± SCIENCE BackgroundWhether artificial intelligence can identify publishable scientific work is untested. We evaluated…± HEALTH Aim Online health information seeking continues to gain popularity. Quality assessment of online information…± HEALTH Age-related macular degeneration (AMD) is a leading cause of irreversible vision loss, yet systemic determina…+ HEALTH Objectives Early diagnosis of autism spectrum disorder (ASD) remains a significant challenge due to the time-…+ SCIENCE The semi-quantitative assessment of brain [18F]FDG PET provides a more objective interpretation and improved…± HEALTH Objectives To develop a machine learning (ML) approach to explore self-reported factors predictive for recove…± HEALTH Objective : Clinical remission (CR) is an ambitious and attainable treatment goal for patients with asthma. T…± HEALTH BackgroundSeveral commercial artificial intelligence (Al) algorithms are available for detecting intracranial…
TruaceTracing the truth around AIWednesday, September 16, 2026
Subscribe
Health·The Trace·Dual reading·Published 2026-09-16

AI-assisted detection of intracranial hemorrhage on emergency non-contrast head CT

Source article: Retrospective comparison of three commercial artificial intelligence algorithms for detection of intracranial hemorrhage (ICH) in the emergency radiology department

Abstract: BackgroundSeveral commercial artificial intelligence (Al) algorithms are available for detecting intracranial hemorrhage (ICH), but independent clinical validation remains limited.PurposeTo compare three commercially available Al algorithms for ICH detection on non- contrast head computed tomography (NCHCT).Material and MethodsIn this retrospective study, 4027 consecutive NCHCT examinations from a large emergency hospital in southwest Sweden were analyzed. Three Al algorithms were applied, with one vendor disclo…

TRV-2026-1107Peer-reviewedPermanent record — cite & verify
Trace impact reading

Contested: both sides are scored from claims and sources, not community votes.

P 71The P score combines the specificity and measured human impact of the grounded problem claim with the strength of this Trace’s cited sources.G 69The G score combines the specificity and measured human impact of the grounded gain claim with the strength of this Trace’s cited sources.
Retrospective comparison of three commercial artificial intelligence algorithms for detection of intracranial hemorrhage (ICH) in the emergency radiology department

NCCT Brain Imaging by Goleisureintl. CC BY 4.0 · https://creativecommons.org/licenses/by/4.0

The quick read

In a retrospective study of 4027 consecutive non-contrast head CT examinations from an emergency hospital in southwest Sweden, researchers compared three commercial AI algorithms for intracranial hemorrhage detection against reports from two radiologists, using two-tier consensus adjudication as the reference standard for 385 positive or discrepant cases.

By the September 2026 publication date, the observed combination of Aidoc with a human reader reached 96.0% sensitivity at 99.4% specificity, comparable to two radiologists, but the result relied on a simulated logical OR model assuming perfect dismissal of false positives, leaving prospective workflow impact, generalizability beyond a single center, and performance of the other two systems uncertain.

Main points
  • Retrospective analysis of 4027 consecutive NCHCT examinations from a large emergency hospital in southwest Sweden, with 3902 evaluable.
  • Reference standard was two-tier consensus adjudication after manual review of 385 positive or discrepant cases, confirming ICH in 176 cases (4.5% prevalence).
  • Three commercial AI algorithms were compared; Aidoc had highest standalone performance at 90.3% sensitivity and 99.0% specificity.
Gain

Combining Aidoc AI with a human radiologist increased sensitivity for intracranial hemorrhage on non-contrast head CT to 96.0% while maintaining 99.4% specificity, detecting cases missed by radiologists alone.

Problem

Performance across three commercial ICH detection algorithms varied substantially, with only one system showing clinically relevant accuracy, indicating limited independent validation for routine emergency use.

The rundown

The study applied three commercial AI algorithms to 4027 consecutive non-contrast head CTs from a large emergency hospital in southwest Sweden; 3902 were evaluable and 3517 were consistently negative by all readers.

All 385 positive or discrepant cases underwent expert manual review with two-tier consensus adjudication as reference, confirming 176 ICH cases and excluding 209, with eight cases missed by both radiologists but flagged by AI.

What this doesn’t fix

Human-AI performance was estimated using an idealized model that assumes perfect dismissal of false positives, and the data come from a single-center retrospective cohort, limiting generalizability to prospective clinical implementation.

Sources

Reader signal

How should this claim be treated?

The debate