Clinical Implementation of Artificial Intelligence in Endoscopy: A Human-Artificial Intelligence Interaction Perspective

Abstract

Artificial intelligence (AI) for gastrointestinal endoscopy has shown remarkable performance in detecting and characterizing lesions. A randomized controlled trial reported that AI significantly reduced the miss rates for gastric neoplasms, but real-world implementation studies have shown inconsistent results. This discrepancy cannot be explained solely by technical limitations. Regardless of the AI capabilities, the visualization quality and systematic inspection remain fundamental prerequisites, and traditional apprenticeship training cannot be replaced by technology. This review examines AI implementation in endoscopy from a human-AI interaction perspective. Two cognitive phenomena are relevant: 'automation neglect,' where experienced endoscopists dismiss AI recommendations due to overconfidence or distrust, and 'automation bias,' where users over-rely on AI outputs, potentially missing unhighlighted lesions. Recent evidence raises concerns regarding deskilling, with studies showing decreased diagnostic performance after exposure to AI. A systematic analysis of 52 human-AI teaming studies showed that none achieved ideal complementarity, and collaboration sometimes decreased accuracy compared to humans alone. AI effectiveness varies according to operator expertise. High-performing endoscopists gain minimal benefit, while those with intermediate experience show the greatest improvement. Nevertheless, excessive false-positive alerts can negate benefits. Strategies to address these challenges include explainable AI, human-centered design, structured education, trust calibration, and expertise-tailored AI systems. Maintaining human expertise remains paramount. AI is a powerful tool, but clinicians must remain the final decision maker. Periodic AI-free practice may be necessary to preserve clinical competence.

Ähnliche Arbeiten