“Machine Learning–Based Prediction of Postoperative Refraction in Cataract Surgery: A Stacking Ensemble Approach”

ABSTRACT Purpose To develop and evaluate a stacking ensemble machine learning model for predicting postoperative spherical equivalent (SEQ) refraction after cataract surgery and to compare its performance with established intraocular lens power calculation formulas. Design Retrospective dataset study using a publicly available, de-identified dataset. Subjects A total of 1,710 eyes from 1,710 patients who underwent cataract surgery with monofocal intraocular lens implantation (Vivinex or SA60AT). Methods Following preprocessing and feature engineering, we developed a stacking ensemble architecture comprising three base learners - Multi-Layer Perceptron (MLP), Support Vector Regressor (SVR) with a radial basis function (RBF) kernel, and Spline Transformer with Linear Regression - and a Ridge Regressor meta-learner. We trained the model on 80% of the data (n=1,368) using 5-fold cross-validation and evaluated it on an independent 20% test set (n=342). We compared performance against six standard IOL formulas: Barrett Universal II, Kane, Haigis, Hoffer Q, Holladay 1, and SRK/T. Main Outcome Measures Mean absolute error (MAE), median absolute error (MedAE), root mean squared absolute error (RMSAE), standard deviation (SD), and percentage of eyes achieving absolute prediction error within ±0.25 D, ±0.50 D, ±0.75 D, and ±1.00 D. Results The stacking ensemble model demonstrated excellent predictive accuracy, achieving a Mean Absolute Error (MAE) of 0.271 D on the independent test set (n=342). The model achieved the lowest MAE among all evaluated methods and demonstrated statistically significant improvement over Barrett Universal II (MAE 0.317 D, p=0.004), Haigis, Holladay 1, SRK/T, and Hoffer Q (all p