Moderate evidence
More
AI spotted 85.4% of urgent head CT findings in one ER, with 37 initial misses
Study title: Artificial Intelligence Triage of Urgent Versus Non-Urgent CT Brain Findings to Support Expedited Emergency Department Disposition: A Retrospective Validation Study
- Evidence type
- Observational study
- Sample
- n = 3,424
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-1238cite or check it
Study title: Artificial Intelligence Triage of Urgent Versus Non-Urgent CT Brain Findings to Support Expedited Emergency Department Disposition: A Retrospective Validation Study
Researchers retrospectively tested an AI model that classifies non-contrast CT brain scans as urgent or non-urgent on 3424 consecutive adult scans from a quaternary emergency department in 2024. Against consultant radiologist reports as reference, the model achieved 85.4% sensitivity, 69.1% specificity and 98.3% NPV, with 64% of encounters identified as true negatives potentially eligible for expedited disposition.
Read the study abstract
Objective Evaluate the diagnostic accuracy of an artificial intelligence (AI) model for identifying urgent computed tomography brain (CTB) findings in consecutive emergency department (ED) patients and estimate the proportion eligible for expedited disposition. Methods We retrospectively analysed 3424 consecutive non-contrast CTB scans from adults presenting to a quaternary ED in 2024. The AI model (Harrison Enterprise CTB) classified each scan as "urgent" or "non-urgent." The reference standard was the index co…

57Gain side
How the two sides scoreClose call
54Problem side
The gain
In a retrospective test on 3424 emergency head CT scans, the AI's non-urgent calls were right 98.3% of the time, and 64% of encounters were potentially eligible for faster disposition.
The problem
In a retrospective test of 3424 head CT scans, the AI initially missed 37 urgent scans (85.4% sensitivity); a single unblinded re-review reclassified 24 as non-urgent, leaving 13 small or subtle findings, and specificity was only 69.1%.
A fixed formula sets both scores from the evidence and its sources. Reader votes don’t move them.
What the scores mean
G is the gain side, out of 100. It goes up when the gain is specific, when it measurably helped people and when the sources behind it are strong.
P is the problem side, out of 100. It goes up when the problem is specific, when it measurably hurt people and when the sources behind it are strong.
In brief
The high NPV supports a safety rationale for a prospective trial to incorporate AI triage into ED workflows to speed disposition, but initial false negatives and reliance on a single-center retrospective design with report-based reference and unblinded post hoc re-adjudication leave uncertainty about real-world safety and generalizability.
Main points
- Study analyzed 3424 consecutive non-contrast CTB scans from adults presenting to a quaternary ED in 2024.
- Reference standard was the index consultant radiologist report alone, classified by two emergency physicians with blinded radiologist tiebreak, all blinded to AI output.
- Initial sensitivity was 85.4% (95% CI 80.0%-89.0%) and specificity 69.1% (95% CI 67.0%-71.0%) with 253 (7.4%) urgent findings.
- Post hoc re-adjudication of 37 false negatives by a single unblinded assessor reclassified 24 as non-urgent, raising sensitivity to 94.3% and NPV to 99.4%.
The full story
The study evaluated Harrison Enterprise CTB on 3424 consecutive adult non-contrast CT brain scans from a quaternary ED in 2024, comparing AI classification of urgent versus non-urgent against consultant radiologist reports classified by emergency physicians blinded to AI output.
Initial results showed 253 urgent cases, sensitivity 85.4% and NPV 98.3%, with 64% true negatives potentially eligible for expedited disposition; post hoc unblinded re-adjudication of 37 false negatives reclassified 24 as non-urgent, leaving 13 small or subtle findings that did not require urgent intervention during the index presentation.
What this doesn’t fix
Retrospective single-center design using radiologist report alone as reference, and post hoc false-negative re-adjudication by a single unblinded assessor with access to medical record, limits generalizability before prospective trial.
Sources
1 on file- Emergency Medicine Australasiadoi:10.1111/1742-6723.70353
Talk it over
The AI's non-urgent calls were right 98.3% of the time, yet it initially missed 37 urgent scans. How would you weigh faster disposition for 64% of encounters against those misses?
AI spotted 85.4% of urgent head CT findings in one ER, with 37 initial misseshttps://truvace.com/item/artificial-intelligence-triage-of-urgent-versus-non-urgent-ct-brain-findings-to-
The AI's non-urgent calls were right 98.3% of the time, yet it initially missed 37 urgent scans. How would you weigh faster disposition for 64% of encounters against those misses?
Answer in the threadThe researchers re-reviewed the 37 misses with a single unblinded assessor, who reclassified 24 as non-urgent. What does that step make you trust or doubt about the 94.3% sensitivity that followed?
Answer in the threadIf you were a patient with a head CT that the AI called non-urgent but which showed one of the 13 small or subtle findings, what would you want the emergency department to have done?
Answer in the thread
If you have more time
What would a prospective trial in this quaternary emergency department need to measure to show whether the 98.3% figure holds for patients going home sooner?
Answer in the threadSpecificity was 69.1%. How might that play out in a busy emergency department you know, and who would handle the extra scans flagged as urgent?
Answer in the thread
ace

What readers say