TIC-XNet: a structured evidence translation framework for interpretable multimodal pediatric tic event detection with...

This study aimed to develop an interpretable multimodal framework for detecting tic events in children with tic disorders by translating model decisions into structured, time-aligned evidence from synchronized video and physiological signals.

TIC-XNet was developed to jointly analyze synchronized video, heart rate, and electrodermal activity signals and to generate structured evidence outputs. Recordings from 417 children with clinically diagnosed tic disorders were collected during structured clinical assessments and home-based observations. TIC-XNet was compared with a prediction-only black-box model and a post-hoc explainable model under matched predictive backbones and identical training settings.

Within the evaluated internal subject-level split, TIC-XNet achieved the best performance on the pooled shared test set, with a window-level AUC of 0.915 ± 0.019, higher event-level recall and precision, fewer missed events, and lower post-buffering prediction latency than the comparator models. Its translated outputs also showed higher decision fidelity, greater stability under perturbation, and closer temporal alignment with expert-annotated tic onsets. Subject-level translated numerical signals were associated with tic severity.

These findings indicate that evidence translation can support more interpretable multimodal detection of tic events in children with tic disorders while remaining compatible with strong predictive performance.