Machine Learning based Predictions of Subjective Refractive Errors of the Human Eye

The aim of this research was to demonstrate the suitability of a data-driven approach to identify the sphero-cylindrical subjective refraction. An artificial deep learning network with two hidden layers was trained to predict power vector refraction (M, J(0) and J(45)) from 37 dimensional feature vectors (36 Zernike coefficients + pupil diameter) from a large database of 50,000 eyes. A smaller database of 460 eyes containing subjective and objective refraction from controlled experiment conditions was used to test for prediction power. Bland-Altmann analysis was performed, calculating the mean difference (eg Delta M) and the 95% confidence interval (CI) between predictions and subjective refraction. Using the machine learning approach, the accuracy (Delta M = +0.08D) and precision (CI for Delta M = +/- 0.78D) for the prediction of refractive error corrections was comparable to a conventional metric (Delta M = +0.11D +/- 0.89D) as well as the inter-examiner agreement between optometrists (Delta M = -0.05D +/- 0.63D). To conclude, the proposed deep learning network for the prediction of refractive error corrections showed its suitability to reliably predict subjective power vectors of refraction from objective wavefront data.