A refractive error prediction model for children and adolescents based on ocular biometric parameters

TL;DR

This study creates a deep learning model to predict refractive errors in children and adolescents using ocular biometric data.

Contribution

A novel deep learning model using LSTM and T-LSTM for predicting refractive error progression in children.

Findings

The model achieved low prediction errors, especially in older children with fewer measurements.

Temporal deterioration of refractive errors was observed over 12 quarters.

Prediction accuracy improved with additional biometric measurements.

Abstract

To develop a predictive model based on ocular biometric parameters and deep learning for studying the progression of refractive errors in children and adolescents.

This longitudinal observational cohort study included 559 children and adolescents (1,118eyes; 252males, 307females) aged 5–18 years, enrolled at Shanxi provincial Eye Hospital (Taiyuan, China) between 2019 and 2023. Refractive error and ocular biometric parameters were prospectively assessed through serial non-cycloplegic automated refraction and LENSAR 900 biometry, with participants undergoing 2–5 follow-up evaluations at irregular intervals. After data cleaning, preprocessing, and augmentation through truncation techniques, Long Short-Term Memory (LSTM) and Time-aware LSTM (T-LSTM) models were used to predict refractive changes over the next three years. Mean Absolute Error (MAE) and Standard Deviation (SD) were used to evaluate model performance, reported as MAE ± SD.

Baseline MAEs measured 0.40 ± 0.39D (sphere) and 0.30 ± 0.46D (cylinder), exhibiting temporal deterioration over 12 quarters (sphere: 0.35 ± 0.42D→0.52 ± 0.37D; cylinder: 0.20 ± 0.21D→0.33 ± 0.30D).Additional biometric measurements reduced errors across cohorts: with five measurements ≤ 6y achieved 0.11 ± 0.10D (sphere)/0.09 ± 0.10D (cylinder); 10-12y />12y attained 0.18 ± 0.13D/0.10 ± 0.08D. The steepest error reduction occurred in older cohorts (10-12y/>12y), suggesting that fewer measurements achieve lower prediction errors with increasing age.

The deep learning model developed in this study, based on ocular biometric parameters, demonstrated high accuracy and stability in refractive error prediction.

Figures