AI for Intracranial Hemorrhage Detection: What the 2026 Evidence Actually Shows

By Junaid Kalia, MD — board-certified neurologist and neurocritical care specialist. Last reviewed July 21, 2026.

An intracranial hemorrhage is one of the few findings on a head CT where the clock is part of the diagnosis. A bleed missed on the worklist for forty minutes is a different clinical event from the same bleed flagged in one. That time pressure is exactly why intracranial hemorrhage (ICH) has become the flagship use case for triage AI, and why the marketing around it runs well ahead of the evidence.

This piece looks at what the peer-reviewed literature actually supports about AI for ICH detection on non-contrast head CT: how accurate these tools are, whether they change the thing that matters (time to the right patient), where they fail, and what "FDA cleared" does and doesn't mean. Every figure below is linked to its primary source.

Key takeaways

• ICH is 10–15% of all strokes but carries disproportionate mortality, so detection speed is clinically decisive.
• Across 58 studies, deep-learning detection reaches a pooled sensitivity of 0.92 and specificity of 0.94.
• A 2024 prospective trial found AI triage did not improve radiologist accuracy or report turnaround, and specificity was slightly lower with it.
• "FDA cleared" for these tools means 510(k) clearance for triage and notification, not a diagnosis and not FDA approval.

Why minutes matter in intracranial hemorrhage

ICH accounts for 10–15% of all strokes, a minority of cases that carries a majority of the harm [1]. One-year mortality runs from 51% to 65% depending on the location of the bleed, and half of the deaths occur in the first two days [1]. The window in which intervention changes the outcome is narrow, and it opens the moment the scan is acquired.

The management side of that window is well codified. The 2022 American Heart Association/American Stroke Association guideline sets out the acute pathway once a hemorrhage is confirmed [2]. What the guideline cannot do is get the study in front of a radiologist faster. That is the gap triage AI claims to fill: not making the diagnosis, but moving the suspicious scan up the worklist so a human reads it sooner.

What the diagnostic-accuracy evidence shows

On raw detection, the evidence is genuinely strong. A 2025 systematic review and meta-analysis of deep learning for ICH on non-contrast CT pooled 58 studies and found a sensitivity of 0.92 (95% CI 0.90–0.94) and a specificity of 0.94 (95% CI 0.92–0.95) [3]. Those are numbers a triage tool can be built on.

Pooled figures from curated datasets tend to flatter, though, so the real-world validations matter more. In one prospective evaluation of 527 consecutive head CTs, of which 79 (15%) contained hemorrhage, a commercial tool held a sensitivity of 0.92 (0.84–0.96) and a specificity of 0.96 (0.94–0.98), and it caught 13 of 14 subarachnoid hemorrhages, the subtype that is both easiest to miss and most dangerous when missed [4]. The positive predictive value in that cohort was 0.82 [4], which is the first hint of the catch: at that precision, close to one in five flagged studies is a false alarm.

Does it actually change the workflow?

Detection accuracy is the easy question. The one that decides clinical value is whether flagging a scan faster produces a faster read and a better outcome. Here the evidence splits.

The optimistic case is real. An early and widely-cited deployment integrated a convolutional neural network directly into the radiology worklist, where it detected suspected ICH and reprioritized those studies ahead of the queue, shortening time to diagnosis in routine operation [5].

Then someone ran the trial. A 2024 prospective single-center study compared radiologist performance before and after a commercial ICH triage system went live [6]. Accuracy was 99.5% without AI and 99.2% with it, a difference that was not significant [6]. Specificity was actually higher without the tool, 99.8% versus 99.3% (p = .004) [6]. And the headline metric, mean report turnaround for ICH-positive exams, barely moved: 147.1 minutes without AI versus 149.9 minutes with it (p = .11) [6]. The authors concluded the system did not improve radiologists' diagnostic performance or turnaround times.

One prospective study at one site is not the last word, and a busy academic center with fast baseline reads is close to the hardest place to show a time benefit. But it is a useful corrective. High standalone accuracy does not automatically convert into a faster or better read, and any vendor claiming a turnaround improvement should be able to show it in a deployment that looks like yours.

Limitations and what to watch for

Three failure modes deserve attention before a purchase, not after.

• Generalizability. The gap between pooled meta-analytic performance [3] and single-site real-world numbers [4] is the whole story of medical-imaging AI. A model tuned on one population, one set of scanners, and one slice thickness can behave differently on yours. Ask for validation data that resembles your scanners and your case mix.
• False positives and alert fatigue. A PPV around 0.82 means roughly one in five alerts is a false alarm [4]. Triage tools are tuned for high sensitivity by design, which pushes precision down; the operational cost lands on the radiologists who learn to trust or tune out the alerts.
• No proven workflow benefit is guaranteed. The one prospective trial to date found none at a fast-reading site [6]. The tool still has to be measured in situ.

None of this argues against ICH triage AI. It argues for buying it on evidence rather than on a sensitivity number pulled from a slide.

What "FDA cleared" actually means here

The regulatory vocabulary matters, because it is often used loosely. These tools reach the market through the FDA 510(k) pathway as computer-assisted triage and notification software (product code QAS), which results in clearance, not approval, and covers triage and prioritization, not diagnosis [7]. The radiologist still reads every study and makes the call. A device labeled "FDA approved" for ICH triage is using the wrong term, and that is worth noting on its own.

Clearance is also specific to one device and one indication. As an example that can be checked in the public record, the FDA 510(k) database lists K241719 as NeuroICH, cleared for its applicant on November 7, 2024 under product code QAS [7]. A clearance number belongs to that device alone. A vendor with several modules does not get to imply the whole suite is cleared because one module is.

How SaveLife.AI approaches ICH triage

We build to the standard described above, and we hold our own claims to it. NeuroICH, the intracranial hemorrhage module of the AI Suite, is FDA 510(k) Cleared under K241719, verifiable in the public FDA 510(k) database [7]. We state clearance module by module and never imply the suite as a whole is cleared. The performance figures and validation detail live in our clinical evidence hub rather than in a headline, so you can read the numbers against their sources.

The honest summary of the field is that ICH detection AI is accurate, that its workflow benefit depends on your baseline and has to be measured, and that the right way to choose one is to verify every claim against evidence you can open. Run that test against every vendor, including us.

Frequently asked questions

Is AI for intracranial hemorrhage FDA approved or FDA cleared?

Cleared, not approved. ICH triage tools go through the 510(k) pathway as computer-assisted triage and notification software (product code QAS), which is a clearance for prioritizing suspected studies, not for making a diagnosis [7]. The radiologist still reads and diagnoses every case.

How accurate is deep learning at detecting ICH on head CT?

In a 2025 meta-analysis of 58 studies, pooled sensitivity was 0.92 and specificity was 0.94 [3]. Real-world single-site performance is broadly similar but with lower precision: a positive predictive value around 0.82 in one 527-scan validation, meaning roughly one in five alerts is a false positive [4].

Does ICH triage AI make radiologists faster?

Not necessarily. A 2024 prospective trial found no significant change in accuracy and no turnaround-time improvement (147.1 vs 149.9 minutes) after a commercial tool was deployed, and specificity was slightly lower with it [6]. Earlier operational reports did show faster time to diagnosis from worklist reprioritization [5], so the benefit is real in some settings and absent in others. Measure it in yours.

Does the AI replace the radiologist?

No. Triage and notification software flags and prioritizes suspected cases so they move up the worklist [7]. Every flagged study is still read by a radiologist, who makes the diagnosis.

Why does intracranial hemorrhage detection get so much attention?

Because the stakes and the time pressure are both high. ICH is 10–15% of strokes but carries a one-year mortality of 51–65%, with half of deaths in the first two days [1]. Anything that reliably shortens the path from scan to treatment has outsized value.

See the evidence in your workflow

Book a 20-minute demo to see NeuroICH triage in a workflow like yours, or read the clinical evidence first.

This article is for educational purposes and is not medical advice. Device performance figures reflect each product's FDA 510(k) record and indications for use; verify them against the primary sources before purchase.

References

1. Rymer MM. Hemorrhagic stroke: intracerebral hemorrhage. Mo Med. 2011;108(1):50-54. PMC6188453
2. Greenberg SM, Ziai WC, Cordonnier C, et al. 2022 guideline for the management of patients with spontaneous intracerebral hemorrhage: a guideline from the American Heart Association/American Stroke Association. Stroke. 2022;53(7):e282-e361. doi:10.1161/STR.0000000000000407
3. Karamian A, Seifi A. Diagnostic accuracy of deep learning for intracranial hemorrhage detection in non-contrast brain CT scans: a systematic review and meta-analysis. J Clin Med. 2025;14(7):2377. PMC11989428
4. Wang C, Jin Y, Shieh S, et al. Real world validation of an AI-based CT hemorrhage detection tool. Front Neurol. 2023;14:1177723. doi:10.3389/fneur.2023.1177723
5. Arbabshirani MR, Fornwalt BK, Mongelluzzo GJ, et al. Advanced machine learning in action: identification of intracranial hemorrhage on computed tomography scans of the head with clinical workflow integration. NPJ Digit Med. 2018;1:9. doi:10.1038/s41746-017-0015-z
6. Savage CH, Tanwar M, Abou Elkassem A, et al. Prospective evaluation of artificial intelligence triage of intracranial hemorrhage on noncontrast head CT examinations. AJR Am J Roentgenol. 2024;223(5):e2431639. PMID: 39230402
7. US Food and Drug Administration. 510(k) premarket notification K241719 (NeuroICH). Decision date November 7, 2024. Product code QAS. FDA 510(k) record

Written by Junaid Kalia, MD — board-certified neurologist and neurocritical care specialist; founder, SaveLife.AI. About the author