Artificial Intelligence in Stroke Care, Where the Evidence Is Strong and Where It Is Not

The strongest evidence for AI in stroke care sits in two places, detecting large vessel occlusions on imaging and accelerating the workflow that gets a patient from the scanner to treatment. Outcome prediction is advancing but has not yet shown it beats the simple scores clinicians already use. And almost all of this evidence comes from well-resourced systems that already have CT scanners, angiography suites, and thrombectomy teams, which is exactly what most of the world does not have. This review separates what is proven from what is promising, and names where the gap between a published result and a deployable tool is widest.

Why stroke is the proving ground for clinical AI

Stroke is unusually well suited to AI because it is an emergency measured in minutes. For large vessel occlusion, the type of ischemic stroke caused by a clot in a major artery, the treatment is mechanical thrombectomy, and every block of delay between arrival and reperfusion reduces the chance of a good recovery. Pooled analyses of the original thrombectomy trials established that faster treatment means better functional outcomes, which makes any tool that shaves minutes off the pathway clinically meaningful rather than cosmetic (JAMA Neurology, individual patient data meta-analysis).

That time pressure is also why stroke became an early target for regulators. The first AI software for radiologic detection of large vessel occlusion received FDA clearance in 2018, and a crowded field of commercial tools now follows, including Viz.ai, Brainomix, RapidAI, and CINA (systematic review and meta-analysis, Neuroradiology, 2025). The question is no longer whether these tools exist. It is how well they work, and for whom.

Diagnostic applications, the most mature evidence

AI detection of large vessel occlusion on CT angiography is the most validated application in stroke care, and the evidence is moderate to strong for the anterior circulation. Large vessel occlusion accounts for roughly a third of ischemic strokes, and a 2025 systematic review found that AI tools identify these occlusions on CT angiography with good accuracy across multiple commercial products (moderate evidence, pooled meta-analysis).

The more useful question is whether AI changes what clinicians actually do. A 2025 multi-reader study of Brainomix e-CTA put 21 NHS clinicians through suspected stroke cases with and without software support and found that AI assistance improved reader performance, with the largest benefit for less experienced readers (moderate evidence, single multi-reader crossover study, BJR Artificial Intelligence). That pattern, AI helping the non-expert more than the expert, matters in any system where a neuroradiologist is not on call at 3am.

Two honest limits sit underneath the headline accuracy. First, performance drops outside the anterior circulation. Posterior circulation occlusions and medium or distal vessel occlusions are harder for current tools and less well validated. Second, the most accessible imaging is the weakest target. Detecting an occlusion on non-contrast CT, the scan available almost everywhere, is far harder than reading it on CT angiography. Tools that attempt large vessel occlusion prediction from non-contrast CT alone, such as the JLK-CTL score validated across Korean stroke centers, are genuinely interesting for settings without CT angiography, but the evidence is still limited and retrospective rather than prospective (limited evidence, multicenter validation, Stroke: Vascular and Interventional Neurology). Beyond occlusion detection, AI is also used to automate ASPECTS scoring of early ischemic change and to flag intracranial hemorrhage, usually bundled into the same imaging platforms.

Predictive applications, promising but not yet superior

Outcome prediction is where the hype most outruns the evidence. Machine learning models have been built to predict 90-day functional outcome, successful reperfusion, and hemorrhagic transformation after treatment, and on paper their discrimination looks respectable. A 2022 systematic review and meta-analysis found conventional machine learning models predicting functional outcome with a pooled AUC around 0.81, with deep learning models slightly lower at 0.75 (limited evidence, Frontiers in Neurology).

The caveat in that same review is the point of this section. The authors concluded that these models did not generally outperform the prognostic scores clinicians already use, that most were developed on small datasets, that external validation was usually weak or absent, and that many carried a high risk of bias (Frontiers in Neurology). A model that matches a simple bedside score while being harder to interpret and impossible to validate locally is not an advance. The literature on predicting hemorrhagic transformation tells a similar story, strong reported accuracy undercut by small samples and single-center designs (limited evidence, systematic review and meta-analysis).

Newer models add explainability frameworks such as SHAP to show which variables drive a prediction, which is a real improvement for clinical trust, but it does not fix the underlying problems of small cohorts and missing prospective validation. For now, predictive stroke AI is a research-grade capability, not a deployable decision tool.

Workflow applications, where the strongest real-world signal lives

The most convincing evidence that AI improves stroke care is not about reading a scan more accurately. It is about moving the patient faster. This is the one area with randomized and large prospective evidence.

The randomized anchor is a 2023 cluster trial across four comprehensive stroke centers, which found that automated large vessel occlusion detection coupled with mobile alerts cut time to thrombectomy initiation by about 11 minutes (moderate evidence, single cluster randomized trial, 243 patients, JAMA Neurology). A 2025 meta-analysis of Viz.ai deployments reinforced the pattern, showing consistent reductions in CT-to-thrombectomy, door-to-groin, and door-in-door-out times. Notably, that same analysis did not find a statistically significant improvement in patient clinical outcomes (moderate evidence for workflow, weaker for outcomes, Translational Stroke Research). Faster is well established. Better is harder to prove.

The largest real-world study to date pushes that outcome question further. Published in December 2025, a prospective observational study used England's national stroke audit covering more than 450,000 patients across all 107 NHS stroke hospitals, and evaluated Brainomix 360 at 26 of them. Thrombectomy rates at evaluation sites doubled, from 2.3% to 4.6%, with the biggest gains at primary stroke centers that do not perform thrombectomy themselves and must transfer patients out. At the patient level, AI review was associated with higher odds of receiving thrombectomy, and patients reviewed with AI were more likely to have a favourable functional outcome at discharge with no increase in in-hospital mortality (moderate evidence, large prospective observational study, The Lancet Digital Health; NHS England summary).

This is the strongest signal that workflow gains can translate into access and outcomes, and it should be read with its design in mind. It is observational rather than randomized, the analysis was conducted with the software manufacturer, and an increase in thrombectomy access at transferring hospitals is partly a story about referral networks, not the algorithm alone. The honest reading is that AI triage measurably widened access to a proven treatment in a system already built to deliver it.

A quieter limitation runs through all of this, the false positive. A tool that alerts on occlusions that are not there erodes trust and creates alert fatigue, and the operating point that maximizes sensitivity for a time-critical disease tends to generate more false alarms. Reported specificity in validation studies is rarely the specificity a busy center experiences at 2am.

The low-resource reality, where the gap is widest

Every application above assumes infrastructure that much of the world lacks, and this is the gap Curely cares about most. The evidence base for stroke AI was built almost entirely in high-income systems with CT angiography on demand and thrombectomy-capable centers within reach. That assumption does not hold across much of Africa.

The numbers are stark. Stroke is the second leading cause of death and disability on the continent, yet some African countries have fewer than one CT scanner per million people, against more than ten per million in high-income regions, and roughly three neurologists per ten million people, against up to 900 per ten million in high-income countries (Frontiers in Stroke, 2025). Only about one in ten African stroke patients receives a guideline-based acute intervention. Access to thrombectomy specifically has been estimated at 0.48% in low- and middle-income countries against 21.56% in high-income countries, with Egypt highest in Africa at 2.8% and Kenya at 0.4% (Mechanical Thrombectomy in Africa review). An audit at Mozambique's national reference hospital recorded a median delay of 20 hours from stroke onset to admission (delays audit, 2025).

When the bottleneck is a missing scanner, a missing neurologist, and a 20-hour delay, an algorithm that triages CT angiography in 90 seconds solves a problem the system does not yet have. This is not an argument against AI in these settings. It is an argument for sequencing it correctly. The applications most likely to transfer are the ones that lower the infrastructure they require, occlusion signals readable from non-contrast CT rather than CT angiography, and AI embedded in telemedicine and hub-and-spoke referral models that extend scarce specialist expertise rather than assuming it is already on site. Those tools exist mostly as early, single-vendor, retrospective work today (preliminary evidence). The honest position is that they are the right research priority for low-resource stroke care, not a finished product.

The takeaway

AI in stroke care has earned a real place at two points in the pathway, finding the occlusion on imaging and getting the patient to treatment faster, with the December 2025 NHS evidence the most credible signal that faster also means better access and outcomes. Outcome prediction is not there yet, and should be treated as a research capability until models show external validation and a clear advantage over existing scores. For systems building stroke care from a lower base, the binding question is not which algorithm to buy. It is whether the imaging, the workforce, and the referral pathway that the algorithm presupposes actually exist, and which AI tools can be designed to need less of them.