Artificial Intelligence Triage of Urgent Versus Non-Urgent CT Brain Findings to Support Expedited Emergency Department Disposition: A Retrospective Validation Study.
Authors
Affiliations (6)
Affiliations (6)
- RPA Green Light Institute, Sydney Local Health District, Sydney, New South Wales, Australia.
- Department of Radiology, Royal Prince Alfred Hospital, Sydney, New South Wales, Australia.
- RPA Virtual Hospital, Sydney Local Health District, Sydney, New South Wales, Australia.
- Sydney Medical School, Faculty of Medicine and Health, the University of Sydney, Sydney, New South Wales, Australia.
- Department of Neurosurgery, Chris O'brien Lifehouse, Sydney, New South Wales, Australia.
- Emergency Department, Royal Prince Alfred Hospital, Sydney Local Health District, Sydney, New South Wales, Australia.
Abstract
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. 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 consultant radiologist report alone, classified independently by two emergency physicians with disagreements resolved by a blinded consultant radiologist tiebreak, all blinded to AI output. In a separate post hoc analysis, false negatives were re-adjudicated by a single unblinded assessor with access to the medical record and any subsequent imaging. Of 3424 scans, 253 (7.4%) had urgent findings on the reference standard. AI sensitivity was 85.4% (95% CI 80.0%-89.0%), specificity 69.1% (95% CI 67.0%-71.0%) and negative predictive value (NPV) 98.3% (95% CI 98.0%-99.0%). An estimated 64.0% of encounters were true negatives potentially eligible for expedited disposition. Of 37 false negatives, independent adjudication reclassified 24 as non-urgent, yielding a post hoc sensitivity of 94.3% and NPV of 99.4%. The 13 remaining cases were small or subtle findings. None of the 13 cases required urgent intervention during the index ED presentation. The AI model demonstrated high NPV for urgent CTB findings, with no false negative requiring urgent intervention during the index ED presentation. Preliminary results provide the safety rationale to proceed with a prospective trial aiming to incorporate this technology into ED workflows.