Cost-effectiveness of biennial screening for diabetes related retinopathy in people with type 1 and type 2 diabetes compared to annual screening

People with either T1DM or T2DM (total 91,393; T1DM 5003; T2DM 86,390) undergoing screening for DR in a systematic community-based programme in Wales, UK between 2005 and 2009. Those included in the study had no evidence of DR at the first screening and had at least one further screening event [27].

Data from the Diabetic Eye Screening Wales (DESW) for all persons with diabetes ≥ 12 years registered with a GP in Wales, which included age, gender, date of diagnosis, type and treatment of diabetes, date of screening, DR grade and visual acuity [12] were transferred to the National Health Service (NHS) Wales Informatics Service for anonymisation and given a unique linking field number [28]. The data were then transferred to the Secure Anonymised Information Linkage (SAIL) databank [29] where it was linked with the Welsh Longitudinal General Practice (WLGP) dataset [30] which contained information on HbA_1c, BP, cholesterol level, treatment for hypertension and dyslipidaemia, diabetes treatment, and smoking status. The WLGP dataset contains data from approximately 80% of primary care practices in Wales. The combination of the DESW and WLGP datasets, along with the requirement of at least 2 screening events (the first of which had no evidence of DR) with clinical information available within the WLGP database and with a confirmed specific diagnosis of T1DM or T2DM (and not ‘diabetes’), limited the population size available for the analysis. Therefore, the number of people included in the combined, final datasets available for analysis to 1232 people with T1DM and 26,812 people with T2DM.

Information was sourced from the literature on transition probabilities for the different levels of DR whilst under the care of hospital eye services (Supplementary Table 1), treatment and service costs within hospital eye services (Table 1), cost of blindness (Table 1) and quality of life (Table 2). Transition probabilities for sight-threatening DR could not be calculated from DESW data as people are referred to hospital eye services (HES) once a sight-threatening level of DR is reached and they do not re-enter screening until treatment has stabilised, therefore, the progression of DR during monitoring and treatment within HES was unavailable for this population. To derive the transition probabilities for the CUA the time to an event of sight-threatening DR was modelled although the exact time the sight-threatening DR developed was unknown because it will have occurred between 2 screening events. Therefore, the data was interval censored, as well as right censored for those where sight-threatening DR did not occur by the last time they were screened but may occur in the future. Due to the interval censoring a Cox proportional hazards model could not be used, therefore parametric models were considered. Weibull, Gompertz and lognormal models were fitted to the data and the Akaike information criterion (AIC) method used to determine the best fit. Weibull modelling was used to estimate the probability of an event of sight-threatening DR occurring for each person, based on their clinical history in the form of a survival function S(t), that is the probability that the retinopathy event has not occurred by time (t). This has survival and hazard functions of the form

Weibull regression analysis was implemented by the routine INTCENS programme in STATA. Risk factors for the progression of DR were added to the Weibull model in a stepwise manor with those not statistically significant removed. The final Weibull model used to generate transition probabilities for moving from one severity level of DR to another including the risk factors HbA_1c, BP, cholesterol level, treatment of hypertension and dyslipidaemia, diabetes treatment and smoking status (Supplementary Table 2).

The transition probabilities derived from the survival analysis were then used to populate a de-novo multi-level Markov model to estimate the impact of extending screening beyond annual and the consequences at different HbA_1c levels were estimated. The model structure and flow of people are represented in Fig. 1. Life tables were used in the analysis. Life table data for the general population obtained from the Office of National Statistics (ONS) was used with adjustments made to reflect the increased mortality risk associated with T1DM or T2DM.

DR in Wales is graded according to the enriched version of the English DR grading protocol [27, 31] where No DR means no lesions of DR are visible on the retina; minimum Background DR (BDR) where microaneurysms, haemorrhages and exudates (not within 1 disc diameter of the fovea) can all be present but do not meet the criteria for Moderate BDR; Moderate BDR where microaneurysms, < 8 blot haemorrhages and exudates (not within 1 disc diameter of the fovea) are present and Referable DR (RDR) consisting of Pre-proliferative DR (PPDR) where > 8 blot haemorrhages, intra-retinal microvascular anomalies (IRMA) and venous beading are present, Proliferative DR (PDR) where new vessels on the optic disc and/or elsewhere, pre-retinal haemorrhages, vitreous haemorrhages, retinal detachment and fibrosis are present and exudative maculopathy.

As the primary care dataset obtained from SAIL was related to GP appointments and not screening events, modelling methods were employed to enable estimation of the clinical values (HbA_1C, systolic and diastolic BP, serum cholesterol, high and low-density lipids, triglycerides, GFR, body mass index (BMI)] which were combined into body mass index [BMI] at the time of DR screening. Multilevel modelling was used for the longitudinal data using a quadratic function of time for the correlation between different measurements from the same subject but constrained to follow a normal distribution across subjects. These coefficients were combined with the times of screening allowing prediction of the values at each screening event.

For the economic analysis the population and distribution of people with diabetes undergoing screening in Wales in 2005–2009, i.e. 5003 people with T1DM and 86,390 people with T2DM—were used to populate the model [27]. The economic analyses estimated the rate of progression from no disease to different stages of DR and the dynamic transition probabilities from the estimates of progression were incorporated into the model (supplementary Tables 3a and 3b).

The impact of extending the screening interval to biennial was compared against usual care which is annual screening for both people with T1DM and T2DM. The analysis was conducted with a lifetime time horizon enabling estimation of the long-term impact on DR and its management resulting from changes to the screening schedule.

As some of the observed DESW screening intervals exceed the target of an annual review at exactly 12 months, the future costs and benefits beyond 12 months were discounted back to present day values in line with the NICE reference case at a standard rate (currently 3.5% with sensitivity analysis at 1.5%) [32].

To assess the relative cost effectiveness of extending screening intervals from annual to biennial a cost-utility analysis (CUA) was used. A Markov model was utilised to facilitate the CUA with a cycle length of 6 months. The impact on health-related quality of life (HRQoL) quantified as health utilities according to the health status characterising DR in its later stages as described by Royle and colleagues in their study were used [33].

Quality-adjusted life years (QALYs) were the outcome measure used to calculate the difference in cost between screening interval options, divided by the difference in their effect. The decremental cost-effectiveness ratio (DCER), an approach used in a situation where the analysis of a health strategy appears to reduce healthcare expenditure by removing or substituting an intervention, was also calculated (supplementary Fig. 1). The decision rule is that a strategy is considered cost-effective if the incremental cost-effectiveness ratio (ICER) is below a threshold of £20,000 per QALY, as used by the National Institute of Health and Care Excellence (NICE) in England and Wales. The reverse is true for a DCER, with values above the threshold of £20,000 per QALY considered cost-effective.

An NHS and personal social services (PSS) perspective was adopted and the model was designed to estimate healthcare resource use and cost of sight-threatening DR treatment and management.

A price year of 2018 was used in the analysis and costs were reported in pound sterling (£). A number of sources were used to identify the relevant UK NHS resources and related costs associated with DR screening, treatment and management of different stages of DR (Table 1).

Analysis of our data revealed the progression of DR to be a dynamic process, i.e. the development of RDR accelerates with increasing severity of DR. To ensure the model also reflected usual practice within DESW, the model pathway was divided into four DR states i.e. No DR, Minimum BDR, Moderate BDR, and RDR. People entered the model at the time of the first screening event. Those who entered the RDR state within the first year of the 2-yearly screening interval were assumed (based on clinical advice) to progress without treatment during the following year. People who stayed within the RDR state were characterised by PDR and maculopathy.

The impact of a delay in the identification of people with RDR, because of longer screening intervals, is driven primarily by the severity of DR diagnosed at the first visit to the hospital eye service (HES) following referral from the screening service. The prevalence for people attending HES was calculated by progressing the current RDR distribution [27], in combination with the transition probabilities reported by Royle et al. [25].

On referral to HES, the RDR pathway followed in Wales incorporated ten clinically relevant health states with differing severity levels of DR and maculopathy. The assumptions made about treatment for those within the RDR category in HES can be seen in Supplementary Tables 4a and 4b [30]. The use of surrogate markers for maculopathy based on non-stereoscopic digital photography may result in over referral to HES for further assessment with a proportion discharged after the first HES appointment back to DESW. To account for this and simulate the known occurrence of false positives from the DESW screening reports and subsequent referrals to HES which prove negative, a linear reduction of one severity step within the RDR states was implemented at a rate of 50%, this adjustment was made to reflect an appropriate ‘knockback’ effect [27].

The RDR distribution was progressed through the model for one year to simulate the unidentified progression of DR (Supplementary Table 5) and related health service treatment costs and the decrement in health utility. The resulting value was adjusted to represent the average difference in cost and health utilities for a single patient within a six-month period. The combination of STDR and maculopathy and the occurrence of screening within the RDR pathway meant that patients often remained untreated for one or both conditions and were assumed to follow the same transition progression as before. People, who received treatment for STDR or maculopathy, were assumed to be ‘stationary’ within the model, thereafter.

To explore the effect of uncertainty on the outcomes of our analyses we conducted both deterministic and probabilistic sensitivity analysis.

People with T1DM without DR at first screening event were younger (22.7 years) with a longer duration of diabetes (7.8 years) and a higher HbA_1c [71.6 mmol/mol (8.7%)] compared to those with T2DM without DR [age 62.3 years, known duration of diabetes 4.2 years, HbA_1c 58.5 mmol/mol (7.5%)] (Table 3). All people with T1DM were treated with insulin, with 1.6% receiving metformin and surprisingly 9.8% were recorded as being prescribed sulphonylureas in addition to their insulin which casts doubt on the accuracy of the designated type of diabetes recorded. However, in the absence of being able to review the clinical data due to the anonymisation procedure no change was made to the database, but this limitation was noted.

Transition through the stages of DR to RDR occurred faster for people with T1DM compared to those with T2DM. The risk of progression analysis indicates that HbA_1c and duration of diabetes were the strongest predictors of progression of DR. The transition probabilities were highly sensitive to the threshold level of HbA_1c for both T1DM and T2DM, with higher levels resulting in an increased rate of DR progression. Supplementary Figs. 2a, 2b, 3a and 3b illustrate the time to the development of RDR for T1DM and T2DM at different levels of HbA_1c and known duration of diabetes. The median time to the development of RDR occurred at 20 years for HbA_1c of 47.5 mmol/mol (6.5%), 13 years at 63.9 mmol/mol (8.0%) and 8 years 74.9 mmol/mol (9.0%) respectively for T1DM and > 20 years, 17 years and 12 years respectively for T2DM. For people with T1DM there was a distinctly slower rate of progression to RDR for those with a shorter duration of diabetes (< 6 years) versus those with a longer duration (≥ 6 years) with little difference in the speed of progression between 6–12 and > 12 years duration. For people with T2DM the speed of progression to RDR was quickest in people with diabetes for > 6 years, intermediate for 3-6 years and slowest for < 3 years duration. The median time to development of RDR for T1DM was < 20 years after < 6 years of diabetes, and 10 years after ≥ 6 years of diabetes. In comparison, the median time to development of RDR in T2DM was > 20 years with a known duration of diabetes of < 3 years, 19 years after 3–6 years and > 15 years after a known duration of diabetes of > 6 years.

The base case results of the analysis showed that extending the screening interval in people with T1DM and T2DM to biennial reduced costs by £37 and £79 per person, respectively and effectiveness by 0.0016 and 0.0011 QALYs respectively. The resulting DCERs of £23,446 and £71,243 per QALY for T1DM and T2DM, respectively are above the threshold of £20,000 per QALY indicating that the biennial screening strategy is cost-effective in both populations.

Deterministic sensitivity analysis was conducted for a number of scenarios using the strongest predictors of progression to RDR i.e. HbA_1c and known duration of diabetes (Table 4). For people with T1DM at Analysis 2: HbA_1c levels of 48 mmol/mol (6.5%) the DCER was £94,696 (analysis 1) falling to £37,646 at 64 mmol/mol [8.0% (analysis 2)], to £23,446 at 72 mmol/mol [8.7% (analysis 3)] and finally to £11,089 at 80 mmol/mol [9.5% (analysis 4)]. When the duration of diabetes was varied those with T1DM for < 6 years had a DCER of £83,856 (analysis 5) which reduced to £23,446 after 6-12 years (analysis 6) and £13,340 after 12 years (analysis 7). These analyses, at a threshold of £20,000, suggest that for those people with T1DM for less than 6 years it would be considered cost-effective to extend the screening interval to biennial (analysis 5), becoming borderline cost-effective between 6 and 12 years (analysis 6) and not cost-effective after 12 years (analysis 7) nor would it be cost-effective at a HbA_1c level between of 80 mmol/mol (9.5% [analysis 4]) or above. In contrast for those people with T2DM extending the screening interval to biennial for all analysis performed the DCER remained cost effective at a threshold of > £20,000 for all modelled HbA_1c levels and diabetes durations.

Additional sensitivity analysis was performed where the discount rate for costs and outcomes was reduced to 1.5%. This reduction reduced the DCER for biennial screening intervals to £14,221 per QALY lost for T1DM and £56,822 per QALY lost for T2DM.

Probabilistic sensitivity analysis was conducted to assess the combined parameter uncertainty in the model. In this analysis, the mean values that were utilised in the base-case were replaced with values drawn from distributions around the mean values. At a threshold of > £20,000 per QALY gained, extending the screening interval to biennial was found to have a 57% and 97% probability of being cost-effective in people with T1DM and T2DM, respectively.