SCH: Computer Vision Algorithms to Detect Tics In Patients with Tourette Syndrome

University Of Pennsylvania, Philadelphia PA

Investigators

Linked publications & trials

Abstract

Tourette Syndrome (TS) onsets in childhood, affects 1% of the population, and causes substantial impairment. Health professionals recommend behavior therapy as the first-line treatment for TS due to its efficacy and adverse effect profile. In behavior therapy, patients learn tic management skills and are assigned “homework” to solidify skill learning. Core skills involve building awareness to tic occurrence and implementing behavioral strategies to inhibit tic expression upon such awareness. Youth who exhibit a treatment response to behavior therapy continue to benefit for 10 years. However, more than 50% of patients do not achieve a treatment response and rely upon FDA-approved medications that have detrimental health effects. A key challenge with behavior therapy is the reliance on a human practice partner for “homework” due to accessibility and accuracy. Our team will create an activity-based recognition system and algorithms to identify and classify tics across activities and sensor viewpoints in patients with TS. This system will evolve into a therapeutic tool (i.e., a “digital practice partner”) that is scalable, accessible, and accurate in detecting tics. This will enable patients to effectively practice behavior therapy skills and achieve optimal long-term outcomes. First, we will develop a multi-view sensor system to observe tics in patients with TS and refine our hierarchical ontology of tics to annotate data collected from our sensor system. This will enable both a clinically interpretable and fine-grained classification of tic and non-tic movements. Second, we will design a novel spatio-temporal CNN called Tic-Net for fine-grained detection of facial and upper body tics in video data from multiple viewpoints, which will rely on facial action unit intensities and interpretable upper body part features that we design, temporal segmentation and detection networks, as well as contrastive and self-supervised learning losses to detect tics without requiring large amounts of annotations. Third, we will design a novel spatio-temporal Transformer architecture called Tic-DETR for fine-grained tic detection, which captures long-range interactions among face action units and/or skeletal joints across multiple views as well as relations between tic instances to produce interpretable detections of tics of varying durations from multiple viewpoints. Finally, we will compare detection outcomes between our algorithm and a human practice partner, evaluate the robustness of algorithms across viewpoints, and assess its clinical interpretability.