Artificial intelligence-assisted endoscopists improve the detection rate of high-risk gastric lesions: A propensity score-matched retrospective study

Table 8 Mediation analysis evaluating biopsy number as a potential mediator between artificial intelligence assistance and high-risk gastric lesions detection

Parameter	Unadjusted model	Adjusted model1
Sample size	15528	15528
a-path: AI assistance → biopsy number	0.093 (SE = 0.018, P < 0.001)	0.107 (SE = 0.017, P < 0.001)
b-path: Biopsy number → HrGLs	0.605 (SE = 0.047, P < 0.001)	0.480 (SE = 0.050, P < 0.001)
Total effect2	0.0039 (0.0015-0.0063)	0.0044 (0.0020-0.0069)
Direct effect2	0.0034 (0.0012-0.0058)	0.0039 (0.0017-0.0064)
Indirect effect2	0.0002 (0.0001-0.0004)	0.0002 (0.0001-0.0003)
Proportion mediated	6.3%	4.9%

¹The adjusted model controlled for age, sex, anesthesia use, and endoscopist experience.

²Total, direct, and indirect effects are reported as risk differences (with 95% confidence intervals in parentheses) on the probability scale (0-1).

a-path coefficients (reported as mean difference ± standard error) represent the average increase in biopsy count in artificial intelligence-assisted vs non-artificial intelligence-assisted procedures. b-path coefficients (reported as log-odds ± standard error) represent the increase in log-odds of high-risk gastric lesions detection per additional biopsy. AI: Artificial intelligence; HrGLs: High-risk gastric lesions; SE: Standard error.