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

Area of Science:

• Gastroenterology
• Medical Artificial Intelligence
• Health Services Research

Background:

• Computer-aided detection (CADe) using artificial intelligence (AI) in endoscopy has shown promise in controlled trials.
• However, these positive results have not been consistently replicated in pragmatic trials, highlighting a gap between efficacy and effectiveness.
• Understanding human-AI interaction is crucial to address this discrepancy.

Purpose of the Study:

• To investigate colonoscopists' perceptions and experiences with AI-powered CADe systems.
• To explore reasons behind the null results observed in a pragmatic CADe trial.
• To gain insights into human-AI interactions in the context of endoscopic procedures.

Main Methods:

• A sequential, mixed-methodology approach was employed, combining surveys and semi-structured interviews.
• Colonoscopists completed surveys before and after using CADe, and again after trial results were disclosed.
• Thematic analysis was used to analyze qualitative data from surveys and interviews, with responses analyzed by baseline adenoma detection rate (ADR).

Main Results:

• High colonoscopist engagement, with 92% and 88% participation in surveys.
• Strong professional identity tied to endoscopic ability (96%) and sustained trust and enthusiasm for AI (82%-87%) despite null trial outcomes.
• Most colonoscopists (62%) were surprised by the null results, with no single explanation emerging from the qualitative data.

Conclusions:

• Colonoscopist trust and enthusiasm for AI/CADe do not explain the null findings in pragmatic trials.
• The study suggests that AI in endoscopy may need to evolve beyond optical recognition capabilities to fulfill its potential.
• Further research into AI's role in endoscopy should consider factors beyond detection accuracy.