An Artificial Intelligence Supervision Program to Improve Video-Observed Therapy for Managing Tuberculosis Treatment:...

Although video-observed therapy (VOT) boasts substantial potential advantages in tuberculosis (TB) treatment management, insufficient staffing of community healthcare workers leads to inefficient and inadequate video review. The aim of the study was to develop an artificial intelligence (AI)-powered program to assist community healthcare workers by reviewing medication-taking videos, and then validate the program's performance and effectiveness. We developed a video automatic review program (VARP) using a development dataset of 8000 medication-taking videos, which was partitioned at the participant level into 5781 videos for model training and 2219 videos for internal testing. We evaluated its performance versus community health workers based on TB specialists' annotations and compared clinical management metrics before and after VARP adoption to verify its effectiveness. VARP combined YOLOv8n-based medication detection with dual-stream RGB and pose-based medication action recognition; the RGB stream used a 3D ResNet-50 backbone, whereas the skeleton stream used a modified 3D ResNet backbone (F₁ = 0.90, 95% CI 0.89-0.91). In an independent post-deployment validation set of 7237 videos, VARP achieved significantly better accuracy (84.7%, 6133/7237 vs. 78.1%, 5653/7237) and specificity (72.7%, 1151/1584 vs. 0%, 0/1584) than community health workers (p < 0.001), with 75.3% (5449/7237) consistent judgments against TB specialists and vastly shorter video processing time (7.72 s vs. 23.12 h, p < 0.001). Compared with standard VOT management (n = 149), the VARP-supported group (n = 158) delivered faster daily monitoring completion (80.4%, 14,842/18,454 vs. 58.0%, 13,124/22,635 within 24 h, p < 0.001) and higher adverse event detection (67.7%, 107/158 vs. 35.6%, 53/149, p < 0.001). AI-powered review of TB medication videos optimizes key community VOT indicators such as review efficiency and assessment consistency, offering a practical, effective solution to address flaws in conventional manual VOT and strengthening routine TB monitoring.