Deep Learning for Automated Tic Detection and Prediction in Tourette Syndrome Using Electromyography and Video

Tourette syndrome is characterized by involuntary motor and vocal tics, which are challenging to monitor objectively. This study evaluated convolutional neural networks (CNNs) and Transformer models on EMG and video data ('Rest', 'Tic', and 'Voluntary Movement') for detection and prediction. We hypothesized that EMG windows <5s would better capture tics and improve detection and prediction accuracy. 685 EMG recordings (n = 16 participants) were each segmented into 1-, 5-, and 10-second windows. Each model was trained on clinical EMG and validated on held-out data, using corresponding expert-annotated video labels as ground truth. We evaluated model performance using the area under the receiver operating characteristic curve (AUROC) with 95% confidence. Detection performance exceeded prediction across models. The temporal CNN achieved the highest detection performance (mean AUROC 0.66 ± 0.05), with optimal results in 5s windows (AUROC up to 0.82 for tic detection). In contrast, tic prediction performance was lower (best 1s-AUROC 0.66 (0.65-0.68)). Contrary to our hypothesis, longer temporal windows (≥ 5s) improved tic detection, suggesting that tic-related EMG patterns benefit from extended temporal context, whereas prediction relied on short-term, 1-second signal dynamics. These findings help define a benchmark and suggest that future progress will depend on multimodal integration, patient-specific adjustments, and advanced temporal modeling.