8th Global Gastroenterology and Artificial Intelligence Summit Separating AI Hype from Clinical Reality | September 2026-13

13 - Chiang_AIGI Evidence_Expectations

Pdf Summary

The document explains how evidence for AI in gastroenterology should be judged for safety, effectiveness, and ongoing monitoring. It emphasizes that the FDA does not apply one universal “AI accuracy threshold.” Instead, evidence is evaluated based on the device’s intended use, clinical claims, consequences of errors, workflow, user, population, and environment. Metrics like sensitivity, specificity, AUROC, PPV, and NPV are only meaningful when tied to a real clinical decision.<br /><br />It outlines the main regulatory pathways: 510(k) for substantial equivalence to a predicate device, De Novo for novel moderate-risk devices needing reasonable assurance of safety and effectiveness, and PMA for higher-risk claims requiring stronger scientific evidence. The pathway depends on risk and intended use, not simply on whether the product uses AI.<br /><br />Physicians are advised to look for independent validation data, representative populations, clear prespecified endpoints, and evidence of safety in real workflows. Important issues include edge cases, failure modes, bias, repeatability, and reproducibility. Safety is framed as a balance of probable benefit versus probable risk, with mitigation strategies such as warnings, training, workflow design, and user controls.<br /><br />Effectiveness is claim-specific: detection, characterization, and workflow support each require different types of evidence. Real-world studies can confirm or refine trial findings, but performance may vary by local workflow, equipment, and patient mix. Finally, AI updates should follow a Predetermined Change Control Plan that defines what may change, how changes are validated, and how impact is assessed. The overall message is that AI tools should be useful, bounded, and continuously monitored across their lifecycle.

Keywords

AI in gastroenterology

FDA regulatory pathways

safety and effectiveness

clinical validation

510(k) De Novo PMA

independent validation

real-world performance

bias and reproducibility

predetermined change control plan

workflow monitoring