Artificial Intelligence-Based Prediction of Subjective Refraction and Clinical Determinants of Prediction Error

Background/Objectives: Subjective refraction is the clinical gold standard but is time-consuming and examiner-dependent. Most artificial intelligence (AI)-based approaches rely on specialized imaging or biometric data not routinely available. This study aimed to predict subjective refraction using only routine, non-cycloplegic autorefraction and keratometric data and to identify factors associated with reduced prediction accuracy. Methods: This retrospective study included 1856 eyes from 1006 patients. A multi-output histogram gradient-boosting model predicted subjective spherical equivalent, cylindrical power, and astigmatic axis. Performance was evaluated on an independent test dataset using R² and mean absolute error, with circular statistics for axis prediction. Prediction failure was assessed using clinically relevant tolerance thresholds (sphere/cylinder ≤ 0.50 D; axis ≤ 10°) and multivariable logistic regression. Results: The model achieved high accuracy for spherical and cylindrical prediction (R² = 0.987 and 0.933; MAE = 0.126 D and 0.137 D). Astigmatic axis prediction demonstrated strong circular agreement (ρ = 0.898), with a mean absolute angular error of 4.65° (median, 0.96°). Axis errors were higher in eyes with low cylinder magnitude (<0.75 D) and oblique astigmatism. In multivariable analysis, steeper keratometry (K2; OR = 7.25, 95% CI 1.62-32.46, p = 0.010) and greater objective cylindrical power (OR = 2.79, 95% CI 1.87-8.94, p = 0.032) were independently associated with poor prediction. Conclusions: A machine-learning model based solely on routine, non-cycloplegic autorefractor and keratometric measurements can accurately estimate subjective refraction, supporting AI as a complementary decision-support tool rather than a replacement for conventional subjective refraction.