Intraoperative Oct-Based Prediction Of 3-Month Icl Lens Separation Vault Using Machine Learning

To train and validate machine learning models to predict 3-month lens separation vault after Implantable Collamer Lens (ICL) implantation, using iOCT- measured intraoperative vault as the key predictor alongside preoperative biometric parameters.

Private Practice Academic Unit, Reinstein Vision, London Vision Clinic, EuroEyes Group, UK.

Retrospective analysis of 399 eyes undergoing V4c EVO/EVO+ ICL implantation. Sixteen variables were evaluated, centred on iOCT-measured intraoperative vault (ARTEVO 800), with preoperative anterior segment metrics from VHF digital ultrasound (Insight 100) and AS-OCT (MS-39). Seven models were trained using an 80/20 train–test split and tuned with 10-fold cross-validation: gradient boosting machine, k-nearest neighbours, regularised linear regression (LASSO/elastic net), polynomial regression, random forest, stacking ensemble, and support vector regression. Performance on the test set (n=80) was assessed using mean absolute error (MAE) and R² between achieved and predicted 3-month vault, with reporting of eyes within ±100 µm and ±300 µm.

Mean central vault was 829±330 µm intraoperatively and 594±218 µm at 3 months. The best-performing models were support vector regression (MAE=109 µm, R²=0.54) and stacking ensemble (MAE=111 µm, R²=0.56). Support vector regression and stacking ensemble predicted the 3-month vault within ±100 µm of the achieved vault for 60% and 56% of eyes, and within ±300 µm for 94% and 95% of eyes, respectively.

Machine learning models incorporating iOCT-measured intraoperative vault achieved clinically useful accuracy for predicting 3-month postoperative vault, supporting intraoperative assessment of lens size suitability and offering a pathway to reduce vault-related risk through informed on-table decision-making.