AI Stroke Scan Triage: Faster Treatment in a Trial, Uneven Gains Across NHS Hospitals

The use case

In a stepped-wedge trial (hospitals switched on one at a time) across four Houston stroke centres between 1 January 2021 and 27 February 2022, a cloud-based algorithm read each CT angiogram of the brain and sent a phone alert to the stroke team when it suspected a blocked large artery¹. That blockage is called a large vessel occlusion, and it is treated by thrombectomy: a catheter is threaded to the brain to pull the clot out. The aim of the software is to shorten the time between a patient arriving and the clot being removed.

The US regulator describes the first authorised product of this kind as a "notification-only, parallel workflow tool": it sends an alert while the standard radiology read carries on, and its findings are "not for diagnostic use beyond notification"³. The clinician who receives the alert must view the full-quality images and evaluate the patient before any care decision³.

Scope note. Here, AI means software that analyses brain scans with machine-learning algorithms, which the regulator describes as an artificial intelligence algorithm³. Older technology and human services do similar work and are not credited as AI: radiologists reading scans, paging and phone calls to specialists, remote teleneurology consultations, and England's wider stroke improvement programmes, which ran at the same time as the software². The Houston trial tested the software together with secure messaging, so its effect cannot be credited to detection alone¹. Pulmonary embolism triage and other scan-alert uses are outside this paper; no evidence on them was read.

Where it is running

NHS England reports (operator-reported, December 2025) that AI decision-support tools were rolled out to all stroke centres in England in Summer 2024, and that every regularly admitting stroke service now has them⁷. Within the evaluation sites, the share of cases read with the software rose from 53% at implementation to 66% in March 2023; the basis of that usage figure was not stated in the portion read².

The first US authorisation dates from 13 February 2018³, which is more than three years old, and the total number of US sites was not established from the sources read. The data in the randomised trial are from 2021 and 2022, so the main randomised result is also more than three years old; no newer randomised result was found in the sources read.

What the evidence shows

Randomised trial. In the Houston trial (measured in a cluster randomised trial, 2021 to 2022), adjusted door-to-groin time, the time from hospital arrival to puncture of the artery, fell by 11.2 minutes (95% CI -18.22 to -4.2). The unadjusted median fell from 100 minutes (IQR 81 to 116) to 88 minutes (IQR 65 to 110)¹. Time from the start of the CT scan to the start of thrombectomy fell by 9.8 minutes (95% CI -16.9 to -2.6)¹. The mechanism is parallel alerting: the stroke team learns of the blockage before the formal scan report is written³. The alert did not clearly speed clot-dissolving drugs: arrival to the drug bolus changed by -3.5 minutes (95% CI -9.1 to 2.2)¹. In the exploratory 90-day analysis, the odds ratio for functional independence was 1.3 (95% CI 0.42 to 4.0), which the abstract reports as no difference¹. An interval that wide, in 243 patients, cannot rule out either benefit or harm. Unadjusted mortality was 31% before and 13% after (P < .001)¹; that figure is not adjusted for case mix and the authors' discussion of it was not read, so it is not treated as an effect of the software.

Single-centre before-and-after studies. At the Miami centre (measured without a control group, 2018 to 2022), thrombectomy within 90 minutes of arrival rose from 26.5% (95% CI 17.9 to 36.8) to 68.1% (95% CI 60.0 to 75.5), and median door-to-groin time fell from 109 to 75 minutes⁴. The share of all ischaemic stroke patients receiving thrombectomy rose from 11.2% to 18.1%⁴. Reperfusion (blood flow restored) was 93.5% before and 90.3% after, and death or hospice was 29.4% before and 25.9% after; the study was not powered for either⁴. At the San Diego centre (before-and-after quality improvement, abstract only), door-to-groin time for direct arrivals out of hours fell from 157 to 95 minutes (P = .009) in a small subgroup of 82 cases overall⁶.

Teleneurology network. Across 14,116 cases, arrival-to-notification of the thrombectomy specialist was 39.5 minutes faster at facilities using the software than at those without (44.13% reduction, abstract), in a retrospective comparison of different facilities⁵. At facilities that cannot perform thrombectomy, the presence of the software correlated with a lower rate of specialist acceptance for thrombectomy, down 10.79 percentage points (23.17% relative)⁵.

England. The independent evaluation of 25 stroke units used a one-group before-and-after design and reported that findings can only be associated with, not attributed to, the software². Thrombectomy rates rose from 2.1% to 4.5% at evaluation sites, against 1.7% to 2.9% at other English sites, a descriptive comparison the authors caution because other sites were also adopting AI². The rise was significant at acute stroke centres (0.9% to 2.4%, p < 0.0001) and not at comprehensive centres (4.3% to 6.4%, p = 0.10)². The rate of clot-dissolving drug treatment was unchanged, at 12.02% before and 12.01% after². NHS England's summary of a later peer-reviewed study (operator-reported; the paper itself was not opened) gives thrombectomy rates doubling from 2.3% to 4.6% at 26 participating hospitals, against 1.6% to 2.6% elsewhere, a 64-minute shorter door-in-door-out time at primary stroke centres, and a higher chance of a favourable functional outcome at discharge with no rise in in-hospital mortality; it gives no confidence intervals⁷.

Regulator test. In the 2018 regulatory review (retrospective, 300 CT angiograms from two US sites, judged against neuroradiologists), sensitivity was 87.8% (95% CI 81.2 to 92.5) and specificity 89.6% (95% CI 83.7 to 93.9). In 44 true cases the alert arrived earlier than standard care in 42, at a mean of 7.32 minutes against 58.72 minutes³. No newer accuracy figure from an independent source was read.

What it takes: cost, integration, governance

Cost. No licence price, cost per scan or cost per hospital was found in the sources read, and the one cost-benefit model in the English evaluation was read only through a tool summary and is left out. Costs of the Houston trial and of the NHS rollout are not reported in the pages read.

Integration and people. The Houston system required CT angiography, a cloud service, a phone app with image viewing and secure messaging, and staff trained to act on alerts¹. In England, the evaluation reports that ambulance availability affects transfer times independently of the software². Under the regulator's conditions, a clinician must always review the full images and the patient³.

Rules. In the United States the first product was classified as Class II (moderate risk) under 21 CFR 892.2080 and is for prescription use only³. In the European Union, the Commission states that the AI Act's high-risk rules for AI in products covered by the Annex I product legislation apply from 2 August 2028, after the AI omnibus entered into force on 27 July 2026⁹. Annex I lists the Medical Devices Regulation (EU) 2017/745¹⁰; this paper reads that to mean certified stroke-triage software would fall under the 2028 date, but the Commission pages do not name stroke tools.

Where it falls short

The challenge for this use case is to ask what would make it look less successful than it is presented, by testing the main claims against their design.

Reported gains in treatment times are largest in the weakest designs and smaller in the one randomised trial. The Miami before-and-after study shows a 34-minute fall in median door-to-groin time⁴, and the teleneurology comparison shows 39.5 minutes⁵, against 11.2 minutes adjusted in the randomised trial¹. Before-and-after designs absorb time trends and growing team experience; the Miami authors say confounding from these cannot be excluded⁴. The randomised trial had four centres, no blinding, and a small sample, and its 90-day independence result is exploratory and compatible with large harm or large benefit¹.

In England the gains are uneven. At acute stroke centres, median door-in-door-out time (the time a patient spends in the first hospital before transfer) was faster at 8 sites (by 2 to 70.5 minutes) and slower at 9 (by 16.5 to 92 minutes), with one site lacking data. Ipswich fell from 184.5 to 114 minutes, while North Devon rose from 150 to 242 minutes². Faster times were associated with more months since implementation (r = -0.547)², which suggests that benefits depend on how well the pathway around the tool is rebuilt, not on installing it. The authors also note that national quality-improvement programmes ran at the same time².

The software shortens one step. Transfers, ambulance availability and staffing remain, and the regulator's design means every alert is checked by a person²³. At the 2018 test sensitivity of 87.8%, roughly one in eight true blockages in that sample was not flagged, and at specificity of 89.6% roughly one in ten scans without a blockage was³. Missed cases are still read by radiologists under standard care³, so the tool adds speed rather than replacing the read.

The teleneurology result raises a second effect. Where the software was present at non-thrombectomy hospitals, fewer patients were accepted for thrombectomy⁵. The abstract does not say whether this reflects fewer inappropriate transfers or missed candidates, so the net effect on patients is unknown.

Finally, much of the published evidence has links to vendors. One author of the Miami study is employed by Viz.ai⁴, the teleneurology authors are linked to the care provider studied⁵, and English sites were recruited partly through the supplier Brainomix². Dedicated searches for withdrawals, recalls, regulator action and lawsuits found none in the sources read; the searches returned mostly journal articles, so the absence of such reports is not proof that none exist.

Phase: In wide use — why

The hypothesis of In wide use is kept. NHS England states that AI decision-support covers every regularly admitting stroke service in England⁷, and evaluation sites used the software on two-thirds of cases by March 2023². In the United States, authorisation dates from 2018³ and several centres have published results, but the national number of sites was not established. The regulator and NHS England both describe the products as AI, but the internal methods of each product were not independently checked. The phase rests mainly on England's rollout.

What to watch next

• 2 August 2028: the date from which the EU AI Act's high-risk rules apply to AI in products covered by Annex I product legislation⁹.
• Release of effect sizes and confidence intervals for functional outcome from the peer-reviewed study of English hospitals, which NHS England summarised without them in December 2025⁷.
• Any randomised trial reporting 90-day functional outcome with enough patients to narrow the wide interval seen in the Houston trial¹.
• Any audit of English door-in-door-out times after the Summer 2024 national rollout, which would test whether the uneven pattern in the evaluation sites persisted²⁷.

Questions to ask before adopting

• Which step in our stroke pathway is the real delay: detection, communication, transfer or treatment, and does a scan alert address it?
• What are the false-positive and missed-occlusion rates on our own scans, not only in vendor or regulator studies?
• Do we have a baseline and a comparison group to separate the tool from other improvement work running at the same time?
• Who reviews every alert at night and at weekends, and what happens when the tool and the clinician disagree?
• What will licence, integration and training cost, and how will we check that patient recovery, not only time, improves?