Machine learning-based prediction of persistent diabetic macular edema using OCT-derived structural biomarkers and clinical features

Persistent diabetic macular edema (DME) remains a leading cause of vision loss in diabetic retinopathy, even with anti-VEGF therapy. About 30–40% of patients respond poorly after standard loading doses, resulting in prolonged disease and increased treatment burden. Early identification of high-risk patients is essential for optimizing therapy and reducing unnecessary interventions. This study aimed to develop and compare machine learning models integrating optical coherence tomography (OCT)-derived biomarkers with clinical features to predict persistent DME after anti-VEGF therapy and identify key predictors. This retrospective study analyzed 339 patients with DME treated with anti-VEGF injections at Changhai Hospital between January 2018 and December 2022. Eight predictive models (M1–M8), including logistic regression, LASSO, XGBoost, random forest (RF), and support vector machine (SVM), were constructed using baseline clinical and OCT data. Model performance was assessed by AUC, accuracy, precision, recall, and F1-score. DeLong’s test was used for AUC comparison, decision curve analysis (DCA) for clinical utility, and SHAP values for model interpretability. Models integrating OCT biomarkers and clinical features outperformed single-modality models. SVM (M8), XGBoost (M6), and RF (M7) achieved the highest validation AUCs (0.865, 0.819, and 0.812, respectively), demonstrating strong discrimination in the internal validation cohort. DeLong’s test confirmed significant AUC improvements for multimodal and regularized models compared with clinical-only models (P < 0.05). SHAP analysis highlighted diabetes duration, ellipsoid zone (EZ)/external limiting membrane (ELM) disruption, and hyperreflective foci (HF) as key predictors. Machine learning models integrating OCT-derived structural biomarkers with clinical features provide superior predictive accuracy for persistent DME compared with conventional models. SHAP-based interpretation consistently identified diabetes duration, EZ/ELM disruption, and HF as key predictors. Among the developed models, SVM achieved the highest AUC in the internal validation cohort, whereas XGBoost and RF provided a favorable balance between predictive performance and interpretability, supporting their potential use in individualized treatment planning.