Machine learning-based prediction of diabetic retinopathy using clinlabomics: a multi-center study

BACKGROUND: Diabetic retinopathy (DR) is a leading cause of vision loss, yet conventional retinal screening remains costly and resource-intensive. This study developed and validated machine-learning (ML) models using routine laboratory data to provide a cost-effective, accessible alternative for DR risk stratification and triage. METHODS: We analyzed data from 750 patients (363 T2DM; 387 DR) and externally validated the findings with 451 additional cases. Fifty hematological and biochemical parameters were screened. Six algorithms were trained via five-fold cross-validation, with XGBoost emerging as the top performer. Model interpretability and feature selection were conducted using SHapley Additive exPlanations (SHAP) and ablation analysis. RESULTS: The XGBoost model achieved high discriminative performance (AUC = 0.87). Feature ablation identified a streamlined set of four key predictors-total cholesterol (TC), blood urea nitrogen (BUN), fibrinogen (FIB), and glucose (GLU)-maintaining an AUC of 0.87. External validation confirmed robustness (AUC = 0.86) with balanced sensitivity (0.73) and specificity (0.80). Decision curve analysis indicated significant clinical utility, while SHAP provided individualized prediction transparency. CONCLUSIONS: Routine laboratory parameters effectively power ML models for DR prediction. The resulting web-based XGBoost tool offers an interpretable, accessible solution for adjunct risk scoring and early triage, particularly beneficial for prioritizing high-risk patients in community and primary-care settings where specialized retinal imaging is unavailable.