Ensemble machine learning prediction model for clinical refraction using partial interferometry measurements in childhood

Purpose: To develop an ensemble machine learning prediction model for clinical refraction in childhood using partial interferometry measurements. Methods: Age, sex, cycloplegic refraction, and partial interferometry data collected within one month were obtained from patients aged 5–16 years, retrospectively. Four ensemble regression models were used to develop prediction models of spherical equivalents (SE) from the collected data. Root mean squared error (RMSE) was used to compare the accuracy among the models. The accuracy of the ensemble models was compared with that of a previously developed multiple linear regression model. Results: 4156 eyes from 1965 patients (50.3% female) were included. Mean age was 8.4 ± 2.3 years and mean SE was −1.01 ± 2.94 diopters. Mean axial length was 23.63 ± 1.41 mm and mean keratometry reading of flat and steep axis was 43.58 ± 1.40 diopters. Developed ensemble models had accuracy of RMSE 0.800 to 0.829 diopters, which was superior to that of the conventional regression model (1.213 diopters). Simulations with the same biometric parameters showed that female sex was associated more with myopia than that of male sex. Long eyes showed dampened increase in the myopic refraction per unit axial length. Conclusions: Refractive errors can be calculated in the childhood using these ensemble models with ocular biometric parameters. Moreover, the models were able to simulate hypothetical relationships between ocular parameters and SE to understand the nature of clinical refraction.