AI on non-contrast CT detects more large vessel occlusions than readers, on low-certainty evidence

Design

Systematic review and Bayesian diagnostic test accuracy network meta-analysis

Population

10 retrospective validation studies, 3,632 patients with suspected anterior circulation stroke

Primary outcome

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

Effect

AI sensitivity 0.78 (0.68 to 0.86) vs experts 0.62 (0.48 to 0.75); specificity 0.88 vs 0.86

This systematic review and Bayesian diagnostic test accuracy network meta-analysis, published 24 September, included 10 retrospective validation studies with head-to-head comparisons: 11 datasets and 3,632 patients with suspected anterior circulation stroke imaged with non-contrast CT.

Unimodal imaging AI had a sensitivity of 0.78 (95% credible interval 0.68 to 0.86) and specificity of 0.88 (0.81 to 0.94). Expert readers had a sensitivity of 0.62 and non-experts 0.60, with similar specificity of about 0.86. AI's sensitivity was 16 percentage points higher than experts (0.03 to 0.29). Multimodal AI incorporating clinical data looked slightly better still, but only through indirect comparison.

In a hospital where CT angiography or a neuroradiologist is not immediately available, a sensitive signal on the first scan could speed escalation. But the evidence is retrospective, rated low certainty, and measures accuracy only — none of it shows faster reperfusion or better outcomes. About one occlusion in five is still missed.

• Use NCCT-based AI as a prompt to prioritise CT angiography or tele-stroke review, not to exclude occlusion.
• Never withhold CT angiography because the AI output is negative.
• Expect about one false alert in eight to ten scans at the reported specificity.
• Audit local AI performance against CT angiography before relying on it for transfer decisions.

Why it matters

Could shorten the delay to thrombectomy referral where vascular imaging or specialist reading is not on site.

Don't overread it

Accuracy in retrospective datasets is not proof of faster treatment or better outcomes.

The statistics, in plain English

A sensitivity of 0.78 means the AI detected about 78 of every 100 occlusions, against about 62 for expert readers looking at non-contrast CT alone. Credible intervals are the Bayesian equivalent of confidence intervals.

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.