Optimal Frequency of Retinopathy Screening in Adolescents With Type 1 Diabetes: Markov Modeling Approach Based on 30 Years of Data

Diabetic retinopathy (DR) is a progressive condition characterized by retinovascular abnormalities, usually detectable by fundoscopic examination or assessment of retinal photographs. Annual screening for DR was recommended when widespread screening to detect sight-threatening DR, requiring ocular intervention, and prevent vision loss, was introduced in the early 1990s. Subsequently, less frequent screening has been supported by some studies, based on relatively low rates of severe DR (1–4). More recently, personalized screening intervals have been advocated, with intervals influenced by diabetes duration, HbA_1c levels, and prior DR status (5–7). Screening intervals for people with type 1 diabetes (T1D) have been critically appraised in relatively few studies (8,9), the largest to date being the interventional Diabetes Control and Complications Trial (DCCT), with its observational follow-up, the Epidemiology of Diabetes Interventions and Complications (EDIC) study. The DCCT/EDIC investigators suggest extending screening intervals to 4 years for those with no DR at the current visit because of low probability of progression to proliferative DR (PDR) or clinically significant diabetic macular edema (DME) (8), and in adolescents suggest that one screening only may be required (10). Of concern, however, it is widely recognized that initial screening in groups at greater risk for DR progression than the DCCT cohort was already less frequent and delayed, especially in younger people with diabetes (11,12). There is nonetheless a growing opinion that DR screening may be performed even less frequently, potentially resulting in cost savings and improving adherence (10).

Over the last 30 years, 7-field stereoscopic fundus photography has been used for DR screening in our diabetes tertiary referral clinic, providing the opportunity to review the natural history of DR transitions in a young population with T1D from childhood up to the age of 25 years. The primary objective was to determine appropriate screening intervals for DR with low probability of progression to a state requiring medical ophthalmological intervention. The secondary objective was to measure clinical modifiers for DR progression in youth with T1D to examine how these affect probabilities of DR progression at the various screening intervals. We used Markov chain as this allows modeling transitions of DR progression as well as regression.

We analyzed data from 2,063 youth with T1D who had their first DR screening at age 10–18 years after at least 2 years of diabetes duration at recommended intervals of 1 to 2 years. All participants attended the Diabetes Complications Assessment Service (DCAS) at the Children’s Hospital at Westmead (CHW) in Sydney, New South Wales, Australia, at least twice over a 30-year period between 1990 and 2019 and were eligible for repeated assessments until the age of 25 years. They were referred from within the CHW Diabetes Clinic and from the broader community in metropolitan and rural New South Wales (13). The diagnosis of T1D was based on the International Society for Pediatric and Adolescent Diabetes guideline definition (14), with diabetes-associated autoantibody testing at diagnosis being standard practice. The study was approved by the CHW Human Research Ethics Committee (HREC no.: 2020/ETH00326) and conducted in adherence with the Declaration of Helsinki.

DR was assessed using 35° angle 7-field mydriatic stereoscopic retinal photography of both eyes using a TFC 50-VT camera (Topcon, Tokyo, Japan), updated to TRC 50DX (Topcon) in 2012. Retinal images were graded from slides between 1990 and 2004 and from digital images using the IMAGEnet R4 system (Topcon) thereafter. All grading was performed at the single center by a consultant ophthalmologist (1990–2005 S.H.) and an orthoptist (2006–2019 A.P.) and adjudicated by consultant ophthalmologists (2006–2014 S.H.; 2014–2019 G.L.), all masked to clinical data. Intragrader concordance was excellent: for repeated assessment of 100 eyes, the weighted κ value was 0.86 (95% CI 0.72–0.99). Intergrader reliability was measured after independent reassessment by two graders of 100 eyes and resulted in a weighted κ value of 0.81 (95% CI 0.70–0.91).

The retinopathy grading protocol was adapted from the Early Treatment Diabetic Retinopathy Study (ETDRS) and The Wisconsin Epidemiologic Study of Diabetic Retinopathy (15–19): grade 10 was no retinopathy; grade 21 was minimal NPDR; grade 31 was mild NPDR; grade 37, 43, and 45 was moderate or moderately severe NPDR; and grade 53 was severe NPDR; grade ≥60 was PDR. DME was defined as presence of hard exudate at the macula with retinal thickening (from stereoscopic retinal images) in field 1 (optic disc centered), field 2 (macula centered), or field 3 (temporal to the macula), according to ETDRS criteria (15,20). Optical coherence tomography data, the gold standard nowadays for DME diagnosis (21), were not available. For analysis, each person visit was stratified into a retinopathy “state” based on grading of retinal images in the worst eye: state 1—grade 10, state 2—grade 21, state 3—grade 31, and state 4—comprising grades 37 to ≥60 or DME. We thus examined grading results of each individual’s repeated retinal photography-based DR status between 10 and 25 years of age to determine DR progression rates. The main outcome was referable retinopathy, defined as that requiring medical referral to an ophthalmologist, according to Australian national guidelines (22). This corresponded to state 4 (termed “the absorbing state” in the Markov chain). Patients who reached state 4 were referred to local ophthalmologists for further investigation and treatment.

Glycemic control was assessed by glycated hemoglobin (GHb) colorimetrically before February 1994 (23). From 1994 to 2009, HbA_1c was measured by high-quality liquid chromatography by using the Variant analyzer (Bio-Rad Laboratories, Hercules, CA) and subsequently with the Adams automated analyzer (Arkray, Kyoto, Japan) from January 2010 onward (Adams = 1.0566 × Variant, R² = 0.98). GHb values were converted to HbA_1c (24) (Diamat = 1.9088 + 0.0043 × GHb; R² = 0.85). DCA 2000 analyzer values were included at interim clinic visits from 1994 (Diamat = 1.0766 × DCA 2000– 0.0871; R² = 0.9206), and all available values for GHb were included to calculate the individual’s mean HbA_1c.

Time-weighted mean HbA_1c was calculated 1) using DCCT methodology (summing first DCAS HbA_1c × diabetes duration at first DCAS) and (mean HbA_1c × years of follow-up duration) and dividing by total diabetes duration (25) and 2) as regular time-weighted (sum of HbA_1c × time between visits divided by duration of follow-up). Mean HbA_1c was recorded as a National Glycohemoglobin Standardization Program (NGSP) percentage and converted to mmol/mol using the International Federation of Clinical Chemistry (IFCC) Standardization formula ([10.93 ∗ NGSP HbA_1c] – 23.50) (24). HbA_1c SD and HbA_1c coefficient of variation (CV), as additional measures of glycemic variability, were calculated using data points from all visits available for the patient.

Cholesterol was measured using a Beckman CX5 (1990–1999), a Dimension RXL (2000– 2005), and a Vitros analyzer (Ortho Clinical Diagnostics, Raritan, NJ) from 2005 onward. Elevated total cholesterol was defined as ≥5.5 mmol/L.

Height (to the nearest 0.1 cm) was measured using a Harpenden stadiometer and weight (to the nearest 0.1 kg) using electronic scales at each clinic visit. BMI was calculated as kilograms per meters squared. These measurements were converted to BMI Z-scores using the 2000 Centers for Disease Control and Prevention reference standards (26). Normal weight was defined as BMI between the 5th and <85th percentile, overweight as BMI between the 85th and 95th percentile (inclusive), and obesity as BMI >95th percentile.

Systolic and diastolic blood pressure (SBP, DBP) were measured after a 5 minute rest in the seated position by auscultation using an appropriately sized cuff, with age and sex-related SD scores calculated according to the new reference. Elevated BP was reported as SBP and/or DBP >90th percentile (on the basis of age, sex, and height percentiles) (27).

Microalbuminuria was defined as albumin excretion calculated from two of three timed overnight urine collections being >20 µg/min.

Demographic data of the cohort were examined by summary statistics. Data were analyzed using R 4.0.5 software. We examined between-visits transitions from one DR state to another and overall rates of progression and regression of retinopathy using retinal images grading. We calculated the incidence of MNPDR or worse and, separately, of DME.

A longitudinal Markov chain model, implemented in R package msm was used for data analysis (28). The model allowed progression from state 1 to 2, state 2 to 3, and state 3 to 4. It also allowed regression from state 2 to 1 or from state 3 to 2. Data were censored at the follow-up age of 25 years and/or when the retinal grade reached state 4, a critical level of DR that required medical ophthalmological attention. The maximum follow-up per individual was thus 15 years. Fragments of R code (cumulative distribution function equation) were imported from available literature (29). We used mixed-effects Cox regression, implemented in R package coxme, to estimate regression coefficients between HbA_1c, diabetes duration, and retinopathy states transitions (30).

The “sojourn time” is the nonprogression time in the same retinopathy state before progression to a higher/worse state. On the basis of cumulative incidence of state 4, the probability of progression to state 4 from each state was calculated for time intervals of 1, 3, and 6 months and at 1, 2, 3, 4, and 5 years from the assessment (Supplementary Table 2). The duration of the time between visits was calculated at fixed intervals (Table 3) and in a manner to limit this probability of progression to state 4 to 1% or 5% (Supplementary Fig. 1).

The effects of cofactors for transition were examined, and likelihood for progression was calculated using these intervals and fixed time intervals. In addition, their effect was calculated as the hazard ratio (HR) of transition to a higher or lower DR state. Putative cofactors examined were baseline HbA_1c, current HbA_1c, mean HbA_1c, time-weighted mean HbA_1c, sex, age, age of T1D diagnosis, diabetes duration, BMI, microalbuminuria, overweight/obese category, ratio of insulin dose to weight, diabetes duration at the first DCAS, and year of the first DCAS assessment (Supplementary Table 3).

An algorithm was developed to determine the recommended interval until the next DR screening, encompassing the most important cofactors (current HbA_1c, diabetes duration, and sex) and the current retinopathy state, with variable probability for progression to state 4. The calculator is available at https://www.ajenkinsdiabetes.org/html/calculators/dr-screening.

Baseline characteristics are summarized in Table 1. The median age at the first assessment was 13.1 (interquartile range 11.9, 14.5) years and at the last assessment was 17.6 (16.2, 19.3) years. Most individuals (59.4%) were diagnosed with T1D before 2000, and 76.4% had an initial DR screening before 2010. The median time between screening visits was 1.5 (1.1, 2.2) years, and follow-up time 4.1 (2.4, 6.5) years. This analysis did not include 1,047 adolescents who only had a single DR screening and 243 (2.75%) who had assessments with ungradable photographs.

A total of 8,598 retinal assessments were used in this analysis. There were 9,823 patient-years of follow-up. The incidence of MNPDR or worse was 8.6 per 1,000 patient-years, and the incidence of DME was 4.9 per 1,000 patient-years.

Between consecutive visits, individuals’ retinal grades progressed in 18.4%, stayed the same in 72.0%, and improved in 9.6% (Supplementary Table 1). Two-step retinal grade progression was observed in only 3.0%, and two-steps regression was noted in 0.7% of visits. Overall, 84 transitioned to state 4 (Supplementary Table 1) between two consecutive visits (1%). Among those 20 individuals whose DR transitioned to moderate or moderately severe NPDR, 4 transitioned to severe NPDR, 1 transitioned to mild PDR, 49 transitioned to DME, and 10 transitioned to moderate or moderately severe NPDR and DME.

The sojourn time spent in each state before progression to a higher grade or improvement to a lower grade (mean [95%CI]) was 4.49 (4.08–4.94) years for state 1, 1.11 (1.01–1.22) years for state 2, and 1.26 (1.04–1.53) years for state 3.

The probabilities of DR state transitions are reported in Supplementary Table 2. For those with no retinopathy (state 1), the probability of progression to severe (state 4) retinopathy was 1.3% at 3 years and 3.3% at 5 years. For those with state 2 DR, the probability was 1.3% at 1 year and 3.3% at 2 years, and for those with state 3 DR, the probability was 12.2% at 1 year and 18.5% at 2 years. Probabilities of transition to this state at 1 and 3 years were from no DR, 0.1% and 1.3%; from minimal DR, 1.3% and 5.1%, and from mild DR, 12.2% and 22.2%, respectively.

The cumulative incidence of state 4 DR was strongly associated with HbA_1c (Fig. 1). A 1% higher HbA_1c was associated with a 23% increased risk for DR progression from state 1 to state 2, and a 10% lower risk for DR regression from state 2 to state 1 (Table 2). The HR of DR progression from state 1 to state 2 and from state 3 to state 4 was significantly increased with an increase of baseline HbA_1c (at first DCAS), mean HbA_1c, or time-weighted HbA_1c (Supplementary Table 3). HbA_1c alone explained 27% of the transitions between no DR (state 1) and severe DR (state 4). The addition of diabetes duration into the model with HbA_1c increased this value to 31% (P = 0.02).

The effect of risk modifiers on the probability of progression to state 4 from each earlier state is reported in Table 3. For example, for a youth with no DR (state 1), if the current HbA_1c is 10% and repeat screening is 3 years later, the probability of DR progression to state 4 is 2.2%. If the screening is extended to 4 years, the probability of DR progression to state 4 will be 3.8%. Similarly, the probability of DR progression to state 4 will be 7.2% if the individual has minimal DR and 26.8% for mild DR and HbA_1c of 10%. If screening is extended to 4 years, those values would be 3.8%, 9.8%, and 30.2%, respectively. Even with no retinopathy at a previous visit, those with a between-visits interval of 15 years will have 14.3% probability of DR progression to state 4 (Supplementary Fig. 1).

The HR of DR regression from state 2 to state 1 and from state 3 to state 2 changes by 7% and 10%, respectively, with an increase in diabetes duration of 1 year at the first DCAS (Supplementary Table 3). The HR of DR progression from state 1 to 2 and from state 2 to 3 was reduced with an increase of calendar year of the first DCAS by 6% and 4%, respectively, but increased for DR transitions from state 3 to state 4 by 6% (Supplementary Table 3). Older age was associated with a 7% higher HR of DR progression (from state 1 to 2) and 11% (from state 3 to 4) and a 13% lower HR of regression (from DR state 2 to 1) (Supplementary Table 3). Longer diabetes duration increased the HR of DR progression from state 1 to 2 and from state 2 to 3 (but not from state 3 to 4) by 5% and 7%, respectively, and reduced that for DR regression from state 2 to 1 by 10% (Supplementary Table 3).

Sex, BMI, and insulin dose (Supplementary Table 3) showed no effect on the HR of DR progression or regression. Older age at T1D diagnosis, HbA_1c SD, and HbA_1c CV only significantly altered DR progression from state 1 to 2, with no effect on DR regression. Elevated BP at baseline was associated negatively with DR regression and (paradoxically) with progression from DR state 3 to state 4 (latter explainable because of BP treatment medication being prescribed when elevated BP detected).

Based on 30 years of clinical data from 2,063 Australian youth with T1D, we have described the rate of change in DR and modulating clinical factors. Using these data, we also developed a risk calculator to estimate the risk of and rate of change in DR status in this population. In youth aged <25 years with T1D, DR progression determined by 7-field stereoscopic fundus photography was generally slow, with low rates of transition from no or early DR to PDR and DME.

In this cohort, with a minimum age of 10 years and a minimum duration of 2 years at the initial DR assessment, the sojourn time with no DR (state 1) was 4.5 years, but only 1.1 years and 1.3 years in state 2 (minimal) and state 3 (mild), respectively. Progression to state 4, referable DR that required ophthalmologic attention, including MNPDR or worse, PDR and DME, only occurred in 4.1% of the cohort (8.6 per 1,000 patient-years).

Our approach in this real-world large cohort of youth with T1D aligns with that of the previously published (8) DCCT/EDIC cohort study in adolescents and adults with diabetes, but a minimum of 3-yearly examinations are supported for most youth, compared with 4-yearly screening intervals suggested by Nathan et al. (8). Similarly, we used Markov chain modeling analysis to calculate the probability of progression or DR regression over various model time intervals and the effect of clinically relevant covariates, including HbA_1c, sex, age, age of diabetes diagnosis, and prior DR status. The model allows one to choose an acceptable risk of missing progression to state 4 DR. If we accept a 1% probability of progression to state 4, then DR screening intervals would be 2.65 years for those with no retinopathy, 0.85 years for those with state 2 (minimal retinopathy), and 1 month for state 3 (mild retinopathy). If we accept a 5% probability of progression, then the screening intervals would be 6.57 years for those with no retinopathy, 2.93 years for those with state 2 DR, and 4 months for state 3 DR (Supplementary Fig. 1). These intervals would be shorter for those with adverse risk factors, such as higher HbA_1c, and longer diabetes duration, all of which can be input into the calculator.

Current guidelines of the International Society for Pediatric and Adolescent Diabetes (31) and the American Diabetes Association (32) recommend commencing DR screening from the age of 11 years and/or after 2–5 years of diabetes duration. The guidelines also recommend repeating screening every 2 years in patients with short diabetes duration, mild NPDR, and target glycemic control, but at shorter intervals, such as annually or even more frequently, in youth with more advanced DR or at high risk for vision loss (31). The American Diabetes Association Standards of Medical Care in Diabetes for children and adolescents recommend a screening interval of every 2 years but less frequent examination every 4 years based on risk assessment (4). The current cohort study provides real-world data for this risk assessment.

Recommendation of a 3-year compared with 4-year screening interval is advisable for youth even with no retinopathy. A 4-year screening interval would be unacceptable (>5% risk of progression) if duration was >15 years. A 3-year screening interval would be acceptable (<5% risk) for youth with no retinopathy. For youth with minimal retinopathy, a 4-year screening interval would not be acceptable if they had high HbA_1c or were the age of 17 years or had longer duration, with probability of progression rising to 14% for the latter. A 3-year screening interval would not be acceptable for more than half from this cohort. This interval would not be acceptable for individuals with mild retinopathy, even with moderately elevated HbA_1c of 8% (risk 13%), and for high HbA_1c, the risk is 30%. Screening intervals should remain yearly for youth with background retinopathy and possibly every 6 months for those with mild retinopathy.

Adolescents can have rapidly progressing retinopathy (33) and can develop DME (34) before the age of 18 years. The current DME incidence of 4.9 per 1,000 patient-years, followed until 25 years, is similar to 2.2 per 1,000 patient-years in the adolescent cohort of DCCT, followed till age of 18 years (10). Our rate of development of DR requiring referral of 8.6 per 1,000 patient-years is similar to that reported from clinical care DR screening of 4.7 per 1,000 years in a younger cohort in the U.K. studied from the age of 12 years, followed for a mean of 3.1 years (35). Our similar but higher rates in the current cohort reflect the longer follow-up. Indeed, the years between 20 and 25 years are likely to confer higher risk, especially in women due to pregnancy.

Poor uptake of screening recommendations is particularly common in those at greatest risk of DR, based on administrative health care data sets (32), and for young adults documented from the U.K. screening programs (11). Hence any extension of the current recommendations for annual and biennial assessments needs to be very carefully considered. Algorithms for individual risk need consideration, and one such algorithm has been developed from the current data set. We have used current HbA_1c and diabetes duration as the clinical variables strongly associated with DR progression. Adding diabetes duration to the model showed significant improvement in the risk calculation.

Study strengths include its clinical relevance, clinical pediatric diabetes setting, the large study size, long duration of follow-up, and detailed and precise DR grading. Nearly half of the studied subjects had duration >10 years at the end of follow-up. Our study used a consistent retinal imaging and grading protocol. The retinal camera was upgraded in 2004 and allowed for digitalization of images.

Adolescents at greater risk of DR may have been referred to our no-cost, comprehensive complications screening service; however, a selection bias for attendance by those most adherent to medical advice may also have occurred. Another weakness is the low rate of progression to advanced DR, in keeping with more modern cohorts. Also, the median length of follow-up is similar to the time required to progress to the higher state from “no retinopathy.” Only 21% were >20 years at follow-up, so recommendations for the age-group of young adults aged 20–25 years are weaker, and this is a time when pregnancy can accelerate retinopathy progression.

In conclusion, our real-world data approach supports extension of DR screening intervals for youth with T1D. However, adherence, patient risk, and cost-savings benefits (which are widely divergent between countries) need to be considered. Personalized screening depends on diabetes duration and glycemia. A pragmatic approach is to extend subsequent screening intervals in youth to 3 years in the absence of retinopathy and remain at annual screening if mild (or worse) retinopathy is detected.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. A.S.J. contributed to the formal analysis, methodology, software, visualization, and interpretation and writing of the manuscript. V.V. and S.H. contributed to the investigation. P.Z.B.-A., M.E.C., G.L., and Y.H.C. contributed to the investigation and to interpretation and writing of the manuscript. J.C. and A.P. contributed to data curation and the investigation. A.P. contributed to data curation and the investigation. E.Y.C. and A.J.J. contributed to the methodology and to interpretation and writing of the manuscript. K.C.D. contributed to conceptualization, investigation, methodology, and interpretation and writing of the manuscript. K.C.D. is the guarantor of this work and, as such, had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Parts of this work were presented as oral presentations at the 80th Scientific Sessions of the American Diabetes Association, virtual meeting, 12–16 June 2020, and the 56th Annual Meeting of the European Association for the Study of Diabetes, virtual meeting, 21–25 September 2020.