Predicting atropine treatment response heterogeneity in pediatric myopia using explainable machine learning: an on-treatment longitudinal monitoring framework

To establish an explainable machine learning framework to monitor longitudinal refractive changes in children receiving atropine therapy and visualize individual treatment-response heterogeneity. This retrospective cohort included 1,545 eyes from pediatric patients treated with 0.1%–1.0% atropine at a tertiary center. Eight cumulative clinical and treatment-related features: baseline spherical equivalent (SE), intraocular pressure (IOP), IOP change, age, sex, treatment duration, cumulative dosage, and average monthly dosage were analyzed. Four machine learning models (Lasso, CART, XGBoost, Random Forest) were compared using 100 repeated runs, and the best-performing model was interpreted with SHapley Additive exPlanations (SHAP) to quantify global and individual feature contributions. Random Forest achieved the lowest average error metrics (RMSE 0.531 ± 0.030; MAE 0.388 ± 0.016). SHAP identified baseline SE as the most influential determinant of final refractive change, followed by followed by cumulative dosage and average monthly dosage. Negative SHAP contributions from higher doses captured clinician-driven titration patterns (confounding by indication) rather than a causal hazard. Case-level SHAP waterfall plots revealed successfully mapped response profiles, distinguishing effective stabilization in mild baseline myopia from pharmacologic saturation in high baseline myopia. Integrating Random Forest with SHAP delivers transparent, high-fidelity monitoring of atropine response in pediatric myopia. Rather than relying solely on static baseline predictors, this explainable framework leverages accumulated clinical process metrics to bridge data science and clinical reasoning. It provides a robust tool for longitudinal treatment tracking, retrospective clinical audit, and early risk stratification.