Effect of a Computer-Aided Device on Endoscopists' Biopsy Practice for Gastric Neoplasms: A Secondary Analysis of a Randomized Controlled Trial.

BACKGROUND

Endoscopic biopsy is fundamental for diagnosing gastrointestinal lesions, yet the effectiveness of AI assistance on endoscopists' biopsy strategies and efficiency for detecting gastric neoplasms remains unclear. This study aims to investigate this effectiveness.

METHODS

This study is a secondary analysis of a randomized controlled trial (AI group: n = 13,440; control: n = 12,561) compared AI-assisted versus standard screening. The AI system marked high-risk and low-risk lesions with red and blue boxes, respectively. The AI group was stratified into red-box and non-red-box subgroups. Outcomes included neoplasm biopsy efficiency and proportion of non-biopsied documented lesions. The patient-level correlation between false-positive alerts and biopsy counts was examined. All analyses were based on the pathological results after centralized pathology review.

RESULTS

The non-biopsy rate was significantly lower in the red-box group than in the non-red box group and the control group (18.00% vs. 43.26%, 18.00% vs. 45.60%, both p < 0.001), suggesting that lesions were biopsied more frequently in the red-box group. Neoplasm biopsy efficiency was higher in the red-box group (8.19% vs. 0.36%, 8.19% vs. 1.77%, both p < 0.001). In negative binomial regression, red-box classification remained associated with a lower non-biopsy rate (IRR 0.393, 95% CI 0.360-0.429, p < 0.001). False-positive alerts showed a weak inverse correlation with biopsy counts (ρ = -0.182, p < 0.001).

CONCLUSIONS

Red-box classification was associated with more targeted biopsy sampling, while false-positive prompts were not associated with indiscriminate biopsies.