Myopia Prediction Using Machine Learning: An External Validation Study

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Machine learning (ML) models were previously developed to predict cycloplegic spherical equivalent refraction (SER) and myopia using non-cycloplegic data under standardized conditions.
• The generalizability of these ML models to diverse clinical settings with varying cycloplegia agents and biometry devices was uncertain.

Purpose of the Study:

• To evaluate the performance of ML models in predicting cycloplegic SER and myopia status in an independent cohort of Chinese students.
• To assess the models' robustness despite variations in cycloplegic agents (tropicamide, cyclopentolate) and biometry devices (IOLMaster 700, SW-9000).

Main Methods:

• An independent cohort of 614 Chinese students aged 8-13 years was studied.
• Autorefraction was performed before and after cycloplegia using either 0.5% tropicamide or 1% cyclopentolate.
• Biometric measures were collected using either an IOLMaster 700 or an Optical Biometer SW-9000.
• ML models (XGBoost, random forest) were evaluated using R², mean absolute error (MAE), sensitivity, specificity, and area under the ROC curve (AUC).

Main Results:

• The XGBoost model demonstrated excellent prediction of cycloplegic SER (R² = 0.95, MAE = 0.32 D).
• Both ML models accurately predicted myopia status (random forest: AUC 0.99; XGBoost: AUC not specified) and myopia prevalence (observed 62.9%; predicted 58.8%-60.6%).
• High sensitivity and specificity were achieved for myopia prediction by both models, even with heterogeneous cycloplegia and biometry data.

Conclusions:

• XGBoost and random forest ML models performed effectively in predicting cycloplegic SER and myopia status using non-cycloplegic data in an independent cohort.
• External validation confirmed that ML models can reliably estimate cycloplegic SER and myopia prevalence despite heterogeneous clinical parameters.
• Further studies in diverse populations are recommended to solidify the utility of ML tools for ophthalmic assessments.