The cost-effectiveness of testing strategies for type 2 diabetes: a modelling study

Article history

The research reported in this issue of the journal was funded by the HTA programme as project number 10/133/01. The contractual start date was in July 2012. The draft report began editorial review in November 2013 and was accepted for publication in September 2014. The authors have been wholly responsible for all data collection, analysis and interpretation, and for writing up their work. The HTA editors and publisher have tried to ensure the accuracy of the authors’ report and would like to thank the reviewers for their constructive comments on the draft document. However, they do not accept liability for damages or losses arising from material published in this report.

Declared competing interests of authors

Professor Kamlesh Khunti has acted as a consultant and speaker for Novartis Pharmaceuticals UK Ltd, Novo Nordisk, Sanofi-aventis, Eli Lilly and Company and Merck Sharp & Dohme Corp. He has received grants in support of investigator and investigator-initiated trials from Novartis Pharmaceuticals UK Ltd, Novo Nordisk, Sanofi-aventis, Eli Lilly and Company, Pfizer, Boehringer Ingelheim and Merck Sharp & Dohme Corp. Professor Kamlesh Khunti has also received funds for research and honoraria for speaking at meetings and has served on advisory boards for Eli Lilly and Company, Sanofi-aventis, Merck Sharp & Dohme Corp. and Novo Nordisk. Professor Melanie Davies has acted as consultant, advisory board member and speaker for Novo Nordisk, Sanofi-aventis, Eli Lilly and Company, Merck Sharp & Dohme Corp., Boehringer Ingelheim, Janssen and AstraZeneca. She has received grants in support of investigator and investigator-initiated trials from Novo Nordisk, Sanofi-aventis, Eli Lilly and Company, Boehringer Ingelheim, Merck Sharp & Dohme Corp. and GlaxoSmithKline.

Copyright statement

© Queen’s Printer and Controller of HMSO 2015. This work was produced by Gillet et al. under the terms of a commissioning contract issued by the Secretary of State for Health. This issue may be freely reproduced for the purposes of private research and study and extracts (or indeed, the full report) may be included in professional journals provided that suitable acknowledgement is made and the reproduction is not associated with any form of advertising. Applications for commercial reproduction should be addressed to: NIHR Journals Library, National Institute for Health Research, Evaluation, Trials and Studies Coordinating Centre, Alpha House, University of Southampton Science Park, Southampton SO16 7NS, UK.

Rising prevalence and burden of diabetes

The number of cases of diagnosed diabetes in the UK increased to 3.2 million in 2013, with an estimated 850,000 people having the disease without knowing it, and as many as 7 million more at high risk of developing it. By 2025, if current trends continue, an estimated 5 million people will have diabetes. 1 Type 2 diabetes accounts for about 90% of cases of diabetes. 2 Of the total cost of diabetes, 75–80% is incurred in the treatment of complications associated with poor preventative care (e.g. poor glycaemic control) and the long duration of diabetes. 3 Identification of individuals at high risk of diabetes (HRD) can at least delay the onset of diabetes, and early intervention for diabetes can reduce or at least delay the onset of complications, which already account for around 10% of the total NHS budget. 4 This percentage is projected to rise to 17% over the next 20 years4 as a result of increasing rates of obesity and an ageing population.

Scope and context of the evaluation

This evaluation is concerned with which blood test to use when screening in order to identify cases of undiagnosed type 2 diabetes. Implicit in any evaluation of screening for diabetes is the concurrent opportunity to identify individuals at HRD and subsequently manage them to reduce their risk. The two main blood tests concerned are glycated haemoglobin (HbA_1c) and fasting plasma glucose (FPG), these being the options recommended by the National Institute for Health and Care Excellence (NICE) for blood testing for diabetes and HRD. 5 A HbA_1c test does not require an overnight fast and measures the amount of glucose that is being carried by the red blood cells in the body. The result indicates an individual’s average blood glucose levels over the previous 2–3 months. A FPG test directly measures glucose levels and is to be taken after an 8-hour overnight fast. A HbA_1c or FPG measurement can be used both for the screening test and, where the first test is in the relevant diabetes range, for the confirmatory diagnostic test. The corresponding HbA_1c and FPG definitions of diabetes are sufficient to make a diagnosis, the performance of HbA_1c or FPG test no longer being assessed with reference to the oral glucose tolerance test (OGTT) as a gold standard test.

The starting context for the cost-effectiveness analysis is an individual attending an appointment at a general practice (GP) centre in England having been offered a NHS health check. From this point, a blood test may be offered either to all individuals or only to those exceeding a diabetes risk score threshold obtained from risk factor data.

The figure could be extended to include rescreening but we decided not to include this, as we have not modelled this as there was insufficient evidence on some key parameters that would be required. These issues are discussed in detail in Chapter 5, Secondary analyses of epidemiological studies.

There is the possibility that individuals offered a test will not accept the offer or not attend, as shown in Figure 1. NHS Health Checks is the Department of Health’s 5-year programme to reduce the vascular risk, especially for cardiovascular disease (CVD), of individuals between 40 and 74 years of age with elevated risk factors for these conditions. The evaluation, therefore, compares the cost-effectiveness of offering screening for diabetes and HRD using a HbA_1c test versus a FPG test at the time of the health check appointment, as an addition to the other standard checks such as cholesterol and blood pressure levels.

FIGURE 1.

Scope of the evaluation. SBP, systolic blood pressure; TC, total cholesterol.

The main objectives of the modelling were as follows:

1. To assess which test is most cost-effective, following the 2012 NICE guidance on the identification of individuals with HRD as shown in Figure 2. Specifically, the recommended cut-off points for referral for preventative intervention (aimed at supporting diet and lifestyle changes) were adopted, that is 6.0% for HbA_1c or 5.5 mmol/l for FPG.

2. To determine if the results and conclusions might be different in cohorts other than the Leicester Ethnic Atherosclerosis and Diabetes Risk (LEADER) cohort, in which the relative prevalence rates of diabetes and HRD differ markedly from those in LEADER. We also simultaneously examined the impact of alternative assumptions regarding uptake of HbA_1c tests and FPG tests.

3. To assess which test is most cost-effective for screening strategies with lower HbA_1c and FPG cut-off points than the NICE recommendations. These were included because some studies have suggested that intervention may be cost-effective at lower HbA_1c cut-off points. 6

4. To determine if it is more cost-effective to use a random capillary glucose (RCG) test or the Leicester Practice Database Score (LPDS – see Chapter 2, Prescreening using the LPDS risk score) to prioritise who should receive the blood test (FPG or HbA_1c).

FIGURE 2.

The NICE flow chart: identifying and managing risk of type 2 diabetes. BMI, body mass index. Figure reproduced with permission from NICE (2012) ‘PH 38 Preventing type 2 diabetes: risk identification and interventions for individuals at high risk’. London: NICE. 5 Available from http://guidance.nice.org.uk/PH38.

Absence of a single gold standard definition of diabetes

There no longer exists a single gold standard definition of diabetes. An individual can receive a diagnosis based on either HbA_1c-based criteria or FPG-based criteria [or 2-hour plasma glucose (2hPG) criteria, but the associated test, the OGTT, is outside the scope of this assessment].

The cohorts identified by the alternative tests only partially overlap; this raises some issues to address around prognosis of individuals with differing diagnoses according to the two tests. Evidence on this is lacking, although we can estimate risk of diabetes and CVD through individuals’ risk factors.

It should be noted that diabetes risk assessment should be undertaken in all eligible 40- to 74-year-olds; it is not conditional on having a high CVD risk.

Overview of approach

The model developed for these analyses is an adaptation of the model used to assess screening strategies as part of NICE’s Public Health Guidance Development work in 2012. 5

The modelling comprises two stages. The first stage entails a model to determine the individual screening outcomes from alternative strategies for screening and diagnosis. To populate this we used individual patient-level data from the Leicestershire-based LEADER study. We used these data to analyse the following questions:

• What proportion of individuals would receive a blood test, taking account of two options to identify those in greatest need of a blood test? These options are the non-invasive LPDS risk score, which uses risk factors held in computer databases to calculate a score relating to an individual’s risk of developing diabetes (see Prescreening using the Leicester Practice Database Score), and a RCG test (see Prescreening using random capillary glucose).

• Taking account of the different uptake rates (base case 20% higher for HbA_1c than FPG), what proportion of individuals would be detected with diabetes or HRD (1) under HbA_1c testing and (2) under FPG testing? HRD is defined as a HbA_1c between 6.0% and 6.4%, or a FPG between 5.5 and 6.9 mmol/l.

The LEADER study did not include information on the RCG finger-prick test, so RCG values were incorporated into the LEADER data set using a mapping between HbA_1c and RCG from the Anglo-Danish–Dutch Study of Intensive Treatment In People with Screen Detected Diabetes in Primary Care (ADDITION)-Cambridge study. This enabled us to answer the following question:

• How do screening outcomes compare using the LPDS versus RCG testing as a prescreening tool?

The second stage entails using an adaptation of the Sheffield Type 2 Diabetes Model that was used to assess screening strategies as part of NICE’s Public Health Guidance Development work. The model is used to simulate the lifetime patient clinical pathways, incidence of complications and associated cost and health utility impacts. The assumptions for the modelling are described in detail later in the report, but the key ones are listed here:

• Uptake rates of HbA_1c tests (given that someone has attended a health check) were assumed to be 20% higher than for FPG tests (based on clinical advice), 95% and 75%, respectively, because of the need to fast for the FPG test.

• The central estimates for the costs of screening tests (including staff time, transport and laboratory costs) were estimated to be £12.18 for a FPG test and £14.40 for a HbA_1c test (see Units costs of prescreening and blood glucose tests for costing).

• Cases of diabetes identified through screening are assumed to be treated at the point of diagnosis through routine care.

• Cases identified as HRD are assumed to be offered a preventative intervention in the form of a less intensive group-based adaptation of the intervention used in the US Diabetes Prevention Programme.

• Cases of undiagnosed diabetes and HRD missed by the screening process (because of either a risk score being below the threshold for blood testing or lack of uptake when offered a blood test) are modelled accordingly.

• Economic modelling was used to estimate the lifetime discounted costs and quality-adjusted life-years (QALYs) and the net monetary benefit (NMB) associated with each screening strategy. The optimal strategy is the one with the highest NMB (assuming 1 QALY is valued at £20,000) among those with an incremental cost-effectiveness ratio (ICER) below £20,000/QALY compared with ‘no screening’ (this is not necessarily the one with the lowest ICER compared with no screening).

• Uncertainty was analysed using one-way sensitivity analysis and probabilistic sensitivity analysis (PSA).

• Evidence from studies other than LEADER, in particular from the University of East Anglia-Impaired Fasting Glucose (UEA-IFG) study, indicates that there may be significant regional variations in the relative prevalence of FPG versus HbA_1c-defined diabetes and HRD. We therefore carried out some additional analyses based on alternative prevalence scenarios.

The model used for this evaluation naturally includes all of the inputs that would be included for an evaluation of screening for diabetes compared with no screening. Models of screening for diabetes necessitate inclusion of a large number of inputs and assumptions, as will become apparent during this section. For this test-specific screening evaluation, however, it was felt to be important that the key inputs are highlighted at the outset, namely those for which differential evidence exists for HbA_1c versus FPG testing. The key inputs and methods are:

• the prevalence of diabetes and HRD, for each of the HbA_1c- and FPG-based definitions, in the LEADER cohort and other cohorts (see Prevalence of diabetes and high risk of diabetes)

• the uptake rates when individuals are offered testing (see Uptake rates of blood tests)

• the approach to mapping screening outcomes to subsequent glycaemic trajectories (see Discordance between the groups of individuals identified by glycated haemoglobin and those identified by fasting plasma glucose)

• the approach to modelling the risks of diabetes and CVD in individuals with HRD, conditional on an individual’s HbA_1c and FPG levels (see Rate of progression from high risk of diabetes and Fasting plasma glucose/glycated haemoglobin at baseline and risk of incident cardiovascular disease).

The costs of the tests are relatively small; therefore, they are not a key driver of relative long-term cost-effectiveness.

Prevalence of diabetes and high risk of diabetes

The Leicester Ethnic Atherosclerosis and Diabetes Risk cohort data set

We used the LEADER data set in all of the base case analyses to estimate the performance of the alternative screening strategies. The LEADER cohort is a combination of two systematic screening programmes conducted in Leicestershire, the ADDITION–Leicester study7 and the STAR (Screening Those At Risk) study. 8 These studies recruited from a population of over 950,000 in the relevant age range, approximately one-third of whom were resident in the City of Leicester. All individuals aged 40–75 years were invited to attend for screening and an OGTT was carried out according to World Health Organization (WHO) 1999 criteria. In addition, those aged 25–39 years and not of white European origin were invited for screening. Simultaneously, a HbA_1c measurement was taken and measured on a correctly aligned assay analyser. The screening was conducted in a general practice centre, in a mobile screening unit or at one of the Leicester teaching hospitals, between February 2002 and August 2009. Those identified by the programme with HRD were offered an annual follow-up. In 2011, 9494 people who had been screened were included in the LEADER database. The data set includes HbA_1c, FPG, 2hPG, family history of diabetes and routine demographics collected on all patients.

According to the 2001 census, 30% of this population classified themselves as belonging to Indian, Pakistani or Bangladeshi ethnic groups. Different strategies were used for participant recruitment in each study. The majority of participants (two-thirds) were screened regardless of risk of diabetes. For the remaining third, eligibility was subject to having a risk factor for diabetes, as recommended by Diabetes UK. Participants were recruited from 40 Leicestershire general practices from a range of deprivation levels.

Determining diagnostic outcome of individuals in the Leicester Ethnic Atherosclerosis and Diabetes Risk data set

The LEADER data set contains the LPDS risk score, results for a single FPG test (and, for some individuals, a second FPG result) and a single HbA_1c test. In clinical practice, two test results are needed to diagnose diabetes but confirmatory test results were not available for all individuals in the LEADER data set. Where the LEADER test result was in the diabetes range, assumptions were necessary to determine whether or not an individual’s confirmatory test would result in a diagnosis of diabetes:

• Not everyone with a first FPG ≥ 7 mmol/l in the LEADER data set had a second confirmatory FPG test. To populate the data set with sampled outcomes of confirmatory testing, we needed evidence on the proportion of confirmatory tests that confirm diabetes. We therefore undertook an analysis of the subset of individuals in the LEADER study in whom two FPG tests were undertaken. The results suggest that approximately 70% of repeat FPG tests confirm diabetes (i.e. FPG ≥ 7 mmol/l). We then sampled whether or not the second FPG result was confirmatory of diabetes, assuming that 70% would be confirmed.

• For HbA_1c testing, we assumed that the result of the first test would be replicated by the confirmatory test. Clinical experts advised that this was a reasonable assumption because HbA_1c has much lower variation between consecutive test results than FPG. We have, however, included the cost of a second confirmatory HbA_1c test where the initial test indicates diabetes, as the second test is required to make a formal diagnosis of diabetes.

For both HbA_1c testing and FPG testing, where the result of the first test is below the cut-off for diabetes but in the range considered HRD, there is not a requirement to carry out a second confirmatory test.

Baseline characteristics and descriptive analyses

The baseline characteristics of the 8147 individuals from the LEADER data set aged 40–74 years for whom data were available for all data fields required for our analysis are shown in Table 1.

TABLE 1 - Baseline characteristics of the LEADER study cohort (n = 8147)

HDL, high-density lipoprotein.

Where FPG ≥ 7.0 mmol/l on first test and sampled confirmatory result ≥ 7.0 mmol/l. If sampled confirmatory result < 7.0 mmol/l then classed as HRD.

Characteristic	Mean	Standard deviation
Age (years)	57.30	9.67
Systolic blood pressure (mmHg)	137.01	19.68
HDL cholesterol (mmol/l)	1.36	0.41
Total cholesterol (mmol/l)	5.53	1.06
HbA1c (%)	5.71	0.62
FPG (mmol/l)	5.21	0.91
	Number	Percentage
Male	3874	47.6
White	6199	76.1
Current smoker	1480	18.2
Diabetes prevalence (HbA1c ≥ 6.5%)	467	5.7
Diabetes prevalence (FPG ≥ 7.0 mmol/l)a	150	1.8
Total diabetes prevalence (either test)	513	6.3
Diabetes with both tests	104	1.3
HRD prevalence (HbA1c 6.0–6.4%)	1487	18.3
HRD prevalence (FPG 5.5–6.9 mmol/l)	1936	23.8
Total HRD prevalence (either test)	2823	34.7
HRD with both tests	600	7.4

Total diabetes prevalence (based upon either HbA_1c or FPG testing) is 6.6%, but only 1.3% test positive on both tests. HbA_1c testing identifies more than three times as many individuals as FPG testing (5.7% compared with 1.8%).

The prevalence of HRD is similar with the two tests, but for the most part they identify different individuals, as, out of a total HRD prevalence of 34.7%, only 7.4% (fewer than one-quarter) are identified with both tests.

Although one of the two studies forming the LEADER cohort had a recruitment criterion of having at least one risk factor, this did not materially increase the risk of the LEADER cohort overall.

Table 2 shows how many individuals from the LEADER cohort belong to non-diabetic, HRD and diabetic subgroups based upon HbA_1c or FPG values. The majority of people (62.8%) are not diagnosed as either diabetic or HRD with either of the two tests (italic text). Text in bold italics indicates percentages of individuals classified as at HRD (but not diabetic) for at least one of the criteria, whereas text in bold indicates percentages of individuals classified as diabetic for at least one of the criteria.

TABLE 2 - Matrix showing numbers and percentages of individuals in each HbA1c and FPG subcategory

Text in italics indicates percentages of individuals not diagnosed as either diabetic or at HRD. Text in bold italics indicates percentages of individuals classified as at HRD (but not diabetic) for at least one of the criteria. Text in bold indicates percentages of individuals classified as diabetic for at least one of the criteria.

FPG subgroup (mmol/l)	HbA1c < 6.0%		HbA1c 6.0–6.4%		HbA1c ≥ 6.5%		Totals
	Number	Percentage	Number	Percentage	Number	Percentage	Number	Percentage
< 5.5	5114	62.8	853	10.5	94	1.2	6061	74.4
5.5–6.9	1067	13.1	600	7.4	269	3.3	1936	23.8
≥ 7.0	12	0.1	34	0.4	104	1.3	150	1.8
Totals	6193	76.0	1487	18.3	467	5.7	8147	100

Figure 3 shows the distribution of HbA_1c values in individuals aged between 40 and 74 years from the LEADER cohort. Individuals with HbA_1c values between 6.0% and 6.4% are at HRD (green bars), whereas individuals with HbA_1c values of 6.5% and above have diabetes (blue bars). Most of the population have HbA_1c values under 6.0% (dark green bars).

FIGURE 3.

Histogram of the distribution of HbA_1c values in people in the LEADER cohort.

Figure 4 shows the distribution of FPG values in individuals aged between 40 and 75 years from the LEADER cohort. Individuals with FPG values between 5.5 and 6.9 mmol/l are at HRD (green bars), whereas individuals with FPG values of 7.0 and over have diabetes (blue bars). Most of the population has FPG values under 5.5 (dark green bars). Comparison with Figure 3 shows clearly that, in the LEADER cohort, fewer individuals are diagnosed with diabetes using the FPG test than with the HbA_1c test.

FIGURE 4.

Histogram of the distribution of FPG values (from single test) in people in the LEADER cohort.

Figure 5 shows HbA_1c values plotted against FPG values for each individual in the LEADER cohort aged between 40 and 75 years. The trend line (black dotted line) illustrates the positive correlation between FPG values and HbA_1c values.

FIGURE 5.

Scatterplot of values of HbA_1c against FPG for the LEADER cohort.

Figure 6 shows HbA_1c values plotted against FPG values for individuals in the LEADER cohort who are diagnosed with diabetes according to either of the two tests. Individuals who are over the cut-off point for diabetes in both tests are represented with blue dots, whereas individuals who are over the cut-off point, in the HbA_1c test or the repeated FPG test, are represented with black and dark green dots, respectively. Overall, the HbA_1c test identifies more individuals with diabetes than does the FPG test.

FIGURE 6.

Scatterplot of values of FPG against HbA_1c for individuals with diabetes (according to either the HbA_1c or FPG test) in the LEADER cohort.

Figure 7 shows HbA_1c values plotted against FPG values for individuals in the LEADER cohort who are at HRD according to either of the two tests. It clearly illustrates that the two tests predominantly identify different individuals, as only the green dots represent individuals who are below the cut-off points for HRD in both tests. Individuals who meet the criteria for HRD with just the HbA_1c test or just the FPG test are represented with black and dark green dots, respectively. Some of the individuals diagnosed as being at HRD based on one of the two tests are diagnosed as having diabetes using the other test. Overall, the FPG test identifies more individuals with HRD than the HbA_1c test. Note that most of these data are dependent only on results from a single FPG test, as individuals diagnosed with HRD using FPG testing will not be eligible for a second FPG test.

FIGURE 7.

Scatterplot of values of HbA_1c against FPG for individuals defined as at high risk of diabetes (either by HbA_1c or FPG) in the LEADER cohort. Fasting plasma glucose values are from a single test: either (i) FPG 5.5–6.9 mmol/l or (ii) FPG ≥ 7.0 mmol/l in first test but < 7.0 mmol/l on confirmatory testing (all values in mmol/l).

Prevalence of diabetes and high risk of diabetes among South Asians of 25–39 years of age in the Leicester Ethnic Atherosclerosis and Diabetes Risk cohort

Table 3 shows that, within the LEADER cohort, there are also significant differences in the relative prevalence of HbA_1c to FPG-defined undiagnosed diabetes in South Asians under 40 years of age.

TABLE 3 - Prevalence of diabetes and HRD among South Asians of 25–39 years of age in LEADER (%)

Diabetes		HRD
HbA1c (≥ 6.5%)	FPG (≥ 7.0 mmol/l)	HbA1c (6.0–6.4%)	FPG (5.5–6.9 mmol/l)
3.3	0.5	10.4	11.2

Other epidemiological evidence relevant to the model

A version of the Sheffield Diabetes Model (see The Sheffield Type 2 Diabetes Model) adapted for screening and prevention assessments was available at the start of the project. The model is described in a 2012 report for NICE. 12 During various prior projects, the model has drawn upon a wide array of evidence from published reviews (including the Waugh 200713 review of screening for diabetes), targeted searches and literature identified through topic experts. As a result, the vast majority of the evidence required for this assessment was already contained within the economic model. It was therefore neither necessary nor practicable to comprehensively search for, review and synthesise the vast volume of literature on all aspects of the epidemiology of diabetes within the scope of this project. We were also aware of an imminent review due to become available during the project, the 2013 update of evidence for screening for diabetes undertaken by Waugh and colleagues for the National Screening Committee. 14 The form, effectiveness and cost of preventative interventions in a real-world setting (see Form of intervention, Initial weight loss and Durability of reduction in risk) relied heavily on a clinical review undertaken for the NICE guidance. 15

The necessary endeavours to obtain and familiarise ourselves with the necessary data fields from the LEADER cohort had already been done during the NICE work. We were also aware of the recent publication of the final version of the LPDS risk score. 16

There were some areas where it was realised that new or updated evidence was required for the model. These include:

1. Revisiting rates of progression from HRD to diabetes, which was the subject of a recent meta-analysis that had become available (see Rate of progression from high risk of diabetes to diabetes)

2. Evidence on the multivariate risks for the incidence of diabetes and CVD according to both baseline HbA_1c and FPG levels (see Adjusting an individual’s risk of diabetes to take account of both fasting plasma glucose and glycated haemoglobin and Fasting plasma glucose/glycated haemoglobin at baseline and risk of incident cardiovascular disease – evidence review). For this evidence, which typically necessitates a large epidemiological study to obtain adequate statistical power, it was decided that a systematic search for and synthesis of such epidemiological evidence would be both time-consuming and inefficient given the time available and existing sources of evidence at our disposal. We therefore identified studies from (1) literature already identified during previous work, (2) studies described in Section 3 of the 2013 update of evidence for screening14 and (3) evidence sources signposted by the clinical members of the team.

Economic evidence from other studies

A version of the Sheffield Diabetes Model (see The Sheffield Type 2 Diabetes Model), the economic model adapted for screening and prevention assessments, was available at the start of the project. Over various previous projects, this model has utilised evidence from previous literature reviews that include the economics of screening and prevention. These include a review of screening for diabetes (Waugh 200713) for the National Institute for Health Research Health Technology Assessment programme, and a review of key cost-effectiveness studies undertaken for the 2012 NICE guidance on risk identification. 5

The model already contained all of the economic parameters required for this assessment; therefore, it was considered unnecessary to undertake any new economic reviews or additional systematic reviews within this project.

Unit costs of tests (see Unit costs of prescreening and blood glucose tests) were drawn from the 2012 NICE economic modelling. 12

Chapter 6 of the 2013 update review of screening for diabetes14 included a review of economic studies since the 2007 review. 13 None of the studies, however, compared the long-term cost-effectiveness of alternative blood tests for diabetes; therefore, they were not of use for comparison with our results.

Defining the prescreening and blood glucose test strategies to be assessed

In this section, we describe the rationale by which we have defined the strategies to be assessed in our study. This covers prescreening methods examined, the blood glucose tests assessed and the thresholds for deeming an individual as at HRD. We then discuss the alternative combinations of these that were arrived at, forming a set of alternative overall screening strategies.

Unit costs of prescreening and blood glucose tests

The full costs of tests include all costs associated with completing the test including nurse or health-care assistant time and laboratory costs. These are shown in Table 8. The costs for a HbA_1c test and a FPG test are for laboratory tests (not POC tests).

There is no standard national source of unit costs for England for the blood tests for diabetes. The most recent costing in the UK is the one undertaken as part of the Vascular Checks modelling work,23 from which we obtained the cost of a FPG test. For this analysis, the most important issue is the difference in cost between a FPG test and a HbA_1c test. The only difference in cost between a FPG test and a HbA_1c test is the laboratory cost and we obtained this information from an estimate of these costs from Professor Kamlesh Khunti (University of Leicester, 2011, personal communication).

The cost for undertaking a RCG test was estimated based on a published study by Chatterjee. 22

Cost estimates were updated as appropriate for inflation. Annual inflation adjustments were obtained from the Hospital and Community Health Services Index reported in Unit Costs of Health and Social Care 2012. 24

Uptake rates of blood tests

Uptake rates for diabetes screening are often reported to be relatively low, with only 61% of patients taking up screening in the pilot diabetes screening programme in England. 25

We have not used this rate because the pilot programme was a research study in which participants were required to give consent, which some individuals may not wish to do, while some individuals may choose not to respond or attend for a variety of reasons.

For this economic evaluation, the appropriate rates need to reflect a setting where someone has already presented at the GP centre for the wider health check (i.e. for cholesterol and blood pressure). This makes the proportion accepting a blood test for HbA_1c or FPG at the same time higher (if an individual is having cholesterol checked anyway, very few people would refuse to have the needle in a very short time longer to draw another sample for the HbA_1c or FPG test).

There is currently no published evidence on rates of uptake of HbA_1c testing and FPG testing within the NHS Health Checks programme, and we are not aware of any evidence from a similar setting elsewhere. Estimates were, therefore, based on discussions with clinical experts. Table 9 shows the estimates of uptake rates. These represent the proportion of people that have already presented at their GP centre for the health check who then agree to have the blood test for diabetes (or HRD).

Two laboratory blood tests are required to make a diagnosis of diabetes. For the purpose of discussing uptake rates, we refer to the initial test as the ‘screening test’ and any second test to confirm if an individual has diabetes as the ‘confirmatory test’. The confirmatory test requires a second visit to the GP centre once the laboratory results for the first test have been sent to the GP centre.

The differential uptake between HbA_1c and FPG testing is uncertain and was therefore explored within sensitivity and scenario analyses as discussed in Deterministic one-way sensitivity analyses and Scenario analyses – alternative prevalence and uptake rates, respectively). This is partly because there is variation across the country in instructions accompanying Health Checks invitations, in particular for practices that use FPG to test for diabetes. Some such practices request that individuals fast before their visit so that the FPG test can be taken at the same time as the blood test for cholesterol whereas others do not, in which case individuals indicated (after a prescreen if used) for a test for diabetes would need to return on a later date. The need for a separate visit would be expected to reduce the uptake of the test.

After discussion with clinical experts, it was considered to be conservative to assume no difference between the uptake of confirmatory HbA_1c testing and uptake of confirmatory FPG testing. The rationale for this is that, given that an individual has been willing and able to attend for a first FPG, the reasons for lower uptake of FPG testing in general at the screening test (i.e. the inconvenience of fasting and/or visiting their GP in the morning) may not apply to that individual. In other words, having attended for a first FPG test, you may be as likely to return for a confirmatory FPG test (if needed) as an individual undergoing HbA_1c testing would return for a confirmatory HbA_1c test (if needed).

We assume 100% availability of data in GP databases to calculate the LPDS for each individual.

Monte Carlo sampling process for determining individual uptake and screening outcomes

For each individual in the model, random sampling was used to determine whether or not the individual accepts the offer of a blood test, based on the evidence and assumptions for uptake probabilities in Table 9.

If the stochastic screening outcome was HRD in the model, further random sampling was used to determine whether or not an individual would take up the offer of an intensive lifestyle intervention to reduce his or her risk of developing diabetes (see Referral for and uptake of preventative interventions in people with high-risk diabetes for evidence on uptake of prevention intervention). For any given individual in the model, the same sampled random numbers were used across the range of strategies to avoid introducing sampling bias.

Mapping individual screening outcomes to initial glycaemic trajectories

Rate of progression from high risk of diabetes to diabetes

The landmark Finnish Diabetes Prevention Study (DPS) and American Diabetes Prevention Programme (DPP) trials were designed to assess the ability of intensive lifestyle interventions to reduce the risk of progression from impaired glucose tolerance (IGT) to diabetes. These trials reported incidence rates of diabetes of 23% over 4 years27 and 11% over 3 years,28 respectively. One review of progression rates has quoted higher annual rates of 5–10%. 29

Rates reported in some research studies may also be inflated to some extent by:

1. selective recruitment of individuals at particularly HRD, for example because of high average baseline BMI levels

2. the combination of the annual frequency of the OGTTs to test for diabetes during follow-up and the between-test variability of FPG and 2hPG measures that make up an OGTT.

The average progression rates in clinical practice may therefore be lower than in these trials, although there is likely to be significant variation regionally within England according to factors such as ethnicity, deprivation and other demographics.

It was decided that the best rates to use were those presented in a meta-analysis published in 2013. This reported progression rates approximately equivalent to 3.5% per year from a baseline HbA_1c level of 6.0–6.4%. 30

Adjusting an individual’s risk of diabetes to take account of both fasting plasma glucose and glycated haemoglobin

In this section, we describe the evidence review that we undertook to identify published literature reporting the independent contributions of FPG and HbA_1c to the risks of developing diabetes.

As already discussed, a HbA_1c test and a FPG test identify only partially overlapping groups of individuals with diabetes and at HRD. The average levels of HbA_1c and FPG of individuals identified as at HRD with HbA_1c testing may differ from the average levels under FPG testing. It cannot, therefore, be assumed that the average risk of diabetes for an individual identified at HRD with a HbA_1c test is the same risk as that for an individual identified using a FPG test. It is, therefore, imperative that risks of developing diabetes take account of individual risk factors, in particular baseline FPG and HbA_1c levels.

Additional literature was required to identify studies that had evaluated diabetes risk conditional on both baseline FPG and baseline HbA_1c. In the model it was necessary to vary the risk of diabetes according to the individual’s FPG and HbA_1c levels. Therefore, it was necessary to estimate the independent effects of these continuous measures on the probability of diabetes from published literature.

Given the time available for this particular topic, a new comprehensive literature search and review was not possible. However, we were able to rely on (1) studies identified in the 2013 evidence review update on screening for diabetes for the HTA,14 (2) studies which we had already identified as part of the work on the 2012 NICE risk assessment work5 and (3) additional studies identified from clinical experts. Studies were included in our review if they reported baseline measures for both FPG and HbA_1c, and an analysis of the risk of diabetes.

We have identified 11 studies that have reported risk or incidence of diabetes by FPG and HbA_1c score, as shown in Table 10. We aimed to identify a multivariate regression model that included FPG and HbA_1c as continuous variables and that could be included within our individual-level simulation model. Table 10 summarises the population that was studied and the definition of diabetes used based on a HbA_1c, FPG and/or 2hPG glucose.

TABLE 10 - Studies reporting analysis of the risk of diabetes conditional on FPG and HbA1c measures

M/C, men/control; M/D, men/diabetic; N/A, not applicable; Q, quartile; W/C, women/control; W/D, women/diabetic.

Study	Population	Basis of diagnosis of diabetes	Baseline HbA1c (%)	Baseline FPG (mmol/l)	Baseline 2hPG (mmol/l)
Law et al. (2010)31	Hong Kong Cardiovascular Risk Factor Prevalence Study, Hong Kong	ADA 2010, FPG ≥ 7.0 mmol/l, 2hPG ≥ 11.1 mmol/l and/or HbA1c ≥ 6.5%	N/A	N/A	N/A
Valdes et al. (2011)32	Asturias study, Spain	FPG ≥ 7.0 mmol/l, 2hPG ≥ 11 mmol/l and/or clinical diagnosis	Q1 3.4–4.8	5.0	5.3
			Q2 4.9–5.1	5.2	5.7
			Q3 5.2–5.4	5.3	5.8
			Q4 5.5–6.9	5.6	6.6
Sato et al. (2010)33	Kansai Healthcare Study, Japan	FPG ≥ 7.0 mmol/l or were taking an oral antidiabetic agent or insulin	N/A	N/A	N/A
Ko et al. (2000)34	The Diabetes and Endocrine Centre of the Prince of Wales Hospital, China	FPG ≥ 7.0 mmol/l	5.78	5.36	7.55
Norberg et al. (2006)35	Vasterbotten Intervention Programme, Sweden	FPG ≥ 7.0 mmol/l or 2hPG ≥ 12.2 mmol/l	M/D: 4.7	M/D: 6.0	M/D: 7.9
			M/C: 4.3	M/C: 5.3	M/C: 6.2
			W/D: 4.7	W/D: 5.8	W/D: 8.4
			W/C: 4.3	W/C: 5.2	W/C: 7.2
Rasmussen et al. (2008)36	ADDITION, Denmark	FPG ≥ 6.1 mmol/l or 2hPG ≥ 11.1 mmol/l	IFG: 5.6	IFG: 5.8	IFG: 6.2
			IGT: 5.9	IGT: 5.3	IGT: 9.1
Selvin et al. (2011)37	ARIC, USA	Definition 1: a single FPG value ≥ 7.0 mmol/l at baseline (visit 2). Definition 2: FPG values ≥ 7.0 mmol/l at two separate examinations	N/A	N/A	N/A
Takahashi et al. (2010)38	Tokyo, Japan	HbA1c ≥ 6.5% or self-reported, or commencement of glucose-lowering treatment	5.4	5.5	N/A
Wang et al. (2011)39	Indian tribes/communities in Arizona, North/South Dakota, and Oklahoma	HbA1c ≥ 6.5% or FPG ≥ 7.0 mmol/l or if on diabetes medications	N/A	N/A	N/A

Law and colleagues31 describe the 8-year incidence of diabetes in a cohort of 530 non-diabetic Chinese individuals. 31 There were 47 diagnoses by 3 years and 81 at 8 years of follow-up. The authors report the hazard ratios from a multivariate Cox regression which includes covariates for HbA_1c, FPG and 2hPG. The results suggest that HbA_1c and FPG are independent predictors of a diagnosis of diabetes: the baseline hazard ratio for HbA_1c was 3.74 [95% confidence interval (CI) 1.98 to 7.04] per 1% HbA_1c and the hazard ratio for FPG was 1.76 (95% CI 1.13 to 2.74) per mmol/l. Unfortunately, the authors do not report mean baseline HbA_1c or FPG in the article.

Valdes and colleagues32 report results from the Asturias study from northern Spain, in which the incidence of diabetes is reported for individuals with high FPG and/or high HbA_1c. 32 The estimated cumulative incidence values and hazard ratios reported in that study are shown in Table 11.

TABLE 11 - Approximate results from Valdes et al.32 comparing diabetes incidence between subgroups

Incidence/hazard ratio	FPG < 5.56 mmol/l		FPG ≥ 5.56 mmol/l
	HbA1c < 5.5%	HbA1c ≥ 5.5%	HbA1c < 5.5%	HbA1c ≥ 5.5%
Cumulative incidence at 6 years (%)	2	7	9	32
Hazard ratio vs. low risk	1	3.5	4.5	16
Hazard ratio vs. high risk	0.063	0.219	0.281	1

The Sato et al. 33 report results from a Japanese study in which study participants consisted of 9116 Japanese men aged 40–55 years with FPG less than 7.0 mmol/l who were not taking an oral antidiabetic agent or insulin at study entry. The study reports the results of a logistic regression which included categories for FPG and HbA_1c. The estimated odds ratios for diabetes for each subgroup are shown in Table 12. The results suggest that both classifications are strong independent predictors of the diagnosis of diabetes.

TABLE 12 - Odds ratios for the risk of diabetes reported in Sato et al.33

HbA_1c of the National Glycated Haemoglobin Standardization Programme is shown in parentheses.

FPG or HbA1c category	Odds ratio	95% CI
FPG ≤ 5.5 mmol/l	1.00	−
FPG 5.6–6.0 mmol/l	3.28	2.57 to 4.18
FPG 6.1–7.0 mmol/l	14.54	11.31 to 18.68
HbA1c ≤ 4.9% (5.3)a	1.00	−
HbA1c 5.0–5.4% (5.4–5.7%)a	1.71	1.26 to 2.31
HbA1c 5.5–5.9% (5.8–6.2%)a	4.50	3.30 to 6.14
HbA1c 6.0–6.4% (6.3–6.7%)a	11.04	7.23 to 16.87
HbA1c ≥ 6.5% (6.8)a	33.58	18.88 to 66.78

Ko and colleagues34 categorised 208 subjects into groups based on their FPG (≥ 6.1, < 6.1 mmol/l) and HbA_1c (≥ 6.1%, < 6.1%). The incidence of diabetes according to the OGTT test is reported after variable duration of follow-up. Since the OGTT is used to define diabetes at follow-up, the results have not been extracted here. This study was not used to estimate risk based on the diagnosis criteria and because it was measured in a Chinese population.

Norberg and colleagues35 report analyses of 468 participants in a Swedish study. They performed a logistic regression to predict the odds ratio of diagnosis of diabetes according to categories of HbA_1c and whether or not the individual met the criteria for IFG at baseline (5.6–6.9 mmol/l). The results are reported in Table 13.

TABLE 13 - Odds ratios for risk of diabetes from Norberg et al.35

HbA1c/IFG	Odds ratio	95% CI
HbA1c < 4.5%	1.0	–
HbA1c 4.5–4.69%	1.2	0.28 to 5.34
HbA1c ≥ 4.7%	16	2.23 to 115.3
IFG	18.8	2.88 to 123.4

Takahashi and colleagues38 report an odds ratio of 1.06 for FPG scores when added to HbA_1c to predict diabetes. They report only cumulative incidence by categories of HbA_1c.

Wang and colleagues39 describe analyses of 4549 American Indian men and women. They developed a logistic model for the risk of diabetes defined according to HbA_1c ≥ 6.5% and FPG ≥ 7.0 mmol/l. The odds ratios for defined high-risk states according to these measures are reported in Table 14.

TABLE 14 - The odds ratio for diabetes given previous glucose tests from Wang et al.39

Risk group	FPG/HbA1c-defined diabetes
	Mean odds ratio	95% CI
IFG	2.34	1.81 to 3.03
6.0% ≤ HbA1c < 6.5%	3.43	2.27 to 5.16

Two studies were identified as particularly useful for our applications:

Selvin and colleagues37 undertook analyses on a large population from the US Atherosclerosis in Communities (ARIC) study, which did not recruit based on glucose tests; therefore, representing a broad range of risk for diabetes. The ARIC study consisted of 12,485 individuals, after excluding individuals who identified their race as other than black or white, those with self-reported diabetes diagnosis, individuals with missing values for key variables, or individuals who were non-fasting. Maximum follow-up of participants was 15 years, and diabetes diagnosis was assessed by either glucose measurements or self-reported diagnosis.

The 10-year risks of FPG-defined diagnosed diabetes were stratified by categories of baseline FPG and HbA_1c, as summarised in Table 15.

TABLE 15 - Ten-year risk of diabetes analysed by HbA1c and FPG subgroups reported in Selvin et al.37

Incidence of diabetes defined by FPG ≥ 7.00 mmol/l or were taking an oral antidiabetic agent or insulin.

FPG category (mmol/l)	HbA1c < 5.7%	HbA1c ≥ 5.7% and < 6.5%	HbA1c ≥ 6.5%
< 5.56	2.65	9.69	20.00
≥ 5.56 and < 7.00	7.19	22.75	48.84
≥ 7.00	30.16	55.06	88.43

Rasmussen and colleagues,36 in the ADDITION-Denmark study, described a European population and analysed the data using continuous variables for FPG and HbA_1c. The ADDITION-Denmark study was a population-based screening and intervention study for type 2 diabetes. This study included analysis of 607 individuals with IFG and 903 individuals with IGT identified as part of the screening programme. The definition of IFG corresponds to that of isolated IFG (5.6 mmol/l ≤ FPG < 6.1 mmol/l and 2hPG < 7.8 mmol/l), whereas IGT included isolated IGT and combined IFG and IGT (FPG < 6.1 mmol/l and 7.8 mmol/l ≤ 2hPG < 11.1 mmol/l). Incident diabetes was defined as one diabetic value of FPG (≥ 6.1 mmol/l) or 2hPG (≥ 11.1 mmol/l). The median follow-up for the groups was 2.5 and 2.1 years, respectively.

Rasmussen et al. 36 reported a statistical model for the hazard ratios for diabetes in those individuals who had IFG (FPG 5.6 mmol/l) at screening, adjusting for their HbA_1c and FPG score in a multivariate model. They report a similar analysis in individuals who met the threshold for IGT at screening. Cumulative risks, progression rates and hazard ratios for progression to diabetes (≥ 6.1 mmol/l) were estimated with a regression model using interval censoring. The results of the regression model are reported in Table 16.

TABLE 16 - Hazard ratios for diabetes in two high-risk groups reported in Rasmussen et al.36

Blood glucose variable	Isolated IFG (5.6 mmol/l ≤ FBG < 6.1 mmol/l and 2hPG < 7.8 mmol/l)		IFG + IGT (7.8 mmol/l ≤ 2hBG < 11.1 mmol/l and FBG < 6.1 mmol/l)
	Mean hazard ratio	95% CI	Mean hazard ratio	95% CI
HbA1c (per 1%)	1.40	1.09 to 1.80	1.23	1.08 to 1.47
FBG (per mmol/l)	3.19	2.33 to 4.37	1.65	1.43 to 1.92
2hPG (per mmol/l)	1.10	1.00 to 1.21	1.26	1.18 to 1.35

These two studies from Selvin et al. 37 and Rasmussen et al. 36 provide useful information about the independent effects of FPG and HbA_1c on diabetes risk. However, individually, the studies provide incomplete data on the independent effects of FPG and HbA_1c on diabetes risk and could not be used directly in the cost-effectiveness model. Selvin et al. 37 described the absolute risk of diabetes for a cohort from the USA. These data may not be generalisable to a UK population in which the incidence of diabetes may be different. Rasmussen et al. 36 report the hazard ratios of HbA_1c and FPG test scores on diabetes risk in a Danish population with either IFG or IGT. In order for these estimates of the hazard ratios to be applied to UK diabetes incidence rates, it would need to be established if they could be extrapolated to individuals who do not meet the criteria for either IFG or IGT. As a consequence of the limitations in both studies, we used the data to estimate parameters for the cost-effectiveness model that were based on the evidence provided by both studies.

We developed a simple simulation model to predict 10-year incidence of diabetes among individuals with baseline FPG and HbA_1c test results. The simulation included parameters to adjust individuals’ risk according to their FPG and HbA_1c test result using alternative hazard ratios. The simulated diabetes incidence rates for subgroups defined by Selvin et al. 37 are conditional on FPG and HbA_1c to enable comparison of the simulated and observed diabetes incidence by subgroup.

The simple simulation model used data from the LEADER cohort to describe individual test results from HbA_1c, FPG and 2hPG. Counts of the number of individuals in the subgroups defined by Selvin et al. 37 are detailed in Table 17. We do not have diabetes incidence data for the LEADER cohort; therefore, the baseline 7-year cumulative incidence of diabetes from the Finnish DPS was used to estimate the annual incidence of diabetes for individuals with the average glycaemic tests scores observed in the Finnish DPS control group. From this baseline risk, we adjusted an individual’s annual risk of developing diabetes according to his or her HbA_1c and FPG test levels and hazard ratio parameters to estimate 10-year risk of diabetes. Hazard ratios parameters were taken from Rasmussen et al. ,36 but we included an additional analysis in which the simulation was calibrated to fit the Selvin et al. 37 data.

TABLE 17 - Counts of individuals in FPG and HbA1c subgroups from the LEADER cohort

FPG (mmol/l)	HbA1c < 5.7%	HbA1c 5.7–6.5%	HbA1c ≥ 6.5%
< 5.56	4236	3029	128
5.56–7.0	421	1032	240
> 7.0	8	68	169

Analyses comparing the predicted cumulative incidence of diabetes in subjects from the LEADER cohort, using the risk equations described in Rasmussen et al. ,36 were conducted. Long-term survival estimates were extracted from the Finnish DPS along with estimates of mean baseline HbA_1c, FPG and 2hPG. The mean estimates were used to estimate deviance from the mean in the LEADER cohort. Four predictive models were tested:

1. hazard ratio for HbA_1c (1.23) and FPG (1.65) from the IGT subgroup

2. hazard ratio for HbA_1c (1.23), FPG (1.65) and 2hPG (1.26) from the IGT subgroup

3. hazard ratio for HbA_1c (1.40), FPG (3.19) and 2hPG (1.10) from the IFG subgroup

4. a modified analysis in which the baseline cumulative incidence from the Finnish DPS was increased to 0.82 to reduce incidence in the LEADER cohort and the hazard ratios HbA_1c (1.4), FPG (1.65) and 2hPG (1.26) applied to investigate what magnitude of parameters were necessary to fit the Selvin et al. 37 study.

The predicted 10-year risks of diabetes according to the subgroups defined in Selvin et al. 37 are illustrated in Figure 8.

FIGURE 8.

Results of four predictive models for development of diabetes. (a) IGT model with only FPG and HbA_1c used to modify risk; (b) IGT model with FPG, HbA_1c and 2hPG to modify risk; (c) IFG model with FPG, HbA_1c and 2hPG to modify risk; and (d) a calibrated model to match with Selvin et al. 37

In all analyses, 10-year diabetes incidence increases for individuals with higher test scores on FPG and HbA_1c. However, the difference in risk between the subgroups varies according to the assumed values for the hazard ratios. The simulated output in Figure 8a illustrates that the Rasmussen IGT model, with only FPG and HbA_1c scores to adjust risk, does not generate sufficient risk differentiation between FPG and HbA_1c subgroups. Greater differentiation in risk is achieved if 2hPG is included from the IGT model, as illustrated in Figure 8b. Estimates from the Rasmussen IFG model demonstrate differentiation of risk between FPG subgroups that more closely reflects estimates from Selvin et al. 37 reported in Table 15. However, this output overestimated risk in individuals with FPG ≥ 7.0 mmol/l. In the calibrated model, changing the HbA_1c hazard ratio from that in Figure 8b and adjusting the baseline risk produced results more similar to Selvin et al. 37 (see Table 15) than found in the other analyses.

Fasting plasma glucose/glycated haemoglobin at baseline and risk of incident diabetes – evidence incorporated into the model

We chose to use the parameter estimates used to generate simulation output Figure 8d in the cost-effectiveness analysis. The final hazard ratios for FPG, HbA_1c and 2hPG are reported in Table 18. These hazard ratio parameters were applied to the baseline risk of diabetes to generate individualised risk of diabetes estimates conditional on FPG, 2hPG and HbA_1c.

TABLE 18 - Hazard ratio parameters applied to screened individuals in the LEADER cohort to estimate diabetes risk conditional on baseline FPG, HbA1c and 2hPG values

Blood glucose variable	Mean hazard ratio	95% CI
HbA1c (per 1%)	1.40	1.09 to 1.80
FPG (per mmol/l)	1.65	1.43 to 1.92
2hPG (per mmol/l)	1.26	1.18 to 1.35

Fasting plasma glucose/glycated haemoglobin at baseline and risk of incident cardiovascular disease: evidence review

This section describes additional literature that was used to adjust risks of CVD to take account of both HbA_1c and FPG. While HbA_1c, systolic blood pressure (SBP), total cholesterol and high-density lipoprotein (HDL)-cholesterol are included within the published UK Prospective Diabetes Study (UKPDS) risk equations for CHD40 and stroke risk,1 FPG is not included.

The purpose of the following review was to identify studies that had evaluated CVD risk conditional on both baseline FPG and baseline HbA_1c. In the model it was necessary to adjust individuals’ risk of diabetes according to their FPG and HbA_1c. Therefore, it was necessary to estimate the independent effects of these continuous measures on the probability of CVD from published literature. Studies were identified from a previous literature review and were included if they reported baseline measures for FPG and HbA_1c and an analysis of the risk of CVD.

We identified two articles that discussed multivariate risk factors for CVD events including both HbA_1c and FPG in the regression model.

The first useful study is from a conference presentation in 200542 reporting the results of a multivariate regression of FPG and HbA_1c scores to the risk of macrovascular complications. The study used data from 3538 participants in the UKPDS to determine if HbA_1c and FPG are associated independently with incident macrovascular complications in type 2 diabetes. The data comprised 766 macrovascular events including fatal or non-fatal myocardial infarction, fatal or non-fatal stroke, non-fatal ischaemic heart disease and sudden death. Cox models were fitted to mean duration of follow-up of 9.7 years and adjusted for post-dietary run-in values for age, sex, ethnicity, HDL and low-density lipoprotein cholesterol, triglycerides, SBP and albuminuria (urine albumin ≥ 50 mg/l) and smoking status at time of diagnosis. FPG and HbA_1c were used in the first analysis as baseline values and in the second analysis as time-dependent variables. Table 19 reports the results of a multivariate analysis of baseline and mean updated HbA_1c. The updated mean HbA_1c and FPG hazard ratios demonstrate the independent contribution of elevated HbA_1c and FPG to the risk of developing macrovascular complications.

The hazard ratios in Table 19 are applied in the model to estimate the probability of macrovascular events given an individual’s FPG and HbA_1c score in the simulation.

The second study is a more recent study from Selvin and colleagues in 2010,43 which analysed the ARIC cohort to observe the risk of diabetes and CVD outcomes by HbA_1c level. We have not directly used evidence from this study within the model, but present it here to show some comparison with the UKPDS. Hazard ratios for the progression to diabetes and CVD outcomes by categories of HbA_1c (< 5%, 5–5.5%, 5.5–6%, 6–6.5%, > 6.5%) were presented. The results show that higher HbA_1c is associated with a higher risk of diabetes and CVD outcomes. The analysis includes an adjusted model for FPG. Although the FPG estimates are not reported, the increased risk of diabetes and CVD outcomes remains significant, after adjusting for HbA_1c.

Table 20 reports the hazard ratios for HbA_1c categories and a continuous risk score per 1% increase in HbA_1c reported in Selvin et al. 43 The results report greater sensitivity to HbA_1c in a prediabetes population than that reported in the diagnosed UKPDS cohort. For example, baseline HbA_1c is not a significant risk factor for CVD disease after adjustment for FPG in the UKPDS, whereas a high HbA_1c at baseline in the ARIC cohort is a large and significant risk factor for coronary heart disease (hazard ratio 1.50) and stroke (hazard ratio 1.55). However, the difference may also be a result of geographical variation. For this reason, and because hazard ratios for FPG have not been reported, we have not used these estimates in the final model.

Fasting plasma glucose/glycated haemoglobin at baseline and risk of incident cardiovascular disease – evidence incorporated into the economic model

The impact of any difference in risk factors such as SBP, age and cholesterol identified by the two tests will be captured within the UKPDS risk equations which include these risk factors. 41,40 The UKPDS CHD risk equation also includes HbA_1c but not FPG.

At any given HbA_1c level, the FPG level for an individual would be expected to be higher for someone identified by a FPG test than by a HbA_1c test (given that the former has implicitly met a FPG threshold criterion). It is therefore necessary to make some adjustment to the risk calculated using the UKPDS.

The most appropriate values to be used in the model were obtained from those in Kim et al. 42 and are shown in Table 21 (see Parameter values and distributions).

Specification of the intervention included in the model to prevent diabetes in people at high risk of diabetes

Time to clinical detection for cases of diabetes which are not screen detected

Our economic model previously assumed that the lead-time between the point of potential screen detection and clinical detection is around 6–7 years13 based on the available evidence at the time.

However, additional evidence was published in 2012 from the Ely study. 46 This study compared the duration of clinically recognised diabetes in individuals who had been offered screening versus those had not been offered screening. The results suggested a lead-time of just 3 years. There are some caveats around this finding, in particular some dilution of the effect of screening through ad hoc opportunistic screening for diabetes and improved detection of risk factors for CVD (including diabetes) within primary care.

We have assumed a mean lead time of 4–5 years, that is between the two ‘extremes’ (which we have assumed to represent the 95% CIs for the true value when undertaking PSA).

Given the HbA_1c level at which an individual is identified symptomatically with diabetes within clinical practice and the lead time before clinical detection, we can plot the (non-linear) upwards HbA_1c trajectory for the individual. We have assumed clinical detection occurs on average at a HbA_1c level of 8%, as per those recruited into the DESMOND (Diabetes Education and Self Management for Ongoing and Newly Diagnosed) study. 47

Benefits of early detection of diabetes

In this section, we summarise the assumptions made regarding the effect on treatment of a diagnosis of diabetes and thereby the benefits of earlier detection.

Economic modelling

The majority of the analyses and uncertainty analysis described in this section are based around the multiethnic LEADER cohort. (For additional scenario-based analyses covering alternative prevalence and blood test uptake rates see Scenario analyses – alternative prevalence and uptake rates.)

Analyses to be undertaken

Figure 9 is a summary of the various analyses undertaken, in particular which base case, scenario and sensitivity analyses are based on prevalence in the LEADER cohort and which are based on alternative prevalence scenarios.

FIGURE 9.

Schematic of analyses undertaken.

The model developed for these analyses is a screening and prevention adaptation of the Sheffield Type 2 Diabetes Model that was used to assess screening strategies as part of NICE’s Public Health Guidance on risk assessment for diabetes. 5 A fully detailed report on that work is available on the NICE website. 12

Briefly, the Sheffield Diabetes Model is an integrated health state simulation model of the natural history of diabetes and the lifetime cost-effectiveness of different treatments for type 2 diabetes. The model replicates patients’ risk of progression through five comorbidities: retinopathy, nephropathy, neuropathy, CHD and cerebrovascular disease. Patients can experience three of the major complications associated with diabetes: neuropathy, nephropathy and retinopathy. The time spent by patients in each state for each co-morbidity is recorded, for example years spent on dialysis, severe vision loss, together with transitions between states.

Total costs are obtained by adding the costs of therapy, the costs of one-off treatments (e.g. cost of amputation) and ongoing treatment of complications (e.g. treatment following stroke). The health benefit, the incremental QALYs, is obtained by applying quality of life measures (such as preference scores from the Harvard web-based database) to the time spent in the various diabetic health states. Cost-effectiveness estimates for potential interventions are obtained by dividing the total costs by the incremental QALYs.

Parameter values and distributions

Parameter values for screening and prevention-related parameters are shown in Tables 21 and 22, respectively.

TABLE 22 - Parameter values and distributions for probabilistic sensitivity analysis: prevention-related parameters

RRR, relative risk reduction.

Parameter	Mean	95% CI	Source	Distribution
Modifiers of effectiveness of preventative intervention (for diabetes)
Specific adjustment (multiplier) to effectiveness of preventative interventions in phenotype identified by tests other than IGT-orientated OGTT	0.70	Assumed 0.5 to 1.0	Assumption	Log-normal
Assumed initial uptake of preventative interventions	55%	Assumed 45% to 65%	Advice from clinical authors (Khunti, Davies)	Beta
RRR per kg lost	16%	13% to 19%	Hamman (2006)53	Log-normal

The health state utilities in the model are shown in Table 23 and unit costs are shown in Table 24.

TABLE 23 - Utility data

SE, standard error.

Health state utility parameter	Mean (SE)	Source	Distribution
Utility for diabetes with no complications	0.785 (0.0530)	UKPDS 6256	Beta
Decrements for complications
CHD	–0.055 (0.0064)	UKPDS 6256	Gamma
CHF	–0.108 (0.0309)	UKPDS 6256	Gamma
Stroke	–0.164 (0.0298)	UKPDS 6256	Gamma
Microalbuminuria	–0.011 (0.009)	Coffey & Associates57	Gamma
Macroalbuminuria	–0.011 (0.009)	Coffey & Associates57	Gamma
Dialysis	–0.078 (0.026)	Coffey & Associates57	Gamma
Post renal transplant	–0.052 (0.0133)	Mount Hood 4 Conference data	Gamma
Neuropathy	–0.065 (0.008)	Coffey & Associates57	Gamma
Amputation	–0.280 (0.0559)	UKPDS 6256	Gamma
Proliferative retinopathy	–0.020 (0.0051)	Mount Hood 4 Conference data	Gamma
Macular oedema	–0.020 (0.0051)	Mount Hood 4 Conference data	Gamma
Severe vision loss	–0.074 (0.0255)	Mount Hood 4 Conference data	Gamma
Weight (per kg)	–0.0025 (0.0011)	Weighted average of published studies58	Gamma

TABLE 24 - Unit costs

Based on 42% of events non-fatal, 58% fatal (UKPDS 6559).

Based on 79% of events non-fatal, 21% fatal (UKPDS 6559).

Unit costs	Mean	Distribution assumptions	Source	Distribution
Acute cost of MI – non-fatal MIa	£6153	Mean –20%, +25%	UKPDS 6559	Log-normal
Acute cost of MI – fatal MIa	£1742	Mean –20%, +25%	UKPDS 6559	Log-normal
Annual cost following MI	£702	Mean –20%, +25%	UKPDS 6559	Log-normal
Acute cost of strokeb
Acute cost of stroke – non-fatal strokeb	£3579	Mean –20%, +25%	UKPDS 6559	Log-normal
Acute cost of stroke – fatal strokeb	£5115	Mean –20%, +25%	UKPDS 6559	Log-normal
Annual cost following stroke	£5892	Mean –20%, +25%	Chambers et al.60	Log-normal
CHF incidence	£3594	Mean – 20%, +25%	UKPDS 6559	Log-normal
CHF state cost	£909	Mean –20%, +25%	UKPDS 6559	Log-normal
Haemodialysis per annum	£36,419	Mean –20%, +25%	UK Transplant61	Log-normal
Peritoneal dialysis per annum	£18,210	Mean –20%, +25%	UK Transplant61	Log-normal
Transplant – first year	£17,689	Mean –20%, +25%	UK Transplant61	Log-normal
Cost of immunosuppression per annum	£5203	Mean –20%, +25%	UK Transplant61	Log-normal
Annual cost of neuropathy	£214	Mean –20%, +25%	Gordois et al.62	Log-normal
Amputation	£12,789	Mean –20%, +25%	UKPDS 6559	Log-normal
Post-amputation costs per annum	£454	Mean –20%, +25%	Palmer et al.63	Log-normal
Major hypoglycaemic episode	£659	Mean –20%, +25%	Heaton et al.64	Log-normal
Retinal photocoagulation	£1073	Mean –20%, +25%	UK National Screening Committee65	Log-normal
Severe vision loss per annum	£425	Mean –20%, +25%	UKPDS 6559	Log-normal
Cost of management/monitoring – clinic visits, glucose tests, and proteinuria and eye screening	£269	Mean –20%, +25%	Calculation	Log-normal
Heart failure (temporary adverse event)	£3426	Mean –20%, +25%	UKPDS 6559	Log-normal
Oedema	£42	Mean –20%, + 5%	UKPDS 6559	Log-normal

Details of coefficients used within the CHD,40 stroke,41 CHD/stroke case fatality54 and congestive heart failure (CHF)55 risk equations used are shown in Tables 25–28.

TABLE 25 - Coronary heart disease risk equation parameter estimates (UKPDS 56)40

SE, standard error.

Interpretation	Estimate (SE)	Distribution
Intercept	0.0112 (0.001)	Normal
Risk ratio for 1 year of age at diagnosis of diabetes	1.059 (0.005)	Normal
Risk ratio for female sex	0.525 (0.054)	Normal
Risk ratio for Afro-Caribbean ethnicity	0.390 (0.102)	Normal
Risk ratio for smoking	1.350 (0.122)	Normal
Risk ratio for 1% increase in HbA1c	1.183 (0.036)	Normal
Risk ratio for 10-mmHg increase in systolic blood pressure	1.088 (0.026)	Normal
Risk ratio for unit increase in logarithm of lipid ratio	3.845 (0.640)	Normal
Risk ratio for each year in duration of diagnosed diabetes	1.078 (0.015)	Normal

TABLE 26 - Stroke risk equation parameter estimates (UKPDS 60)41

SE, standard error.

Parameter	Interpretation	Estimate	SE	Distribution
q0	Intercept	0.00186	0.0004	Normal
β1	Risk ratio for 1 year of age at diagnosis of diabetes	1.092	0.013	Normal
β2	Risk ratio for female sex	0.700	0.109	Normal
β3	Risk ratio for smoking	1.547	0.237	Normal
β4	Risk ratio for atrial fibrillation	8.554	2.963	Normal
β5	Risk ratio for 10-mmHg increase in systolic blood pressure	1.122	0.042	Normal
β6	Risk ratio for unit increase in lipid ratio	1.138	0.053	Normal
d	Risk ratio for each year in duration of diagnosed diabetes	1.145	0.026	Normal

TABLE 27 - Coronary heart disease and stroke case fatality parameter estimates (UKPDS 66)54

SE, standard error.

Parameter	Estimate	SE	Distribution
CHD
Age	0.048	0.024	Normal
HbA1c (per 1%)	0.177	0.012	Normal
SBP	0.140	0.061	Normal
Time to event, from diabetes diagnosis	0.104	0.042	Normal
Stroke
SBP	0.246	0.078	Normal
Previous stroke	2.21	0.545	Normal

TABLE 28 - Congestive heart failure risk equation coefficients (UKPDS 68)55

SE, standard error.

Parameter	Estimate	SE	Distribution
Age	0.093	0.016	Normal
HbA1c	0.157	0.057	Normal
SBP	0.114	0.056	Normal
λ	–8.018	0.408	Normal
γ	1.711	0.158	Normal

Comparison of pathways and outcomes of the screening process

Table 32 shows the screening yield and resource implications of each screening strategy evaluated.

The third column of Table 32 shows the percentage of 40- to 70-year-olds attending health checks who would be offered a blood test, taking account of those indicated as at low risk for diabetes at the prescreening stage and, therefore, not offered a blood test.

The fourth column shows the proportion of cases of undiagnosed diabetes that would be detected given any prescreening and assuming 100% uptake of blood tests.

The fifth column shows the proportion of cases of undiagnosed HRD that would be detected given any prescreening and assuming 100% uptake of blood tests.

The sixth column shows the proportion of health check attendees whose diabetes risk score is calculated and who, where applicable, take up any offers of an initial and confirmatory blood test.

The seventh column shows, for each screening strategy:

• upper figure (within the bracket): the proportion of 40- to 74-year-olds eligible for risk assessment that would receive a diagnosis of diabetes, taking account of rates of uptake of blood tests

• lower figure: the prevalence of undiagnosed diabetes among 40- to 74-year-olds eligible for risk assessment.

The eighth column shows:

• upper figure: the proportion of 40- to 74-year-olds eligible for risk assessment that would receive a diagnosis of HRD, taking account of rates of uptake of blood tests

• lower figure: the prevalence of undiagnosed HRD among 40- to 74-year-olds eligible for risk assessment.

Clearly, one of the main results is that the numbers of people detected with diabetes is strongly influenced by the glucose test, as the prevalence of HbA_1c-defined diabetes is 5.7%, compared with a prevalence of 1.8% for FPG-defined diabetes. The sensitivities of the testing strategies are broadly similar when comparing HbA_1c and FPG. For example, for the two strategies LPDS 4.75/HbA_1c 6.5 and LPDS 4.75/FPG 7.0, the sensitivity for detecting diabetes is 94.2% and 96.0%, respectively, while sensitivity for HRD is 84.0% and 81.5%, respectively (all assuming 100% uptake of tests).

It is the proportion of cases detected with HRD that differs most across the strategies. The two NICE-recommended strategies which use a risk score (‘LPDS 4.75/HbA_1c 6.0’ and ‘LPDS 4.75/FPG 5.5’) are estimated to detect 15.8% and 15.9%, respectively, of the total 40–74 years eligible population as at HRD. For the two NICE-recommended strategies that do not use a risk score, strategy ‘HbA_1c 6.0%’ would detect HRD in slightly more people, at 17.6%, while strategy ‘FPG 5.5’ would detect HRD in considerably more, at 23.1%. These levels of detection are from an underlying prevalence of HRD of 18.3% for HbA_1c 6.0–6.4% and 23.8% for FPG 5.5–6.9 mmol/l. Other potential future strategies, with lower thresholds than recommended by NICE for defining HRD, would detect even more cases of HRD, for example ‘LPDS 4.75/HbA_1c 5.8’ would detect 28.7%, ‘LPDS 4.75/FPG 5.2’ would detect 26.6%, and ‘LPDS 4.75/HbA_1c 5.9% or FPG 5.4’ would detect 25.7%.

Screening cost per case detected results

The screening cost per case of detected diabetes or HRD is shown in Table 33. This includes only the costs incurred up to the point of obtaining a definitive diagnosis, that is the cost of prescreening, blood tests for diabetes/HRD and associated staff and laboratory costs.

The cost per person of diabetes detected is lower when screening with a HbA_1c test than with a FPG test because of the higher prevalence of undiagnosed diabetes (5.7% vs. 1.8%). When cases of HRD are included, the cost per case detected is lower for FPG because the test is cheaper (and, when no risk score is used, a greater number of cases of HRD are identified). The cost per person eligible for screening is lower for a FPG test because of its lower test cost.

This measure is of limited usefulness to decision makers as an indicator of value for money because:

1. It depends on the arbitrary definition of HRD. For example, progressive lowering of the cut-off point for defining HRD will inevitably reduce the cost per case detected because it results in identification of more cases. However, these additional people are at progressively lower risk of diabetes; therefore, they will, on average, have less capacity to benefit from the prevention intervention.

2. When cases of HRD are included in addition to diabetes, the cost per person attending health checks becomes very small, as shown in Table 33. As a result, any difference in the short-term screening cost of HbA_1c testing versus FPG testing is unlikely to be a key driver of overall long-term cost-effectiveness, which takes account of many other elements that are important.

Results for strategies recommended in National Institute for Health and Care Excellence guidance, based around the Leicester Ethnic Atherosclerosis and Diabetes Risk cohort

Figure 10 is a recap of the analyses undertaken.

Table 34 shows results comparing HbA_1c testing with FPG testing assuming that any prescreening is undertaken with the LPDS. Results for the comparison of whether or not LPDS or RCG is the best tool for prescreening are shown later in Use of an RCG test or the Leicester Practice Database Score to prioritise who should receive the blood test.

Table 34 shows the total costs and QALYs of each strategy in the third and fourth columns. The next two columns show the incremental costs and QALYs of the NICE-recommended strategies compared with the ‘no screening’ strategy. The seventh column shows the ICER of each strategy compared with ‘no screening’. The next column, ‘net monetary benefit’ of a strategy, reports the monetary value of the expected total QALYs less the total costs (including screening cost), valuing 1 QALY at £20,000. The NMB is a useful measure, as the most cost-effective strategy is easily identified by the highest NMB. The most cost-effective strategy is also the one with the highest incremental NMB versus ‘no screening’, shown in the ninth column. The final column shows the probability that a strategy is the most cost-effective out of the set of ‘comparable strategies’ (i.e. those within each subsection of the table, such as ‘NICE-recommended strategies (diabetes and HRD) – with use of risk score’.

The main findings from this analysis are as follows: if a LPDS cut-off point of 4.75 is used to determine which individuals receive a blood test, then screening everyone at the Health Check using a HbA_1c test and an intervention cut-off point of 6.0% is more cost-effective than screening everyone at the Health Check using a FPG test with a cut-off point of 5.5 mmol/l. The incremental costs and QALYs of HbA_1c testing compared with FPG testing are a saving of £12 (£66 – £78) and 0.0220 (0.0513–0.0293), respectively; therefore, HbA_1c testing appears to marginally dominate FPG testing. PSA indicates a 98% probability that HbA_1c 6.0% is more cost-effective than FPG 5.5 mmol/l.

While both of these strategies would identify around 16% of individuals as at HRD, more would be identified with diabetes in HbA_1c testing than FPG testing (4.4% vs. 1.2%).

Probabilistic sensitivity analysis around the long-term cost-effectiveness for the National Institute for Health and Care Excellence-based strategies using the Leicester Ethnic Atherosclerosis and Diabetes Risk data set

Probabilistic sensitivity analysis process

Probabilistic sensitivity analysis was undertaken on each screening strategy with 60 PSA sample runs, each containing 8147 patients (i.e. nearly 490,000 patients in total).

To establish what number of PSA runs would be enough to ensure that results were stable, we investigated the stability of example model runs with pairs of strategies, checking whether or not the difference in net benefit between the strategies was stable enough to be meaningful. Figure 11 shows that the results become stable after a relatively small number of PSA sample runs (£20,000 cost/QALY threshold assumed).

FIGURE 11.

Cumulative mean difference and 95% confidence intervals for the mean difference in net benefit between strategies r4.75/HbA_1c 6.0 and r4.75/FPG 5.5.

Cost-effectiveness acceptability curve for glycated haemoglobin versus fasting plasma glucose for National Institute for Health and Care Excellence-recommended strategies

The chart in Figure 12 shows the likelihood of each of the two NICE strategies (LPDS 4.75/HbA_1c 6.0, LPDS 4.75/FPG 5.5) being the more cost-effective of the two strategies at alternative willingness-to-pay thresholds. It can be seen that, even at low willingness-to-pay thresholds, such as £6000 per QALY, HbA_1c is 95% likely to be more cost-effective than FPG and remains extremely likely (97–98%) to be cost-effective at higher thresholds up to and beyond the usual NICE threshold of £20,000 per QALY.

FIGURE 12.

Likelihood of each of the two NICE strategies (with a risk score) being the more cost-effective at alternative acceptability thresholds.

Cost-effectiveness plane for glycated haemoglobin versus fasting plasma glucose for National Institute for Health and Care Excellence-recommended strategies

Figure 13 shows the cost-effectiveness plane for strategy LPDS 4.75/HbA_1c 6.0 versus LPDS 4.75/FPG 5.5. Each dot represents the incremental cost and QALY results for each of the 60 sample runs of the PSA. The error bars show the 95% confidence intervals for the incremental costs and QALYs. The green ellipse indicates the range within which 95% of the data points are expected to lie (so there is 95% certainty that the true incremental costs and QALYs be within this range). The blue line is the £20,000 cost/QALY willingness-to-pay threshold; therefore, any dots below and to the right of the green line indicate that strategy LPDS 4.75/HbA_1c 6.0 is more cost-effective than LPDS 4.75/FPG 5.5. As 98% of the dots lie below and to the right of the green line, this indicates that LPDS 4.75/HbA_1c 6.0 is 98% likely to be the more cost-effective of the two strategies, as per Cost-effectiveness acceptability curve for glycated haemoglobin versus fasting plasma glucose for NICE-recommended strategies.

FIGURE 13.

Cost-effectiveness plane for strategy LPDS 4.75/HbA_1c 6.0 vs. LPDS 4.75/FPG 5.5.

Scenario analyses to assess the effect of alternative evidence on prevalence and alternative assumptions for rates of uptake of blood tests

Scenario analyses were undertaken only around the NICE guideline-based strategies that include a risk score, that is LPDS 4.75/HbA_1c 6.0 versus LPDS 4.75/FPG 5.5, because these strategies were more likely to be sensitive to alternative prevalence of diabetes and HRD than ISO-resource strategies (see Table 35).

Marginal net benefit figures greater than zero indicate that HbA_1c is more cost-effective than FPG.

For the majority of scenario combinations of prevalence and uptake, HbA_1c testing is very or highly likely to be more cost-effective than FPG testing.

The exceptions are where HbA_1c-based prevalence of undiagnosed diabetes is much lower than LEADER and at a similar level to FPG-based prevalence, as in the UEA-IFG study, but it still depends on the relative prevalence of HRD and relative uptake rates as per Figure 14. These exceptions can be broken down into two cases:

1. If the prevalence of HbA_1c-based HRD is very low compared with that for FPG (as in UEA-IFG), then FPG testing is more likely to be cost-effective than HbA_1c testing, unless there is a very large difference in uptake of the tests (at least of the order 35%).

2. If the prevalence of HbA_1c-based HRD is lower than that for FPG but higher than in UEA-IFG, then only if there is a small difference in uptake rates (less than 20%) is it likely that FPG testing is more cost-effective than HbA_1c testing.

FIGURE 14.

Marginal net benefit for prevalence and uptake scenarios.

Deterministic, one-way (non-prevalence-related) sensitivity analysis results using Leicester Ethnic Atherosclerosis and Diabetes Risk data set

A set of eight deterministic sensitivity analyses were undertaken for the main comparison of HbA_1c-based versus FPG-based testing.

Results are presented comparing the two NICE-recommended strategies that include a risk score and separately comparing the two corresponding strategies with no risk score. An additional four were carried out just in the context of the NICE-recommended strategies that include a risk score.

These analyses test how results change under different assumptions or parameter values in the model. The results shown in Tables 36 and 37 all indicate that HbA_1c testing appears to be more cost-effective than FPG testing.

National Institute for Health and Care Excellence guidance strategies: results obtained without using a risk score

As per Table 38, if no risk score is used, the incremental costs and QALYs of using HbA_1c 6.0% versus FPG 5.5 mmol/l are a saving of £30 (£75–£105) and 0.0224 (0.0566–0.0342), respectively; therefore, HbA_1c testing appears to dominate FPG testing. PSA indicates that there is a 95% probability that HbA_1c 6.0% is more cost-effective than FPG 5.5.

With no risk score, more cases of HRD are identified with FPG screening, which partially offsets the benefits of HbA_1c screening identifying more cases of undiagnosed diabetes (5.7% vs. 1.8%).

When comparing all four NICE-related strategies (i.e. with and without a risk score) together with no screening, screening everyone using HbA_1c 6.0 provides the most QALYs, and when measured on the net benefit scale assuming a value for 1 QALY of £20,000, also has the highest expected net benefit. There are, however, disadvantages to not using a risk score (see Chapter 5, Statement of principal findings) even if there were the capacity to undertake blood tests on everyone.

Results for alternative cut-off points for preventative intervention to the National Institute for Health and Care Excellence recommendations (‘ISO-resource’ strategies)

Table 39 shows that, for any pair of ‘ISO-resource’ strategies (i.e. where the proportion of individuals identified as at HRD is the same for the HbA_1c strategy as for the FPG strategy, as explained in Chapter 2, Derivation of final set of screening strategies to model), the HbA_1c-based strategy is more cost-effective than the FPG-based strategy. This is clear from the fact that HbA_1c has both a higher NMB and a higher incremental net benefit than ‘no screening’.

Figure 15 aims to help to compare many strategies visually. It presents the results in the form of a cost-effectiveness plane, the x-axis showing incremental QALYs compared with no screening and the y-axis showing incremental cost compared with no screening. The bold dashed black diagonal line represents the acceptability threshold value of £20,000 per QALY gained, that is points below and to the right of this line would be considered cost-effective compared with ‘no screening’. This is equivalent to a strategy having a higher expected NMB than ‘no screening’.

FIGURE 15.

Incremental costs and QALYs vs. ‘no screening’ for each strategy.

Use of a random capillary glucose test or the Leicester Practice Database Score to prioritise who should receive the blood test

Once it had been established that HbA_1c is more cost-effective than FPG, a separate analysis was undertaken to compare the cost-effectiveness of using an RCG test with using the LPDS risk score as a first step to prioritise individuals for further testing for diabetes or HRD using a HbA_1c test (see Table 40).

The RCG and LPDS cut-off points for this comparison were chosen such that a similar proportion of individuals (with HbA_1c ≥ 6.0%) would be identified. The comparisons were made in the context of (1) the NICE HbA_1c threshold of 6.0% for HRD and (2) a lower HbA_1c threshold of 5.7% for HRD.

If blood tests cannot be undertaken in all eligible individuals because of capacity or budget constraints, the LPDS is highly likely to be more cost-effective for prescreening than using an RCG test at a HbA_1c cut-off point of 6.0% for HRD. At lower HbA_1c thresholds for HRD, the choice appears to be much less certain.

There are a couple of factors that could not be taken account of within the modelling:

1. The LPDS might also identify individuals with slightly higher non-invasive risk factors for progression to diabetes. Unfortunately, there does not exist, to our knowledge, a risk equation that captures the variables included within the LPDS, plus HbA_1c and FPG. Without this, it is not possible to account for these risk differences within the modelling.

2. An important part of the two-stage process of prescreening with a risk score, followed by a blood test if indicated, is that the risk score provides feedback to individuals on modifiable risk factors which may motivate them to make lifestyle changes to reduce their risk of developing diabetes. Such effects have not been quantified for use in the LPDS and therefore cannot be captured within the modelling.

Statement of principal findings

HbA_1c testing is highly likely to be more cost-effective than FPG testing for screening for diabetes and HRD based on (1) the multiethnic LEADER cohort and (2) most alternative prevalence and uptake scenarios assessed. Although the absolute differences between tests in health gains per individual are small, the PSA shows that this conclusion is 95–98% certain in a region with similar relative (HbA_1c- to FPG-defined) prevalence of undiagnosed diabetes and HRD as in the multiethnic Leicestershire-based LEADER cohort.

Scenario analyses indicate that, where the relative prevalence of HbA1c- to FPG-defined undiagnosed diabetes and HRD is lower than in the LEADER cohort, for example as in the UEA-IFG study, the relative cost-effectiveness depends on the size of the difference in uptake rates of screening for a HbA_1c test versus a FPG test.

Given that only a few UK-based studies appear to exist that contain both baseline HbA_1c and baseline FPG levels from a screening setting (or recruitment into a diabetes prevention programme), it is not possible to suggest to what extent the LEADER cohort is representative of the UK as a whole.

The prevalence patterns shown in Table 4 (see Chapter 2, Other UK cohorts providing estimates of prevalence of diabetes and high risk of diabetes using both glycated haemoglobin and fasting plasma glucose), together with the results from the scenario analyses (around alternative prevalence and uptake rates), would seem to suggest that HbA_1c testing is likely to be more cost-effective than FPG testing in most localities; the exceptions are for combinations of prevalence and uptake that closely match one of the scenarios that reports a negative incremental net benefit in Table 35 in the results section.

Based on the LEADER cohort, sensitivity analyses around parameters unrelated to prevalence and uptake suggest that the conclusions are generally insensitive to a range of alternative assumptions around such parameters.

Lowering the thresholds for HRD and preventative intervention does not change the finding that HbA_1c testing appears more cost-effective than FPG testing. This same finding was found for several ‘ISO-resource’ comparisons.

Owing to capacity or budget constraints, in most localities prescreening is likely to be necessary for the purpose of identifying which individuals should receive a blood test. In some areas with a very high diabetes risk profile (e.g. multiethnic localities), it may be more practical to carry out blood tests in everyone. In such cases, HbA_1c testing remains highly likely to be more cost-effective than FPG testing. Note that this does not mean that the risk scoring should not be done at all (because drawing people’s attention to modifiable risk factors is likely to be beneficial).

Use of the LPDS risk score is likely to be more cost-effective than using an RCG test for prescreening. It should also be noted that a risk score may have additional effects not captured within the modelling – the calculation of the score may draw attention to high-risk factors for developing diabetes in the future (rather than just providing a normal/abnormal blood test result). This might prompt individuals to make lifestyle changes to reduce their risk, that is there may be an intervention effect even if their HbA_1c or FPG is below the corresponding HRD threshold for intensive intervention.

Strengths and limitations

This analysis represents, to our knowledge, the first economic analysis comparing the long-term cost-effectiveness of HbA_1c testing with FPG testing to screen for diabetes and HRD. A particular strength is the allowance of an individual to have a different screening outcome (in terms of diabetes, HRD or NGT) according to the differing diagnostic definitions of HbA_1c and FPG tests rather than based on the historical gold standard OGTT test.

Comparison with related studies

It should be noted that the total and incremental cost and QALY figures reported in this document may be an order of magnitude different from previous work on screening and prevention, including the economic modelling for the 2012 NICE guidance on risk assessment. 12 This is because of different discount rates used as a result of the recent NICE recommendation to use 1.5% per annum for public health-related evaluations, rather than 3.5%. 68

Implications

Unanswered questions and recommendations for further research

About the Sheffield Health Economics and Decision Science (HEDS) research group

The HEDS research group is part of the School of Health and Related Research (ScHARR), a multidisciplinary group within the Faculty of Medicine. Within the HEDS group, there is a range of specialties, including decision-analytic modelling and technology appraisal. The ScHARR Technology Assessment Group undertakes reviews of the clinical effectiveness and cost-effectiveness of health-care interventions for the NHS Research and Development (R&D) Health Technology Assessment (HTA) Programme on behalf of a range of policy-makers, including the NICE. Health economists and mathematical modellers work with systematic reviewers to undertake cost-effectiveness analyses using decision-analytic modelling for local and national health-care decision-making bodies. Members of HEDS also carry out and publish research papers on the role and methods of, and quality assurance in, modelling.

Contributions of authors

Mike Gillett, Research Fellow (Health Economics): principal investigator and conducted the economic modelling.

Alan Brennan, Professor of Health Economics and Decision Modelling: supervised the economic modelling.

Penny Watson, Research Associate (Health Economics): assisted with analysis of the LEADER data set and literature reviewing.

Kamlesh Khunti, Professor of Primary Care Diabetes and Vascular Medicine: provided advice on clinical and policy issues.

Melanie Davies, Professor of Diabetes Medicine and Honorary Consultant: provided advice on clinical and policy issues.

Samiul Mostafa, Clinical Research Fellow/SpR Diabetes Medicine: provided advice on clinical and policy issues.

Laura J Gray, Senior Lecturer of Population and Public Health Sciences: provided advice on LEADER data set.

The views expressed in this report do not necessarily reflect those of the NHS HTA programme or the UK Department of Health.

Disclaimers

This report presents independent research funded by the National Institute for Health Research (NIHR). The views and opinions expressed by authors in this publication are those of the authors and do not necessarily reflect those of the NHS, the NIHR, NETSCC, the HTA programme or the Department of Health. If there are verbatim quotations included in this publication the views and opinions expressed by the interviewees are those of the interviewees and do not necessarily reflect those of the authors, those of the NHS, the NIHR, NETSCC, the HTA programme or the Department of Health.

Glycated haemoglobin: limitations and practical use

The standardisation of HbA_1c testing has much improved over recent years. Even so, recent data from an external Quality Assessment sample found a 4.3% coefficient of variation between laboratories, with results varying from 40 to 60 mmol/mol, and that the analytical variation for HbA_1c, though reducing, was still greater than for FPG. 17

The principal remaining concerns are around:14

• other conditions, such as haemoglobinopathy, that affect HbA_1c levels

• a growing body of evidence that some ethnic groups have naturally higher HbA_1c levels (although the clinical significance of this is not known)

• the effect of ageing on HbA_1c levels. 73

Fasting plasma glucose: limitations and practical use

Glucose testing using a FPG test is currently still subject to less analytical (laboratory) variation than HbA_1c testing. However, FPG testing suffers from much greater non-laboratory variation (biological variation and preanalytical variation). 74

Intra-individual biological variation has been reported to be between 5.7% and 8.3%, with inter-individual biological variation up to 12.5%. 74

As stated in Chapter 2, Determining diagnostic outcome of individuals in the Leicester Ethnic Atherosclerosis and Diabetes Risk data set, where the result of the first blood test is below the cut-off for diabetes but in the range considered HRD, there is not a requirement to carry out a second test.

For FPG in particular, this raises some concern about reliability of a HRD label in such cases, especially given the high biological variation and problems with stability of glucose levels within clinical practice. It is often argued that, even if such individuals test ‘false-positive’ based on a single FPG test, they will still obtain a degree of benefit from lifestyle intervention. This argument may not be economically sound, as the benefits of an intensive lifestyle intervention in lower risk patients may not justify the cost.

Ideal methods to avoid error before testing by stabilising glucose, such as placing tubes in ice water immediately after collection and/or separating cells from plasma within minutes, are impractical in a clinical setting. Across the UK, practices probably range from samples being kept in a fridge to being left on a shelf until collected and transported to a laboratory around 1pm. As glucose levels decrease in test tubes by 5–7% per hour because of glycolysis,74 a sample with a true blood glucose value of 126 mg/dl would fall to 111 mg/dl and 98 mg/dl after 2 hours and 4 hours, respectively, at room temperature. This is shown in Figure 16, alongside an alternative scenario where decay occurs at half the rate, that is 3% per hour.

FIGURE 16.

Decline in FPG over time under alternative decay assumptions.

The duration and implications of this overall lead time between a FPG sample being drawn and being processed at a laboratory are unclear. Qualitatively at least, we understand that fairly similar practices probably occurred in epidemiological and intervention studies from which the evidence for the modelling is sourced, although whether or not this was to the same extent is not known to us. It does, however, suggest that, where the lead time is long, some individuals may not be treated for diabetes when they should be.