Preoperative prediction of intraoperative complications in cataract surgery: a machine learning approach

Intraoperative complications during cataract surgery, while uncommon, can significantly impact visual outcomes and identifying high-risk patients preoperatively improves surgical planning and supports personalized care. This study presents an early application of advanced machine learning (ML) models for predicting multiple intraoperative complications using routine preoperative data.

We analyzed 3,957 cataract surgeries performed at a tertiary hospital and developed four ML models, XGBoost, Random Forest, CatBoost and TabPFN, to classify cases into three categories: no complications, posterior capsule rupture/vitreous loss and other complications. To address class imbalance due to the predominance of uncomplicated cases, random undersampling was applied during training. Model performance was assessed using area under the curve (AUC) and F1-score. SHAP (SHapley Additive exPlanations) analysis was used to interpret feature importance.

TabPFN achieved the highest AUC score (0.62), followed closely by Random Forest and CatBoost (0.61 each), with Random Forest also demonstrating the highest F1-score (0.46). SHAP analysis highlighted age, tamsulosin use and surgeon experience as key predictors.

This study provides a proof of concept for the feasibility of using machine learning for preoperative risk modeling of intraoperative complications in cataract surgery. While predictive performance was modest, validation on larger or external datasets is encouraged to further assess its potential role in surgical planning.