Exploring human–artificial intelligence interactions in a negative pragmatic trial of computer-aided polyp detection

Abstract

Background and Aims: The progress of artificial intelligence (AI) in endoscopy is at a crossroads. The positive results of randomized controlled trials of computer-aided detection (CADe) have not been replicated in multiple pragmatic CADe trials, including ours. This gap between efficacy and effectiveness remains to be understood. We surveyed and interviewed our trial's colonoscopists to gain insight into human-AI interactions. Methods: We used a sequential, mixed-methodology design. After the trial, we administered Survey 1, focusing on attitudes and beliefs before and after trying CADe. The trial's null results were disclosed, and we then administered Survey 2 and conducted open-ended interviews, focusing on reactions to the null results. Responses were analyzed overall and by baseline adenoma detection rate (ADR) tertile. We identified key themes using thematic analysis and qualitative software. Results: Nearly all colonoscopists responded (22 and 21 of 24 [92% and 88%] for Surveys 1 and 2, respectively). Most (96%) regarded endoscopic ability as critical to their professional identity. Large majorities conveyed trust in and enthusiasm for AI before and after trying CADe (82%-87%) and desired to have CADe available (72%). Nearly two-thirds (62%) were surprised by the null results. There were few differences by ADR. No unifying explanation for the null results emerged from surveys or individual interviews. Colonoscopists expressed a range of expectations for AI in endoscopy. Conclusions: Lack of enthusiasm or mistrust of AI/CADe do not explain our pragmatic CADe trial's null results. AI may need to target dimensions beyond optical recognition to realize its promise in endoscopy.

Ähnliche Arbeiten