Multicenter evaluation of AI-based intraocular lens formulas and interpretable machine learning for cataract refractive prediction

Abstract Accurate intraocular lens (IOL) power selection remains challenging in cataract eyes with long axial length, altered corneal geometry, or previous corneal refractive surgery. This multicenter retrospective study evaluated conventional, post-refractive, and AI-based IOL formulas and developed interpretable machine learning models for postoperative refractive prediction. The development cohort included 1203 eyes from West China Hospital; the external validation cohort included 336 high-myopia eyes from Chengdu Aidi Eye Hospital, with 257 eyes available for cleaned counterfactual target-achievement analysis. Formula and model performance were assessed using mean absolute error, root mean square error, median absolute error, and predefined refractive error thresholds. In eyes without prior corneal refractive surgery, AI-based formulas outperformed conventional formulas, with Hill-RBF showing the lowest internal prediction error. Locally developed linear regression, support vector regression, and random forest models achieved comparable internal accuracy, but external validation showed reduced portability. SHAP analysis identified target refraction, axial length, and IOL power as dominant predictors. Post-refractive eyes, very long eyes, and flat corneas remained difficult. These findings support cautious, locally validated use of AI-based formulas and interpretable machine learning in refractive cataract surgery.