AI-flagged gastric lesions were biopsied more often and yielded more neoplasia

Design

Secondary analysis of a randomised controlled trial of AI-assisted endoscopy

Population

About 26,000 patients undergoing upper GI endoscopy screening (13,440 AI, 12,561 control)

Primary outcome

Neoplasm biopsy yield and non-biopsy rate of documented lesions

Effect

Non-biopsy 18.0% (red box) vs 45.6% (control); neoplasm yield 8.19% vs 1.77%

This secondary analysis drew on a randomised trial of AI-assisted versus standard upper endoscopy screening, with about 13,400 patients in the AI arm and 12,600 in the control arm. The AI marked lesions it judged high risk with a red box and low risk with a blue box. Pathology was centrally reviewed.

Lesions in red boxes went unbiopsied 18% of the time, against 43% for other lesions in the AI arm and 46% in the control arm. The share of biopsies that found neoplasia was 8.19% for red-box lesions, against 0.36% for non-red-box lesions and 1.77% in the control arm. False-positive alerts were not linked to more biopsies overall.

The concern with computer-aided detection is that it may drive endoscopists to biopsy everything it flags. Here, the high-risk flag appeared to focus biopsies rather than multiply them. This is a secondary analysis, and the effect on the number of cancers found was not reported in the abstract.

• Computer-aided detection at gastroscopy appeared to focus biopsies on higher-risk lesions in this trial.
• False-positive alerts were not associated with indiscriminate biopsy.
• Treat AI flags as a prompt for closer inspection, not as a diagnosis.
• Keep systematic photo-documentation and inspection time regardless of AI support.

Why it matters

It eases the fear that AI alerts will flood pathology with unnecessary biopsies.

Don't overread it

This secondary analysis looked at biopsy behaviour, not at whether more cancers were detected.

The statistics, in plain English

An incidence rate ratio of 0.39 means red-box lesions were left unbiopsied at about 40% of the rate of other lesions, after adjustment. Because red-box lesions were chosen by the AI rather than randomised, the comparison within the AI arm is observational.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.