JMIR Publications

ABSTRACT

Background:

Clinical diagnosis of tic disorders (TD) is often delayed by the fluctuations of symptom and the subjective nature of assessments. While automated methods have been proposed to bridge this gap, traditional machine learning models frequently lack robustness in unconstrained environments and fail to capture the complex, multi-site tics characteristic of severe TD.

Objective:

This study aimed to develop and evaluate a fine-tuned large vision-language model (LVLM) framework for the automated, interpretable detection of pediatric motor tics, addressing the clinical need for objective and efficient symptom assessment in unconstrained environments.

Methods:

A cohort of 168 pediatric participants (83 with tic disorders and 85 healthy controls) was enrolled for 12-hour continuous vEEG monitoring. We fine-tuned the Qwen3-VL model using 3- to 5-minute video segments. The framework utilized Chain-of-Thought (CoT) prompting to generate both binary classifications and textual diagnostic justifications. A strict patient-wise split strategy was employed to ensure zero identity overlap between the training and test sets.

Results:

A total of 168 pediatric participants (83 with TD and 85 healthy controls; mean age 8.3 ± 2.3 years) were enrolled, providing 559 expert-validated video segments for analysis. The framework achieved 89.1% accuracy and an area under the curve (AUC) of 0.950. On the independent test set, recall was 100% (25/25 positive clips), although this estimate should be interpreted in light of the limited sample size. Model-generated justifications showed substantial agreement with expert annotations (κ = 0.84). Performance was maintained across the ward recordings collected under routine monitoring conditions.

Conclusions:

These findings suggest that a vision-language framework can capture clinically relevant tic features beyond conventional geometric descriptors. By combining high sensitivity with text-based diagnostic justifications aligned with clinical severity dimensions (e.g., frequency and intensity), the approach may support automated video review and preliminary symptom quantification in controlled monitoring settings.