Interpretable machine learning to predict functional visual outcomes after the anti-VEGF loading phase for macular...

Functional visual outcomes in macular edema (ME) secondary to retinal vein occlusion (RVO) after the anti-vascular endothelial growth factor (VEGF) loading phase are highly variable. We aimed to develop and validate an interpretable prediction model integrating baseline clinical features and quantitative optical coherence tomography (OCT) biomarkers to identify RVO-ME patients at high risk of poor functional visual outcomes.

This study retrospectively included 196 patients with RVO-ME. Adopting a rigorous temporal validation design, cases from 2021 to 2024 were assigned to the training set (n = 156), while those from 2025 served as an independent test set (n = 40). A least absolute shrinkage and selection operator (LASSO) regression was utilized to select key features from 15 candidate variables to facilitate the development and comparison of six ML models, such as XGBoost, random forest, and logistic regression. Finally, SHapley Additive exPlanations (SHAP) analysis was employed to provide a transparent interpretation of the decision-making logic for the optimal model. Poor functional visual outcomes were defined as post-loading BCVA < 0.5 decimal (>0.3 logMAR).

Initially, LASSO regression identified five key predictors: baseline BCVA, age, RVO subtype, subretinal fluid (SRF) cross-sectional area, and length of disorganization of retinal inner layers (DRIL). The XGBoost model was selected as the best-performing model, showing strong performance in the test set (AUC, 0.898; 95% CI, 0.797-1.000; sensitivity, 0.917; specificity, 0.812) at the default probability threshold of 0.5. SHAP analysis identified baseline BCVA as the primary prognostic driver and suggested a non-linear relationship with prognosis, with a critical risk threshold at approximately 1.14 logMAR (95% CI: 1.07-1.20). Additionally, a heterogeneous, non-linear interaction pattern was observed between advanced age and severe baseline visual impairment. SHAP analysis identified SRF cross-sectional area and DRIL length as independent contributors to model predictions after accounting for other included features.

XGBoost showed encouraging performance for predicting poor functional visual outcomes in patients with RVO-ME after the anti-VEGF loading phase and maintained high sensitivity in the independent temporal test cohort. Furthermore, SHAP analysis suggested a non-linear relationship between baseline BCVA and prognosis, with a breakpoint at approximately 1.14 logMAR, and highlighted the independent predictive value of quantitative DRIL and SRF in determining functional visual recovery. Overall, this study provides a proof-of-concept for interpretable prediction in RVO-ME and may help identify patients at higher risk of poor short-term functional outcomes.