A machine learning approach to determine refractive errors of the eye

Abstract

Purpose : In order to account for the neural contributions to perception, refractive errors of the eye are traditionally assessed by a subjective measurement, or by using mechanistic models, like image quality metrics. The aim of this research was to explore the suitability of a machine learning approach to identify the perceptually optimal sphero-cylindrical refraction correction, when compared to subjective measurements.

Methods : The data of 460 eyes from 230 subjects with a mean age of 33.1 ± 11.5 years (range: 18 to 78 years) were included in the study. The subjective assessment of the non-cycloplegic refractive errors under monocular conditions were performed by two optometrists using a digital phoropter (ZEISS Visuphor 500, Carl Zeiss Vision GmbH, Germany) and SLOAN optotypes. Wavefront aberrations were approximated by Zernike polynomials up to the 7^th radial order (i.Profiler plus, Carl Zeiss Vision). A multilayer perceptron (MLP) with two hidden layers was trained to predict the sphero-cylindrical refraction correction from 37 dimensional feature vectors (36 Zernike coefficients + pupil diameter). All data was used for training and testing purposes via 10-fold cross validation. Bland Altman analysis was performed to investigate the mean differences and the 95% Limits of Agreement (LoA) between refractive components (M, J₀ and J₄₅) from MLP predictions and independent subjective refractions.

Results : Using the machine learning approach to predict the power vectors of refraction, the mean squared error (MSE) of the prediction (M: MSE = 0.17 D²; J₀ MSE = 0.03 D²; J₄₅ MSE = 0.03 D²) resided in a similar range as the resolution commonly used for subjective prescriptions (M: 0.14 D²; J₀ and J₄₅: 0.02 D²). 95% of the predictions were contained in the interval of ±0.83 D (M), ±0.35 D (J₀) and ±0.34 D (J₄₅) respectively. These intervals are comparable to those previously observed for both mechanistic models and subjective assessments of refraction. The bias with respect to subjective measurements is trivially close to zero for all three vector components (ΔM = 0.00 D, ΔJ₀ = 0.01 D and ΔJ₄₅ = 0.02 D).

Conclusions : The neural network based prediction of the refractive correction lead to comparable values of the power vectors of refraction, when compared to the subjective measurements.

This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.