AI detected large vessel occlusion on plain CT more sensitively than readers

Design

Bayesian diagnostic-test-accuracy network meta-analysis, GRADE-rated

Population

3,632 patients across 10 retrospective validation studies

Primary outcome

Detection of anterior-circulation large vessel occlusion on noncontrast CT

Effect

Imaging AI sensitivity 0.78 (0.68 to 0.86) vs expert readers 0.62; specificity similar

Noncontrast CT is usually the first scan in suspected stroke, but spotting a large vessel occlusion on it is hard outside specialist centres. This diagnostic-accuracy network meta-analysis pooled 10 retrospective validation studies and 3,632 patients to compare AI with human readers for anterior-circulation large vessel occlusion on noncontrast CT.

Unimodal imaging AI had a sensitivity of 0.78 (95% credible interval 0.68 to 0.86) and specificity of 0.88, against expert-reader sensitivity of 0.62 and non-expert 0.60 at similar specificity. The league-table advantage in sensitivity was 0.16 (0.03 to 0.29) over experts, with no clear difference in specificity. Certainty was low, and every study was retrospective.

The authors frame AI as a bounded human-in-the-loop prompt — flagging cases for expert review, CT-angiography prioritisation, tele-stroke consultation or transfer — not as an autonomous reader. Crucially, the evidence is accuracy-based: it does not yet show faster reperfusion or better functional outcomes, which is what a stroke pathway ultimately needs.

• Network meta-analysis of 10 retrospective validation studies, 3,632 patients.
• Imaging AI sensitivity 0.78 (95% credible interval 0.68 to 0.86) vs expert readers 0.62, at similar specificity.
• Sensitivity advantage over experts 0.16 (0.03 to 0.29); no clear specificity difference.
• Use AI as a prompt for expert review, CT-angiography prioritisation or transfer, not as an autonomous reader.
• Evidence is accuracy-based and does not yet show faster reperfusion or better outcomes.

Why it matters

It suggests AI could shorten the recognition step in stroke pathways where specialist reading is not immediately available.

Don't overread it

All studies were retrospective and certainty was low; this is diagnostic accuracy, not evidence of faster treatment or better recovery.

The statistics, in plain English

A sensitivity of 0.78 versus 0.62 means AI missed fewer occlusions than readers in these datasets; low certainty and an all-retrospective evidence base mean the real-world margin could be smaller, and specificity was no better.

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.