Predicting Visual Acuity after Retinal Vein Occlusion Anti-VEGF Treatment: Development and Validation of an...

Area of Science:

• Ophthalmology
• Medical Informatics
• Machine Learning

Background:

• Accurate prediction of visual acuity post-treatment for retinal vein occlusion with macular edema (RVO-ME) is crucial for effective anti-VEGF therapy.
• Current machine learning (ML) models for ophthalmic prognostication often lack interpretability and clinical applicability in RVO management.

Purpose of the Study:

• To develop and validate an interpretable ML model for predicting visual acuity changes in RVO patients undergoing anti-VEGF treatment.
• To systematically compare ML algorithms for RVO-ME visual acuity prediction and enhance clinical transparency.

Main Methods:

• Retrospective analysis of 259 RVO patients treated with anti-VEGF therapy.
• Feature selection using the Boruta algorithm and evaluation of eight ML algorithms, with Extreme Gradient Boosting (XGBoost) identified as optimal.
• Interpretability analysis using Shapley Additive exPlanations (SHAP) values.

Main Results:

• The XGBoost model achieved an AUC of 0.91, with 0.83 accuracy, 0.88 sensitivity, 0.73 specificity, and 0.87 F1 score.
• Key predictors included baseline visual acuity, systolic blood pressure, age, diabetic retinal inner layer dysfunction (DRIL), and RVO subtype.
• SHAP analysis highlighted baseline visual acuity, systolic blood pressure, and age as the most influential prognostic factors.

Conclusions:

• The developed interpretable ML model demonstrates high predictive performance (AUC > 0.9) for visual acuity outcomes in RVO-ME patients.
• The model's clinical transparency through SHAP visualization facilitates practical implementation in guiding anti-VEGF treatment decisions.
• Further validation in multicenter cohorts is recommended to enhance generalizability for personalized RVO management.