Design: Bayesian diagnostic test accuracy network meta-analysis of 10 retrospective studies
Population: 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 vs expert 0.62; difference 0.16 (95% CrI 0.03–0.29)

This Bayesian diagnostic network meta-analysis compared AI and human readers for detecting anterior circulation large vessel occlusion on non-contrast CT. It included 10 retrospective validation studies with 3,632 patients, all with head-to-head comparisons.

Imaging-only AI had 78% sensitivity (95% credible interval 68–86%) and 88% specificity, against 62% sensitivity for expert readers and 60% for non-experts, both with about 86% specificity. The sensitivity gain over experts was 16 percentage points (3–29). Multimodal AI using clinical data looked similar but rested on indirect evidence.

The certainty is low and all data are retrospective. The use case is narrow: a prompt for faster CT angiography or tele-stroke review in hospitals where CTA and specialist reads are delayed. It is not a replacement for CTA.

Do not rely on non-contrast CT, read by AI or human, to exclude large vessel occlusion — CTA remains the test.
Where CTA is delayed, an AI flag on plain CT can justify prioritising CTA or early transfer discussion.
Expect AI to miss about one in five occlusions and to over-call about one in eight.
In spoke hospitals without on-site neurology, AI flags may speed tele-stroke consultation.

Why it matters

It points to a practical role for AI in hospitals without rapid CTA, where treatment delays are longest.

The statistics, in plain English

Sensitivity of 78% means AI flagged 78 of every 100 true occlusions; experts on plain CT flagged 62. Low GRADE certainty and retrospective data mean real-world performance may be worse, and no study showed faster reperfusion or better outcomes.