Machine Learning–Based Prediction Of Lasik Console Inputs For Aspheric Planning (Q-Factor, Defocus, Astigmatism): A Translational Methods Study

To develop an artificial intelligence (AI)-based framework to determine optimal clinical console settings for aspheric excimer laser surgery. This addresses the noisy, nonlinear, high-dimensional mapping between programmed inputs and delivered corneal shape.

Given the weak, context-dependent cross-coupling between programmed Q-factor and delivered asphericity or induced defocus, we benchmark machine-learning regressors against traditional linear model to predict console-ready Q-factor, defocus, and astigmatism. We quantify out-of-sample accuracy and calibration while characterizing residual behavior (bias, dispersion, tail risk) to establish the analytic validity required prior to prospective clinical evaluation.

Retrospective translational analysis of anonymized LASIK treatments from a single center and a single surgeon in France. Derivation cohort (July 2017–Sept 2023): TENEO 317 excimer aspheric ablations with FEMTO LDV Z2 femtosecond flap creation. External validation cohort (Sept 2023–Aug 2024): temporal/device shift using a second TENEO 317 unit under matched protocols. Supracor and aberrometry-guided treatments were excluded.

Treatment planning was cast as an inverse-identification multi-output regression. To bypass quantization-induced error floors, targets were resolved at 0.01 D/1°. Z-standardized features underwent dimensionality reduction (L1-Lasso, Pillai’s trace), optimized via hyperparameter-tuned forward nested cross-validation to constrain degrees of freedom. An orthonormal $2\times 2$ Hadamard rotation was applied to 4D Zernike-basis power-vector aberrometric states to attenuate common-mode noise. Benchmarks included Bayesian, Gradient Boosted Bagging, Vanilla MultiLayer Perceptron, and Linear model. Analytic validity was interrogated through standardized heteroscedastic tail-risk diagnostics and higher-order residual moments (skewness, kurtosis).

Derivation included 2,448 eyes; external validation 147 eyes (temporal/device shift). Aberrometer noise floor for defocus was Sw​=0.15 D. Programmed defocus linearly predicted outcome (R2=0.97), but Q-factor explained only 51% of spherical aberration change and weakly predicted defocus cross-coupling (R2=0.12).

The Bayesian Dynamic Multi-head (BDMH) network, integrating latent-variable representations and heteroscedastic noise modeling, performed best; on external validation, the Mean Absolute Error (MAE) is 0.22 D for defocus (R2=0.98) and 0.21 for Q-factor (R2=0.81). The framework yielded consistent accuracy on a separate hardware unit, with predictive mean error (0.22 D) remaining near the instrument’s noise floor (0.15 D).

This proof-of-concept establishes the analytic validity of AI for refractive planning, suggesting a potential framework to model the nonlinear reciprocity between asphericity and defocus. Continuous targets mitigate quantization-induced information deficits (0.25 D, 5° step). Despite retrospective limits, joint optimization in a structured latent space captures complex optical couplings inaccessible to nomograms.

Future scaling should leverage federated learning (privacy) and foundation models (generalization), pivoting evaluation from point-performance toward epistemic robustness via conformal prediction. This ensures safety through human-in-the-loop evaluation and structural causal modeling to formalize the actuator-response map.