Abstract
BACKGROUND AND AIMS: Artificial intelligence (AI) may detect endoscopically active ulcerative colitis (UC) and streamline clinical trial enrollment. Here, we investigate the performance of a commercially available AI model, autoinflammatory bowel disease (AutoIBD)-UC (Virgo Surgical Video Solutions, Inc, San Francisco, Calif, USA), to detect moderate-to-severe UC in colonoscopy videos collected as part of routine clinical practice.
METHODS: AutoIBD-UC was applied to consecutive endoscopy videos between August 30, 2024, and October 17, 2024. Each video received an AutoIBD score (0-1); videos with scores above a predefined confidence threshold were flagged as possible moderate-to-severe UC. Flagged videos and medical records were reviewed to confirm UC diagnosis and grade Mayo endoscopic subscore (MES). Concurrently, research staff retrospectively reviewed all endoscopy reports to identify 3 cohorts: (1) 20 MES ≥2, (2) 20 non-UC inflammation, and (3) 20 screening colonoscopies. Sensitivity and specificity were calculated for all videos screened by AI. Median AutoIBD scores were compared by MES and disease extent.
RESULTS: AutoIBD-UC was applied to 2273 endoscopy videos. Twenty-one videos were flagged, 10 of which had MES ≥2. AutoIBD-UC flagged 9 of 20 videos in cohort 1 and all MES 3 videos. More procedures with extensive (4 of 8) or left-sided disease (5 of 7) were flagged compared with those with proctitis (0 of 5). Median AutoIBD-UC scores differed between MES 2 (0.437; interquartile range [IQR], 0.270-0.571) and MES 3 (0.690; IQR, 0.607-0.765; P = .006), and by disease extent: proctitis (0.224; IQR, 0.193-0.270), left-sided (0.571; IQR, 0.554-0.603), and extensive (0.506; IQR, 0.408-0.598; P = .021). The sensitivity of AutoIBD-UC was 0.47, and the specificity was 0.99.
CONCLUSIONS: AutoIBD-UC accurately detects moderate-to-severe UC with moderate sensitivity and high specificity. Larger studies are necessary to evaluate AutoIBD-UC's utility as a recruitment aid compared to standard practice.