ETD-XAI: An explainable two-stage EEG-biochemical framework for pediatric tic disorders

Tic disorders in children are mainly assessed by behavioral observation and rating scales, with limited interpretable physiological biomarkers. EEG-based modeling and vitamin D-calcium-phosphorus indicators may provide complementary information, but segment-level analysis limits subject-level interpretation. This study develops an interpretable, physiology-informed framework for pediatric tic disorder assessment.

We propose the explainable two-stage assessment framework with XAI (ETD-XAI) for pediatric tic disorder assessment that integrates subject-level EEG neurophysiological features with vitamin D-calcium-phosphorus metabolic indicators and clinical variables. The framework follows clinical decision logic by sequentially performing tic disorder screening and YGTSS-based severity grading, while using training-fold-restricted explainability for feature contribution estimation, modality-aware weighting, cross-modal interaction modeling, and post hoc interpretation.

ETD-XAI achieved feasible performance in both cross-validation and temporally independent validation. For tic disorder screening, AUROC was 0.92 ± 0.02 in five-fold cross-validation and 0.90 in independent validation. For severity grading, weighted AUROC was 0.90 ± 0.03 and 0.87, respectively. SHAP analysis identified β/γ-band EEG activity, complexity-related features, and serum 25-hydroxyvitamin D as the main contributors, with feature contributions showing severity-related gradients and age-stratified consistency.

ETD-XAI demonstrates the feasibility of interpretable multimodal modeling for subject-level, physiology-informed assessment of pediatric tic disorders. However, it should be regarded as a feasibility-oriented research framework rather than a validated diagnostic tool, and larger multicenter cohorts, external validation, and prospective clinical evaluation are required before any clinical use can be considered.