Introduction: To develop and externally validate machine learning (ML) models for predicting treatment response and adverse events in children with attention-deficit/hyperactivity disorder (ADHD) receiving acupuncture combined with conventional first-line care. Methods: In this multicentre retrospective study, demographic, clinical, behavioral, and laboratory data from 809 children with ADHD aged 6–12 years were used to develop and evaluate ten ML models. Treatment response was defined as a ≥30% reduction in the Swanson, Nolan and Pelham–IV total score 8–16 weeks after treatment initiation, and adverse events were recorded from parent reports. Model development used repeated 10-fold cross-validation, with evaluation in an internal testing cohort and two external cohorts. Discrimination, calibration, and clinical utility were assessed, and model interpretability was examined using SHapley Additive exPlanations. Results: The extreme gradient boosting model demonstrated the best discriminative performance for predicting treatment response and was retained for further analyses. For treatment response prediction, it achieved an AUROC of 0.83 in the internal testing cohort and showed stable external performance, with AUROCs ranging from 0.78 to 0.87 across both external cohorts. For adverse event prediction, discriminative performance was moderate, with external AUROCs ranging from 0.73 to 0.85. Model interpretability analyses suggested that multiple clinical feature domains contributed to model predictions. Conclusion: Using routinely collected clinical data, ML-based models showed reproducible discrimination for predicting treatment response and adverse events in children with ADHD receiving acupuncture as adjunctive therapy. They may support individualized benefit–risk stratification in integrative treatment settings, although prospective assessment of clinical impact is warranted.
Li et al. (Mon,) studied this question.