Multicenter Evaluation of Diagnostic Circulating Biomarkers to Detect Sight-Threatening Diabetic Retinopathy

This cross-sectional study investigates the use of circulating serum biomarkers to distinguish individuals without diabetic retinopathy from those with sight-threatening diabetic retinopathy.

T here are approximately 537 million people living with diabetes globally, the majority of whom have type 2 diabetes. 1 Timely identification and prompt treatment of sight-threatening diabetic retinopathy (STDR) will reduce the risk of visual impairment in people with diabetes. Therefore, regular retinal screening is recommended, but global coverage of systematic DR screening is a formidable challenge. 2,3 Approximately 80% of people with diabetes live in low-and middle-income countries where resources are restricted. 1 Visual impairment owing to STDR is also more common in minoritized ethnic groups. 1 Therefore, a paradigm shift in DR screening strategy is required.
An alternative and more cost-effective risk-based strategy for early identification of STDR may be possible if circulating biomarkers could be used to triage those at risk of STDR for retinal screening. Previous discovery and verification studies have reported a selection of circulating biomarkers with good diagnostic power to identify those with STDR. 4 These need to be validated in a large cohort in a clinical setting before being proposed as a viable triage tool for retinal screening. 4 In this study, we aimed to evaluate previously verified blood biomarkers using enzyme-linked immunosorbent assay (ELISA) for their potential usefulness as indicators of STDR.

Methods
This study was approved by the National Research Ethics Service in the UK and the institutional review board of Vision Research Foundation and Aravind Medical Research Foundation in India. All patients provided written informed consent, and the study followed the Declaration of Helsinki. Participants did not receive any compensation or incentives to participate. This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines.

Study Design and Participants
This multicenter cross-sectional study was conducted from October 22, 2018, to December 31, 2021, to evaluate previously verified circulating biomarkers for STDR in adults 40 years and older with type 2 diabetes. The study was conducted in parallel in the UK and India. In the UK, the study population was recruited at 3 outpatient ophthalmology clinics, with 2 in London, UK, and 1 in Norwich, UK. Participant race and ethnicity information was gathered based on self-reported data and classified according to their relevant office for national statistic classification. Race and ethnicity categories included South Asian, other Asian (ie, as categorized by the UK 2011 Census, an individual who is not Bangladeshi, Chinese, Indian, or Pakistani is classified as other Asian), Black, White, and other (ie, any individual ethnicity that is not named in the UK 2011 Census). Prior studies have shown that racial and ethnic minoritized groups are disproportionately affected by STDR. 5 In India, study participants were recruited at 2 centers and included exclusively people of Indian origin residing in both urban and rural regions in Tamil Nadu, India. Participant sex was determined based on self-reported data.

Masking
Routinely collected blood parameters such as glycated hemoglobin A 1c (HbA 1c ), estimated glomerular filtration rate (eGFR), and lipids were measured in hospital laboratories. The samples were anonymized before transfer to the laboratories. All laboratory staff were masked to the clinical diagnosis and data and were assigned a study cohort.

Preparation of Serum Samples
Laboratory researchers in both countries followed the same storage and processing instructions. Blood samples were collected in serum separator tubes (BD vacutainer [BD]), allowed to stand for 60 to 120 minutes, then centrifuged at 1100g to 1300g for 10 minutes. Serum was then aliquoted, frozen, and stored at −80°C until sample processing in each laboratory. For the UK cohort, all 13 markers were analyzed in a single University College London laboratory whereas for the Indian cohort, 6 markers were analyzed at the Aravind Eye Hospital, Madurai, India, and 7 at the Vision Research Foundation, Chennai, India.

Enzyme-Linked Immunosorbent Assay
Serum samples were thawed on ice and serially diluted to the required, previously identified, concentration. ELISA procedures, including the establishments of standard curves, were carried out in accordance with manufacturer's instructions (eTable 1 in the Supplement). Final absorbance values were measured on spectrophotometers. Dilutions of known protein standards (as provided by the manufacturers) were included on each plate. From this, standard curves were derived using 4-parameter logistic regression, in GraphPad Prism. Absolute values for each experimental sample were interpolated from these curves.

Outcome Groups
Group 1 with no diabetes was compared with group 2, which served as the control for this study and consisted of people with type 2 diabetes for at least 5 years and no evidence of any DR. Individuals with STDR consisted of group 3, individuals with type 2 diabetes and nonproliferative DR (NPDR) with diabetic macular edema (DME) or group 4, individuals with type 2 diabetes and PDR with or without DME. DR and DME were diagnosed using fundus photographs and optical coherence tomography, respectively.

Sample Size
A sample size of 45 participants with diabetes and no DR was sufficient for modeling STDR outcomes as this reliably distinguished between observed specificity estimates of 60%, 75%, and 90%, having respective 95% CI half-widths of 14.3%, 12.7%, and 8.8%. A sample size of 50 participants allowed for an anticipated 10% of missing data. Assuming that 135 of 150 participants (90.0%) recruited with STDR have complete data, observed sensitivity estimates of 60%, 75%, and 90% for detecting STDR would also be well distinguished, having 95% CI half-widths of 5.1%, 7.3%, and 8.3%, respectively. Sample groups with full data of sizes 135 and 45 participants also provide informative 95% CI half-widths of 0.09, 0.07, and 0.04 for observed areas under the receiver operative characteristic (ROC) curve of 0.6, 0.75, and 0.9, respectively.

Statistical Analyses
Data were summarized by outcome group and by country, using mean (SD) or median (IQR) for continuous variables and number (%) for categorical variables. Statistical differences between no DR (group 2) and no diabetes (group 1) as well as between no DR (group 2) and individual STDR groups 3 and 4 were compared using the t test or Mann-Whitney U test for continuous variables and χ 2 test or Fisher exact test for categorical variables. Probability weighted comparisons were made between no DR and combined STDR group. Potential multicollinearity was assessed using the variance inflation factor. Weighted logistic regression was used to identify variables that improve detection of STDR relative to no DR, in addition to a (base) model consisting of known risk factors for STDR: age, disease duration race and ethnicity (in the UK), and HbA 1c . Owing to the low prevalence of STDR in people with diabetes, recruiting a cohort that is representative of the general population in the proportion with STDR would require a large sample size to ensure there were sufficient events. Therefore, we oversampled STDR groups in our study and used weighted logistic regression to mitigate differences in the sample and population prevalence induced by case-control sampling. 6 The diabetes population proportions of the groups were derived as follows: group 2 (no DR), 93.96%; group 3 (NPDR with DME), 5.28%; and group 4 (PDR), 0.76% for the UK from the English Diabetes Eye Screening Programme 7 and group 2 (no DR), 94.80%; group 3 (NPDR with DME), 4.45%; and group 4 (PDR), 0.75% for South India from the South India part of the Statistical and Economic Modeling Study of Risk-Based Stratified and Personalized Screening for Diabetes and its Complications in India (SMART) India study, 8 after discounting the NPDR with-out DME group, not recruited in this study. Probability weights w i for each disease group, were calculated as τ i /y i , the ratio of the population proportion (τ i ) and the sample proportion (ȳ i ). Variables that reached statistical significance in the adjusted analysis were introduced into a forward stepwise selection routine with an entry criterion of α = 0.1, with age, disease duration, race and ethnicity (in the UK), and HbA 1c forced into the models. Log e transformation was used for variables serum creatinine, C-reactive protein, triglycerides, A2m, 8-OHDG, ApoA1, ApoB, ApoC3, ApoB/ApoA1, LRG-1, and cystatin C for logistic regression analysis, to avoid estimating misleading associations from influential outlying observations. Odds ratios (ORs), therefore, are interpreted in relation to relative percentage increase in the biomarker rather than per absolute increase in the measured units for other biomarkers.
The final models were summarized using ORs with 95% CIs and the ROC curve. Sensitivity analysis was carried out to compare the area under the curve (AUC) of closely related biomarkers serum creatinine and cystatin C. In addition, cystatin C levels were compared between STDR (groups 3 and 4) and no DR (group 2) in patients with normal kidney function (eGFR ≥90 mL/min/1.73 m 2 ) as elevated cystatin C levels may be attributable to diabetic kidney disease rather than STDR. All 2-tailed P values < .10 were considered statistically significant. Statistical analysis was carried out in RStudio, version 3.6.3 (RStudio) 9 and Stata MP, version 15 (StataCorp). 10  Table 1). In the UK, mean (SD) age of participants was 63.0 (10.5) years and 148 (54.0%) were men, whereas in India, mean (SD) age was 58.4 (8.5) years and 171 (64.8%) were men.

Descriptive Analysis
The demographic, clinical, and biomarker data of the UK and India cohort, by outcome group, are shown in Table 1, and weighted descriptive analyses comparing No DR and STDR groups used for modeling are shown in eTable 2 in the Supplement. There were significant differences in biomarker values between the UK and India across nearly all disease groups ( Figure 1 and eFigure 2 in the Supplement); therefore, we continued as planned to analyze the markers by country as allowed in the sample size calculation but did not further pool the data. Box plots summarizing biomarker values by ethnic group for the UK are presented in eFigure 3 in the Supplement.   Analysis of Expression Patterns: UK Adjusted logistic regression results for demographic, clinical, and biomarker data are shown in Table 2. In the adjusted analysis, the following biomarkers were found to be elevated in STDR (groups 3 and 4) relative to the group with no DR (group 2) in the UK: cystatin-C (OR, 1.12 μg/mL; 95% CI, 1.02-1.23 μg/mL; P = .02) and LRG-1 (OR, 1.06 μg/mL; 95% CI, 1.00-1.11 μg/mL; P = .03).

Multicollinearity Between Biomarkers
In those variables that reached significance in adjusted analysis (age, duration, HbA 1c , cystatin C, and LRG-1), the variance inflation factor ranged from 1.1 to 1.2 in the UK. In India, the variance inflation factor ranged 1.1 to 1.6 for variables age, duration, HbA 1c , cystatin C, afamin, A2m, Lp-PLA2, and CFB, which were variables considered for inclusion in the multimarker panel, indicating that multicollinearity is unlikely to cause problems for our analysis.

Combination of Biomarkers to Improve Discriminatory Power
For the UK model, the forward routine with selected additional biomarkers LRG-1 and cystatin C together yielded an AUC of 0.807 (95% CI, 0.734-0.879) on top of variables in the base model. In India, additional biomarkers included cystatin C and CFB, with a combined AUC of 0.735 (95% CI, 0.647-0.822) (eTable 2 in the Supplement, Table 3, and Figure 2).

Diagnostic Test Accuracy and Applications in a Prospective UK and India Population
STDR detection relative to no DR in 1000 individuals in the UK and India populations using the biomarker prescreening strategy is shown in Table 3. In a UK population with estimated prevalence of STDR relative to no DR of 6%, if 1000 people with no DR or STDR undergo retinal screening, 60 out of 1000 people will be identified as having STDR. If the model including cystatin C without LRG-1 is used to prescreen these 1000 patients, 274 patients would be classified as high risk of STDR, and 47 out of 60 patients with STDR will be detected with a test that maximizes the Youden index. The remaining 13 patients with STDR would need to be identified as per current practice. In India, retinal screening on 1000 people with no DR or STDR will identify 52 individuals with STDR based on the prevalence of STDR relative to no DR of 5.2%. If cystatin C was used to test these patients first, a total of 115 out of 1000 will be triaged for retinal screening and 29 out of 52 STDR will be identified, using a targeted threshold of more than 8.6%. At 80% sensitivity (30% specificity), we would need to screen 704 patients out of 1000 to detect 42 out of 52 STDR cases, with approximately 97% of low-risk patients correctly identified as having no STDR.

Sensitivity Analysis
As both cystatin C and serum creatinine had comparable AUC, cystatin C (log) was substituted with serum creatinine (log). In the UK, only 205 patients with 137 STDR events had available data to use serum creatinine instead of cystatin C in the models. The base model with serum creatinine instead of cystatin C, yielded an AUC of 0.779, and when combining with LRG-1, the AUC was 0.791 (eFigure 4 in the Supplement). Using the same sample of 205 patients yielded an AUC of 0.799 in the model using cystatin C instead of serum creatinine. In India, replacing cystatin C with serum creatinine reduced the AUC to 0.661 in a model with the known risk factors and serum creatinine, and combined with CFB, the AUC was 0.731 compared with 0.735 in a model with cystatin C and CFB. As cystatin C is also a marker of kidney disease, we conducted a subgroup analysis in patients with normal kidney function, ie, eGFR of 90 mL/min/1.73m 2 or greater (63 in the UK [33 STDR events] and 61 in India [38 STDR events]). In patients with an eGFR of 90 mL/min/1.73m 2 or greater, the presence of STDR correlated with elevated cystatin C levels in the UK (median, 1.1; IQR, 0.74-1.5 vs median, 1.7; IQR, 1.0-1.9; weighted Mann-Whitney U test, P = .004; age-and duration-adjusted weighted logistic regression, P = .05) and similarly in India (median, 1.1; IQR, 1.1-1.3 vs median, 1.4; IQR, 1.1-1.5; weighted Mann-Whitney U test, P = .21; age-and durationadjusted weighted logistic regression, P = .04).

Discussion
This cross-sectional study evaluated previously verified circulating biomarkers for DR to establish if these markers could identify STDR from no DR in clinical practice. Many of the chosen markers did not perform as well as previously reported. 11,12 However, our results suggest possible use of circulating cystatin C as a triage test, as well as the known risk factors of age, race and ethnicity, duration of diabetes, and HbA 1c 13,14 as a potentially valid method to identify individuals with STDR.
In the UK cohort, only cystatin C and LRG-1 showed significant differences between no DR (group 2) and STDR (groups 3 and 4). Comparing the STDR and no DR groups in India, CFB, A2m, afamin, cystatin C, and Lp-PLA2 were significantly dif-ferent at the α = 1% threshold. Owing to interlaboratory variations in some of the biomarkers investigated, we opted to include only the most stable biomarker in the strategy with nearacceptable discrimination in both study cohorts.
When cystatin C levels were added to age, duration of diabetes, race and ethnicity, and HbA 1c in the UK, the analysis showed acceptable performance, with an AUC of 0.779 compared with an AUC of 0.735 without cystatin C. In India, the model with age, duration, and HbA 1c had performance closer to chance, with an AUC of 0.552, and adding cystatin C had close to acceptable performance with an AUC of 0.696. The validity of this strategy was similar in the UK and India; therefore, results suggest that this biomarker was useful in both countries. 15 Both models achieved adequate sensitivity while preserving test specificity to an extent; however, the UK models had higher specificity across all tested thresholds. This might be partly attributable to the duration of diabetes not being as accurate in detecting STDR in India unlike the UK, where this was one of the strongest indicators of STDR. Nonetheless, our results show stability across both data sets. Several studies done globally have demonstrated an association between cystatin C and DR. [16][17][18][19][20]

Strengths and Limitations
This study had several strengths. First, both the UK and India models achieved close to acceptable discrimination. Second, the models relied mostly on routinely collected data, except for cystatin C. However, a point-of-care cystatin C biosensor is feasible to produce, will not require skilled workers to perform, and may be accessible as a web-based application for self-monitoring by patients. 21 Third, in contrast to many other published studies, which often use mass spectrometry of plasma, 4,11,12 we opted for ELISA analyses of serum to remain close to what is routinely available in a health care setting. 22 Fourth, data for cystatin C, used in our final model, did reflect that seen in the literature. 18,19 However, there were several limitations. First, we were unable to pool the samples from UK and India owing to differences in the distribution of some biomarkers. More studies are required to understand intercountry variations in the profiles of these biomarkers. Second, further validation in a cohort representative of the real-world prevalence of STDR and not oversampled for cases is needed. Third, as these markers were analyzed in stored and freeze-thawed serum only, further validation of our prescreening strategy should be undertaken in a community screening setting using finger-prick blood test. Notably, similar low-cost point-ofcare HbA 1 c kits are already available for community screening. 23 Fourth, validation of a cystatin C point-of-care kit is needed before use in a health care-or communityscreening setting. Fifth, patients with NPDR but without DME were not recruited into this study, and therefore, statistical models may not generalize to all patients with diabetes or represent population-level risk of STDR. Sixth, serum cystatin C has also been shown to be an early marker of diabetic kidney disease, 24 but our results showed that cystatin C remained elevated in STDR in a subgroup of patients with normal eGFR. United Kingdom India DME indicates diabetic macular edema; NPDR, nonproliferative diabetic retinopathy; PDR, proliferative diabetic retinopathy. a P values were generated from the Mann-Whitney U (Wilcoxon rank sum) test comparing biomarkers in the United Kingdom and samples from India. b P Յ .0001. c P >.05.    Area under the receiver operating characteristic curve (AUC) and 95% CIs are presented. LRG-1 indicates leucine-rich α2 glycoprotein 1. a In the United Kingdom, a total of 215 patients with 146 sight-threatening diabetic retinopathy (STDR) events were used to derive the final models. The base model includes age, diabetes duration, race and ethnicity (South Asian, Black, other) and hemoglobin A 1c (HbA 1c ). b In India, a total of 208 patients with 155 STDR events were used to derive the final models. The base model includes age, diabetes duration, and HbA 1c .