Replication data for "The USE-AI study: a multicenter, single blind proof-of-concept study on the reliance on AI in gastrointestinal endoscopy"

The dataset consists of a curated collection of short, high-resolution gastrointestinal endoscopic video clips designed to simulate realistic diagnostic decision-making scenarios in clinical practice. Specifically, it includes 90 capsule endoscopy and colonoscopy videos, each lasting approximately 15–20 seconds, representing three common gastroenterological diagnostic tasks: the identification of small bowel lesions with high bleeding potential, the detection of inflammatory lesions associated with Crohn’s disease, and the evaluation of ulcerative colitis severity using the Ulcerative Colitis Endoscopic Index of Severity (UCEIS).Each video is associated with clinically validated annotations and ground-truth diagnostic labels derived from internationally accepted scoring systems such as the Saurin classification, the Lewis score, and UCEIS.
The dataset was created with the objective of investigating how clinicians interact with artificial intelligence-based decision support systems in visually intensive diagnostic tasks, rather than focusing exclusively on algorithmic performance. In particular, its purpose is to enable the empirical evaluation of different Human-AI Collaboration Protocols (HAICPs) and to assess how variations in interaction design between human decision-makers and AI systems influence diagnostic accuracy, appropriate reliance on AI recommendations, and cognitive biases such as automation bias and self-anchoring bias. This makes the dataset suitable for studying human-AI collaborative decision-making dynamics in medical imaging contexts, especially in scenarios where diagnostic performance depends on both visual interpretation and trust calibration.
Video clips were selected from a clinical digital archive and independently reviewed by a panel of nine expert endoscopists who established the diagnostic ground truth through consensus agreement of at least eight out of nine reviewers. Each video was subsequently augmented with visual cues generated by a simulated AI decision support system with predefined performance characteristics (80% accuracy, 86% sensitivity, and 75% specificity), designed to replicate the behavior of real-world computer-aided diagnostic tools. Participants in a multicenter web-based study evaluated the videos in a within-subject experimental setup, first providing an unaided diagnosis and then revising their judgment after exposure to AI-generated annotations