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

• Published in: Journal of optometry 2022, Vol.15 (Suppl 1), p.S22-S31
• Main Authors: Hernández, Carlos S., Gil, Andrea, Casares, Ignacio, Poderoso, Jesús, Wehse, Alec, Dave, Shivang R., Lim, Daryl, Sánchez-Montañés, Manuel, Lage, Eduardo
• Format: Article
• Language: English
• Subjects: Artificial Intelligence; Machine learning; Portable autorefractor; Subjective refraction; Wavefront aberrometry
• ISSN: 1888-4296; 1989-1342
• DOI: 10.1016/j.optom.2022.03.001

To assess the performance of machine learning (ML) ensemble models for predicting patient subjective refraction (SR) using demographic factors, wavefront aberrometry data, and measurement quality related metrics taken with a low-cost portable autorefractor. Four ensemble models were evaluated for predicting individual power vectors (M, J0, and J45) corresponding to the eyeglass prescription of each patient. Those models were random forest regressor (RF), gradient boosting regressor (GB), extreme gradient boosting regressor (XGB), and a custom assembly model (ASB) that averages the first three models. Algorithms were trained on a dataset of 1244 samples and the predictive power was evaluated with 518 unseen samples. Variables used for the prediction were age, gender, Zernike coefficients up to 5th order, and pupil related metrics provided by the autorefractor. Agreement with SR was measured using Bland-Altman analysis, overall prediction error, and percentage of agreement between the ML predictions and subjective refractions for different thresholds (0.25 D, 0.5 D). All models considerably outperformed the predictions from the autorefractor, while ASB obtained the best results. The accuracy of the predictions for each individual power vector component was substantially improved resulting in a ± 0.63 D, ±0.14D, and ±0.08 D reduction in the 95% limits of agreement of the error distribution for M, J0, and J45, respectively. The wavefront-aberrometry related variables had the biggest impact on the prediction, while demographic and measurement quality-related features showed a heterogeneous but consistent predictive value. These results suggest that ML is effective for improving precision in predicting patient's SR from objective measurements taken with a low-cost portable device.