Axial and choroidal changes as key predictors of myopia control in red-light therapy: Evidence from machine learning models.

BACKGROUND

Repeated low-level red-light (RLRL) therapy has emerged as a promising non-invasive intervention for myopia control. However, the predictive factors underlying its efficacy remain insufficiently explored.

METHODS

This multicenter cohort study included 538 pediatric patients who underwent RLRL treatment with a minimum follow-up of one year. Baseline ocular parameters and dynamic changes in axial length (AL) and choroidal thickness (ChT) over three months were collected. Multiple feature selection approaches-LASSO regression, Boruta, recursive elimination, and multivariate regression-were applied. Seven machine learning algorithms were trained, and their performance was evaluated using the area under the curve (AUC), sensitivity, specificity, and F1 score. SHAP and LIME analyses were utilized for interpretability.

RESULTS

Logistic regression and gradient boosting models demonstrated the highest discriminative ability. XGBoost achieved optimal performance (AUC: 0.90-0.92; accuracy: 87.7-88.2 %; F1 score: 0.72-0.86). Across one- and two-year prediction models, six stable predictors were identified: age, baseline AL, anterior chamber depth (ACD), RNFL thickness at the temporal inferior quadrant (TI), ChangeAL, and ChangeChT. SHAP analysis revealed that ChangeAL was the dominant short-term predictor, whereas ChangeChT was most influential for long-term outcomes. Older age and greater choroidal thickening were consistently associated with a protective effect against myopia progression.

CONCLUSIONS

We developed and validated an interpretable machine learning model that accurately predicts short- and long-term outcomes of RLRL therapy in children. ChangeAL and ChangeChT serve as key dynamic biomarkers for treatment monitoring. These findings provide a foundation for personalized clinical decision-making in myopia management.