Development of a cost-effectiveness model for optimisation of the screening interval in diabetic retinopathy screening

Article history

The research reported in this issue of the journal was funded by the HTA programme as project number 10/66/01. The contractual start date was in May 2012. The draft report began editorial review in December 2014 and was accepted for publication in May 2015. 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

Peter Scanlon has received fees from Bayer Plc (Newbury, UK), Allergan Ltd (Marlow, UK), Roche (F. Hoffmann-La Roche Ltd, Basel, Switzerland) and Boots Opticians Ltd (Nottingham, UK) for consultancy work or advisory board participation and part of his salary is supported by the NHS Diabetic Eye Screening Programme in Public Health England. Stephen Aldington has part of his salary supported by the NHS Diabetic Eye Screening Programme in Public Health England. Sobha Sivaprasad has received fees from Novartis Pharmaceuticals UK Ltd (Frimley, UK), Bayer, Allergan and Roche for travel grants and research grants and advisory board participation.

Copyright statement

© Queen’s Printer and Controller of HMSO 2015. This work was produced by Scanlon 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.

Introduction

Diabetes affects over 3 million people in the UK. The treatment and complications of diabetes cost over 10% of the NHS budget. 1 The International Diabetes Federation (IDF) published data2 in 2006 which showed that diabetes affects 246 million people worldwide, with 46% of all those affected in the 40–59 years working age group. By 2009, IDF estimates showed that 285 million people were living with the disease, with 90% of the burden caused by type 2 diabetes. 3 According to the IDF, reasons for these rises were driven by increases in life expectancy combined with sedentary lifestyles and increasing levels of obesity. In England, quarterly returns to the Department of Health4 show that the number of people identified by general practitioner (GP) practices as having diabetes had risen from approximately 1.4 million people in 2003 to 2.4 million5 in the third quarter of 2011, the numbers having been rising by 120,000 per annum since accurate data from the screening services became available in 2008. The number in the UK as a whole exceeded 3 million in 20136 and, at the same time, a further 850,000 people were estimated to have undiagnosed type 2 diabetes mellitus (DM) by the charity Diabetes UK. 6

Diabetic retinopathy (DR) results from damage to the small blood vessels in the back of the eye, which generally occurs after about 10 years of diabetes and is the leading cause of blindness in the working age population around the world,7,8 except in countries such as Iceland9 and England,10 where screening programmes have become established. Two recent studies11,12 have shown that since 1985 people with diabetes have experienced lower rates of progression to proliferative DR (PDR) and severe visual loss, probably reflecting improvements in diabetes care. It is practical and affordable to treat13,14 sight-threatening diabetic retinopathy (STDR) if detected early using screening with retinal photography. DR is treated effectively with laser photocoagulation, although this has been found to be cost-effective only if retinopathy is detected before irreversible damage takes place. 13,15,16 Therefore, in order for DR treatment to be cost-effective, diagnosis has to be timely. Published evidence has shown that screening for STDR is highly cost-effective. 17–20

Since 2003, when the English NHS screening programme for DR was introduced, cases of loss of sight have reduced. Offering annual DR screening to increasing numbers of patients puts immense strain on resources and may not be necessary. National Institute for Health and Care Excellence (NICE) recommendations exist but there is no evidence-based consensus as to the optimal frequency of testing for DR. The Scottish Intercollegiate Guidelines Network (SIGN)21 reported that patients with diabetes with no detected retinopathy could be screened every 2 years, with all others being screened at least annually. There have been a number of studies22–25 that have modelled the possible effect of moving from annual to 2–3-year intervals in patients with type 1 and type 2 diabetes using incidence data from screening programmes in Iceland, Liverpool and Norwich. In 2007, Olafsdottir and Stefansson26 reported the 10-year experience of biennial eye screening in patients with diabetes without retinopathy from Iceland, concluding that this seems to be safe and effective.

In 2000, Vijan et al. 27 reported on a cost–utility analysis of screening intervals for DR in patients with type 2 DM in an American population and concluded that annual retinal screening for all patients with type 2 DM without previously detected retinopathy may not be cost-effective, and tailoring recommendations to individual circumstances may be preferable.

An alternative modelling approach by Mehlsen et al. 28 used clinical data from 5365 patients who had undergone 23,324 examinations at the Department of Ophthalmology, Aarhus University Hospital and concluded that a subset of known risk factors for development and progression of DR should be used to construct a decision model for optimising screening intervals for DR.

The current studies have not included factors such as the economic consequences of visual loss other than legal blindness, the effect in ethnic or racial minority groups (the studies had mostly been in patients of northern European extraction), and that long intervals between appointments may lead to difficulties in maintaining follow-up with patients. 29 A study30 in the Coventry area of the UK found that people of South Asian ethnicity with diabetes were diagnosed at a younger age, had higher glycated haemoglobin (HbA_1c), systolic and diastolic blood pressure, and total cholesterol, and greater prevalence of DR and maculopathy. The Diabetic Retinopathy In Various Ethnic groups (DRIVE) UK study, comprising patients in screening programmes in West Yorkshire and South London, showed that DR and visual impairment were more prevalent in people of African-Carribean and South Asian origin. 31,32

In Gloucestershire, we have undertaken a preliminary analysis33 of screening data and know that we can produce accurate and quality-assured screening results from the system because data already transferred from GP practices have been of high quality. We have also been the pilot site for the national system development and roll-out of software [General Practice to Diabetic Retinopathy Screening (GP2DRS)] for transfer of demographic and risk factor data from GP practices to screening software systems. It is vital that any modelling uses data that can be obtained easily from routine clinical care.

Grading of retinal photographs is not an exact measurement and, hence, intraobserver and interobserver variability occurs in the grading. 34 Screening and monitoring programmes for chronic conditions generate both ‘false-positive’ (FP) tests and ‘false-negative’ (FN) tests. In the retinal screening programme, FP tests correspond to people detected by the retinopathy screening service as requiring referral to hospital, but who are subsequently found not to have retinopathy requiring treatment. FN tests correspond to people with retinopathy who are missed by the screening service. FP and FN tests are an inevitable consequence of inexact measurement.

Glasziou et al. 35 showed that, even when there is no reference standard to determine which positive tests in a database are FP, the proportion that are FP can be estimated from a model of the variability of the test; similarly, so can the proportion of negative tests that are FN be estimated. The model has been subsequently applied to blood pressure monitoring36 and it has been developed further for monitoring HbA_1c37 and microalbuminuria in diabetes. 38 In general the model needs to allow for variation between people, for average rate of change and variation in rate of change between people, and for the error rate in any individual test; we have recently published full details of the methodology. 39 These previous papers used a normal distribution model for the error structure, but we have conducted pilot methodological work showing that a logistic error structure can also be used successfully to model dichotomous outcomes.

As described previously, there is currently a well-documented epidemic of diabetes both in the UK and worldwide. With the increase in diabetes, the economic burden that DR places on society is high. 40,41 In the UK, it was estimated in 2006 that the lifetime costs of dealing with retinopathy can be up to £237,000 per person in the working age group, and that up to 50% of these costs are accounted for by the loss of productivity as a result of blindness or vision impairment. 42 Therefore, if 1000 cases of DR could be avoided in the working age population, the potential savings to the government could be in excess of £237M. 43 These compelling statistics encouraged the government to include annual DR screening in the National Service Framework for diabetes. 44

Conservative estimates of the costs of screening a patient with diabetes have been made of £26 per patient (although costs may be higher if screening is carried out in GP surgeries, and the estimates ignore costs to the patient in terms of travel costs and time off work). Although this cost might appear small, there are about 2.6 million people with diabetes in England alone. Assuming an uptake of 71% of those on the screening register (80% of those offered), these costs are of the order of £44M annually, without taking into account the centralised costs of the management of the national programme. This is a very sizeable financial resource, especially with the number of diabetes patients currently increasing at a rate of over 100,000 cases a year,4 which will drive up screening costs even further. In addition, any improvement in uptake would, obviously, have a corresponding impact upon these costs.

The current study uses currently available data from screening software systems and GP systems. The English NHS Diabetic Eye Screening Programme (DESP) has contracts with all the major GP software suppliers, including Egton Medical Information Systems (EMIS) Ltd (Leeds, UK), In Practice Systems Ltd (London, UK), iSOFT (formerly Torex; CSC Ltd. London, UK) and Microtest Ltd (Bodmin, UK). The Health and Social Care Information Centre GP Systems of Choice contracts make available the demographic transfer of data to screening programmes to invite people with diabetes for screening and to transfer risk factor data. These contracts use a consent model that has been agreed with the Patient Information Advisory Group and the National Information Governance Board. 45 With the imminent availability of this data to screening programmes in England there is an opportunity to provide cost-effective screening to patients which takes into account their absolute risk of developing STDR.

If tailored screening intervals were found to be clinically effective and cost-effective through this research, the logistical difficulties of delivering a national screening programme at a time of an epidemic of diabetes would be reduced, without increasing risks for any individual patients of losing vision. In addition, any potential savings might also give an opportunity to concentrate resources on persuading those most at risk to attend.

In addition, the evidence on the cost-effectiveness of different screening frequencies has been shown to be mixed. 46 A recent systematic review on the economic evidence of DR screening found three studies addressing the issue of screening frequency,46 one performed in the USA and two in the UK.

Vijan et al. 27 examined the cost-effectiveness of differing DR screening intervals for type 2 diabetes patients in the USA. The authors employed a Markov model using quality-adjusted life-years (QALYs) as the main outcome measure, with costs being assessed from a third-party payer perspective. Using the definition of $50,000 per QALY gained as a cost-effective intervention, the authors reported that annual retinal screening for all type 2 diabetes patients was not warranted on the basis of cost-effectiveness.

A UK study by Brailsford et al. 47 also found similar results to that of Vijan et al. 27 The study, which again used decision modelling techniques to assess the cost-effectiveness of screening intervals in diabetes patients (both types 1 and 2), found that a 30-month screening interval was the most cost-effective option. However, in contrast to these two studies, another UK study by Davies et al. 48 found that screening diabetic patients less than once a year would not be any more cost-effective than every year. Davies et al. 48 also based their study on a decision analytic model populated using results obtained from the published literature. A finding common in both UK studies was that screening of type 1 diabetes patients was more cost-effective than screening of type 2 diabetes patients.

There were, however, important limitations with both of the UK studies. Both studies used sight years saved as their main outcome measure rather than a more generalisable health outcome, such as QALYs, that can be readily compared across interventions and disease areas, aiding the decision process. 49 In addition, the use of QALYs captures the full impact of the disease, in this case sight loss or blindness, on patients’ lives. The two studies also failed to include in their models the additional costs to the health-care service of patients losing sight or going blind. Finally, data in these two studies were derived from a wide range of sources and diabetic populations. As a result, numerous assumptions had to be made as to how best to synthesise the available data.

Our research plan was divided into four phases

Phase 1: Risk factor approach (see Chapter 5).

Phase 2: Extension to ethnically diverse populations (see Chapter 6).

Phase 3: Monitoring interval approach (see Chapter 7).

Phase 4: Cost-effectiveness of differing screening intervals in DR screening (see Chapter 8).

Planned inclusion/exclusion criteria

The criteria for inclusion in these analyses were people with diabetes aged 12 years or above who have been screened by the participating programmes of the English NHS DESP. We included patients who had assessable retinal images from two or more screening episodes (SEs) and for whom we had sufficient clinical information. The exclusion criteria as set by the English NHS DESP50 include the following:

1. patients with no perception of light in each eye

2. patients with a terminal illness

3. patients with physical disabilities making photographic screening impossible

4. patients with learning or mental disabilities making photographic screening impossible

5. patients who are currently under the care of an ophthalmologist and for whom a report on retinal status has been provided to the screening service

6. patients who have chosen to opt out of screening.

Proposed outcome measures

To develop a risk-based algorithm for screening interval, cost-effectiveness and adverse events.

Other outcomes: key recommendations for further research

Research governance

The Nominated Sponsor of this research application is Gloucestershire Hospitals NHS Foundation Trust.

A Project Advisory Group was formed to oversee the progress of the research, and included an independent chair (Professor John Sparrow), the lead researcher (Professor Peter Scanlon), the co-applicants (Mrs Irene Stratton and Mr Steve Aldington), an independent statistician (Ms Catey Bunce), an independent consultant ophthalmologist (Miss Clare Bailey), a lay representative (Mr Mike Whatmore), and the research and development lead (Dr Julie Hapeshi) for the host Trust. Four of the Advisory Group members (John Sparrow, Catey Bunce, Clare Bailey and Mike Whatmore) formed the Independent Trial Steering Committee.

Proposed time period for retention of relevant documentation

The data from this study are pseudo-anonymised and will be kept for 8 years. It is likely that if any data are required after that time then a new extraction would be required.

Patient and public involvement

Mike Whatmore, a lay member of the Diabetes Research Network writing group ‘Screening for Diabetic Retinopathy’, was involved in the original discussions on both reasons for non-attendance in DR screening and screening intervals for DR and commented on the original application. A Diabetes UK Gloucestershire lay member, Mike Larkin, read and commented on the application. Mike Larkin has had type 1 diabetes since 1977 and has had DR since 1988. Mike Larkin is currently a member of a professional and lay group within Diabetes UK, setting the research strategy for the charity for the next 5 years. Feedback on the original application was also obtained from the Thames Valley Diabetes Local Research Network Patient and Lay Involvement Group during the development of the application.

Mike Whatmore serves as a member of the Project Advisory Group and the Independent Trial Steering Committee. Two other Gloucestershire lay members have regularly provided help and advice to the project at the different stages.

In our proposal we sought the views of the Warwick Diabetes Research and Education User Group on the original application and we were hoping to continue to consult with them but, unfortunately, as the Coventry and Warwickshire Screening Programme were unable to continue with the research project, it was not appropriate to continue using the resources of that user group.

Lay members contributed to this study report (in particular Mike Whatmore and Mike Larkin) and to the dissemination of results both formally and in their multiple contacts with health professionals and Diabetes UK.

Timeline changes

• Phase 1: Gloucestershire risk factor approach analysis commenced on 1 May 2012 and completed on 31 October 2012 as planned (Figure 1).

• Phase 2: extension to an ethnically diverse population commenced on 1 July 2012 in South London and the first data sets were received in May 2013 and further data sets received until 1 April 2014.

FIGURE 1.

Screening intervals for DR original study design.

Coventry and Warwickshire withdrew in August 2013.

The first DR screening data sets were successfully received from Nottinghamshire in August 2013 and further data sets were received until 1 April 2014.

Analysis continued until mid-November 2014.

• Phase 3: the Oxford monitoring interval approach commenced on 1 November 2012 and was completed on 27 April 2013.

• Phase 4: the health economic modelling commenced on 1 July 2012 and continued until mid-November 2014.

Rationale

Rates of incidence and progression of retinopathy are greater in people with a longer duration of diabetes and with poor glycaemic control (higher HbA_1c) and may also vary with other factors such as blood pressure and diabetes type. It has been shown in Denmark that it is possible to identify those at higher absolute risk of STDR,59 and this suggests that it may be possible to identify subgroups of those with diabetes in whom the screening interval can be extended without risk of STDR developing before the next screening visit. This Danish model was constructed using published data from several clinical trials for which patients were recruited in the 1970s and 1980s and followed into the 1990s. The Danish model may not be optimal for patients diagnosed in the past decade as these patients tend to be more obese and to be diagnosed somewhat earlier in the disease process as opportunistic screening becomes more frequent.

Data set

The Gloucestershire Diabetic Eye Screening Programme (GDESP) includes 27,000 people aged 12 years or older with diabetes who are eligible for annual screening. Our database includes:

• longitudinal data on photographs from 2005 to the present

  • digital photographs on two fields, with mydriasis (eyes dilated) to maximise image quality

  • consistent grading team

  • quality control, including repeat grading and external quality assurance

  • patients invited for screening annually but may attend less frequently

• risk factor data through separately funded ‘GP2DRS’ project

  • risk factor data were available from GP electronic records (EMIS) on 4400 screened patients by September 2010

  • by the end of December 2011, we had received risk factor data on over 12,000 patients.

Given that this data set is much larger than needed for this project, we used a data-splitting approach. We had originally planned to randomly select 50% of patients for the model derivation data set (‘derivation data’) and to reserve the other 50% for model validation purposes (‘validation data’). However, the reviewers on our original application suggested that a non-random method might be used for constructing the samples for the risk model and then testing it. We therefore altered the method of splitting the database to a systematic method using the postcode of the GP to define two areas (Group A and Group B) with similar sized populations using the population data from the Office for National Statistics for Local Government Districts (Table 2). Group A includes three more urban larger towns (Gloucester, Tewkesbury and Cheltenham) and the remaining areas that make up Group B are more rural.

Methods

Exploratory data analysis was used to look for outliers in all variables. The population was divided into derivation and validation sets based on the postcode of the GP practice. Screening data were used from 2005, the date at which retinal cameras were upgraded to modern camera backs and grading was stable. For each SE, the most recent clinical data items were identified and the baseline SE was defined as the first at which all required variables were present. Patients were included in the cohort if they had one of three categories of DR at initial eye screening: no DR in either eye; mild DR in only one eye; or mild DR in both eyes. Cox proportional hazard models were used with interval censoring to identify risk factors for progression. The parameter estimates from the derivation set were applied to the patients within the validation set and the risk of progression to referable DR is reported by centile of risk. In those patients with at least three SEs and clinical variables at the second of these visits we compared this model with our previously published risk model. 39

The multivariate modelling used the following parameters:

• outcome (either of):

  • referable DR (grade R2 pre-proliferative or R3 proliferative)

  • referable diabetic maculopathy (grade M1)

• exposures:

  • age, sex, smoking, BMI

  • previous retinal photographs

  • duration of diagnosed diabetes

  • type of diabetes

  • HbA_1c

  • systolic and diastolic blood pressure

  • measurements of renal function (urine albumin, serum creatinine)

  • lipid measurements

• analysis:

  • Cox proportional hazards model with interval-censoring (as retinopathy status can only be observed at screening visits not at the point at which STDR actually develops)

  • stepwise selection was used to develop a risk score for referable retinopathy/maculopathy at subsequent visits.

We also developed a model for risk of progression to maculopathy alone and the risk of progression to pre-proliferative DR and PDR alone.

Results

In total, 84,148 screening events were available, of which 75,491 were in 2005 or later. However, grading results were stored separately and once matching was carried out, and those with missing grading results omitted, there were 72,077 episodes and, of these, 15,410 had two or more SEs.

The data extraction from the GP practices yielded 3.3 million data items. Results recorded within the data set are shown in Tables 3 and 4.

In exploratory analyses, models were fitted using all variables from Table 3, but urine albumin was subsequently dropped as the data were not available for all patients, and its inclusion led to the overall loss of 12% of data from patients and 20% of the follow-up time.

The grading levels recorded by the programme are the same as those recorded in the NHS DESP in England and are shown in Table 5 with an approximate Early Treatment Diabetic Retinopathy Study (ETDRS) grade.

The model fitting was explored using patients for whom age, duration of diabetes, blood pressure, HbA_1c, glomerular filtration rate, smoking status, diabetes type and BMI were available. In total, 7012 people with diabetes were included in the derivation set and 5778 people with diabetes were included in the validation set (Table 6).

We investigated the effect of clinical variables having adjusted for the number of eyes with mild, non-proliferative DR (NPDR) to look at the univariate effects (Table 7).

The results (Table 8) of the Cox proportional hazards model show that the presence of background DR in both eyes at a screening visit has the highest hazard ratio (HR) of 7.13 [95% confidence interval (CI 5.84 to 8.70)] with background DR in one eye having a HR of 2.56 (95% CI 2.05 to 3.20), and HbA_1c of 1.28 (95% CI 1.23 to 1.34), and duration of diabetes (per 5 year increase), total serum cholesterol (per 1 mmol/l) and serum creatinine (per 10 µmol/l) all having lower HRs.

Hence, the risk score for a patient with mild NPDR in both eyes is given by:

The risk score for a patient with mild NPDR in one eye is given by:

The risk score for a patient with no retinopathy is given by:

The model was built using these risk factors. The ROC curve for overall endpoints in the validation set is shown in Figure 2.

FIGURE 2.

Receiver operating characteristic curve for validation data set. AUC, area under the curve.

Estimates of the ROC curve at given time points were obtained using the SAS macro TDROC (time dependent receiver operating characteristic) written by Mithat Gönen of Memorial Sloan Kettring Cancer Centre60 which generates a time-dependent ROC curve at a given time point for a censored outcome and an ordinal or continuous predictor. CIs were obtained using boot-strapping. Figure 3 shows the Kaplan–Meier estimates of time to referral by quintile of risk in the validation data set. At 4 years, 25% of those in the top quintile have been referred and in the lowest quintile fewer than 2% have been referred.

FIGURE 3.

Validation data set: time to referable screening result by quintile of risk (i.e. Group 1 contains those whose risk score is 0–20th centile, Group 2 is 21–40th centile, Group 3 is 41–60th centile, Group 4 is 61–80th centile, Group 5 is 81–100th centile).

The characteristics that are included in the model of the highest and lowest risk groups in the validation set are shown in Table 9.

The characteristics that are not included in the model of the highest and lowest risk groups in the validation set are shown in Table 10. There is very little difference in systolic and diastolic blood pressure between the highest and lowest risk groups in the validation set. Better control of blood pressure in the current era is likely to be the reason why blood pressure does not influence the model.

Figure 4 shows that almost all of the PDR is detected in the highest quintile of risk. The data are also shown in Table 11, which shows that no PDR is detected in the lowest two deciles of risk. The progression to STDR is shown in Table 12.

FIGURE 4.

First referable retinopathy detected by decile of risk in validation set.

The risk estimates are demonstrated in Figure 5.

FIGURE 5.

Model fitting for risk of progression to STDR. (a) No DR, HR compared with patient with no DR, 5 years of diabetes, HbA_1c = 50 mmol/mol; (b) mild NPDR in one eye, HR compared with patient with no DR, 5 years of diabetes, HbA_1c = 50 mmol/mol; and (c) mild NPDR in both eyes, HR compared with patient with no DR, 5 years of diabetes, HbA_1c = 50 mmol/mol.

Although the aim of the project was to examine the risk of progression to STDR for maculopathy M1 or for retinopathy (R2 or R3) collectively, we have examined the model fit for these endpoints separately.

These analyses (Tables 13 and 14) show that for progression to retinopathy R2 or R3, the presence of background retinopathy in both eyes or in just one eye is a more important risk factor than in progression to maculopathy M1. Cholesterol does not enter the model for progression to R2 or R3 but systolic blood pressure does enter the model. The inclusion of systolic blood pressure in the model of progression to R2 or R3 reflects results from earlier studies which analysed the progression of retinopathy alone (not including maculopathy) using scales such as the ETDRS scale.

Further analyses were carried out to assess the performance of the model proposed in our earlier paper based on the outcomes of two consecutive SEs. 39

The derivation data set for these analyses consisted of those patients who had two SEs with no referable DR and at least one further SE.

We looked at the performance of three models:

• two baseline SEs and clinical risk factor (CRF) data

• two baseline SEs

• one SE and CRF data.

In the Tables 15a–15c we have calculated the parameter estimates and HRs for the different models in all patients who had at least three SEs and clinical data available. In the derivation set, there were 162 events in 5774 people, and in the validation set, there were 131 events in 4873 people.

The following figures (Figures 6 and 7) show progression to referable DR by risk group. It is not possible to use quintiles of risk because most of the patients are in the low-risk group (no DR in each of two SEs). Hence, the groups have been defined to have the same number of patients as in the two SEs model.

FIGURE 6.

Progression to referable STDR by risk group in the three models. (a) Risk stratification in validation set using results of two SEs and clinical data; (b) risk stratification in validation set using results of two SEs; and (c) risk stratification in validation set using one SE and clinical information.

FIGURE 7.

The proportion of each grade of retinopathy by risk group in those found with referable retinopathy at the first SE where it was detected. (a) Referable retinopathy found by risk group: two screens plus clinical data; (b) referable retinopathy found by risk group: two screens; and (c) referable retinopathy found by risk group: one screen plus clinical data.

The bar charts in Figures 6 and 7 show the proportion of each grade of retinopathy by risk group in those found with referable retinopathy at the first SE where it was detected. There is no significant difference in the proportion of each grade of retinopathy between the three models.

Figure 8 and Table 16 show very little difference in the area under the curve (AUC) of the ROCs between the three models.

FIGURE 8.

Model comparison.

The AUC and 95% CI from boot-strapping are shown in Table 16.

Does the inclusion of cholesterol into the model improve the fit?

Although the addition of variables into the model may improve the fit using conventional estimations of fit, these may have little impact when it comes to defining groups for allocation of screening intervals. In order to examine this we have fitted the model including and excluding total cholesterol.

The distribution of risk for those who go on to develop STDR and those who do not develop STDR in the validation sets is shown in Figure 9.

FIGURE 9.

Ranked centile of linear risk predictor for those with and without STDR. Centile rank with cholesterol vs. centile rank without cholesterol. (a) With STDR; and (b) without STDR.

Reclassification among people who progressed to STDR and those who did not during follow-up is shown in Tables 17a–17d.

We now sum groups 1 to 3 because these groups have very similar progression rates to STDR.

Reclassifications for subjects with and without events are summarised below.

When cholesterol is added to the model, those whose ranking does not change remain on the diagonal, those who move into a higher group are in shaded light green and those who move into a lower group are shaded dark green.

The net gain in reclassification proportion can be shown using the reclassification methods of Pencina et al. 61

For 12 subjects who progressed to STDR, classification improved using the model with high-density lipoprotein, and for 19 people it became worse. This may be estimated using formulae of Pencina et al. 61 with the net gain in reclassification proportion of 0.120, significantly greater than zero (p-value < 0.001). The net gain in reclassification proportion for subjects who did not experience an event was not significant; 174 individuals were reclassified down and 173 were reclassified up (p-value = 0.957). The net reclassification improvement (NRI) was estimated at 0.121 and was highly significant (p-value < 0.001).

The NRI is estimated as:

Looking at the changes in those with events:

And in those without events:

So, the addition of cholesterol to the model does not improve net reclassification.

Classical risk factors for progression of diabetic retinopathy

The model used in this study was intended to determine the risk of referable DR, which included the risk of developing maculopathy as well as the risk of development of pre-proliferative DR or PDR. Some classical risk factors did not enter the model (e.g. blood pressure) because, although they were significant for the development of pre-proliferative DR or PDR, blood pressure was not significant in the development of maculopathy and so did not have a significant influence on the development of referable retinopathy. A total of 68% of the referable events were for maculopathy. The overall blood pressure was also better controlled than in the era of the UK Prospective Diabetes Study (UKPDS) study and the ETDRS study, and patients who attend regularly for screening do have better control than those who do not.

Rationale

It is good practice to verify statistical modelling results in external data, especially given that the Gloucestershire data are atypically high quality and collected on an ethnically homogenous population. 62 In this project we verified the results of Phase 1 modelling, described in Chapter 5, in external data sets of greater diversity.

Data

• Data from the East Anglian Programme including risk factors and grading results.

• Data from the South London screening service including risk factors and grading results.

• Data from the Nottingham screening service including risk factors and grading results.

We tested the risk score and algorithm against outcome retinopathy grades in the data sets.

Expected outcome

To develop a validated algorithm for recalling patients of white, Asian or African-Caribbean ethnicities for retinopathy screening, for use by the NHS Diabetic Eye Screening Programme and the 84 individual screening programmes across England.

Potential limitations

The Gloucestershire programme has a well-established quality-assured screening service with a stable group of experienced graders. Hence, it may not be typical of all screening programmes. Even with controls in place, it is likely that some misgrading will have occurred. We have, therefore, estimated when this may have occurred in Chapter 5 and utilised this in Chapter 7.

Gloucestershire has a small ethnic minority population. The main ethnic minority groups in Gloucestershire are Indian/British Indian (0.7%), and black/black British (0.8%), although the percentage of people from an ethnic minority group in the population with diabetes in Gloucestershire is nearer 5% because of the higher prevalence of diabetes in these groups. East Anglia has a similar low prevalence of ethnic minority groups. However, the data from South London and Nottingham used data sources that include a higher ethnic minority population from UK Asian, African-Caribbean and other white (i.e. non-British) Caucasian groups.

Results

The minimal model includes the following parameters:

• mild NPDR in both eyes at screening visit, HR 6.98 (95% CI 5.75 to 8.47)

• mild NPDR in one eye at screening visit, HR 2.50 (95% CI 2.01 to 3.12)

• HbA_1c (per 10 mmol/mol increase), HR 1.30 (95% CI 1.25 to 1.35)

• duration of diabetes (per 5 year increase), HR 1.20 (95% CI 1.16 to 1.24).

The model was validated in the following external data sets:

• data from 17,634 people with diabetes in the East Anglian Programme (A) who have grading results from two SEs and risk factor data

• data from 1223 people with diabetes in the South London Screening Service (B) who have grading results from two SEs and risk factor data

• data from 1083 people with diabetes in the Nottingham Screening Service (C) who have grading results from two SEs and risk factor data.

Data from programmes A, B and C are shown in Table 18.

The validation was carried out using the three parameter model (baseline retinopathy, duration of diabetes and HbA_1c), and the proportions progressing to STDR by quintiles of risk within each programme were estimated and are shown in Figure 10. Descriptions of the clinical information and rates of progression to STDR within each risk quintile by programme are shown in Table 19.

FIGURE 10.

Validation in three English programmes. (a) Programme A; (b) programme B; and (c) programme C. Patients grouped into quintiles of risk within each programme (i.e. Group 1 contains those whose risk score is 0–20th centile, Group 2 is 21–40th centile, Group 3 is 41–60th centile, Group 4 is 61–80th centile, Group 5 is 81–100th centile).

The proportion of patients in these three programmes who developed STDR or had no STDR on the date of censoring is shown in Figure 11 by quintile of risk. The component levels of STDR are shown. This demonstrates that the severity of STDR is associated with risk estimation, as was seen in the Gloucestershire validation data set (see Figure 4 and Table 11).

FIGURE 11.

Severity of DR at event or censoring. (a) Programme A; (b) programme B; and (c) programme C.

Model fit has been assessed using ROC curves and is shown in Figure 12.

FIGURE 12.

ROC for model with clinical information and one SE. (a) Programme A; (b) programme B; and(c) programme C.

Validation of model by ethnic group

The participants in programme A are 98% Caucasian, and so this programme cannot be used to examine the effect of ethnicity.

Programme B had too few people with clinical information from primary care in the data set to fit the model with one SE and clinical information and so it was only possible to validate the model that uses two SEs to stratify patients by baseline risk. The risk estimator from this model was applied and ethnic group categories fitted in a Cox proportional hazards model. Results are shown in Figure 13 and Table 20.

FIGURE 13.

Validation in programme B using the risk stratification based on two SEs. Progression to STDR. (a) All risk groups; (b) high risk; (c) intermediate risk; and (d) low risk.

Overall, the HR is over 50% higher for those of African or Caribbean ethnicity and 24% higher for those of other white ethnicity; in this category, many patients are Brazilian and Portuguese but the coding does not enable this to be broken down further.

The ethnicity coding in programme C is unreliable. The data for those with three SEs, with the first two episodes showing no referable retinopathy, are shown in Table 21. Over half the people in the cohort (54.8%) have no ethnic group recorded. There are 1257 people coded as African, yet the 2009 Office for National Statistics population estimates have 3100 people of African ethnicity in the whole area, and the African and Caribbean populations are about the same size. It appears that this coding may well be erroneous. Certainly, this level of uncertainty means that further analysis would be unreliable.

Conclusion

We have used data from three large, heterogeneous DESPs in England to validate the clinical model developed in Phase 1 (see Chapter 5).

• Within each of these three English screening programmes, the risk model discriminates well between those at very low risk and those at very high risk of progression to STDR.

• The model would be suitable for development of personalised screening intervals.

• Further validation in other data sets with robust ethnicity information would be useful.

Rationale

The reasons for the current annual monitoring interval are unclear and may be more historical than evidence-based. In recent years, a statistical methodology has been developed for studying the time intervals in monitoring and screening programmes and their consequences for patients and service providers. In this project we extended that methodology to DR screening to confirm (or otherwise) the appropriateness of annual screening as the default for retinal photography.

Data set

We used the same longitudinal data on digital photographs and risk factor data as the risk-factor approach of Chapter 5.

Overview of method

We used a statistical method that has been used extensively to model the course of chronic diseases, including DR,63 and to evaluate screening programmes. 64 This project advances on the risk-factor approach of Chapter 5 (Phase 1) by allowing classification error in retinopathy grades by using a hidden (unobserved) Markov model to estimate simultaneously the progression of disease and the potential for screening to misclassify the true state of DR. The factors contributing to the risk score of Phase 1 were included as explanatory variables.

Retinal screening gives a composite categorical measure (retinopathy grade and presence or absence of maculopathy). We converted this into seven states or levels relating to the early stages of DR and degrees of sight-threatening or referable disease. We applied the modelling methods to the GDESP data to obtain estimates of disease progression and the probability of misclassification. We calibrated the rates of FP and FN tests to estimates reported in the literature.

We estimated rates of true- and false-positive and -negative tests under the following scenarios:

• annual testing (base case)

• at other fixed intervals including 6-monthly, and every 2, 3 and 5 years.

The impact of varying levels of non-attendance was considered based on estimated attendance rates from our data and, in sensitivity analyses, from the literature. Simulations were cross-checked against other calculation methods, and further checked against the observed data to verify the internal validity of the model.

Expected outcome

An evidence-base for the existing annual monitoring interval or a recommended alternative to annual monitoring as the default.

Introduction

Diabetic retinopathy and maculopathy are complications of diabetes that can lead to vision loss. Regular screening using fundus photography is recommended for all people with diabetes with referral to specialist care for optimal treatment. Recommendations for the frequency of screening for eye disease and subsequent monitoring for further progression in clinical guidelines are inconsistent. NICE recommends65 annual screening, whereas the American Diabetes Association recommend66 annual screening initially but suggest that less frequent examinations (2–3 years) may be considered following one or two normal eye examinations. The SIGN guidelines21 suggest annual screening in general but every 2 years for people with no evidence of retinopathy. Australian guidelines67 for the management of diabetes suggest 2-yearly screening in people with no retinopathy, but annual screening for higher-risk groups and more frequent monitoring after retinopathy has been detected. Canadian guidelines68 differentiate between diabetes type and suggest annual screening for people with type 1 diabetes and 1–2-yearly screening for those with type 2 diabetes. The IDF guidelines69 recommend 1–2-yearly intervals for those with no disease, yearly intervals for those with minimal unchanged retinopathy and 3–6-monthly intervals if there is worsening since the last examination.

Given that screening using digital photography is imperfect and subject to various sources of variation, there is potential for people to be over-referred to specialist ophthalmology services in the absence of sight-threatening disease, or under-referred, when screening fails to detect referable level disease. In addition, imperfect methods can give the impression of change over time (progression or deterioration) when, in fact, the condition is stable. The consequences of such errors are numerous and wide ranging. When people are over-referred to specialist eye clinics, health-care resources (time and money) are wasted that could be better used to treat people with consequential disease, and the extra appointments are inconvenient for patients and may cause unnecessary anxiety. Under-referrals may mean a window of opportunity is missed for optimal interventions and could result in a poorer prognosis. Finally, misclassification can lead to more frequent and unnecessary monitoring which subsequently increases the risk of over-referral or, conversely, less frequent monitoring and unnecessary delays in detection of true change.

The aim of this study was to use routinely collected screening data to model the natural history of DR and maculopathy and the rates of correct and incorrect assignments using a statistical model. These estimates of disease progression and rates of over- and under-referral can be used to compare different screening intervals in cost-effectiveness models (Chapter 8).

Methods

The criteria for screening and grading of digital photographs are as follows. At each screening assessment, visual acuity was assessed using logarithm of the minimum angle of resolution (logMAR) charts and two colour digital retinal photographs of two standard 45-degree fields (macula and disc centred) per eye after dilatation of the pupils. Photographs are then graded by trained assessors in a central location for the presence of maculopathy or DR and the severity of DR. The English screening programme grades retina images39 as R0 if there is no detectable DR and as M0 if there is no evidence of maculopathy. Images are graded R1 if there is at least one microaneurysm and/or retinal haemorrhage. R2 (pre-proliferative) identifies the presence of multiple haemorrhages and/or definite intraretinal microvascular abnormality (IRMA) and/or venous beading and/or reduplication. R3 level (PDR) indicates the presence of neovascularisation. M1 (maculopathy) identifies the presence of two-dimensional photographic markers of diabetic maculopathy, specifically an exudate within 1-disc diameter of the centre of the fovea, a group of exudates within the macula, or any microaneurysm or haemorrhage within 1-disc diameter of the centre of the fovea, but only if associated with a best visual acuity of worse than 0.3 logMAR (equivalent to Snellen 6/12). 39

For statistical analysis, we amalgamated retinopathy and maculopathy grades for both eyes into a univariate outcome with seven levels: (1) no DR in both eyes (R0M0 R0M0); (2) background in one eye, no detectable DR in the other eye (R1M0 R0M0); (3) background retinopathy in both eyes (R1M0 R1M0); (4) pre-proliferative DR or PDR in one eye, no detectable DR or background in the other (R2+M0 R0/1M0); (5) pre-proliferative DR or PDR in both eyes (R2+M0 R2+M0); (6) maculopathy in one eye and any DR (M0 M1); and (7) maculopathy in both eyes and any DR (M1 M1).

Model structure

We treat the seven levels of retinopathy and maculopathy described above as states in a Markov model. We used a continuous-time hidden Markov model to estimate simultaneously the transition rates (intensities) between states and the probabilities of misclassification. Markov models have been used extensively to model disease progression and to evaluate screening strategies. A hidden Markov model accounts for the fact that the true disease state is not always reflected by the test. In this context, the true disease state could be thought of as the potential grade determined by seven field stereo-photography or similar gold-standard tests for DR corresponding to each screening grade.

Model structure and assumptions were informed by what was known about DR, the evaluation needs and discussions with clinical experts, health economists, statisticians and epidemiologists involved in this project. This consisted of an iterative process as the conceptual framework was revisited following the analysis of data and the health economics evaluation. The following modelling assumptions about disease progression are required: (1) as there is good evidence to suggest that background retinopathy can develop and subsequently disappear,70 movement back and forth was permitted between the states corresponding to no detectable retinopathy and background retinopathy (levels 1, 2 and 3); (2) once disease has progressed to at least pre-proliferative retinopathy (level 4 upwards), remission to level 3 or below is assumed to be impossible; (3) we assumed that eyes cannot develop pre-proliferative retinopathy without first developing background retinopathy; hence, there is no direct link from state 1 to pre-proliferative retinopathy or proliferative retinopathy states (4 and 5); (5) for model parsimony, and because advanced retinopathy or maculopathy leads to treatment, we do not consider progressions from either level 5 (pre-proliferative or proliferative retinopathy in both eyes) or state 7 (maculopathy in both eyes). Figure 14 is a graphical representation of the model for underlying disease progression.

FIGURE 14.

Graphical representation of the model for progression of diabetic eye disease. Arrows from one state to another represent instantaneous transitions to be estimated. Absence of arrows indicates that instantaneous progression is not possible.

We did not impose modelling assumptions on the probabilities that any given state is misclassified, during retinopathy grading, as another. The model therefore estimates a 7 × 7 matrix of instantaneous transition probabilities with forced 0 entries representing the modelling assumptions listed above, and an unrestricted 7 × 7 matrix representing misclassification probabilities.

The model was fitted using the msm function71 in R version 3.0.1 (The R Foundation for Statistical Computing, Vienna, Austria). 72 The final model was derived by adding potentially explanatory variables separately and then together in turn in order to find the model with the lowest deviance (–2 log likelihood). The final model adjusted for duration of diabetes at the first SE, baseline HbA_1c and serum cholesterol level. We explored adjusting the model for type of diabetes, but the low number of people with type 1 (≈900) in the screening cohort made the model too unstable. In the case of blood pressure, age and sex, these did not significantly improve the fit of the model and so were not included in the final model.

Estimated over- and under-referrals attributable to screening

We used the model to calculate the FP and FN rates for referable disease, defined as pre-proliferative or proliferative retinopathy or any maculopathy, at a screening visit. In a post-hoc secondary analysis, we also summarised these rates under varying assumptions about case mix to enable comparison with previously published studies.

We then considered the FP and FN rates over time in screening programmes with 6 months, 1 year, 3 years or 5 years between screening visits. Using the fitted intensity and error matrices, we calculated the expected number of people with referable eye disease, the number of over- and under-referrals and the ratio of false to true referrals at follow-up visit, stratified by observed state at baseline visit. Results are presented for two cohorts, one with risk factor levels (duration of diabetes, HbA_1c, cholesterol) comparable to the Gloucester cohort, and a ‘high-risk’ cohort group with a longer (10 years) duration of diabetes, poorly controlled glycaemia (baseline HbA_1c of 65 mmol/mol) and elevated cholesterol (6 mg/l). Results are presented in terms of screening 10,000 people to aid interpretation.

Results

From a total of 14,810 people, 68,992 examinations results were extracted from the screening service database. The modelling data set consisted of 65,839 grades from 14,187 people. Observations were excluded if retinopathy or maculopathy grade were missing from either eye or were obviously duplicate entries, and people were excluded if they only had one useable observation or did not have baseline HbA_1c, serum cholesterol or duration of diabetes recorded. The median number of examinations was 5 [interquartile range (IQR) 3–6] and the median interval between examinations was 1.04 (IQR 0.99–1.17) years. Characteristics of the cohort at baseline are given in Table 22.

Table 23 shows the raw unadjusted counts of transitions across all individuals in the cohort. These represent the number of times one grade has been followed by another. Row percentages are also given and can be interpreted as empirical probabilities of observing changes in consecutive examinations.

The estimated transition intensity matrix is shown in Table 24.

The matrix of estimated error probabilities is given in Table 25. Here, off-diagonal elements represent probabilities of grading errors, and diagonal elements probabilities of correct grading.

The duration of diabetes and baseline HbA_1c and serum cholesterol were independent predictors of true progression of DR. To ensure model convergence, covariates effects were standardised and constrained to be equal for each of the 11 forward transition intensities and all three backward transition intensities. Table 26 shows the proportional HRs on forward and backward transition intensities for the three covariates.

The numbers of referrals (for maculopathy and retinopathy) and the numbers which represent true, over- and under-referrals and the ratio of false to true referrals were calculated using the estimated intensity and error matrices. Table 27 shows the results for people with no detectable retinopathy at baseline and the true and false referrals when the next screen occurs 6 months, 1, 3 or 5 years later. Table 28 shows the results for people with background retinopathy in one eye and Table 29 shows the results for people with background retinopathy in both eyes at intervals of 3, 6 and 12 months. Tables 30–32 show equivalent results for high-risk patients.

Higher-risk patients

Conclusion

Our results show that annual screening for DR is associated with a disproportionate number of FPs or over-referrals, and extending the interval of screening reduces the occurrence of over-referrals. The number of FPs far outweighs the number of FNs even though the specificity (1 – FP rate) of the screening test is high. This occurs because the population being screened consists largely of people with no detectable DR or background DR and, hence, FP error rates are ‘amplified’ by screening. 73 Waiting longer to rescreen reduces the number of errors as the balance of non-referable disease to referable-level disease starts to tilt towards the latter rather than the former. Screening less frequently is likely to reduce significantly the cost of screening and reduce the number of unnecessary referrals to specialist eye services. However, longer intervals of screening will mean inevitable delays in referring a small number of people with true STDR for either closer monitoring or therapy. Our modelling suggests that these FNs will mostly comprise people with pre-proliferative DR or maculopathy, and, although this requires referral to ophthalmology, it is not considered as urgent as PDR. The implications for such delays in people with pre-proliferative disease or maculopathy are not fully understood but could be detrimental.

Introduction

No firm evidence or evidence-based consensus exists as to the optimal frequency of testing for DR. Although previous studies have assessed the cost-effectiveness of differing intervals for DR screening, the evidence is mixed. The aim of this work was to determine if assigning diabetic patients to differing DR screening intervals using a risk estimation model is cost-effective compared with the current English national screening programme (i.e. annual screening for all eligible people with diabetes aged 12 years and over).

Objective

Diabetes is one of the most common chronic diseases, placing a great economic burden on society owing to its increased health-care expenditures and lost productivity. Although the treatment of diabetes on its own is costly, its complications are the major contributors to health-care costs. Among the main diabetes-related complications is DR, which has been shown to be, until recently, the leading cause of blindness in the working age population10 in the UK. DR is treated effectively with laser photocoagulation, although this has been found to be cost-effective only if retinopathy is detected before irreversible damages take place. 13,15,16 Therefore, in order for DR treatment to be cost-effective, diagnosis has to be timely, with published evidence showing that screening for STDR is highly cost-effective. 17–20,83 In the UK, results from published studies also highlight the cost-effectiveness of screening for DR,84 with James et al. 85 showing that a screening programme using retinal photography was cost-effective compared with opportunistic screening.

In 2003, a national screening programme was introduced in England, recommending digital imaging as the preferred method of retinal photography for screening and mandating that local programmes respond to local needs with different models of care,86 with eye screening to be performed at time of diabetes diagnosis and repeated annually thereafter. Annual DR screening was also recommended in NICE guidelines. 65,87

However, despite these recommendations, no firm evidence or evidence-based consensus exists as to the optimal frequency of testing for DR. SIGN reported that patients with diabetes with no detected retinopathy could be screened every 2 years, with all other patients being screened at least annually. 21 Given the rising numbers of people identified by primary care practices as having diabetes, a rise of over 100,000 people annually,5 and lower rates of progression to PDR and severe visual loss,11,12 reflecting in part improvements in general diabetes care, there is an argument for modifying DR screening recommendations. Therefore, the aim of this work is to determine whether assigning diabetic patients to differing DR screening intervals using a risk estimation model is cost-effective compared with the current English national screening programme (i.e. annual screening for all people with diabetes).

Existing research

A published systematic review assessing economic evidence surrounding differing DR screening frequencies showed that the evidence on cost-effectiveness was mixed. 46 The review only identified three studies, one performed in the USA and two in the UK.

The UK study by Brailsford et al. 88 found similar results to the US study. The study, which also used a decision analytic model to assess the cost-effectiveness of DR screening in both type 1 and type 2 diabetes patients, found that a 30-month screening interval was the most cost-effective option. However, in contrast to these two studies, Davies et al. 48 found that screening diabetic patients less than once a year would not be any more cost-effective than screening them every year. The study also based its work on a decision analytic model populated using results obtained from the published literature. A finding common in both UK studies was that screening type 1 diabetes patients was more cost-effective than screening type 2 diabetes patients.

There were, however, important limitations to both of the UK studies. Both studies used sight years saved as their main outcome measure, rather than a more generalisable health outcome, such as QALYs, that can be readily compared across intervention and disease areas, and for which willingness-to-pay thresholds exist to determine cost-effectiveness. The two studies also failed to include in their models the additional costs to the health-care service of patients losing sight or going blind, which have recently been shown to be considerable and non-negligible. 89 Finally, data in the two UK studies were derived from a wide range of sources and diabetic populations. As a result, both studies had to undertake numerous assumptions as to how best to synthesise the available data.

The US study examined the cost-effectiveness of differing DR screening intervals for type 2 diabetes patients in the USA. 27 The authors employed a Markov model using QALYs as the main outcome measure, with costs being assessed from a third-party payer perspective. Using a threshold of US$50,000 per QALY gained, or less, as an indication of a cost-effective intervention, the authors reported that annual screening for all type 2 diabetes patients was not warranted on the basis of cost-effectiveness.

An additional US study, published after the review by Jones and Edwards,46 found that annual screening was costly and added little benefit compared with biennial eye evaluation. 90 In addition, the authors found that the cost-effectiveness of biennial screening depended on the ability of the screening programme to detect other eye conditions. If a screening programme was unable to detect such conditions, then a simple telemedicine programme using acuity eye charts or screening questions became the most cost-effective alternative.

Methods

Interventions under study

For all the screening strategies under study, DR screening consisted of two digital photographs on two fields, with mydriasis (i.e. eye dilatation) to maximise image quality. We therefore modelled the cost-effectiveness of the following screening intervals for all patients with diabetes, including current practice in the English national screening programme, in which all patients with diabetes are screened annually:

1. DR screening every 6 months

2. DR screening every year (current practice)

3. DR screening every 2 years

4. DR screening every 3 years

5. DR screening every 5 years.

In addition, the cost-effectiveness of these screening intervals was assessed in patient subgroups at different risks of developing STDR or maculopathy. We used two different risk stratification models (see Chapter 5), one based on the results of two consecutive initial SEs and another based on the results of one SE in conjunction with CRF data (Table 33).

TABLE 33 - Demographic characteristics of patient subgroups at differing risks of developing STDR based on the results of two consecutive SEs or one SE in conjunction with CRF data

Risk group	Initial screening results	Mean HbA1c (mmol/mol)	Mean cholesterol (mg/l)	Mean age (years)	Mean duration of diabetes (years)
Two SEs
Low	No DR in both eyes in two consecutive episodes (R0M0 R0M0)	51.1	4.25	66.8	5.9
Medium	Background in one eye, no detectable DR in the other eye (R1M0 R0M0) in two consecutive episodes	53.8	4.17	66.77	7.3
High	Background retinopathy in both eyes (R1M0 R1M0) in two consecutive episodes	57.9	4.11	63.96	13.0
One SE and CRF data
Low	No DR in both eyes (R0M0 R0M0) or background in one eye, no detectable DR in the other eye (R1M0 R0M0)	48.7	4.4	64.4	2.9
Low-medium	No DR in both eyes (R0M0 R0M0) or background in one eye but no detectable DR in the other eye (R1M0 R0M0) or background retinopathy in both eyes (R1M0 R1M0)	58.9	4.6	63.9	5.0
Medium–high	Same as above	58.3	4.5	63.5	5.4
High	Same as above	67.9	4.5	60.4	14.8

Common to all screening strategies, a diabetic patient was offered diabetic screening at the specified time intervals which could result in receiving screening or not, conditional on the screening uptake (Figure 15). Patients who were identified at screening to have pre-proliferative retinopathy (i.e. R2), PDR (i.e. R3) or diabetic maculopathy (M1) were referred to Hospital Eye Services (HES) for further tests and/or assessment by an ophthalmologist. Not all referred patients chose to attend their clinical appointment.

FIGURE 15.

Screening pathway for diabetic patients offered DR screening.

If diagnosis of DR (R2 or R3) or diabetic maculopathy (M1) was confirmed in at least one eye in the HES, then, depending on clinical need, patients would either be monitored by HES every 6 months, treated with either laser photocoagulation or antiangiogenic therapy, or referred back to the annual screening programme (Figure 16). Once treatment occurred the patient was referred back to the screening programme. Furthermore, patients who did not receive screening, patients who received screening but were not referred to the HES, and patients who were referred but did not attend their appointment were offered screening in the next time interval.

FIGURE 16.

Clinical pathways of confirmed cases of STDR (R2, R3 or M1) attending referral in HES.

Model structure

A decision analytic model was developed to evaluate the costs, (quality-adjusted) life expectancy and cost-effectiveness of the different DR screening strategies under evaluation. Given the natural history of DR progression with recursive events, the most appropriate type of model was judged to be a Markov model which was developed in Microsoft Excel (Microsoft Corporation, Redmond, WA, USA). The model was used to simulate the natural history of the diabetic population across seven health states representing retinopathy and maculopathy grades for both eyes (Figure 17):

1. no DR in both eyes (R0M0 R0M0)

2. background in one eye, no detectable DR in the other eye (R1M0 R0M0)

3. background retinopathy in both eyes (R1M0 R1M0)

4. pre-proliferative or proliferative retinopathy in one eye, no detectable DR or background in the other (R2/3M0 R0/1M0)

5. pre-proliferative or proliferative retinopathy in both eyes (R2/3M0 R2/3M0)

6. diabetic maculopathy in one eye and any DR (M0 M1)

7. maculopathy in both eyes and any DR (M1 M1).

FIGURE 17.

Model structure capturing the natural history component.

Three absorbing states were defined: maculopathy in both eyes (M1 M1); pre-proliferative or proliferative retinopathy in both eyes (R2/3M0 R2/3M0); and death. We assumed that if patients progressed to these three states, they would remain in that health state for the rest of their lifetime. Model structure and assumptions were informed by what was known about DR, the evaluation needs and discussions with clinical experts, health economists, statisticians and epidemiologists involved in this project. Hence, the construction of the model consisted of an iterative process where the agreed conceptual framework was revisited given new findings in the data and published literature and validity of its outputs. Such discussions also helped to identify key uncertainties in the model structure to be explored in the sensitivity analyses.

The time horizon of the analysis was lifetime, and the population moved between health states according to defined transition probabilities. The transition probabilities between the seven DR health states were obtained from the work presented in Chapter 7. A cycle length of 6 months was considered appropriate given the natural history of diabetic patients and the screening strategies under evaluation. Half-cycle correction was performed and different cycle lengths were trialled (monthly, 3 months) to ascertain the potential impact on the relative costs and effects before deciding on the 6-month cycle.

The model simulated the transition of a cohort of diabetic patients through the health states over time, to estimate expected costs and outcomes. For example, at the start of a given cycle patients without DR in both eyes (R0M0 R0M0) could die, develop maculopathy in one eye (M1 M0) or, alternatively, develop background retinopathy in one eye (R1M0 R0M0). If the cycle coincided with a screening interval, the cohort could then attend screening or not (conditional on screening uptake), be referred or not to the HES (conditional on accuracy of the screening test), and attend or not the referral appointment (conditional on the uptake of assessment) (Figure 15). If patients attended the HES, they could either be monitored every 6 months, treated with either laser photocoagulation or antiangiogenic therapy, or referred back to the annual screening programme (Figure 16).

Given that slit lamp biomicroscopy is considered to be the ‘gold-standard’ for diagnosing DR, we assumed that HES would correctly identify the true state of DR and maculopathy (i.e. 100% sensitivity and specificity). Hence, only individuals with maculopathy (M1) or pre-proliferative/proliferative retinopathy (R2–3) were considered for treatment or monitoring by HES. The remaining individuals were referred back to the annual screening programme.

Patients monitored by the HES were allowed to continue progressing if they had not yet reached an absorbing state. Once patients were treated it was assumed that they would remain in the same retinopathy and maculopathy grades for the remainder of the simulation. This was based on evidence from clinical trials showing that the majority of patients remaining in the same grades despite improvements in visual acuity. 91 As discussed below (see Treatment effectiveness), the benefit of treatment was captured as changes to visual acuity.

Finally, the model adopted the perspective of the UK NHS and personal social services. All costs and effects were discounted beyond the first year of simulation using an annual discount rate of 3.5%, based on current UK government recommendations. The price year was 2012–13 and, when necessary, costs were inflated using the UK health sector pay and prices inflation factor. 92

Results

Diabetic retinopathy screening for patients with diabetes

Table 42 presents the baseline characteristics of the simulated cohort of 1000 people with diabetes. In this cohort, 58% of patients had no DR in either eye, 23% had background retinopathy in one eye and a further 19% had background retinopathy in two eyes.

TABLE 42 - Baseline characteristics of the simulated cohort

True DR grade	Patients
R0M0 R0M0, n (%)	580 (58)
R1M0 R0M0, n (%)	230 (23)
R1M0 R1M0, n (%)	190 (19)
Mean age, years	66.6
Sex: female, n (%)	440 (44)
Mean duration of diabetes, years	7.3
Mean cholesterol, mg/l	4.22
Mean HbA1c, mmol/mol	52.6

For our simulated cohort, screening every 2, 3 and 5 years would result in a reduction of 5375 (95% CI 5332 to 5411), 7166 (95% CI 7111 to 7213) and 8617 (95% CI 8552 to 8672) screens, respectively, compared with annual screening (Table 43). By screening every 2 years, as opposed to every year, a total of 143 (95% CI 120 to 162) fewer patients incorrectly diagnosed at screening as having R2–3 or M1 (i.e. true negatives) would be referred to HES. These reductions in numbers increase to 190 (95% CI 159 to 215) and 229 (95% CI 191 to 259) when screening every 3 or 5 years, respectively.

TABLE 43 - Number of cases screened, referred, assessed and treated over the lifetime of a cohort of 1000 people with diabetes

Outcomes	Screening every 6 months	Annual screening	Screening every 2 years	Screening every 3 years	Screening every 5 years
Number of screens	21,990	11,253	5878	4087	2636
Number of patients referred following screening
True negatives	584	298	155	108	69
True positives	303	277	234	203	157
Number of patients assessed by HES following screening	492	323	222	178	132
Number of patients treated following screening	70	65	57	50	40
Number of patients treated over lifetime	135	124	106	92	72

By contrast, screening every 2 years, as opposed to annually, would reduce the number of patients with true R2–3 or M1 grades (i.e. true positives) referred to HES, over the lifetime of the cohort, from 277 to 234 people, a reduction of 42 (95% CI 35 to 50) cases. Increasing the screening intervals to 3 years and 5 years would further reduce the number of true positives being referred by 74 (95% CI: 62 to 87) and 119 (95% CI: 100 to 139), respectively. This would result in 8 (95% CI 7 to 10), 15 (95% CI 12 to 17) and 25 (95% CI 21 to 29) fewer true cases of R2–3 or M1 being treated following screening if screening intervals were to be extended to 2, 3 and 5 years, respectively, compared with annual screening (Table 43). Over a lifetime, when compared with annual screening, there would be a reduction of 18 (95% CI 15 to 21), 32 (95% CI 27 to 37) and 52 (95% CI 44 to 59) true cases of R2–3 or M1 being treated if screening intervals were to be extended to 2, 3 and 5 years, respectively.

With annual screening, the average discounted screening cost per patient was £273 for annual screening, compared with £144, £101 and £67 for screening every 2, 3 and 5 years, respectively. Mean costs associated with assessment at HES of referral of R2–3 or M1 suspected cases was also higher, at £114, when screening annually than when screening at 2, 3 and 5 years (£70, £52 and £36, respectively), with a similar pattern of costs being observed for treatment of R2–3 and M1. However, as the screening interval increases, mean health-care costs (excluding screening, assessment and treatment) increase from £14,354 when screening every year to £14,377 at 2 years, £14,393 at 3 years and £14,413 at 5 years. Combining all health and social care costs included in the model, mean discounted costs are £20,672 when screening is performed annually, £20,490 when it is performed every 2 years, £20,433 when performed every 3 years, and £20,391 when performed every 5 years (Table 44). The discounted average QALYs gained by patients were 8.3710 when screening every year, 8.3692 when screening every 2 years, 8.3680 when screening every 3 years and 8.3663 when screening every 5 years.

TABLE 44 - Mean discounted costs and mean discounted outcomes of differing screening intervals

Costs and outcomes	Screening every 5 years	Screening every 3 years	Screening every two years	Annual screening	Screening every 6 months
Total costs, £ (95% CI)	20,391 (19,668 to 21,136)	20,433 (19,707 to 21,176)	20,490 (19,771 to 21,229)	20,672 (19,947 to 21,405)	21,050 (20,303 to 21,824)
Screening, £	67	101	144	273	529
HES assessment, £	36	52	70	114	193
Treatment of R2–3 or M1, £	24	35	48	81	141
Other health-care costs, £	14,413	14,393	14,377	14,354	14,337
Social care costs, £	5852	5851	5851	5850	5850
Total QALYs (95% CI)	8.3663 (4.3731 to 12.067)	8.3680 (4.3743 to 12.070)	8.3692 (4.3753 to 12.071)	8.3710 (4.3768 to 12.074)	8.3723 (4.3779 to 12.0755)

After combining costs and outcomes in an incremental cost-effectiveness analysis, and at a £30,000 per QALY threshold, the most cost-effective screening interval was to screen patients every three years (Table 45). The probability of screening every 3 years being the most cost-effective option at £30,000/QALY was estimated to be 46%. Screening every 2 years was associated with an ICER of £45,684 per QALY gained compared with screening every 3 years. Screening every 6 months was estimated to cost an additional £288,497 per QALY gained compared with annual screening.

TABLE 45 - Cost-effectiveness of differing screening intervals

Screening strategies	Difference in costs, £ (95% CI)	Difference in QALYs (95% CI)	ICER, £	Probability that intervention is the most cost-effective at £30,000/QALY, %
Every 5 years	–	–	–	41
Every 3 years	42 (26 to 58)	0.0019 (0.0004 to 0.0029)	26,156	46
Every 2 years	57 (41 to 73)	0.0012 (0.0003 to 0.0023)	45,684	13
Annually	182 (141 to 226)	0.0018 (0.0005 to 0.0034)	98,085	0
Every 6 months	378 (306 to 457)	0.0013 (0.0003 to 0.0024)	288,497	0

Figure 18 reports the CEAC and EVPI per patient associated with the different screening options. The population EVPI over 5 years was estimated to be £86M at the £30,000 per QALY gained threshold. This suggests that undertaking additional major commissioned research work to reduce further decision uncertainty is likely to be of significant benefit given the large scale of the population affected by the decision.

FIGURE 18.

Cost-effectiveness acceptability curve and EVPI per patient.

Figure 19 reports the parameters according to their impact on the variance of incremental costs and QALYs from the comparison of screening every 3 years with every 5 years. Variance in the estimated incremental costs was dominated by the variation in the following variables: costs of assessment following referral to HES; treatment costs; hospital costs; and effectiveness inputs. Variance in the incremental QALYs was dominated by the variance in the regression model associating utility with visual acuity, effectiveness inputs (mostly the effectiveness of laser/injection for M1) and natural history of progression.

FIGURE 19.

Analysis of covariance analysis of proportion of sum of squares for incremental QALYs saved and incremental costs explained by the uncertainty in the model inputs.

Table 46 reports the results of a range of other sensitivity analyses. Changes in the monitoring interval by HES from 6 months to 12 months did not alter the decision. The results were also robust to the proportion of individuals with M1 or R2–R3 that are referred to further monitoring every 6 months if assessed by HES but not treated. Given the difference in effectiveness across treatments, depending on which type of treatment was received there was a significant impact on the results. For example, increasing the proportion of M1 cases receiving injection rather than laser treatment among those treated resulted in screening every 2 years falling under the £30,000 per QALY threshold.

TABLE 46 - Sensitivity analysis scenarios: impact of differing assumptions on ICERs

Sensitivity analysis scenarios	Screening every 3 years vs. every 5 years	Screening every 2 years vs. every 3 years
Base case	£26,156	£45,684
Screening uptake: 81%	£27,060	£47,805
Screening uptake: 91%	£31,175	£56,503
Screening uptake: 71%	£23,305	£40,008
Hospital costs only (excluding social care costs)	£26,724	£46,582
50% shift from visual impairment to reduced vision and from reduced vision to good vision (affects nursing home admission)	£26,296	£46,194
Only cases with reduced vision shift to good vision (affects nursing home admission)	£26,644	£46,495
Proportion of M1 cases treated with injection rather than laser: increase of 30%	£16,515	£31,403
All M1 cases are treated with injection or a combination of laser and injection	£12,438	£24,624
Probability of being referred to monitoring every 6 months of 0% instead of 78%	£29,796	£45,566
Probability of being referred to monitoring every 6 months of 50% instead of 78%	£26,539	£44,520
Probability of being referred to monitoring every 6 months of 100% instead of 78%	£25,880	£47,635
Monitoring interval by HES of 12 months instead of 6 months if R2–3 or M1	£25,709	£45,117

Conclusion

Diabetic retinopathy screening is widely recommended and has shown to be cost-effective compared with no screening. 9,10,12 However, a less clear understanding exists on the most appropriate screening interval, with published cost-effectiveness evidence showing conflicting results. 48

Our cost-effectiveness analysis shows that for patients with diabetes and no pre-proliferative DR or PDR or maculopathy, the current national screening programme of screening patients annually is not cost-effective. Results from our study show that for this group of patients, screening every 3 years would be the most cost-effective option. Annual screening, by contrast, was associated with an incremental cost-effectiveness ratio of £98,085 compared with the next less effective DR screening strategy (i.e. screening every 2 years). At current NICE thresholds of £20,000 to £30,000 per QALY gained, annual screening would not be considered cost-effective. Our results are similar to those by Brailsford et al. ,88 which concluded that a 30-month screening interval was the most cost-effective option.

Irrespective of how patients were stratified in terms of their risk of developing STDR/maculopathy, our results suggest that the cost-effectiveness of the screening programme could be further improved by targeting patients using differing intervals for patients at different risks of developing STDR/maculopathy. Stratifying both based on the results from two consecutive SEs or based on the first single SE in conjunction with CRF data produced very similar results. In both risk stratification models, screening patients at low risk every 5 years, and those at high risk every 2 years would be the most cost-effective DR screening strategy.

There is, however, considerable uncertainty in the evidence informing the model, particularly the natural history of disease progression, association between utility scores and visual acuity and the effectiveness of treatment in diabetic maculopathy. There is also considerable uncertainty concerning the annual marginal costs attributable to treatment and assessments in the HES. This is reflected in the very large EVPI values of £86M to £458M, conditional on the expected lifetime of the current technology for screening. The EVPI gives the maximum amount of funding required for further research to reduce the decision uncertainty about which of the screening intervals is the most cost-effective option. The large population of diabetic patients that will be affected by making the wrong decision, together with considerable uncertainty in the evidence available, warrant caution about the model results. Hence, our results suggest undertaking additional major commissioned research work to further reduce the decision uncertainty.

There are several limitations to consider. First, there was very limited information concerning the proportion of patients with R2–3 or M1 grade that are referred for further monitoring by the HES. This led to assumptions being made concerning both the interval of monitoring and the proportion of patients being referred to it. However, sensitivity analyses showed the model results to be fairly robust to variations in these assumptions. The small number of patients in these health states also prevented the separation of R2 and R3 states. Hence, pre-proliferative and proliferative retinopathy were grouped together as both are grades prompting referral to the HES and are sight-threatening states. The absorbing states used, R2/R3 M0 and Rx M1 in both eyes, meant that no onwards state progression was explicitly modelled between and from these. However, such states lead to treatment, which seems to have limited effect on the retinopathy grade91 and, therefore, disease progression was implicitly modelled in terms of costs, visual acuity and utilities, which were found to increase with age/duration of diabetes. Furthermore, the lack of movements between the absorbing states is likely to have limited impact as few patients were identified as making such a transition.

Second, owing to the cohort nature of the model and the focus on simulating the progression of true states, it was not possible to evaluate strategies based on which states were observed at consecutive screening rounds. An alternative would have been to analyse the Markov model as an individual-based model (i.e. microsimulation) so that we could keep track of each individual’s history. 105 This would have allowed tailoring the frequency of screening during the lifetime of a cohort of patients conditional on the observed grades.

Although a great proportion of the data used were derived from patient-level data in patients with diabetes, we were not able to obtain all model parameters from this group of patients and, therefore, we had to obtain some of the model parameters from the literature. These included the risk of nursing home admission given losses in vision, health state utility values, and treatment effectiveness. As a result, the populations on which these studies based their results might be different from those in which the majority of the data were obtained (i.e. those attending the Gloucestershire DESP). For example, the data on the effectiveness of laser photocoagulation for treatment of STDR were obtained from evidence from the 1980s and early 1990s. In addition, the results are conditional on the model structure and the evidence used to inform it. For example, we used the results from the study by Lloyd et al. ,97 which assessed, using the EQ-5D, the utility associated with visual acuity. However, as with the other three studies identified, this association was only available for the best eye (in terms of visual acuity). Therefore, for modelling purposes, we assumed that the best seeing-eye was always the first to be treated. Internal validity was performed and was complemented by the exploration of different parameter and structural scenarios in the sensitivity analysis.

Routinely collected data were used to assess the impact of DR grade and visual acuity on health-care resource usage and costs. Therefore, it was not possible to determine whether use of health-care services was directly a result of vision loss or whether it was a result of other diabetes-related complications (such as stroke or heart attacks), given that poorly controlled diabetes is associated with both vision loss and other diabetes-related complications. 95 We performed additional analyses to assess whether the association between DR grade/visual acuity and annual care costs was confounded. Using the diagnoses codes included in patients’ HES records, we were able to estimate patients’ Charlson Co-morbidity Index (CCI) based on 22 clinical conditions present within a patient (including myocardial infarction, stroke, acquired immunodeficiency syndrome and tumours). 106 When we adjusted for CCI, the significant and positive association between DR grade/visual acuity and annual care costs was still observed.

Finally, our analysis is from the perspective of the NHS and personal social services. As a result, other important cost categories not relevant to the perspective adopted were not included in the analysis. For example, we excluded some important economic considerations for people with visual loss, such as reductions in earnings through reduced or lost productivity, and the economic consequences of reduced mobility associated with vision loss (for example, people with moderate vision loss not being able to drive).

In conclusion, the work presented here suggests that it may be good value for money to use differing intervals for patients at different risks of developing STDR/maculopathy. Further research is needed to make more informed decisions.

Mike Whatmore’s contribution

I have been an independent member of the HTA Trial Steering Committee from its inception in October 2012, and actively involved at both live meetings and updated with all necessary interim documents by e-mail. Also, during this first meeting, a presentation was made of the data set information to be used for all areas involved in the study, and classification for reference for diabetic retinopathy.

In November 2012, I received an e-mail with the HTA Interim Progress Report attached, which included details of how the studies had been carried out to date and how problems encountered with the lack of data from one of the original expected sources, in South London, had been overcome and an alternative set of appropriate data had been accessed from the Nottingham Screening Service team.

At the second, live meeting, on 31 March 2014, the latest data obtained was presented, showing that progress was continuing to be made, and indicating the potential for considerations to be made to the current screening intervals. I made comments about showing the practical use of extending the screening intervals, and stated that I thought any changes would require good communication between all affected parties, viz. the Screening Service and G.P.’s, and, the G.P. and patient; supporting the reasons for making any changes to the current screening interval with accurate, precise results and providing examples to reinforce the justification.

In November 2014, I received a draft copy of the proposed HTA ‘Final’ Report, which asked for any comments I may wish to make, and I pointed out some minor errors within the text of the document.

I have read and understood the majority of the details contained in the Final Report (some of the statistical documentation, in chapter 5, I found confusing and did not fully understand) and accept the suggested recommendations to amend the intervals for diabetic eye screening. I would hope that monitoring of these changes, if applied, continues in the future and further safe, practical, outcomes are acted upon.

Mike Larkin’s contribution

Mike Larkin has met with the Lead Researcher regularly during the period of this research.

To date, my input has been reading and reviewing the drafts and providing a perspective on the work. I think my feedback covers two distinct areas:

1. Feedback on the report as a lay reader – here I have suggested a different plain English summary that attempts to take out the obvious and simplify the kernel of the research to help other non-researchers understand the project.

2. Patient view on the benefits of extending screening intervals. In a free to use system, patients don’t necessarily see benefit in less access. My comments are intended to acknowledge this and should be part of future discussion.

I think that overall the report moves a great way to better understanding suitable screening intervals – something that will be essential in the forthcoming years as we look to address the Type 2 epidemic.

1. Whilst there is acknowledgement that Type 1 patients are likely to receive closer monitoring (from the risk factor regarding length of condition) it’s dangerous to assume that both types of diabetes will affect the eyes the same. I feel that Type 1 as a condition is far more aggressive than Type 2 (for the same period of time). HbA_1c levels are certainly good indicators for long term control, but are by their nature, averages. They don’t show if a person has had extreme highs over the 3 month period that the test is measuring. It’s for this reason that condition type should be differentiated in any recommendations on screening interval.

2. The reference to the lack of evidence around economic consequences of visual loss to the patient is important. It would be useful if this area could be considered in more depth. Before a person is able to officially register sight impairment, productivity and quality of life will, I believe, already have been significantly impacted. From a personal perspective, losing my sight to a point when I can no longer drive would have a substantial impact as mobility at home and work is restricted. Whilst, from a clinical sense, a person may have acceptable vision, the lack of personal mobility is economically and emotionally dramatic. The report sections centred on economics has focused entirely upon the costs to the NHS and not attributed any value to individual reduction of productivity.

Recommendations

This project recommends that:

1. software to identify patients with diabetes in primary care and to provide or extract promptly demographic and risk factor data to screening programmes is required

2. patient groups and stakeholders be consulted before any screening interval changes are introduced

3. all diabetic eye screening programmes record the date of diagnosis of diabetes and have access to recent HbA_1c results, in addition to the result of grading from the previous SE. Ethnicity and type of diabetes should also be recorded

4. the risk stratification models should be validated in other eye screening programmes (inside and outside England) to identify risk groups

5. misclassification of photographic grading (i.e. grading error rates) needs to be evaluated in all local eye screening programmes

6. extended screening intervals should only be implemented by local programmes where grading misclassification rates have been evaluated and are shown to be acceptably low (once the threshold has been identified as outlined in Future work)

7. the knowledge arising from this project about progression and misgrading matrices are used to inform quality assurance processes in screening

8. data from patients with mild NPDR (R1M0) in just one eye are reported independently of patients with R1M0 in both eyes in DESP reports

9. the effectiveness and cost-effectiveness of this model is tested in other populations.

Future work

This project has identified the following areas for future research:

1. investigate strategies to improve knowledge about risk of DR to the national programme, screening programmes, patients, HES, primary care and diabetes services. This would include investigating grading misclassification rates in local programmes and determining acceptable ranges

2. investigate strategies to increase uptake in those with a poor attendance record

3. determine whether extending the screening interval in patients in low-risk groups alters their behaviour over self-care of their diabetes or their attendance at subsequent screening events

4. investigate whether using the more granular information from feature-based grading (implemented in England from 2013) will enable better identification of those at risk of progression to STDR and improve the sensitivity and specificity of screening

5. data analysis and modelling to identify patients with diabetes who might safely be screened only once (‘screen once and forget’)

6. evaluate the clinical effectiveness and cost-effectiveness of this model in other populations outside Gloucestershire

7. evaluate the use of automated grading to reduce grading misclassifications and further improve the reliability of grading as a tool to provide personalised screening intervals.

Contributions of authors

Peter H Scanlon (Ophthalmologist and Clinical Lead, Gloucestershire) is the lead author for this publication and has written the first and subsequent drafts with input from other authors.

Stephen J Aldington (Research Manager, Gloucestershire Retinal Research Group) project-managed the data collection from the various sites, wrote project updates for submission and commented on various drafts of the publication.

Jose Leal (Senior Researcher, Health Economics Research Centre, University of Oxford) conducted the Markov modelling described in Chapter 8 and wrote Chapter 8 in conjunction with Ramon Luengo-Fernandez.

Ramon Luengo-Fernandez (Senior Researcher, Health Economics Research Centre, University of Oxford) conducted the health economic work and wrote Chapter 8 in conjunction with J Leal.

Jason Oke (Statistician, University of Oxford) conducted the statistical research using the Oxford Monitoring Interval Approach and wrote Chapter 7.

Sobha Sivaprasad (Ophthalmologist, King’s College London) commented on the drafts of the report.

Anastasios Gazis (Diabetologist and Clinical Lead, Nottingham) commented on the drafts of the report.

Irene M Stratton (Senior Statistician, Gloucestershire Retinal Research Group) conducted the statistical research involved in the Risk factor approach described in Chapter 5 and the Extension to an ethnically diverse population described in Chapter 6 and has contributed to various drafts of the report.

Acknowledgement to the Research Design Service

We have consulted the Research Design Service South West (Gloucester Office) at various stages of the application and conduct of this research and they have been extremely helpful.

Health economic support for the application has been provided by the Research Design Service South Central (Oxford Office).

Trial Steering Committee

We are particularly grateful to the members of the Independent Trial Steering Committee (Professor John Sparrow, Dr Catey Bunce, Miss Clare Bailey and Mike Whatmore) for their advice and guidance.

Acknowledgement to individuals who are not authors on this publication

We have very much appreciated the support of the following individuals:

1. Dr Richard Stevens and Professor Andrew Farmer from the Department of Primary Care, University of Oxford, contributed to the original application and gave advice during the conduct of the research but have not been able to contribute to the writing and editing of the report because of competing commitments.

2. Ms Clare Connor worked unceasingly to provide data from the South London Population from the Diabetic Eye Screening Programme based at St Thomas’ Hospital. Obtaining these data required a huge amount of effort from a very dedicated individual and our results in ethnic minority groups have been massively improved by her efforts. Clare also provided data for the health economic survey.

3. Mr Steve Chave for his efforts in obtaining the data set of risk factors in Gloucestershire when he managed the GP2DRS pilot in Gloucestershire and for organising the MIQUEST extractions from GP surgeries.

4. Mr Mike Olson for his support of the MIQUEST extractions in South London and Nottingham.

5. Mr Richard Cragg for his efforts in obtaining the data from the Nottingham Screening Programme.

6. Mr Andrew Tyrell, Data Warehouse Manager at Gloucestershire Hospitals NHS Foundation Trust, for providing the Hospital Episode Statistics records (including inpatient, outpatient and accident and emergency episodes for 12,790 patients identified through the GDESP).

7. Miss Sarah Cooper, Programme Manager in Nottingham, who provided data for the health economic survey.

8. Mr Malcolm Grey and Dr Andrew Brown, Programme Executive and Clinical Lead, respectively, for the Staffordshire Diabetic Eye Screening Programme, who provided data for the Health Economic Survey.

9. Mr Leon Gardner, Programme Manager for the Gloucestershire Diabetic Eye Screening Programme, who provided data for the Health Economic Survey.

10. Mrs Christel Jenkins and Mr Colin Jones, Eye Screening Programme Manager and Clinical Lead, respectively, for the Norfolk and Norwich Diabetic Eye Screening Programme, for the health economic survey.

11. Mr Phil Kirby and Mr Greg Russell from Health Intelligence for their generosity and efforts in providing the East Anglia data set for analysis in this study.

12. Mrs Kate Fenton, HTA Project Manager on behalf of the NIHR, for her constant support, guidance and encouragement.

Acknowledgement to lay members who have contributed to this research

We would like to thank Mr Mike Whatmore, who was involved in the original discussions on both reasons for non-attendance in DR screening and screening intervals for DR and who commented on the original application and the final report. Mike Whatmore has also served as a member of the Project Advisory Group and the independent Trial Steering Committee.

We would also like to thank another Gloucestershire lay member, Mr Mike Larkin, who read and commented on the application and on the final report.

We are grateful for the input into the original application from the Warwick Diabetes Research and Education User Group (WDREUG) and the Thames Valley Diabetes Local Research Network Patient and Lay Involvement Group.

Contributions to reports, posters, presentations and publications

The main output from this research will be data to inform decisions about the most appropriate DR screening intervals to be implemented across England.

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.

Patient Information Advisory Group approval of consent model

The National Screening Programme for Diabetic Retinopathy (‘NSC DR’) approached the Patient Information Advisory Group (PIAG) for advice on whether or not it was necessary to make a full application for Section 60 support. A decision was taken in December 2004 by PIAG and has been minuted as follows:

Diabetic retinopathy screening

The PIAG considered a request from the National Retinopathy Screening Programme for clarification about whether Section 60 support was required for the programme as information would need to be shared across several PCTs, hospital clinics and general practices. There had been reluctance on the part of some data controllers to release patient information to the screening programme because of confusion about this. The Advisory Group agreed that call and recall for retinopathy screening was part of the care pathway. As such, consent to sharing relevant data could be implied from information about how patient information is used by the retinopathy screening programme, being provided to patients, and by making it clear patients had the right to opt out. There was therefore no requirement to apply for Section 60 support.

The PIAG was also asked for their views on a consent model for the transfer of risk factor data within the GP2DRS project, a national project of electronic transfer of demographic and risk factor data between general practices and screening services in England.

The PIAG agreed that all patients should have the following letter sent to them prior to screening attendance, that there would be implied consent for risk factor data transfer if they attended for screening and did not inform the programme staff or general practice staff that they did not want risk factor data transferred.

The data used in this study will not identify individual patients. It will be pseudo-anonymised so that grading data can be matched with risk factor data and then the analysis will be anonymised to analytical staff.

(PDF download)