AI in Colonoscopy: Detection Gains Versus Skill Loss
New evidence suggests that while artificial intelligence may help doctors spot more polyps during colonoscopies, it might simultaneously erode the manual skills of experienced gastroenterologists.
Artificial intelligence tools designed to assist in colon cancer screening are delivering mixed results in real-world clinical settings. While these systems often increase the total number of polyps found, recent studies indicate they do not necessarily improve the detection of the most dangerous types of growths, known as advanced adenomas. This discrepancy raises serious concerns among medical professionals about the long-term impact of relying on automated alerts during delicate procedures.
The core issue is not just about finding more tissue samples, but about maintaining the high standard of care required for effective cancer prevention. If the technology distracts clinicians or reduces their need to actively search for subtle signs of disease, the overall benefit to patient health may be diminished. Experts are now calling for a more cautious approach to integrating these tools into routine screening programs.
Detection Gains Do Not Equal Better Outcomes
Research published in the United European Gastroenterology Journal highlights a critical trade-off in using computer-aided detection (CADe). In high-performing screening programs, the addition of AI helped identify small and hard-to-see lesions but failed to reduce the rate of missed advanced neoplasia. This means that while the statistical count of detected polyps increases, the clinical effectiveness in preventing severe cancer cases does not improve proportionally.
Furthermore, there are signs of what researchers term 'deskilling.' Observational data suggests that endoscopists who frequently use AI may perform worse on procedures where the technology is turned off. This implies that over-reliance on automated alerts can weaken a doctor’s natural ability to spot abnormalities, creating a dependency that could be problematic if the system fails or is unavailable.
Trust and Cognitive Load Shape Usage
The effectiveness of these tools depends heavily on how much trust clinicians place in the software. A commentary in Gastroenterology notes that this trust is not uniform. Early-career doctors tend to react to every alert, while mid-career practitioners use the system as a safety net. However, many senior experts disengage from the tool entirely, citing the burden of false positives.
False activations are a significant source of frustration. Some studies report that the system may flag non-cancerous areas roughly 26 to 27 times per procedure. This constant noise creates cognitive load and distraction, leading some clinicians to deactivate the software altogether. As reported by GN technics/ai (en-US), this pattern of disengagement undermines the potential benefits of the technology and creates inconsistent care across different medical practices.
Need for Rigorous Clinical Validation
Current evaluation methods for these AI tools are often inconsistent, making it difficult to determine their true value. Experts argue that clinical metrics must be anchored to specific, prespecified outcomes that matter to patient health, such as cancer incidence and mortality, rather than just the number of polyps found. Without rigorous validation against these meaningful benchmarks, the adoption of AI in gastroenterology remains risky.
The path forward requires designing workflows that account for human behavior and skill levels. This includes targeted training, monitoring of false-positive rates, and ensuring that the technology supports, rather than replaces, clinical expertise. Until these issues are resolved, the widespread use of AI in endoscopy faces a significant hurdle in balancing technological promise with patient safety.
