Artificial intelligence in upper gastrointestinal endoscopy: what detection rates cannot tell us

We read the article by Lee et al.1 with great interest. They found that real-time artificial intelligence (AI) improved lesion detection during upper gastrointestinal endoscopy, particularly for extremely small lesions (<0.5 cm). This is useful for obtaining real-world evidence from low-resource settings. We would like to raise a few points. First, two of the outcomes (benign and malignant detection rates) depend on the biopsy results and not on what the endoscopist sees. A large randomized trial showed that biopsy and pathology results can be affected by sampling error and by disagreement between pathologists.2 In addition, the number of cancers found was small (26 vs. 10), so the lack of a difference in malignant detection rate may simply reflect a small sample, not a true lack of effect. Second, fewer biopsies were performed after AI. This could mean better and more selective decisions; however, it could also mean that some lesions were skipped because they appeared harmless. Third, this was a single-center study conducted by one endoscopist who compared an earlier period with a later period. This type of design cannot fully separate the effects of AI from other changes over time, such as endoscopists gaining more experience. Fourth, background stomach conditions, such as atrophy and intestinal metaplasia, strongly affect gastric cancer risk.3 If these background conditions differed between the earlier and later patient groups, this alone could explain part of the change in detection rate. Reporting a background risk score, such as the Kyoto score, would help to clarify this. Fifth, in colonoscopy studies, most of the extra polyps found by AI are very small (<5 mm), and AI tends to be less helpful when the endoscopist is already highly skilled.4 The endoscopist in this study was very experienced, with more than 50,000 prior examinations. This may explain why the overall improvement was modest, and why the malignancy detection rate did not change. However, these points do not reduce the value of this study. We hope that future multicenter studies will address these issues using centralized pathology reviews, reporting background risks, and separating AI detection performance from biopsy outcomes.