(PS12-98) Developing and Evaluating a Novel Video-based Tic Detection Algorithm for Tourette’s Syndrome

Comprehensive behavioral therapy for tics (CBIT) is the frontline treatment for Tourette’s Syndrome (TS). Two core components of CBIT include awareness training (i.e., building awareness around tic occurrence) and competing response training (i.e., countering tic expression using behavioral strategies). Our research shows that adherence to implementing these two skills is associated with a positive treatment response to CBIT (Essoe et al., 2021), which is maintained for up to 10 years after treatment concludes (Espil et al., 2021). However, many patients struggle with tic awareness exercises outside of formal sessions due to the limited availability of practice partners. In response, we combined computer vision and machine learning methodologies to develop a video-based tic detection algorithm. Here, we examined the average precision (AP) of the tic detection algorithm across patients’ clinical characteristics and explored whether specific activities influence the AP of the algorithm. 

Fifteen individuals with TS (M_age = 30.40 years) completed a battery of clinical assessments and three different 15-minute activities that elicited varying tic frequencies (High, Low, Baseline) across four sequential visits (180 minutes per participant). Activities were video recorded and precisely annotated to differentiate tics from non-tic movements. Given the common nature of facial tics across all patients with TS (McGuire et al., 2013), our approach to training and testing the algorithm focused on facial tics and used Leave One Subject Out (LOSO) methodology. This generated AP values of tic detection for each activity across the four visits. Correlations, t-tests, and a repeated measure analysis of variance (RM-ANOVA) examined the relationship between AP and clinical characteristics, demographic factors, and tic-activities (High, Low, Baseline).  

The RM-ANOVA found no significant effect for visit (Visits 1-4; p > .05) or activity (High, Low, Baseline; p > .05), which demonstrates that the tic detection algorithm performed comparably across all visits and activities. When examining the effect of clinical characteristics on tic detection algorithm performance, global severity (CGI-S; r = -0.53 to -0.67, p < .05) was significantly correlated with algorithm AP; however, tic severity across other clinician and self-/parent-report scales did not (p > .05). While t-tests found large between group effects between AP and depression, comorbid conditions (e.g., depression, OCD, ADHD, and anxiety; d = -0.49 to -1.07) did not achieve statistical significance. Similarly, both age and biological sex did not significantly correspond with AP (p > .05).

Findings highlight the algorithm’s ability to detect tic occurrence—across time, activities, and patients—a critical initial step to helping patients achieve greater tic awareness in CBIT. Given the difference in associations for AP between global severity versus tic severity, further research is needed to explicate these findings and test algorithm accuracy in other body regions.