PURPOSE: To evaluate the predictive performance of traditional intraocular lens (IOL) power calculation formulas (e.g., SRK/T, Haigis, Hoffer Q, and Holladay I) compared to advanced regression models, including classical linear models, regression splines, and random forest regression, in predicting postoperative refraction following cataract surgery. DESIGN: Retrospective, comparative analysis of IOL power calculations. SUBJECTS: The study included 886 eyes from 631 patients who underwent cataract surgery with monofocal aspherical IOL implantation. METHODS: Biometric measurements were obtained using optical biometry (IOLMaster 700), and postoperative refraction was assessed at least 4 weeks after surgery. Formula constants for 5 IOL formulas (SRK/T, Haigis, Hoffer Q, Holladay I and Castrop V1) were optimized using root mean squared error (RMSE). Regression models (classical linear model, regression splines, and random forest regression) were trained on 4 datasets categorized by axial length (AL)
normal, short, long, and random. Model performance was assessed using mean absolute error (MAE), RMSE, and prediction error variance, for both in-sample and out-of-sample predictions. MAIN OUTCOME MEASURES: The primary parameters measured were MAE, RMSE, and prediction error variance. RESULTS: Regression models outperformed traditional IOL formulas in in-sample prediction error. Overall, linear regression models performed similarly to traditional formulas with respect to out-of-sample prediction error. The lowest out-of-sample prediction error (MAE = 0.279, RMSE = 0.359) was achieved with a model where effects of some covariates (R2, AL, CCT) were modelled as nonlinear via regression splines. This model outperformed all traditional formulas, and the Castrop formula, which had the lowest errors among the formulas (MAE = 0.284, RMSE = 0.359). Random forest regression showed strong in-sample performance but poor out-of-sample generalizability due to overfitting. CONCLUSIONS: Regression models which allow for nonlinear effects, e.g. based on regression splines, provide a promising alternative to traditional IOL formulas for predicting postoperative refraction. Linear regression and random forest regression models can reduce in-sample error, however, their clinical utility is currently limited by out-of-sample performance. Future work should focus on improving generalizability and integrating machine learning models into clinical practice to enhance refractive outcomes, especially for eyes with atypical anatomy.