AI Stroke Alerts in the ED: Faster Transfers, Silent Misses

The 66-year-old welder came in by ambulance with a right gaze preference, left arm drift, and a NIHSS of 14. His wife puts last known well at 70 minutes before arrival. The CTA finished twelve minutes ago, the tenecteplase bolus went in five minutes ago, and the stroke app on your phone has not alerted. The transfer center at the thrombectomy hospital is on the line asking whether he has a large vessel occlusion.

Bottom line: use the automated LVO alert to start the team, and never use its silence to stop the workup. When the exam points to a large vessel occlusion, get a human read of the CTA right away, and treat a missing alert as no information, especially for M2 and posterior circulation occlusions.

What the Houston trial tested

Martinez-Gutierrez and colleagues ran a cluster randomized stepped-wedge trial at 4 thrombectomy-capable stroke centers in the Houston area from January 1, 2021, to February 27, 2022 [1]. The hospitals switched on Viz LVO in a randomized order, one at a time. After activation, noncontrast CT and CTA images went automatically to a cloud algorithm. Clinicians and radiologists then received phone alerts for a possible LVO, along with mobile image viewing and a secure group message thread. Average processing time was under 5 minutes [1].

The trial enrolled patients who arrived through the ED with an occlusion of the internal carotid, middle cerebral, anterior cerebral, basilar, intracranial vertebral, or posterior cerebral artery and who went on to thrombectomy [1]. Of 443 such patients, the investigators excluded 158 outside transfers, 39 in-hospital strokes, and 3 large-core trial patients. That left 243 for analysis, 140 before activation and 103 after [1]. The median NIHSS was 17, and 46% received IV thrombolysis [1].

Median door-to-groin time fell from 100 to 88 minutes. The adjusted model estimate was −11.2 minutes (95% CI, −18.2 to −4.2) [1]. CT-to-groin time fell by 9.8 minutes, and door-to-needle time did not change significantly (−3.5 minutes; 95% CI, −9.1 to 2.2) [1].

Subgroups split the effect. Patients who did not receive thrombolysis gained 14.3 minutes (95% CI, −24.5 to −4.2). Patients who did gained 6.6 minutes, and that interval crossed zero (95% CI, −19.2 to 3.0) [1]. Anterior circulation occlusions gained 12.9 minutes, and patients with NIHSS above 10 gained 12.3 [1].

Ninety-day functional independence was 32% before activation and 42% after, a difference that was not significant (adjusted OR 1.3; 95% CI, 0.42 to 4.0) [1]. Mortality was 31% versus 13%. That comparison was unadjusted, and the paper reports no odds ratio for it [1]. Symptomatic hemorrhage was 5% versus 2%, also not significant [1]. All 4 hospitals shared one health system, one EHR, and rotating physicians, and the intervention was unblinded [1]. The authors concluded that the software improved in-hospital workflows and treatment times, with no significant change in 90-day function [1].

What the software misses

The trial's definition of LVO included the basilar, vertebral, anterior cerebral, and posterior cerebral arteries [1]. Accuracy studies often use a narrower definition. Karamchandani and colleagues reviewed 3,851 consecutive code stroke CTAs from 2021 across the Atrium Health network, which spans urban and rural hospitals. They compared Viz LVO against fellowship-trained neuroradiologist reads [2]. Their primary analysis counted only ICA and M1 occlusions as LVOs. M2 branches, the ACA, the basilar artery, and high-grade stenosis did not count [2].

For ICA and M1 occlusions, sensitivity was 78.2% and specificity was 97%. The software missed 48 of 220 occlusions and generated 109 false alerts, for a positive predictive value of 61% [2]. When proximal M2 occlusions were added to the definition, sensitivity fell to 60.6%: 141 of 358 occlusions produced no alert [2]. Misses came from branch occlusions, segmentation errors, and short or hidden occlusions. False alerts came from poor contrast timing, segmentation errors, and stenosis read as occlusion [2]. The authors described the tool as highly specific and moderately sensitive [2].

That gap matters for ED decisions because the 2026 AHA/ASA guideline still calls EVT reasonable for dominant proximal M2 occlusions [3].

Where the time savings are largest

The same Houston group studied 7 primary stroke centers that do not perform thrombectomy [4]. In 115 patients with LVO who were transferred for possible EVT, median door-in-door-out time fell from 210 minutes before activation to 133 minutes after. The adjusted estimate was −106 minutes (95% CI, −165 to −48) [4]. EVT after transfer rose from 41.2% to 62.9%, but the adjusted odds ratio of 2.13 (95% CI, 0.88 to 5.13) was not significant [4]. Mortality (18.8% vs 17.1%) and symptomatic hemorrhage did not differ, and onset-to-groin time did not improve [4]. This was a pre-post registry study within a single health system [4].

What the 2026 guideline asks of the ED

The guideline sets clear expectations, and the software serves them:

• Brain imaging: protocols should get it done as fast as possible, for example within 25 minutes (COR 1) [3].
• Vascular imaging: CTA or MRA of the cervical and intracranial vessels as fast as possible for suspected LVO within 24 hours, without waiting for a creatinine (COR 1) [3].
• Transfer hospitals: primary stroke centers that rely on transfer must be able to rapidly perform and interpret CTA or MRA, and should have agreements that shorten door-in-door-out times (COR 1) [3].
• EVT within 6 hours: for ICA or M1 occlusion with NIHSS of 6 or more, prestroke mRS 0 to 1, and ASPECTS 3 to 10 (COR 1) [3].
• Dominant proximal M2: EVT is reasonable within 6 hours with NIHSS of 6 or more, prestroke mRS 0 to 1, and ASPECTS of 6 or more (COR 2a) [3].
• Basilar occlusion: EVT within 24 hours with NIHSS of 10 or more, baseline mRS 0 to 1, and PC-ASPECTS of 6 or more (COR 1) [3].

The guideline cites the Houston trial for faster team activation. It also lists AI in stroke recognition and treatment as a knowledge gap that needs more research [3]. None of its recommendations depend on an alert firing.

Building the alert into ED workflow

Four rules keep the software working for the clinician:

• Read on clinical suspicion. Any patient within 24 hours of last known well with a severity scale or exam that suggests LVO gets a human CTA read at once, whether or not the alert fired.
• Time the human read. Track CTA-to-read time separately from CTA-to-alert time, so a fast alert cannot hide a slow read.
• Audit both directions. Review every confirmed LVO with no alert and every false alert each month. Track M2 and posterior circulation misses separately.
• Measure your baseline. The Houston centers started near 100 minutes door-to-groin and 210 minutes door-in-door-out [1,4]. A department already well below those numbers has less time to gain.

Key Takeaways

• In the only randomized trial, automated LVO alerts cut median door-to-groin time from 100 to 88 minutes, with no significant change in 90-day function.
• For ICA and M1 occlusions, sensitivity was 78.2%. Once proximal M2 occlusions were included it fell to 60.6%, so about 4 in 10 of those occlusions produced no alert.
• At non-thrombectomy hospitals, door-in-door-out time fell by an adjusted 106 minutes, which makes the transfer ED the biggest winner.
• Audit monthly: every confirmed LVO with no alert and every false alert, with M2 and posterior circulation misses tracked separately.

FAQ

Does AI stroke detection software improve patient outcomes?

In the one randomized trial, automated LVO detection cut door-to-groin time by an adjusted 11.2 minutes. Ninety-day functional independence was 42% versus 32%, which was not statistically significant (adjusted OR 1.3; 95% CI, 0.42 to 4.0) [1]. The time savings are proven, and an outcome benefit has not been shown.

How accurate is Viz LVO for detecting large vessel occlusion?

In 3,851 code stroke CTAs read against neuroradiologists, sensitivity was 78.2% and specificity was 97% for ICA and M1 occlusions [2]. With proximal M2 occlusions included, sensitivity fell to 60.6% [2]. Positive predictive value was 61% in the primary analysis.

Should emergency physicians wait for the AI alert before calling for transfer?

No. The 2026 AHA/ASA guideline expects transfer hospitals to rapidly perform and interpret CTA themselves and to shorten door-in-door-out times through transfer agreements [3]. In the largest accuracy study, the software missed 48 of 220 ICA and M1 occlusions [2].

Dr. Chet's Take

I have watched stroke transfers from two chairs, as a HEMS medical director whose crews move LVO patients between hospitals and as a TelEmergency medical director supporting rural emergency departments that do not have a neurologist in the building. This article gets the core argument right. The software is a messenger. It shortens the time it takes for the right people to learn that a patient needs a thrombectomy hospital. It does not examine the patient, weigh his prestroke function, or talk to his family about a helicopter. The value is greatest where expertise is thinnest, at the small hospital where the emergency physician is the only physician in the building at night.

That being said, I think the article understates the human problem. When an alert lands in a group thread, everyone on the thread assumes someone else is making the call. That is diffusion of responsibility with a push notification attached. The honest answer is that I do not know whether the minutes saved come from the algorithm or from the fact that the alert finally put the radiologist, the neurologist, and the interventionalist in one conversation. If the communication is doing the work, a department without the software can get part of the benefit by fixing who calls whom and when. Rural hospitals also pay for the software from the same budget that staffs the CT scanner at night, and a fast alert on a scan that started forty minutes late helps nobody.

If you are leading an emergency department that sends stroke patients out, pull your last twenty LVO transfers and time each step from door to departure. Then find where the minutes went. The read is one step. The callback, the acceptance, and the ride are three more, and the software touches only the read and the callback. Share that timeline with your transfer center and your flight program, because they own the minutes after your door. If you are the attending on the next one, call the thrombectomy center the moment you see the occlusion, before the alert and before the formal read if you are confident. The emergency physician who sees the occlusion owns the clock until the patient leaves the department.

— Chester Shermer, MD, FACEP | Emergency Medicine, 25+ Years Clinical Experience | State Surgeon

Continue Your Training

If you're an emergency physician (or any clinician treating patients daily) trying to understand how AI will actually impact your clinical practice — not just the hype — I put together a free practical guide. You can download the AI in EM Survival Guide here.

Browse All Courses at Global MedOps Command

Relevant Reading on Global MedOps Command:

How to Avoid Becoming an AI Casualty — Dr. Shermer's guide to navigating AI tools in clinical and operational settings without compromising judgment or patient outcomes.

Emergency Department Efficiency Playbook — Practical systems for throughput, triage optimization, and operational efficiency.

Read more from Dr. Shermer on Medium

Connect with Dr. Shermer: LinkedIn — Chester "Chet" Shermer, MD, FACEP.

References

1. Martinez-Gutierrez JC, Kim Y, Salazar-Marioni S, et al. Automated large vessel occlusion detection software and thrombectomy treatment times: a cluster randomized clinical trial. JAMA Neurol. 2023;80(11):1182-1190. PubMed 37721738
2. Karamchandani RR, Helms AM, Satyanarayana S, et al. Automated detection of intracranial large vessel occlusions using Viz.ai software: experience in a large, integrated stroke network. Brain Behav. 2023;13(1):e2808. PubMed 36457286
3. Prabhakaran S, Gonzalez NR, Zachrison KS, et al. 2026 guideline for the early management of patients with acute ischemic stroke: a guideline from the American Heart Association/American Stroke Association. Stroke. 2026;57(8):e316-e436. PubMed 41582814
4. Le NM, Iyyangar AS, Kim Y, et al. Machine learning-enabled automated large vessel occlusion detection improves transfer times at primary stroke centers. Stroke Vasc Interv Neurol. 2024;4(3):e001119. PubMed 41584488

Related simulation training: EM-Sim · EMS-MedSim · Books by Dr. Shermer

Related reading: When an AI Model Fails Silently: Detection and Response Protocols for Emergency Physicians