Abstract
Background and Study Aim: Colonoscopy holds the highest volume of all endoscopic procedures, allowing for large colonoscopy databases to serve as valuable datasets for quality assurance. We aimed to build a comprehensive colonoscopy database for quality assurance and the training of future AIs. Materials and Methods: As part of a cluster-randomized controlled trial, a designated, onsite medical student was used to acquire procedural and patient-specific data, ensuring a high level of data integrity. The following data were thereby collected for all colonoscopies: full colonoscopy vides, colonoscope position (XYZ-coordinates), intraprocedural timestamps, pathological report, endoscopist description, endoscopist planning, and patient-reported discomfort. Results: A total of 1447 patients were included from the 1st of February 2022 to the 21st of November 2023; 1191 colonoscopies were registered as completed, 88 were stopped due to inadequate bowel cleansing, and 41 were stopped due to patient discomfort. Of the 1191 completed colonoscopies, 601 contained polypectomies (50.4%), and 590 did not (49.6%). Comparing colonoscopies with polypectomies to those without the withdrawal time (caecum to extubating the scope) was significantly longer for all parts of the colon (p values < 0.001), except the transverse colon (p value = 0.92). The database was used to train an AI, automatically and objectively evaluating bowel preparation. Conclusions: We established the most thorough database in colonoscopy with previously inaccessible information, indicating that the transverse colon differs from the other parts of the colon in terms of withdrawal time for procedures with polypectomies. To further explore these findings and reach the full potential of the database, an AI evaluating bowel preparation was developed. Several research partners have been identified to collaborate in the development of future AIs.