Extreme phenotype-derived machine learning reveals susceptibility and resilience signatures for severe diabetic retinopathy.

Background: The mechanisms underlying why some individuals with diabetes develop severe diabetic retinopathy (DR), whereas others remain free of retinal complications despite long-standing disease, remain incompletely understood. We aimed to identify systemic susceptibility and resilience signatures associated with severe DR using an extreme phenotype-derived machine learning framework. Methods: An extreme phenotype cohort was established comprising 712 individuals with diabetes mellitus, including 437 patients with proliferative diabetic retinopathy (PDR; susceptible phenotype) and 275 patients with diabetes duration ≥10 years without retinopathy (resilient phenotype). Clinical and biochemical variables were integrated to develop interpretable machine learning models. The optimal model was further interpreted using SHapley Additive exPlanations (SHAP). An independent community-based diabetic cohort (n=673) was used to evaluate the distribution of susceptibility signatures in real-world populations. Results: LASSO regression identified 21 phenotype-associated features for model development. Among the evaluated algorithms, LightGBM demonstrated the strongest ability to discriminate susceptible and resilient phenotypes, achieving an area under the receiver operating characteristic curve (AUC) of 0.90 (95% CI, 0.84-0.95) in the internal validation cohort. SHAP analysis identified urinary albumin excretion rate (UAER), diabetes duration, serum creatinine, total protein, age, and hypertension duration as the dominant phenotype-defining features. Notably, UAER exhibited a pronounced nonlinear association with susceptibility scores, suggesting a close link between renal microvascular injury and vulnerability to severe DR. When applied to the community cohort, the susceptibility signature showed limited discrimination in the overall population (AUC = 0.54) but became progressively enriched among individuals with greater metabolic burden, reaching an AUC of 0.71 in participants with fasting blood glucose ≥9.0 mmol/L. Conclusion: Using an extreme phenotype-derived machine learning framework, we identified systemic susceptibility and resilience signatures associated with severe diabetic retinopathy. Renal dysfunction, albuminuria, glycemic burden, and disease duration emerged as key phenotype-defining characteristics. These signatures became increasingly enriched in metabolically stressed individuals, supporting the concept that severe diabetic retinopathy arises through the interaction between intrinsic biological susceptibility and cumulative metabolic exposure. This framework may provide new insights into disease heterogeneity and facilitate future precision risk stratification strategies in diabetic eye disease.