Large Vision-Language Models for Pediatric Motor Tic Detection: Case-Control Diagnostic Study

Clinical diagnosis of tic disorders (TDs) is often delayed by fluctuating symptoms 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, multisite tics characteristic of severe TDs.

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

In this multicenter study, 168 pediatric participants (n=83, 49.4% with TDs and n=85, 50.6% healthy controls) underwent 12-hour continuous video-electroencephalography monitoring. We fine-tuned the Qwen3-VL-8B model using 3- to 5-minute video segments via low-rank adaptation. The framework used chain-of-thought prompting to generate both binary classifications and textual justifications. A strict patient-wise split ensured zero identity overlap between training and test sets. Diagnostic precision was evaluated using 95% CIs calculated with the Clopper-Pearson exact method.

In total, 168 pediatric participants (age: mean 8.3, SD 2.3 years) were enrolled, providing 559 expert-validated video segments for analysis. On the independent test set (n=55), the framework achieved 89.1% (49/55; 95% CI 77.8%-95.9%) accuracy and an area under the curve of 0.950 (95% CI 0.892-1.000), significantly outperforming the landmark-based deep neural network (accuracy 74.1%; P=.004); recall reached 100% (25/25 positive episodes; 95% CI 86.3%-100.0%). Model-generated justifications showed substantial agreement with expert annotations (κ=0.784, 95% CI 0.625-0.943; P<.001).

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