AI colonoscopy finds more 5 mm polyps and essentially nothing else

Design

systematic review and frequentist network meta-analysis with prediction intervals, GRADE assessed, PROSPERO registered

Population

13 randomised trials, 4,156 adults undergoing colonoscopy, 8 AI systems

Primary outcome

mean polyp detection counts stratified by size

Effect

5 mm or less: SMD 0.21 (95% CI 0.07-0.35), prediction interval -1.12 to 1.54, I squared 86.6%; 6-9 mm: 0.02 (-0.02 to 0.06); 10 mm or more: 0.01 (0.00-0.02)

Thirteen randomised trials with 4,156 participants compared eight AI detection systems against standard colonoscopy, with polyp counts stratified by size. No trial compared one AI system against another, so every cross-platform ranking here is indirect.

For diminutive polyps of 5 mm or less, AI gave a standardised mean difference of 0.21 (95% CI 0.07 to 0.35) - modest, and accompanied by heterogeneity of 86.6% and a 95% prediction interval running from -1.12 to 1.54. For polyps of 6 to 9 mm the difference was 0.02 (-0.02 to 0.06), and for those of 10 mm or more, 0.01 (0.00 to 0.02). GRADE certainty was very low for the diminutive finding and low for the others.

This reframes what the technology is doing. Earlier reviews pooled all polyp sizes and reported an overall detection benefit; separating by size shows that benefit is almost entirely diminutive lesions, which are the ones least likely to progress and most likely to add surveillance burden, procedure time and histopathology cost. For a unit deciding whether to buy an AI system, the question this poses is whether more 5 mm polyps is the outcome they were paying for. The finding does not argue against AI-assisted colonoscopy; it argues against the case that has been made for it.

• Ask what outcome an AI procurement case is actually promising - detection counts are not adenoma-related mortality
• Expect more diminutive polyps and plan for the histopathology and surveillance workload that follows
• Do not choose between platforms on this evidence - no head-to-head trial exists
• Keep withdrawal time, bowel preparation quality and adenoma detection rate as the audit measures that matter
• Treat a resect-and-discard policy question as separate; this analysis does not address it

Why it matters

It shows the headline benefit of AI colonoscopy in earlier pooled reviews came from the polyp size that matters least.

Don't overread it

Detection counts are a surrogate - no trial here measured interval cancer or mortality.

The statistics, in plain English

The prediction interval is the number to read here. A confidence interval of 0.07 to 0.35 describes the average effect across these trials; the prediction interval of -1.12 to 1.54 describes what the next trial might find, and it spans harm and large benefit. When the two diverge that far, the pooled average is unreliable, and heterogeneity of 86.6% says why. The effects for larger polyps, at 0.02 and 0.01, are close enough to zero that their statistical significance is a function of sample size rather than of clinical meaning.

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.