Prediction of manifest refraction using machine learning ensemble models on wavefront aberrometry data

Area of Science:

• Ophthalmology
• Biomedical Engineering
• Data Science

Background:

• Accurate subjective refraction (SR) is crucial for optimal vision correction.
• Traditional autorefractors have limitations in predicting precise eyeglass prescriptions.
• Low-cost portable autorefractors offer accessibility but require enhanced predictive capabilities.

Purpose of the Study:

• To evaluate machine learning (ML) ensemble models for predicting patient subjective refraction (SR).
• To utilize demographic factors, wavefront aberrometry data, and measurement quality metrics from a portable autorefractor.
• To assess the performance of ML models in improving prediction accuracy compared to autorefractor outputs.

Main Methods:

• Four ensemble ML models (Random Forest, Gradient Boosting, XGBoost, and an Assembly model) were trained and tested.
• Models predicted individual power vectors (M, J0, J45) using age, gender, Zernike coefficients, and pupil metrics.
• Performance was evaluated using Bland-Altman analysis and prediction error percentages against SR.

Main Results:

• All ML models significantly outperformed the autorefractor's predictions.
• The custom Assembly model achieved the best predictive performance.
• Substantial reductions in error were observed for M (±0.63 D), J0 (±0.14 D), and J45 (±0.08 D).

Conclusions:

• ML ensemble models effectively enhance the precision of subjective refraction prediction.
• Wavefront aberrometry data proved most impactful for prediction accuracy.
• ML offers a viable solution for improving objective measurements from low-cost portable devices.