- 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 (0.68–0.86) vs experts 0.62 (0.48–0.75); specificity 0.88 vs 0.86
This Bayesian network meta-analysis included 10 retrospective validation studies (3,632 patients) comparing expert readers, non-expert readers and AI in detecting anterior circulation large-vessel occlusion (LVO) on non-contrast CT — the scan usually done first in suspected stroke.
Imaging-only AI had a sensitivity of 0.78 (95% credible interval 0.68–0.86) and specificity of 0.88. Expert and non-expert readers had sensitivities of 0.62 and 0.60, with similar specificity (0.86). AI's sensitivity advantage over experts was 0.16 (0.03–0.29). Certainty was rated low.
The practical idea is AI as a prompt: flagging a likely LVO on the first plain scan to prioritise CT angiography or a tele-stroke call. The authors are explicit that this is accuracy data only — nothing yet shows faster reperfusion or better outcomes.
- Do not rely on plain CT alone to exclude large-vessel occlusion — clinicians missed about 4 in 10.
- Get CT angiography promptly in suspected stroke with a significant deficit, whatever the plain CT shows.
- If your centre uses AI stroke software, treat a positive flag as a reason to escalate, not a diagnosis.
- Where CT angiography is not available, use the AI flag and clinical severity to guide transfer discussions.
Why it matters
In centres without immediate CT angiography or stroke expertise, an early flag could shorten the path to thrombectomy.
Don't overread it
Retrospective accuracy studies only — no evidence yet that AI flags speed treatment or improve outcomes.
The statistics, in plain English
Sensitivity of 0.78 means AI detected 78 of every 100 occlusions; experts caught 62. A credible interval is the Bayesian equivalent of a confidence interval.
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.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for emergency & critical care, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free