Optimization of machine learning-based prediction models for myopia development in a long-term longitudinal cohort of Chinese children

Abstract

Purpose : To compare the predictive reliability and accuracy using different types of machine learning-based models for the individual myopia prediction in a long-term follow-up cohort of Chinese children.

Methods : Data of the first-born twins followed annually between 2006 and 2015 from the Guangzhou Twins Eye Study were analyzed. Five types of machine-learning algorithms, including a random forest model, a support vector machine (SVM), a gradient boosting decision tree (GBDT), a neural network and a multivariate adaptive regression splines (MARS) were conducted with bootstrap sampling and validation. The main outcomes were the R² and mean absolute error (MAE) in predicting spherical equivalent (SE) at last visit, and area under curve (AUC) of the presence of high myopia (SE<-6.0D)at age of 18 years. Refraction data and age at examination were used to develop the algorithms.

Results : A total of 1063 subjects were included, with a mean age at baseline of 10.5+2.2 years, and a prevalence of high myopia of 7.4% at last visit. For predicting refractive error at 1-year follow up, all the five models showed similar R²s (0.9446 to 0.9578), while for predicting refraction at 7 years from baseline, the random forest model showed a lower R² than the other four models (0.7163), as well as a greater MAE (1.1044). In SVM model, an R² of 0.7454 and a lowest MAE of 1.0161 at 7-year prediction were presented. The AUCs for predicting high myopia onset at age of 18 years was highest in MARS (0.9763) and lowest in random forest model (0.9699).

Conclusions : We propose that the random forest model was not the optimal machine-learning method in predicting long-term refraction development in children, while SVM or MARS showed high reliability and accuracy and may provide a more precise individual risk assessment.

This is a 2020 ARVO Annual Meeting abstract.