive management or further investigation. Further external validation in external data sets and recalibration for different populations and time periods is required. An economic evaluation is needed to establish the appropriateness of using the prognostic model in practice.

Conclusions

We present validated prognostic models for estimating the risk of all-cause mortality and coronary outcomes based on clinical parameters that are commonly available in all people with SCAD. These models can be implemented alongside further medical investigations to support medical decision-making. Further independent evaluation is required in different settings, including different EHR systems, health-care organisations and geographic locations to guide use in clinical practice. A risk calculator is available online (www.caliberresearch.org/model).

Introduction

Stable coronary arterial disease patients vary widely in their risk of subsequent AMI or coronary death (about 10-fold, between top and bottom deciles of risk). 52 Understanding how drug use, procedures, primary care consultations and hospital admissions vary according to baseline risk may contribute to clinical decision-making for SCAD patients and policy-making.

However, existing studies on resource use and costs in SCAD have important limitations. Studies have been based on populations with disease and do not reflect contemporary and standard clinical practice; for example, they use selected small samples from trials undergoing non-standard care. 335 Despite the need to understand the long-term course of SCAD, previous studies have taken AMI as a starting point336 and use overly simplistic models. We are not aware of any published study using data collected as part of usual care that has evaluated resource utilisation and costs according to baseline cardiovascular risk, and longer-term and lifetime implications of SCAD have not been reported accurately elsewhere.

The majority of models estimating the costs and health effects of CVD have focused on primary prevention,337,338 and have made predictions over relatively short time horizons. 339 Hence, they are unable to capture lifetime costs and health effects, are based on selected samples (potentially biasing baseline risk and cost estimates, and limiting their generalisability) and fail to model all relevant end points and their interdependence. 335

Population-based contemporary linked EHRs could be used to provide reliable estimates of the lifetime cost and quality-adjusted life expectancy of patients with this important chronic condition. Using CALIBER linked data, we set out to (1) determine health-care utilisation and the associated costs in the first year with SCAD and in the year following a first non-fatal event (AMI, ischaemic or haemorrhagic stroke), (2) identify predictors of costs in the first year of SCAD and (3) model patients with SCAD, estimating their baseline risk of experiencing further CVD events and then predicting both costs and key health outcomes over the lifetime of these patients stratified by their baseline CVD risk. This model will allow us both to understand the cost implications of managing this growing population under current standards of care and to evaluate the cost-effectiveness of new treatment strategies, potentially differentiated by risk group.

Methods

Model

A state transition model was developed to capture the natural history of SCAD patients based on previous models in CVD335 and clinical advice. The model (Figure 48) focuses on predicting the first non-fatal CVD event that SCAD patients are likely to experience as well as CVD and non-CVD mortality. The model has a lifetime time horizon and uses a 90-day cycle length to capture the time-varying and age-dependent nature of risks, costs and health-related quality of life. Costs and health-related quality of life were attached to model states and were adjusted for patient covariates at baseline as well as age and time post event in the model.

FIGURE 48.

State transition model capturing the natural history of SCAD in CALIBER patients. MI, myocardial infarction. Reproduced from Asaria et al. 53 Copyright © 2016, BMJ Publishing Group Ltd and the British Cardiovascular Society. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Results

Lifetime costs and outcomes

There was large variation in CVD risk between the lowest- and highest-risk groups, with an absolute difference in risk of 40% over 5 years. The risk of clinical events was positively correlated with age and higher levels of CVD risk factors such as hypertension and diabetes, as well as with higher prevalence of CVD comorbidities. There were no obvious trends in the key modifiable CVD risk factors such as cholesterol, HDL and heart rate in this population of clinically managed patients.

The modelled progression of CVD over time by risk group is shown in Figure 49. Higher-risk groups were predicted to have much higher levels of CVD mortality than lower-risk groups, who were predicted to remain event free for a much longer period.

FIGURE 49.

Progression of disease over time by baseline CVD risk group among CALIBER patients with SCAD. (a) Lowest risk; (b) decile 2; (c) decile 3; (d) decile 4; (e) decile 5; (f) decile 6; (g) decile 7; (h) decile 8; (i) decile 9; and (j) highest risk. MI, myocardial infarction. Reproduced from Asaria et al. 53 Copyright © 2016, BMJ Publishing Group Ltd and the British Cardiovascular Society. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.

Estimated costs over time for stable coronary artery disease patients

A patient with SCAD representative of the lowest-risk group (5-year cardiovascular risk of 3.43% and a life expectancy of 26.9 years) would have expected undiscounted costs of £9181 over 5 years, of which 57% would be CVD related, and undiscounted lifetime costs of £107,052, of which 60% would be CVD related. A representative patient of the highest-risk group (5-year cardiovascular risk of 43.7% and a life expectancy of 5.3 years) would have expected undiscounted costs of £21,872 over 5 years, of which 72% would be CVD related, and lifetime undiscounted costs of £38,606, of which 72% would be CVD related.

Figure 50 shows the predicted total and CVD-related costs over a 25-year period for average characteristics of risk groups 1 (lowest risk), 4, 7 and 10 (highest risk). Higher-risk patients with SCAD have higher initial costs, which are overtaken by the lower-risk patients as a result of greater mortality in the higher-risk groups, resulting in less time to accrue costs.

FIGURE 50.

(a) Expected total health-care costs and (b) CVD-related health-care costs over time for patients in cardiovascular risk groups 1 (lowest-risk group), 4, 7 and 10 (highest-risk group) in CALIBER patients with SCAD. Reproduced from Walker et al. 54 © 2016 Walker et al. 54 Published by Oxford University Press on behalf of the European Society of Cardiology. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.

Policy implications

Using these results, we calculated the maximum price that the NHS should be willing to pay for new treatments targeted at each risk group that reduce CVD hazards by between 10% and 40% (Figure 51). This maximum price increases with both increasing baseline risk and with the higher relative risk-reduction impact of treatments.

FIGURE 51.

Theoretical maximum price of new treatments by risk for cardiovascular events among CALIBER patients with SCAD. MI, myocardial infarction. Reproduced from Asaria et al. 53 Copyright © 2016, BMJ Publishing Group Ltd and the British Cardiovascular Society. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.

Discussion

In this study we address a fundamental gap in knowledge: what is the clinical, health service and economic burden associated with SCAD? Using data from a large, contemporary and representative population of UK patients with SCAD, we have shown that substantial health-care costs are likely to be incurred as a result of improved ACS survivorship and the ageing population. Moreover, 5-year and lifetime costs vary according to CVD risk, which may be predicted from baseline patient-level data.

Stable coronary artery disease patients might be considered to have ‘fallen off the radar’ of clinical interest: they are no longer in cardiac rehabilitation (mainly offered to those immediately after AMI or CABG), no longer on dual antiplatelet therapy (maximum current recommendation for therapy is 12 months) and, in most cases, discharged from specialist care, with prescribing adherence to secondary preventative medication waning. SCAD guidelines give little guidance about how such patients should be followed up in health care or how they should be appropriately risk stratified. 233,329 Our results clearly highlight the unmet need and the shortfalls of current approaches: they experience high use of primary care and frequent hospitalisations, incurring considerable ongoing costs. The number of primary care consultations, a mean of 10.76 per patient in the first year with SCAD, is higher than previous estimates for the average population (5.5 per year). 347 In the first year with SCAD, 37% of patients were hospitalised (26.9% for a CVD). This is markedly higher than recent estimates for the general population in one area of the UK (14.9%)348 and higher than that in the general population for similar age groups (23.4% and 21.2% of 60- to 74-year-old males and females, respectively, based on HES data).

Health-care utilisation is insufficient as a metric of the impact on the health-care system. It is also important to consider the costs associated with those resources to the NHS. Average costs of £3127 in the first year with SCAD are markedly higher than those in patients with no chronic conditions (£293) but comparable with other chronic conditions (e.g. diabetes £3036). 348 High costs in the first year following a non-fatal event (£9239) reflect the health-care resources required to treat that event.

We provide here, for the first time, estimates of lifetime health-care costs. In low-risk patients (5-year CVD risk of 3%), remaining lifetime health-care costs were £117,000 and remaining lifetime QALYs were 19.11 years; in the highest-risk group (5-year CVD risk of 44%), remaining lifetime health-care costs were £43,000 and remaining lifetime QALYs were 3.34 years. This is important because it illustrates and quantifies the general phenomenon that the highest costs per patient occur in the highest-risk patients and are concentrated towards the end of life.

Importantly, we provide new information on the price that the NHS should be willing to pay for new treatments to reduce CVD risk in SCAD patients. A new treatment giving a 20% hazard reduction in cardiovascular events would be cost-effective if priced < £61 per year in the lowest-risk group and £657 per year in the highest-risk group. The relatively high cost and modest efficacy of the newer treatments currently completing Phase III trials suggests that they should be targeted at the highest-risk patient groups in order to be cost-effective.

Our model is the first comprehensive lifetime model based on long-term EHR data modelling the full range of CVD end points and accounting for the interdependence of CVD risks in the SCAD population. We have demonstrated that such data can be used to model the long-term costs and health effects of chronic conditions. Furthermore, the size of such data sets (number of patients and the number of years of follow-up) and their scope (number of end points and risk factors captured) enable increasingly biologically plausible models of such conditions to be built.

The model allows policy-makers to quantify and understand both the health and the cost burden of SCAD, and serves as a basis for evaluating the cost-effectiveness of new treatments aiming to reduce CVD risk in this population. Our results suggest that, for the vast majority of SCAD patients, it is likely that low-cost interventions to improve adherence to existing secondary prevention drugs should be prioritised over high-cost new treatments. It is also notable from our results that, even among the groups with the highest CVD risk, more patients are predicted to die of non-CVD-related than CVD-related causes, highlighting the vital role of primary care in the holistic management of both cardiovascular and non-cardiovascular risk for these patients.

As the data sets underpinning this analysis are further enriched in terms of both length of follow-up and breadth of linked data sets, it will become feasible for models such as ours to be extended to capture the trajectory of prognostic risk factors over time. This would allow us to go beyond a focus solely on treatments that directly impact on CVD risk to address broader questions around the cost-effectiveness of interventions that target these underlying risk factors.

This study has limitations. First, outpatient appointment costs cannot currently be linked from HES, so we were unable to include them. As a result, all estimates are underestimates of the total health-care costs of this population. Second, the estimation of costs and consequences involved extrapolation over longer time periods than is currently available. This extrapolation is reflected in the uncertainty around the model predictions. Third, health-related quality-of-life data were not recorded in CALIBER and were drawn from external studies. Finally, changes in prognostic risk factors over time were not explicitly modelled; instead, the equations underpinning our model are informed by the baseline values of these risk factors.

Conclusions

Patients with SCAD account for significant use of health-care resources and incur substantial health-care costs, which can be predicted by individual 5-year CVD risk profiles. We provide a model for estimating the lifetime health-care costs and health outcomes of patients with SCAD. This model lends itself to informing decisions on commissioning, regulatory approval and the ongoing evaluation of new interventions. It also demonstrates the value of utilising linked EHRs for economic evaluation purposes. Improved ACS survivorship and an ageing UK population will require targeted resource allocation in anticipation of burgeoning demand.

Background

Disease registries have been invaluable in collecting information about a specific disease and its treatment in order to drive audit and research, thus improving the quality of care. AMI is exceptional among diseases in being one of the most ‘registered’, and in being a disease for which health-care interventions have significantly improved prognosis over the years.

Objectives

To evaluate missed opportunities for care in hospital and after discharge for AMI using CVD registries in the UK and other countries.

Methods

Analyses of data from national AMI registries in the UK, Sweden and the USA.

Results

(1) International comparisons identified significant differences in use of surgical interventions and use of some secondary prevention medications in the UK. Short-term mortality after AMI was markedly higher in the UK than in Sweden. (2) More than half of hospitalised STEMI patients eligible for nine evidence-based care components did not receive all components. Greater numbers of missed opportunities for care were associated with higher mortality. (3) Only 23% of smokers received smoking cessation interventions in the 3 months following discharge. Those who ceased smoking experienced fewer cardiac events than those who continued smoking.

Conclusions

Despite guideline recommendations, major opportunities to improve cardiovascular outcomes for ACS patients are being missed, during hospital stays and after discharge. International comparisons may identify key areas that could be targeted to improve outcomes in the UK.

The studies in this chapter relate to patients presenting for the first time with ACS (Figure 52).

FIGURE 52.

Overview of the 33 studies in our programme: presentation with ACS.

Introduction

Our comparative analysis of NSTEMI management in the UK, Sweden and the USA highlighted apparent underuse of invasive procedures and beta-blockers in the UK. How these potential missed opportunities for care influenced prognosis, however, could not be determined owing to the absence of studies that simultaneously examine care and outcomes. Additional uncertainty was introduced by differences in the hospitals contributing data to the national registries, with those in the USA being a self-selected subset of hospitals contributing on a voluntary basis, in contrast to the UK and Sweden where data contribution is mandated across all hospitals. Indeed, the UK and Sweden are the only countries in the world with continuous national ACS registries. These are linked to mortality records and therefore offer a unique opportunity for comparative effectiveness research between whole health systems, incorporating both treatment and outcome data. Until now, studies have not standardised the mortality of patients in one country by the case mix in another, resulting in international comparison studies between health systems.

Our objectives in the present study were (1) to evaluate the validity of comparing data from these two nationwide clinical registries; (2) to compare changes over time in the proportions of patients receiving effective interventions during hospitalisation and at discharge between 2004 and 2010; (3) to compare crude and case-mix-standardised 30-day mortality between the two countries, and in clinically important subgroups; (4) to estimate and compare time trends in case-mix-standardised mortality in Sweden and the UK; and (5) to explore the contribution of clinical care to any mortality difference.

Methods

The study population was drawn from all hospitals that provide AMI care in Sweden (n = 86) and the UK (n = 242), and contribute data on consecutive patients to Swedish Web-system for Enhancement and Development of Evidence-based care in Heart disease Evaluated According to Recommended Therapies/Register of Information and Knowledge about Swedish Heart Intensive care Admissions (SWEDEHEART/RIKS-HIA) and MINAP, respectively. Patients with a first AMI admission between 1 January 2004 and 31 December 2010 were included. Fields in SWEDEHEART/RIKS-HIA and MINAP were compared, leading to selection of 17 case-mix measures and 14 in-hospital treatment measures (Table 27). The primary clinical outcome was all-cause mortality within 30 days after hospital admission, as recorded in Swedish and UK national death registries (data available for 99.3% and 98.7% of patients, respectively).

Results

Eligible patients with AMI included 119,786 from Sweden and 391,077 from the UK (see Table 27). There was a lower proportion of STEMI in Sweden (32% vs. 40%) but indices of infarct severity were similar in Sweden and the UK. In patients with STEMI, primary PCI was more common in Sweden (59% vs. 22%) compared with the UK, where fibrinolysis was used more frequently. In both countries, use of primary PCI increased with time, whereas use of fibrinolysis decreased (Figure 53).

FIGURE 53.

Use of reperfusion. (a) Use of primary PCI; (b) use of fibrinolysis to treat STEMI in Sweden and the UK, by year. Reproduced from Chung et al. 55 Copyright © 2014 Chung et al. This is an open-access article distributed under the terms of CC BY. Published by Elsevier Ltd. This document may be redistributed and reused, subject to certain conditions.

For patients surviving to hospital discharge, Swedish patients were more likely than UK patients to be prescribed dual antiplatelet therapy (68% vs. 61%) and beta-blockers (89% vs. 78%), but less likely to be prescribed ACEIs or ARBs (56% vs. 82%) and statins (80% vs. 93%). The cumulative 30-day mortality was 7.6% (95% CI 7.4% to 7.7%) in Sweden and 10.5% (95% CI 10.4% to 10.6%) in the UK (Figure 54).

FIGURE 54.

Kaplan–Meier curves for cumulative mortality at 30 days after admission with AMI in Sweden and the UK. Reproduced from Chung et al. 55 Copyright © 2014 Chung et al. This is an open-access article distributed under the terms of CC BY. Published by Elsevier Ltd. This document may be redistributed and reused, subject to certain conditions.

In the UK, 30-day mortality was consistently higher in strata defined by measures of infarct severity (troponins, heart rate and BP), age, sex, calendar year, smoking and diabetes status. Standardising the 30-day mortality of UK patients using the Swedish case mix reduced forecast mortality to 7.7% (95% CI 7.3% to 8.2%) with a standardised mortality ratio of 1.37 (95% CI 1.30 to 1.45), corresponding to an estimated 11,263 (95% CI 9620 to 12,827) more deaths in the UK between 2004 and 2010. However, standardised mortality ratios decreased progressively from 1.47 (95% CI 1.38 to 1.58) in 2004 to 1.20 (95% CI 1.12 to 1.29), p-value for trend 0.01 in 2010 (Figure 55). Propensity-matched analyses gave similar findings.

FIGURE 55.

Thirty-day mortality of UK patients admitted in each study year, standardised according to the Swedish case-mix model. p = 0.01 for linear trend across years for the relative risks. Reproduced from Chung et al. 55 Copyright © 2014 Chung et al. This is an open-access article distributed under the terms of CC BY. Published by Elsevier Ltd. This document may be redistributed and reused, subject to certain conditions.

Discussion

We found greater cumulative 30-day mortality after AMI in the UK (10.5%) than in Sweden (7.6%). Case-mix standardisation attenuated the mortality difference. The findings suggest that > 10,000 early deaths would have been prevented or delayed had UK patients experienced the care of their Swedish counterparts. We believe that the comparison between Sweden and the UK was valid because, for both countries, we captured consecutive AMI admissions across all hospitals, using common data definitions, and we showed that the diagnosis of AMI was comparable, with similar troponin values and similar propensity to make a diagnosis.

Variation in the way patients were managed in the UK and in Sweden may have contributed to the 30-day mortality differences. Use of primary PCI for STEMI patients in the UK lagged behind Sweden; the rate in 2010 in the UK (53%) was similar to that of Sweden in 2005. Primary PCI is more effective than fibrinolysis and was incorporated into US and European guidelines in 2004 and 2005, respectively; however, the UK did not have a national policy for primary PCI until 2008. This probably explains the rapid increase in primary PCI at that time, although only in 2011–12 did rates exceed 90%. 350 The use of evidence-based secondary prevention medication showed a mixed pattern, with statins and ACEIs/ARBs being more commonly prescribed in the UK, whereas beta-blocker use on discharge was more common in Sweden. The mortality gap between Sweden and the UK decreased over time, from a standardised mortality ratio of 1.47 in 2004 to 1.20 in 2010, consistent with a narrowing of gaps in treatment differences. Other unmeasured aspects of health care, which probably contributed to the mortality gap, include the dose, timing and adherence to drugs, differences in operator experience, shared and specialty care pathways, use of decision support tools and organisational culture. 351

This study comparing the delivery of health care in different health-care systems has demonstrated differences in short-term outcomes for AMI patients; however, many patients remain at increased risk beyond this period. We analysed linked EHR and disease registry data in four countries {England (CALIBER), Sweden [Health Outcomes, Life Expectancy, Resource Use and Costs in Patients With a History of Myocardial Infarction and Additional Risk Factors for Atherothrombosis (HELICON)], the US (Medicare) and France [Echantillon Généraliste des Bénéficiaires (EGB)]} to investigate longer-term CVD risk among 140,880 patients surviving to the first anniversary of their heart attack. Our study found that cumulative risk of AMI, stroke, CVD death, all-cause mortality and hospitalised bleeding remained high for 3 years across all four countries, and predictors of risk had similar strong positive associations with the outcomes in each country. This work further highlights the power and generalisability of prognosis research that utilises the EHRs across different health-care systems. 352 Linkage to national EHRs will extend comparisons to include ambulatory care before and after admission for heart attack, non-fatal events and longer-term outcome.

Policy initiatives are required to identify, understand and reduce gaps in treatments and outcomes between health systems. We suggest that progress towards this goal is achievable, as mortality differences are narrowing and are, therefore, reversible. The study suggests a novel approach of between-country comparison; to date, quality outcome initiatives have emphasised within-country, within-system metrics. For Sweden, these results provide further evidence of the value of sustained, system-wide quality improvement initiatives, including public reporting of mortality outcomes at hospital level. For the UK, these results suggest the value of learning from systems that appear to be better performing.

International harmonisation of detailed measures of quality of care in clinical registries, such as pathways of care, is needed to identify and evaluate additional patient-level health-care factors that are not measured here that might explain differences between countries. Outcomes should be compared with those in other countries, such as France, Hungary, Poland and the USA, through assessment of national registries.

Limitations include the possibility of unmeasured confounders and the potential for patients not captured in national registries – who are at higher risk – to distort the findings. However, such missed cases may be less common in Sweden, where nationwide registries have a more established history than in the UK, suggesting the mortality difference may actually be wider.

Conclusions

In conclusion, clinically important differences in the care and outcomes of AMI are observed between Sweden and the UK. International comparisons of care and outcome registries may inform new research and policy initiatives to improve the quality of health systems.

Introduction

There are many guideline-recommended acute interventions for patients with ACSs. Monitoring adherence to guidelines is important in order to continually improve the quality of clinical care, and the MINAP registry is invaluable in this respect. In this study we used MINAP to evaluate the association between cumulative missed opportunities for care (CMOCs) and mortality in patients hospitalised with STEMI.

Methods

The MINAP study population was drawn from 121,567 consecutive patients aged ≥ 18 years presenting for the first time with STEMI to hospitals in England and Wales from 1 January 2007 to 31 December 2010 and surviving to be discharged. Mortality data (30-day and 1-year all-cause) were obtained from the ONS. We identified nine guideline-recommended care opportunities recorded in MINAP along the pathway of STEMI care: (1) the recording of a pre-hospital ECG; (2) use of acute aspirin; (3) timely coronary reperfusion (within 120 minutes or thrombolysis within 60 minutes of the call for help); (4) the individual prescriptions of aspirin; (5) the individual prescription of thienopyridine inhibitor; (6) the individual prescription of beta-blockers; (7) the individual prescription of an ACEI or ARB; (8) the individual prescription of statins; and (9) enrolment into a cardiac rehabilitation programme at the time of discharge from hospital. 353–355 A CMOC score was calculated for each patient, based on the number of these nine components that were not provided. To overcome instability due to small numbers, patients were categorised into CMOC groups of 0, 1, 2, 3 and ≥ 4 missed care opportunities.

Results

A total of 112,286 patients were discharged alive from hospital following STEMI, of whom 62,267 were eligible for all nine care interventions. Of the 31,187 (50.1%) eligible for all nine care interventions who had complete data, 49.4%, 36.4%, 11.5%, 1.9% and 0.8% had CMOC scores of 0, 1, 2, 3 and ≥ 4, respectively (Table 28). No patient eligible for all care opportunities missed more than eight components. Patients with higher CMOC scores more frequently were older and female. Patients also had a greater burden of comorbidity, higher GRACE risk scores and less frequently received coronary angiography.

Figure 56 shows that acute aspirin therapy was the least frequent contributor to CMOC, whereas pre-hospital ECG and timely reperfusion were commonly included in CMOC. Both of these upstream interventions independently predicted further downstream missed opportunities [pre-hospital ECG: incident rate ratio (IRR) 1.64, 95% CI 1.58 to 1.70; and timely reperfusion: IRR 9.94, 95% CI 9.51 to 10.40].

FIGURE 56.

Profile of each missed opportunity along the pathway of STEMI care, by CMOC group for patients eligible for all nine care opportunities in MINAP. Reproduced with permission from Simms et al. 56 © 2015, © SAGE Publications.

Hospitals with ≤ 30 cardiology beds had significantly higher CMOC scores than those with > 50 beds. Hospital volume of STEMI was negatively associated with CMOC score but there was no significant association between CMOC score and whether or not hospitals had on-site coronary angiography/PCI facilities.

Risk-adjusted mortality rates rose with increasing CMOC (Figure 57). Kaplan–Meier survival estimates stratified by CMOC also showed a positive association between CMOC and mortality (Figure 58), and this was confirmed by the risk-adjusted Cox analysis (Figure 59).

FIGURE 57.

Crude 30-day and 1-year mortality rates and 95% CIs by number of ineligible care for patients eligible for nine evidence-based care components in MINAP.

FIGURE 58.

Kaplan–Meier unadjusted survival estimates for patients with STEMI, stratified by CMOC group, for patients eligible for all nine evidence-based care components in MINAP. Reproduced with permission from Simms et al. 56 © 2015, © SAGE Publications.

FIGURE 59.

Forest plot of adjusted risk of 30-day and 1-year mortality, by CMOC group, for patients eligible for all nine evidence-based care components in MINAP. Reproduced with permission from Simms et al. 56 © 2015, © SAGE Publications.

Discussion

This study shows that missed opportunities during the patient journey extend beyond primary care and are common following hospital admission in patients with STEMI. More than half of this high-risk group failed to receive at least one of nine components of care recommended in national guidelines. We found a significant relationship between CMOC score and early mortality, which remained significant after 1 year. The data suggest that CMOC represents an important target for improving quality of care and outcomes.

Patients with higher CMOC scores were at greater baseline risk and had more comorbid conditions. Failure to provide upstream interventions – pre-hospital ECG and timely coronary reperfusion – significantly predicted downstream missed opportunities. The association between CMOC score and mortality remained strong, despite adjustment for case mix and invasive treatments, suggesting that it was driven more by hospital factors than the risk characteristics of the patients. Thus, our data identified underprovision of designated cardiology beds and larger catchment populations as hospital factors significantly associated with CMOC. However, hospital associations with CMOC were complex, with low-volume centres reporting more missed components of the care pathway. Overall, our findings suggest that adequately resourced, high-throughput, high-volume hospitals (heart attack centres) with appropriate facilities may offer superior care and may, therefore, deliver better outcomes. Outcomes research specifically designed to follow care into the primary care setting is warranted to quantify the full implication of CMOCs.

Conclusions

Of patients with STEMI eligible for nine evidence-based care components, over half did not receive all components. Missed opportunities for care were significantly associated with early- and longer-term mortality. Outcomes for STEMI would likely improve if CMOC was reduced and physicians provided all components of the pathway of care to their patients.

Introduction

Our analysis of missed opportunities for care along CALIBER’s patient journey has focused largely on deviations from guideline recommendations in the management of patients with CVD, both in primary care and hospital settings. However, registry-based research offers an alternative means of evaluating care through the comparison of management strategies in different countries. To date, there have been few international comparisons of care for patients with NSTEMI, and none that reflects the recent changes in clinical management recommended in international guidelines.

We have established a collaboration with investigators in Sweden and the USA in order to compare patterns of in-hospital treatment and use of interventional diagnostic and therapeutic procedures among patients admitted with NSTEMI across three national clinical registries: MINAP in the UK, SWEDEHEART/RIKS-HIA in Sweden and Acute Coronary Treatment and Intervention Outcomes Network Registry (ACTION) in the USA. 7,358 SWEDEHEART/RIKS-HIA and MINAP cover all hospitals providing care for ACS in Sweden and the UK, respectively; ACTION is a voluntary registry that includes patients in a large but self-selected group of hospitals in the USA. Data validity checks in SWEDEHEART/RIKS-HIA, MINAP and ACTION have demonstrated agreement rates of 96.1%, 89.5% and 89.7%, respectively. 7,86,359

Methods

The study population was drawn from all hospitals providing AMI care in England and Wales (236 hospitals, 137,009 patients), Sweden (74 hospitals, 45,069 patients) and a voluntary subset of hospitals, most with the capability to perform PCI, in the USA (500 hospitals, 147,438 patients). We included patients admitted between 1 January 2007 and 31 December 2010, and aged ≤ 30 years with first recorded NSTEMI. The distribution of case-mix (demographics, past history and presentation characteristics) and treatment variables was compared in the UK, Sweden and the USA. Recognising the difference between nationwide and voluntary registries, we conducted a sensitivity analysis, comparing results of NSTEMI patients attending PCI hospitals only. We defined PCI hospitals as those performing at least 24 procedures in a calendar year.

Results

Patients were youngest in the USA, where rates of smoking, diabetes, hypertension, prior heart failure and prior AMI were higher than in Sweden or the UK (Table 29). Prior use of medications, PCI and CABG was also higher in the USA. Angiography and PCI were performed more often in the USA and Sweden than in the UK, but the rate of increase in these procedures from 2007 to 2010 was greatest in the UK (Table 30). Similarly, PCI was performed more in the USA. Prescription of dual antiplatelet therapy was similar across the countries. UK physicians prescribed beta-blockers at discharge least often but statins and ACEIs/ARBs most often.

Discussion

This comparison of ongoing national registries showed important differences in the application of guideline-recommended treatments in the UK compared with in Sweden and the USA, despite close similarities in case mix. The data suggest missed opportunities for care on a national scale reflected in the UK’s less aggressive use of interventional procedures for NSTEMI, although the gap with Sweden and the USA narrowed over time. In contrast, secondary prevention medications (except beta-blockers) were more commonly prescribed in the UK than in Sweden or the USA.

Clinical trials have previously shown wide variation in practice among countries or groups of countries in the management of AMI patients. 360–362 However, management patterns within a clinical trial do not necessarily reflect management in usual clinical practice. 363 Clinical registries address some of the selection bias and are valuable resources for comparison and for assessing trends in treatment. 363–367 The finding of similarly high PCI rates in an unselected (SWEDEHEART/RIKS-HIA) and a selected (ACTION) registry makes the differences seen in other aspects of NSTEMI care more informative. It suggests that US procedure rates are not being driven by fee for service and that Sweden’s increased emphasis on system-wide quality improvement is bearing fruit. 368,369 The lower rates of interventional management in the UK provide further evidence of slow implementation of emerging technologies, although the steeper increase in in-hospital procedure rates between 2007 and 2010 suggests that concerted efforts to increase PCI capability have been effective. 370

The lower use of beta-blockers in the UK than in the USA and Sweden was accompanied by higher use of statins and cannot, therefore, be driven by economic factors or by delayed implementation (the evidence for beta-blockers in AMI pre-dates that for statins). One possibility is that UK patients had more perceived contraindications, such as COPD, although our CALIBER study (see study 27 below) has shown that beta-blockers are beneficial and safe in patients with COPD who present with AMI. 59

Further studies are warranted to better explain patterns of clinical practice within and outside the UK, such as rates of coronary angiography and PCI.

Conclusions

Differences exist in the acute management of NSTEMI among patients in national registries from the UK, Sweden and the USA. Specifically, the use of invasive procedures such as angiography and PCI and of beta-blockers on discharge was higher in the USA and Sweden than in the UK. Differences are becoming smaller but need to be understood if areas for improvement are to be identified.

Introduction

In mapping the CVD journey in CALIBER, we had the unique opportunity to examine patient care in the community after discharge from hospital, recognising that sustained lifestyle modification and secondary prevention treatment is a key guideline recommendation in patients who have experienced AMI. Perhaps the most important lifestyle modification is smoking cessation, which substantially reduces the risk of recurrent AMI and coronary death. 371 Recommendations are for GPs to offer smoking cessation advice and specific treatment to all smokers after ACS,330 but the extent to which this opportunity to reduce mortality is missed has not been clarified. To answer this question, we used linked CALIBER records to analyse the extent of smoking cessation advice and pharmaceutical interventions for smoking cessation provided by GPs after AMI and to describe patient outcomes stratified by smoking status.

Methods

We studied 4834 patients aged ≥ 30 years discharged alive from hospital after AMI between 1 January 2002 and 30 September 2009. Patients were followed to death, transfer out of the general practice or the last data collection date. Smoking interventions recorded in CPRD were classified as follows: (1) prescribed pharmacological interventions for smoking cessation; (2) GP smoking cessation advice or referral to stop-smoking clinic (without pharmaceutical interventions); and (3) no GP intervention.

Major clinical outcomes were explored in patients who smoked prior to their ACS, stratified by quitting status. Outcomes included death, recurrent AMI, stroke, heart failure and major adverse cardiac events, defined as death, recurrent ACS, CABG or target lesion revascularisation.

Results

Characteristics of patients are shown in Table 31. Of the 4834 patients, 33% were non-smokers, 44% were ex-smokers and 20% were smokers.

In the 3 months following discharge, over three-quarters of smokers received no smoking intervention, despite 99% of patients having had some contact with their GP (Table 32). The 320 (33%) smokers who quit in the first 3 months after ACS experienced significantly lower rates of death and major adverse cardiac events than non-quitters (Figure 60).

FIGURE 60.

Relative rate of major clinical outcomes comparing smokers who quit in the first 3 months after ACS compared with people who continued to smoke in CALIBER. MACE, major acute coronary events. Reproduced from Boggon et al. 58 © 2014 SAGE publications. This article is distributed under the terms of the Creative Commons Attribution License (www.creativecommons.org/licenses/by-nc/3.0/), which permits non-commercial use, reproduction and distribution of the work without further permission provided the original work is attributed as specified on the SAGE and Open Access page (www.uk.sagepub.com/aboutus/openaccess.htm).

Discussion

This CALIBER study provides a unique longitudinal perspective on smoking cessation interventions after hospitalisation for AMI. The results show that only a minority of patients received interventions for smoking cessation and a large number continued to smoke. This missed opportunity for care is important because patients who continued to smoke had a significantly worse prognosis than those who quit.

In the first 3 months after ACS, when motivation to stop smoking was likely to be high, only 23% of smokers received smoking cessation interventions. The benefits of smoking cessation interventions are well established372–375 and the demonstrated failure of adherence to guideline recommendations for smoking cessation interventions no doubt contributed to the high proportion of smokers who failed to quit after ACS and who were thereby exposed to an increased risk of adverse outcomes.

To further understand the association between interventions and cessation and the effects on long-term clinical outcomes, future research into the promotion of GP smoking cessation interventions in both larger prospective epidemiological studies and RCTs is warranted.

This study has limitations. Smoking cessation advice may have been provided without being recorded in the patient record. 376 GP recording of smoking status and of advice against smoking may have changed following the introduction in 2004 of the Quality and Outcomes Framework contract, which offers financial incentives for the recording of smoking status and advice. 377,378

Conclusions

This CALIBER study provides evidence that smoking is common among those patients hospitalised for ACS. However, despite NICE guideline recommendations, few GPs offer smoking cessation advice, referral to smoking clinics, or pharmacological intervention in the 3 months following ACS discharge. Given the lower rates of major clinical outcomes, including mortality and major adverse cardiac events, among those who quit smoking following ACS, there remains a significant opportunity for GPs to actively engage smokers in cessation interventions.

Background

Guidelines recommending use of secondary prevention drugs after AMI are based on results from clinical trials. However, trials do not provide evidence of long-term effectiveness of treatments or impact of interactions between them. Analysis of EHRs may help address these questions.

Objectives

To examine the use, timing, potential drug interactions and long-term effectiveness of secondary prevention medications after AMI.

Methods

An analysis of CALIBER linked data.

Results

(1) Among patients who had experienced an AMI, prognosis was most favourable in patients given a discharge prescription including all five classes of secondary prevention drugs. Use of fewer drugs was associated with increased mortality, particularly within the first 30 days and where beta-blockers, ACEIs and statins were omitted. (2) Use of a beta-blocker during and before AMI admission in patients with COPD was associated with reduced mortality. (3) Among AMI patients receiving clopidogrel plus aspirin, use of a PPI was associated with increased risk of AMI and death, as was use of drugs not predicted to interact with clopidogrel. A case series found no evidence that PPIs increased risk of death. (4) Discontinuation of clopidogrel after AMI is common and is associated with adverse outcomes.

Conclusions

‘Real-world’ data can provide important information beyond that generated by clinical trials to guide prescribing decisions.

The studies in this chapter relate to patients presenting for the first time with ACS (see Figure 52).

Introduction

Guideline recommendations for secondary prevention drugs after AMI include aspirin, clopidogrel, beta-blockers, ACEIs and statins. 353,395,396 For all but clopidogrel, the recommendation is for life-long treatment, on the assumption that benefits are additive and efficacy remains undiminished in the longer term. However, none of the trials was designed to measure the additive effects of treatment or the relative efficacy of the various drugs. Moreover, follow-up in these trials was rarely longer than 3 years and often shorter. It is open to question, therefore, whether indefinite treatment is necessary or whether some of the recommended drugs might be discontinued, with potential benefits for patient safety and health-care costs.

A number of observational studies have suggested that the benefits of secondary prevention drugs after AMI are additive, mortality declining with the number prescribed. However, these studies are likely to be confounded by healthy user bias and are too small to answer the question of comparative effectiveness of specific drug combinations, with most enrolling fewer than 6000 patients. 397–401 The MINAP registry is substantially larger and has provided, for the first time, the opportunity to evaluate the comparative effectiveness of aspirin, clopidogrel, beta-blockers, ACEIs and statins individually and in all combinations.

Methods

We included patients aged ≥ 30, with a first STEMI or NSTEMI in the period 1 January 2004 to 6 August 2009 (n = 316,648). MINAP records if patients were discharged on aspirin, clopidogrel, an ACEI or ARB (conflated as ACEI in this study), a beta-blocker and a statin. Primary outcomes were survival from discharge to 30 days; survival from 30 days after discharge to 1 year; and survival beyond 1 year.

Results

The study group comprised 287,615 patients discharged from hospital with diagnoses of STEMI (39%) or NSTEMI (61%). Rates of omission of secondary prevention drugs were 10% and 11% for aspirin and statins, increasing to 21% and 26% for ACEIs and beta-blockers, respectively, and to 45% for clopidogrel (Table 33). Patients in whom any of the secondary prevention drugs were omitted were older than the mean age of the group, more frequently female and more likely to have a discharge diagnosis of NSTEMI.

The Kaplan–Meier landmark analysis showed that in each of the three time periods (first 30 days, 31 days to 1 year, beyond 1 year) there was progressive worsening of crude and adjusted survival as the number of omitted secondary prevention drugs increased (Figure 61). Only the group with all five secondary prevention drugs omitted deviated from the dose–response relationship, exhibiting a survival trajectory intermediate between two and three omitted drugs.

FIGURE 61.

Kaplan–Meier mortality curves stratified by number of omitted secondary prevention drugs prescribed at discharge at 30 days and 1 year, and prevalence of omission compared with no omitted drugs among patients with a first STEMI or NSTEMI in MINAP.

Adjusted HRs for mortality varied according to which drug was omitted from the discharge prescription, tending to be higher for omission of either a beta-blocker, an ACEI or a statin compared with omission of aspirin or clopidogrel (Table 34).

Hazard ratios for drug omissions diminished sharply beyond the first 30 days, although this was less marked for beta-blockers or statins. When patients were grouped by the number of omitted drugs, Cox analysis showed a dose–response relationship of increasing hazard for 30-day mortality but, within each group, HRs varied according to which drugs were omitted (Figure 62a). Compared to patients discharged on all five secondary prevention drugs, those with single omission of clopidogrel showed a similar mortality hazard (HR 1.00, 95% CI 0.92 to 1.07), whereas those with single omission of aspirin showed a small excess hazard (HR 1.18, 95% CI 1.07 to 1.30). Single omissions of beta-blockers, ACEIs or statins were associated with a greater hazard, all with point estimates > 1.33. This pattern repeated itself as the number of omitted secondary prevention drugs increased, with combinations of omission that included aspirin and clopidogrel tending to be associated with a lower hazard of 30-day mortality than combinations that included beta-blockers, ACEIs or statins. The pattern was similar for mortality between 30 days and 1 year but, beyond 1 year, associations between specific combinations of discharge drug omissions and hazard of death were less clear (Figure 62b).

FIGURE 62.

Adjusted HRs and 99% CIs for death at 30 days and beyond 1 year associated with omission of combinations of secondary prevention medications among patients with a first STEMI or NSTEMI in MINAP. (a) Death at 30 days; and (b) death beyond 1 year.

Discussion

The relationship between the number of secondary prevention drugs omitted from the discharge prescription of patients with AMI and subsequent mortality depends on which drugs are omitted. Effect sizes were greatest in the first 30 days and were most evident for omission of beta-blockers, ACEIs or statins, which increased the hazard of death by > 40%. HRs tended to decrease in the longer term, particularly for omission of ACEIs, but for omission of beta-blockers and statins the hazard of death remained > 27% after 1 year of follow-up. Omission of aspirin from the discharge prescription caused only a small increase in the hazard of death, the point estimate dropping below 10% after the first 30 days, whereas omission of clopidogrel had little discernible effect on mortality.

The present study confirmed an association between the number of prescribed secondary prevention drugs and mortality,397–401 which persisted after full adjustment for all covariates. However, the size of this national cohort allowed us to analyse for the first time the effects of all possible combinations of drug omissions on mortality. The data showed important class effects, omissions of antiplatelet agents having less effect on the dose–response relationship than other secondary prevention drugs, particularly beta-blockers and statins. This was most evident for mortality in the first 30 days, when treatment effects were least likely to be distorted by non-adherence and the influence of new comorbidities, but the pattern remained perceptible at 1 year and beyond.

We found that single omission of aspirin from the discharge prescription of patients who received all other secondary prevention drugs was associated with a 30-day mortality increase of about 13%, falling to < 5% beyond the first year. Moreover, combinations of omitted drugs that included aspirin tended to increase the hazard of death less than combinations including beta-blockers and statins. The data reflect aspirin’s short-term protective effect, demonstrated in the Second International Study of Infarct Survival (ISIS-2) trial,403 but also highlight the caution that others have counselled404 in generalising randomised trial data beyond parameters of patient selection and treatment duration. Thus, patients randomised to aspirin in ISIS-2 received just 6 weeks of treatment and, although sustained benefits for this group have been reported,405 our own landmark analysis shows significant diminution of benefit in the longer term.

Unlike aspirin, omission of beta-blockers or statins from the discharge prescription was associated with a marked increase in the hazard of death in the first 30 days, with only minor reductions in the effect size during the first year and beyond. Our findings provide a strong case for long-term treatment with these drugs after AMI. Omission of ACEIs also increased the early hazard of death but, unlike beta-blockers and statins, the effect size diminished sharply beyond 1 year, when combinations of discharge prescriptions that omitted ACEIs were associated with a < 10% increase in the hazard of death. Our study provides no mechanistic insights, but it is possible to speculate that the pattern of diminishing hazard we found with omission of ACEIs from the discharge prescription reflects the protection they provide against left ventricular remodelling, which has its main impact during the first few months after AMI. 406

We were unable to demonstrate any association of clopidogrel omission with increased mortality and, like aspirin, combinations of omitted drugs that included clopidogrel tended to increase the hazard of death less than combinations including beta-blockers and statins. This is consistent with clinical trials that have shown reductions in composite end points for patients randomised to clopidogrel without clear effects on death from cardiovascular or non-cardiovascular causes. 407 However, in the current era of primary PCI, there is good evidence that clopidogrel can protect against stent thrombosis in the first year after AMI, although guideline recommendations to discontinue treatment beyond the first year appear borne out by our findings.

Combinations of drugs that are recommended lifelong but have not been evaluated in RCTs should be investigated in observational studies for associations with fatal and non-fatal outcomes over the remaining life expectancy.

Conclusions

Our findings remind clinicians and guideline groups that, although randomised trials are the most robust means of evaluating the efficacy of specific treatments, information about comparative effectiveness is available only from trials that randomise patients to two or more different agents. Examples include ISIS-2, which showed comparable effectiveness for streptokinase and aspirin,403 and Gruppo Italiano per lo Studio della Sopravvivenza nell’Infarto Miocardico (GISSI)-3, which showed effectiveness limited to lisinopril, not transdermal nitrate. 408

Our study highlights the variable consequences of specific drug omissions and reminds clinicians of the need to set thresholds for omitting beta-blockers or statins from the discharge prescription at a high level in order to avoid substantial long-term increases in the hazard of death. Our study also shows that the diminution of effectiveness apparent for all secondary prevention drugs in the longer term appears more marked for aspirin and ACEIs. Nevertheless, small reductions in the hazard of death beyond 1 year persist for both these drug classes, and our data do not provide a basis for recommending their discontinuation.

This study confirms that in patients with AMI, a discharge prescription that includes all five classes of secondary prevention drugs is associated with the most favourable prognosis, but suggests that beta-blockers and statins contribute the greatest protection against death, particularly in the longer term. The effectiveness of all drug classes tends to diminish with time but this is particularly marked for aspirin and ACEIs.

Introduction

Patients with COPD have lower survival after AMI than patients without COPD. 406,409–412 Historical concerns that beta-blockers might be harmful in COPD probably contribute to their underuse in AMI, even though these concerns have been challenged by recent evidence. 413 Hence, the underuse of beta-blockers in COPD patients with AMI may be an important missed opportunity to improve cardiovascular outcomes. We used the CALIBER linked data sets to quantify the association between COPD and mortality after AMI to investigate whether or not the use and timing of prescription of beta-blockers in patients with COPD after a first AMI was associated with improved survival, and to identify factors related to the use of beta-blockers in patients with COPD.

Methods

We included patients with COPD experiencing their first AMI from 1 January 2003 to 31 December 2008 as recorded in MINAP. Patients were defined as having COPD if the Read-coded diagnosis was recorded in CPRD before or after AMI. Two or more exacerbations per year were defined as frequent exacerbations. 414 Information on COPD and drug prescriptions was obtained from CPRD. The primary outcome was all-cause mortality after AMI in patients with COPD stratified by beta-blocker prescription during hospital admission. We used Cox proportional hazards models to produce adjusted HRs to determine whether or not mortality differed by prescription of a beta-blocker.

Results

We included 1063 COPD patients with a MINAP record of AMI during the study period. Over half (55%) were never prescribed a beta-blocker and 23% were on a beta-blocker before AMI. The remainder (22%), who were prescribed a beta-blocker at the time of AMI, were younger and less likely to have a history of hypertension, cerebrovascular disease, PAD, heart failure, dyslipidaemia or angina before their AMI. They were also more likely to have infrequent exacerbations and were less likely to have been prescribed diuretics before their AMI. Of the beta-blockers prescribed during hospital admission, 93.8% were cardioselective.

After adjustment for history of smoking and sex, and stratification by age, patients with COPD prescribed a beta-blocker during hospital admission for AMI and those prescribed a beta-blocker before an AMI had better survival than those never prescribed a beta-blocker (HR 0.45, 95% CI 0.34 to 0.60; p < 0.001; and 0.72, 95% CI 0.57 to 0.90; p = 0.004, respectively) (Figure 63).

FIGURE 63.

Kaplan–Meier estimated survival and 95% CIs after AMI in patients with COPD prescribed a beta-blocker before MI, during hospital admission for myocardial infarction or never, excluding deaths on day of diagnosis, in MINAP. MI, myocardial infarction. Reproduced from Quint et al. 59 Copyright © 2013, British Medical Journal Publishing Group. This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by-nc/3.0/), which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited.

After adjustment for age, sex, history of smoking, and frequency of exacerbations, survival analysis with cardiac and non-cardiac death as the outcome showed an improved survival in those prescribed a beta-blocker than those not prescribed a beta-blocker during the hospital admission for AMI (HR for cardiac deaths 0.57, 95% CI 0.38 to 0.86; p = 0.03; and non-cardiac deaths 0.49, 95% CI 0.32 to 0.75; p = 0.01). Among patients prescribed a beta-blocker during hospital admission, 203 (87.1%) were still taking it at 6 months, 185 (79.4%) at 1 year and 142 (60.9%) at 2 years.

Discussion

This is the largest study of the effect of beta-blockers on survival after AMI in patients with COPD. 414–416 Use of a beta-blocker after a first AMI was associated with lower subsequent mortality, a benefit that persisted after adjustment for age and cardiovascular comorbidities. Only 38.6% of people with COPD received a beta-blocker during the hospital admission for AMI, however, indicating an important missed opportunity for improving care. Despite earlier studies reporting improved survival after AMI in COPD patients prescribed beta-blockers,415 the reluctance to prescribe remains a worldwide phenomenon that cannot be attributed to problems with tolerability, our own data showing that nearly 80% of patients prescribed a beta-blocker were still taking it at 1 year.

The lower mortality in COPD patients prescribed a beta-blocker was due to a reduction in both cardiac and non-cardiac deaths, consistent with previous suggestions that their benefits extend beyond cardiac effects. 416,417 There is evidence that beta-blockers do not worsen lung function in patients with COPD,416,417 so severity of disease or lack of spirometry should not deter their prescription in AMI, particularly use of cardioselective agents. A RCT is justified to examine the effectiveness of beta-blockers after AMI in patients with COPD.

The main limitation of our study is that residual confounding, including confounding by contraindications, cannot be excluded, and there is also the risk of healthy user bias.

Conclusions

Our data suggest that beta-blockers should be used more widely in patients with COPD who have had an AMI. Low rates of use represent an important missed opportunity for treatment. Safety to date is good in these patients, but further evaluation of the safety of beta-blockers in this high-risk group might be required to change current prescribing practice.

Introduction

In patients with AMI receiving dual antiplatelet therapy with aspirin and clopidogrel, the risk of gastric bleeding can be reduced by coprescription of a PPI. However, PPIs inhibit cytochrome P450 2C19, the enzyme that converts clopidogrel to its active form. This interaction reduces clopidogrel activity as measured by platelet activation studies, but whether or not this has important clinical effects has been hard to determine, with conflicting evidence from outcome studies. 382–393 We used CALIBER linked data sets to conduct two observational studies to investigate this possible drug interaction, the first using a traditional cohort design, and the second using a self-controlled case series design to eliminate the fixed confounding between people that can affect case–control or cohort designs. 418

Methods

Patients were selected if they were recorded in the CPRD from 1 January 2003 onwards and had at least 12 months between first registration and first recorded prescription for clopidogrel. Patients also had to be concurrently prescribed aspirin as recorded in the database. Follow-up was censored at the earliest of stopping treatment with clopidogrel, stopping treatment with aspirin, occurrence of an outcome of interest, death, transfer out of the practice, last data collection date for the practice or 31 July 2009. Follow-up time was then classified according to whether or not the patient was receiving a PPI as recorded in the database. Treatment with aspirin was assumed to be continuous but for clopidogrel and PPI exposure, where possible, we calculated the anticipated length of each prescription based on pack size and dosing instructions. We undertook an observational cohort study and a self-controlled case series. The primary outcome was a composite end point of all-cause mortality or incident AMI.

Results