Abstract

Background. Local profiling of the aqueous humor (AH) captures key pathobiological axes of diabetic retinopathy (DR) – hypoxia, inflammation, and neuro glial distress. The combination of markers from these biological axes within a multifactorial model may enhance the accuracy of risk stratification and provide a rationale for subsequent individualized patient management. Aim: to quantify hypoxia inducible factor 1α (HIF 1α), interleukins (IL 1β, IL 4, IL 6, IL 8), and neuron specific enolase (NSE) in AH and assess their association with DR severity; and to evaluate the analytical utility of a multivariable predictive model for stratifying patients with type 2 diabetes mellitus (T2DM). Materials and methods. We examined 110 patients with T2DM, grouped by the International Clinical Diabetic Retinopathy (ICDR, 2003) scale: DR0 (n=15), mild nonproliferative DR (NPDR1; n=40), moderate NPDR (NPDR2; n=25), severe NPDR (NPDR3; n=12), and proliferative DR (PDR; n=18). Controls were 25 age and sex matched individuals without T2DM/DR. AH (0.1 mL) was obtained during phacoemulsification; HIF 1α, IL 1β, IL 4, IL 6, IL 8 (pg/mL) and NSE (mg/ mL) were measured by ELISA. Statistical analysis: EZR v1.54; nonparametric tests; multivariable regression with stepwise selection; multiclass One vs All (OVA) classification using the Youden index; and clinically relevant binary phenotyping. Results. Stepwise selection identified three independent predictors of DR severity – NSE and IL 8 (direct associations) and IL 4 (inverse association). Model performance was adequate: R2adjust=0.84, F=219.5, p<0.001; VIF<3 for all coefficients. In multiclass OVA classification, stage specific thresholds of the model output Y (arbitrary units) were: control 2.27; DR0 2.48; NPDR1 3.14; NPDR2 4.52; NPDR3/PDR 6.07; overall accuracy 74% (95% CI: 66%-82%). Binary stratification (NPDR1+NPDR2 vs NPDR3+PDR) using Y=2.71 improved overall accuracy to 80% (95% CI: 72%-86%); for the NPDR3+PDR group, sensitivity 87.5% and specificity 100% were achieved. Although HIF 1α, IL 1β, and IL 6 showed stage dependent trends, they did not enhance the model’s aggregate predictive performance once NSE/IL 8/IL 4 were included. Conclusion. A three marker AH model based on NSE, IL 8, and IL 4 can predict DR severity and yields both multiclass and binary thresholds suitable for clinical application. A threshold of Y≥2.71 can serve as a risk triage criterion to intensify follow up and trigger targeted optical coherence tomography angiography (OCTA) monitoring in patients at high risk of severe DR