Deep Learning Based Prediction of Myopia Control Effect in Children Treated With Overnight Orthokeratology

Abstract

Objectives: To develop and validate a deep learning–based model for predicting 12-month axial length (AL) elongation using baseline factors and early corneal topographic changes in children treated with orthokeratology (Ortho-K) and to investigate the association between these factors and myopia control impact. Methods: A total of 115 patients with Ortho-K were enrolled. Influential baseline factors that have a statistically significant correlation with 12-month AL from medical records were selected using Pearson correlation coefficients. Simultaneously, the height, area, and volume of the defocus region were directly calculated from the corneal topography. Then, the prediction model was developed by combining multiple linear regression and deep neural network and evaluated in an independent group (83 patients for developing the algorithm and 32 patients for evaluation). Results: Age (r=−0.30, P<0.001), spherical equivalent refractive (SE; r=0.20, P=0.032), and sex (r=0.19, P=0.032) were significantly correlated with the AL elongation while pupil diameter, flat k, steep k, horizontal corneal diameter (white to white), anterior chamber depth, and cell density were not (P>0.1). The prediction model was developed using age, SE, and corneal topographic variation, and the validation of the model demonstrated its effectiveness in predicting AL elongation. Conclusions: The AL elongation was accurately predicted by the deep learning model, which effectively incorporated both baseline factors and corneal topographic variation.