AI stroke detection software reads CT angiograms for large vessel occlusions and alerts the stroke team within minutes. In a Houston cluster-randomised trial, it cut door-to-groin time by 11.2 minutes. Those minutes help patients only if the thrombectomy pathway can use them: ambulances for transfer, an on-call interventional team and a free angiography suite.
A 68-year-old woman arrives at a small community hospital at 2 a.m. She cannot move her right arm and her speech is garbled. The CT angiogram is done within half an hour. Before a radiologist has opened the images, an app on the stroke neurologist’s phone at the regional thrombectomy centre buzzes: suspected large vessel occlusion, left middle cerebral artery. The neurologist reviews the images on her phone and calls the community hospital. Everyone agrees: this patient needs a thrombectomy.
Then the clock starts again, and this time software cannot help. The patient needs an ambulance crew able to transfer her. The thrombectomy centre needs an interventional team out of bed and an angiography suite that is not already in use. Each of those steps can take longer than the time the AI just saved.
That is the argument of this piece. AI stroke triage is one of the most mature forms of clinical AI, with real, measured gains in speed. But speed-to-alert only matters if the rest of the pathway can absorb it. In many regions, the minutes AI saves are handed straight to a queue for a transfer or a team that is not yet there.
How does AI detect large vessel occlusion strokes?
Large vessel occlusion, or LVO, is a blockage of one of the major arteries supplying the brain. These strokes cause severe disability and are the ones most likely to benefit from mechanical thrombectomy, in which a catheter is threaded into the artery to remove the clot. Benefit falls steeply with delay. Pooled analyses of the landmark thrombectomy trials, including the HERMES collaboration’s analysis in JAMA in 2016, showed that faster reperfusion means better outcomes.
AI triage software analyses CT angiogram images as soon as they are acquired, flags a suspected LVO and pushes an alert, with images, to the stroke team’s phones. The FDA authorised the first such product, Viz.ai’s ContaCT, through the de novo pathway in February 2018, creating a regulatory category for computer-aided triage. Many similar tools have followed.
It is worth knowing how that category is regulated. A 2022 analysis in Scientific Reports of AI devices cleared by the FDA and Japan’s regulator found that US computer-aided triage devices were cleared on the basis of standalone software testing. In plain terms, regulators checked whether the software could find the occlusion on images. They did not require evidence that patients did better.
Does AI stroke triage improve patient outcomes?
The best evidence is a cluster-randomised stepped-wedge trial published in JAMA Neurology in November 2023 by Juan Carlos Martinez-Gutierrez and colleagues. It switched on automated LVO detection with secure group messaging at four comprehensive stroke centres around Houston, Texas, in random order between January 2021 and February 2022.
Among 243 patients who arrived through the emergency department and were treated with thrombectomy, median door-to-groin time before the AI was 100 minutes. After the AI went live, door-to-groin time fell by 11.2 minutes, and the time from starting the CT scan to starting thrombectomy fell by 9.8 minutes. In an exploratory analysis there was no statistically significant difference in functional independence at 90 days, with an odds ratio of 1.3 and a wide confidence interval from 0.42 to 4.0. The trial was not large enough to answer that question.
Eleven minutes is clinically meaningful for a time-dependent treatment. But notice who the trial excluded. Patients transferred from outside hospitals, 158 of them, were not in the analysis. That was a reasonable design choice for a clean trial. It also means the best randomised evidence says little about the part of stroke care where delays are usually longest: getting a patient from a hospital without thrombectomy to one that has it.
Where do the minutes saved by stroke AI actually go?
Think of the thrombectomy pathway as a relay with five legs: recognition, imaging, decision, transfer and treatment. AI speeds up the handoff between imaging and decision. That is valuable. It does not touch the other legs.
Transfer. For patients who arrive at a hospital without thrombectomy, the time between arriving and leaving, often called door-in-door-out, is frequently dominated by waiting for transport. An alert that arrives twenty minutes earlier saves nothing if the ambulance is dispatched at the same time it would have been anyway.
Team availability. At night and at weekends, interventional teams are often on call from home. Earlier alerts can wake them earlier, which is the point. But if the team’s travel time is longer than the saved interval, or if a second patient is already on the table, the patient waits.
Suite capacity. Angiography suites are shared with cardiology and other interventions in many hospitals. More identified candidates means more competition for a finite room.
Workforce. The number of trained interventionalists limits how many thrombectomies a region can perform around the clock. AI can increase the number of patients recognised as candidates. It cannot increase the number of people qualified to treat them.
This is the pattern we described in predicting hospital capacity is the easy part. Detection is the cheap step. Response capacity is the expensive one.
A worked example: one patient, two regions
Hypothetical, with illustrative timings. Our patient arrives at the community hospital at 2:00 a.m. The CT angiogram finishes at 2:25.
In Region A, before AI, a teleradiologist reads the scan at 2:50 and calls the thrombectomy centre at 2:55. With AI, the alert reaches the stroke neurologist at 2:30. Twenty-five minutes saved.
Region A has a dedicated critical-care transfer service that commits to dispatch within 15 minutes of a request, and an in-house interventional team overnight. The patient leaves at 2:50 and arrives at 3:30. Groin puncture happens at 3:45. Without AI, it would have been about 4:10. The 25 minutes survived the whole journey.
Region B uses the same software. The alert still reaches the neurologist at 2:30. But transfers depend on the general ambulance service, which is stretched overnight; the crew arrives at 3:40, the same time it would have without the AI, because the request only joined a queue earlier. The interventional team is called at 2:35 and is ready at 3:20, then waits. Groin puncture happens at 4:40. Without AI, it would have been about 4:45. Five minutes saved.
Same algorithm, same scan, same alert time. In one region the AI bought 25 minutes of brain; in the other, it bought five and a lot of waiting.
Isn’t every minute worth it, whatever else is broken?
The strongest counterargument is that “time is brain” applies to every minute. Even five minutes saved has value. AI triage is relatively cheap compared with building new thrombectomy centres or hiring interventionalists, and it improves communication, including the images-on-phone feature that lets specialists make remote decisions. It may also help patients who would otherwise be missed altogether, when a scan read by a non-specialist overnight fails to flag an occlusion. Waiting to fix transfer systems before buying AI would mean patients lose minutes that software could save today.
All of that is fair, and no one should argue against buying AI stroke triage on this basis. The point is about expectations and priorities. Health systems often present AI triage as a stroke strategy. It is one component. When regions buy the software and report faster alert times while door-in-door-out times and overnight team availability are unchanged, they risk believing a problem is solved when only its cheapest part has moved. There is also the question, which we explored in our debate on whether AI should make triage decisions, of who is accountable when an alert is missed or wrong.
What this means for stroke networks, hospitals and policymakers
For stroke network leaders, measure the whole relay. Track time from CT to alert, alert to transfer request, request to ambulance departure, door-in-door-out, arrival to groin puncture, and outcomes, by hospital and by hour of day. The weakest leg is where the next investment should go, and it will rarely be the software.
For community hospitals without thrombectomy, the AI alert is a trigger for a pre-agreed transfer protocol, not a notification. Negotiate transfer commitments with ambulance providers based on the alert, rather than waiting for a formal radiology report.
For thrombectomy centres, earlier alerts should translate into earlier team activation. That requires clear rules on who acts on an AI alert at 2 a.m. and what level of confidence triggers a call-in, balanced against false alarms that wake a team for nothing.
For policymakers and payers, fund pathways, not tools. Paying for AI detection while leaving interhospital transfer under-resourced buys speed on the cheapest leg. Regional planning of thrombectomy capacity, transfer services and interventional workforce matters more than the choice of software vendor.
For the public, the most important minutes are still the first ones. Recognising the signs of stroke, such as face drooping, arm weakness and speech difficulty, and calling emergency services immediately, saves more time than any algorithm.
Frequently asked questions
What is a large vessel occlusion stroke?
A large vessel occlusion is an ischaemic stroke caused by a blockage in one of the major arteries supplying the brain, such as the internal carotid or middle cerebral artery. These strokes tend to be severe. They are the main candidates for mechanical thrombectomy, a procedure that removes the clot with a catheter, and outcomes depend heavily on speed.
How accurate is AI at detecting large vessel occlusion?
Commercial LVO tools generally report high sensitivity and specificity in validation studies, but performance varies by product, occlusion location and scan quality. They are cleared as triage aids, not diagnostic replacements, so a radiologist or stroke physician still reviews the images. More distal occlusions are typically harder for software to detect than proximal ones.
Does AI stroke software reduce time to treatment?
Yes, in the best trial so far. A 2023 cluster-randomised trial at four Houston stroke centres found door-to-groin time fell by 11.2 minutes after AI-enabled LVO detection and group messaging went live. The trial was too small to show a difference in 90-day functional outcomes and excluded patients transferred from other hospitals.
Why can’t every stroke patient get a thrombectomy quickly?
Thrombectomy requires a specialised centre, an interventional team and an angiography suite. Many patients first reach hospitals without these and must be transferred, which can take hours. Overnight staffing, ambulance availability and geography all limit access, so regional organisation of stroke care matters as much as speed of diagnosis.
What should I do if I think someone is having a stroke?
Call emergency services immediately. Common signs include sudden face drooping, arm or leg weakness, especially on one side, and difficulty speaking or understanding speech. Note the time symptoms started. Do not wait to see whether symptoms pass, because stroke treatments are most effective when given as early as possible.
AI stroke triage is a genuine advance that saves real minutes. The question for every stroke network is not whether to buy it, but whether anyone is waiting at the other end of the alert.
Related on the GHC blog: Paying for Outcomes Rents Behaviour. It Doesn’t Buy It. and Primary Care Is the Cheapest Hospital Bed, But Only If You Close One.
In your region, what is the slowest leg of the thrombectomy relay, and has AI changed it?
