A Video-Based Artificial Intelligence Platform for Measuring Esophagogastroduodenoscopy Duration and Suggesting Real-Time Landmark Documentation

Abstract Background and Study Aims The European Society of Gastrointestinal Endoscopy (ESGE) sets key quality indicators for esophagogastroduodenoscopy (EGD), including a minimum procedural time of 7 min and photodocumentation of landmarks. We evaluated AI-derived procedural times, physician variations, and the impact of AI suggestions on landmark documentation rates. Methods This retrospective multicenter study included 78,336 EGDs performed by 213 physicians (01.01.2024–31.12.2025). A video-based AI platform (Theator Inc., Palo Alto, USA) measured procedural time, provided real-time landmark suggestions, and measured landmark documentation. Two blinded experts performed confirmatory validation of 500 procedures, with concordance defined as agreement within 2 s. Results Median procedural time was 3.30 min (IQR: 2.27–4.65). AI timestamps were concordant with expert review in 98.5% of cases (95% CI: 97.1–99.3%). Among 203 physicians performing at least 50 procedures, procedural time varied from 1.35 to 7.56 min (a 5.6-fold difference), with only 6.1% of procedures meeting the ESGE-recommended 7-min threshold. AI landmark suggestions reduced zero-landmark procedures by 42.6% (15.8 vs. 27.5%; P < 0.001), with improvements across all seven anatomical sites (P < 0.001 for each). Physicians in the slowest total EGD time quintile had fewer normal examinations than those in the fastest quintile: 19.8 versus 31.7% (P = 0.002), an association that persisted when physicians were restratified by biopsy-free procedural time. Conclusions Video-based AI accurately measured EGD procedure times and improved landmark documentation rate with AI-generated suggestions. A median procedural time of 3.3 min reveals a significant gap in quality standards. Automated, unattended measurement offers a scalable route to auditing upper endoscopy quality.