Machine Learning-Based Prediction of LASIK Console Inputs for Aspheric Planning (Q-factor, Defocus, Astigmatism): A Translational Methods Study.

Abstract

Purpose: To frame aspheric laser refractive planning as the supervised prediction of console-programmable inputs (Defocus, Astigmatism, and Q-factor) and to benchmark competing regression models; this is a translational methods proof-of-concept, not a clinical effectiveness study.

Methods: An anonymized, retrospective, single-platform dataset of 2,448 complete-case treatments was analyzed. Multi-output regressors (linear and nonlinear) were trained and compared using prespecified metrics (R², MAE/MSE) and residual-distribution visualization/calibration. Actuator-response checks related programmed inputs to changes in Defocus (Z₂⁰) and primary spherical aberration (Z₄⁰). External validation used a temporally later, device-shift cohort (n = 147).

Results: Linear regression predicted Defocus and astigmatism well (eg, Defocus R² = 0.98) but degraded for Q-factor (R² = 0.47), whereas nonlinear models improved Q-factor error and calibration. Actuator-response analyses showed strong coupling for Defocus input (R² = 0.97), moderate coupling of Q-factor to ΔZ₄⁰ (R² = 0.51), and a weak Q→Defocus cross-effect (R² = 0.12). On external validation, the best model generalized: Defocus MAE 0.22 D (R² = 0.98) and Q-factor MAE 0.21 (R² = 0.81).

Conclusions: Supervised nonlinear multi-output models achieve lower error and better calibration for Q-factor than linear baselines, supporting a metric-driven pathway toward more reliable control of low-order refractive targets and primary asphericity. Potential clinical implications include tissue sparing, improved contrast, and near-vision gains. Prospective, human-in-the-loop evaluation with safety and patient-reported endpoints is warranted.