AI on the first CT can flag the large-vessel stroke before the angiogram

Design

Bayesian diagnostic-accuracy network meta-analysis of retrospective validation studies

Population

10 studies, 11 datasets, 3,632 patients with suspected stroke

Primary outcome

Sensitivity and specificity for anterior-circulation large-vessel occlusion on non-contrast CT

Effect

Imaging AI sensitivity 0.78 (0.68-0.86), specificity 0.88; readers sensitivity 0.60-0.62; AI +0.16 sensitivity

A Bayesian diagnostic-accuracy network meta-analysis pooled 10 retrospective validation studies (3,632 patients) comparing ways to detect anterior-circulation large-vessel occlusion on non-contrast CT, the scan usually obtained first in suspected stroke, against expert readers, non-expert readers and AI.

Imaging AI reached a sensitivity of 0.78 (95% credible interval 0.68-0.86) and specificity of 0.88, against sensitivities of 0.62 for expert and 0.60 for non-expert readers, with similar specificity, so AI caught about 16 more occlusions per 100 than experts without losing specificity. A clinically informed multimodal model performed slightly better but rested on sparse, indirect data.

The practical role is as a human-in-the-loop prompt: an AI flag on the plain CT can trigger earlier CT angiography, tele-stroke review or transfer discussion, particularly outside specialist centres. The evidence is retrospective and accuracy-based, so it does not yet show faster reperfusion or better function, only that AI reads the first scan more sensitively.

• Imaging AI detected large-vessel occlusion on plain CT with sensitivity 0.78 versus about 0.60 for readers.
• Specificity was similar across AI and human readers.
• AI caught roughly 16 more occlusions per 100 than expert readers.
• Use an AI flag to prompt earlier CT angiography, tele-stroke review or transfer.
• It is most useful where specialist neuroimaging reading is not immediately available.

Why it matters

It offers a way to shorten the delay to recognising a treatable occlusion where expert CT reading is not at hand.

Don't overread it

Retrospective, low-certainty accuracy data; it does not establish faster reperfusion or better functional outcomes, and AI is an adjunct to, not a replacement for, expert review.

The statistics, in plain English

A sensitivity of 0.78 still misses about a fifth of occlusions, so a negative AI read does not exclude one; because the studies were retrospective, these are accuracy figures, not proof the pathway improves outcomes.

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.