AI-Driven Personalized Predictions and Interventions for Myopia Progression
Overview
This study developed a Transformer-based Myopia Progression Predictive Model (MPPM) to accurately forecast myopia progression and individualized treatment effects over a 10-year period in children. The model integrates natural progression and intervention modules trained on large longitudinal datasets, enabling precise predictions of spherical equivalent and axial length changes under various myopia control treatments.
Background
Myopia is a significant global health concern, especially prevalent in Asia, with high myopia increasing risks of severe ocular complications. Current myopia control interventions such as low-dose atropine, orthokeratology, peripheral defocus spectacles, and repeated low-level red light therapy have limitations including side effects, costs, and accessibility issues. Accurate long-term prediction of myopia progression and individualized treatment benefit assessment remain unmet clinical needs. Transformer-based AI models offer a promising approach to address these challenges by modeling long-term refractive changes and treatment responses.
Data Highlights
Dataset
Cohorts
Use
Measurements
WMU Pediatric Myopia Dataset
Non-intervention, Atropine, Ortho-K, PDS, RLRL
Train/Test NPM and IPM
SE (all visits), AL (partial, imputed)
DCH Pediatric Myopia Dataset
Non-intervention cohort
External validation NPM
SE, AL (partial, imputed)
RLRL Investigator Initiated Trial (IIT)
Control and intervention arms
External validation NPM and IPM
SE, AL
Key Findings
The Natural Progression Module (NPM) accurately predicts 10-year myopia progression in untreated children using sex, age, and prior SE and AL data.
The Intervention Progression Module (IPM) forecasts individualized treatment effects for Atropine, Ortho-K, PDS, and RLRL therapies by incorporating causal machine learning techniques to reduce confounding bias.
Machine-learning-based imputation effectively reconstructed missing axial length measurements, enhancing model training and prediction accuracy.
External validation on independent datasets confirmed the robustness and generalizability of both NPM and IPM modules.
The Transformer architecture enabled modeling of long-range temporal dependencies critical for long-term myopia progression prediction.
Clinical Implications
The MPPM provides clinicians with a powerful tool to predict individual myopia progression trajectories and quantitatively estimate treatment benefits, facilitating personalized intervention planning. This approach can optimize resource allocation, minimize unnecessary treatment exposure, and potentially improve long-term visual outcomes in pediatric myopia management.
Conclusion
The Transformer-based MPPM represents a significant advancement in personalized myopia care by enabling accurate long-term progression predictions and individualized treatment effect estimations. Its application may enhance early intervention strategies and reduce the burden of high myopia-related complications.
References
Wenzhou Medical University Dataset and Study Design
Dazhou Central Hospital Pediatric Myopia Dataset
Investigator Initiated Trial of Repeated Low-Intensity Red Light Therapy (ChiCTR2200066365)
