Pre-Hospital Stroke Triage: The Prediction Is Improving, the Destination Question Is Not Settled

The problem

Starting point. The anchor paper reviewed software that reads a brain scan after a stroke patient reaches hospital and alerts the stroke team to a blocked artery¹. In a four-hospital randomised trial in Houston (January 2021 to February 2022, 243 patients), adjusted door-to-groin time, the time from arrival to puncture of the artery for clot removal, fell by 11.2 minutes (95% CI -18.22 to -4.2), while the exploratory odds ratio for 90-day functional independence was 1.3 (95% CI 0.42 to 4.0), reported as no difference². The US regulator describes the first such product as notification-only³. This paper adds a different question: whether AI could help decide, before any hospital scan exists, which hospital the patient is taken to.

Why the destination matters. Clot removal, called thrombectomy, is available only at some hospitals. A patient taken to the nearest stroke centre that cannot perform it may need a second ambulance journey. Software that acts after arrival, as in the anchor paper, cannot shorten that journey.

How large the sorting task is. Most suspected strokes do not need thrombectomy. In Kawasaki City, Japan, 207 of 1,653 transported patients (12.5%) had a large vessel occlusion, a blockage of a major brain artery (April 2023 to March 2024, retrospective analysis, abstract)⁷. In two Dutch ambulance regions, 538 of 3,320 ambulance-activated stroke codes (16%) were either such a blockage or a brain haemorrhage (July 2018 to October 2019, cohort data analysed with cross-validation, abstract)⁶. The task is to pick out a minority of patients, quickly, from information available at the roadside or in the ambulance. The evidence found comes from Spain, the Netherlands, Japan, Germany, Australia and the United States.

What already works

Scales and models that score patients. Ambulance crews already use simple scoring scales to flag a likely blockage. In the Dutch data, balanced accuracy (the average of sensitivity and specificity) on routine variables was 82.1% for a machine-learning model, 79.0% for logistic regression and 74.5% for the RACE scale, a bedside stroke scale; the model reached sensitivity of 76.7% and specificity of 92.1% (abstract)⁶. The authors describe the gains as limited and call for external validation and prospective implementation studies⁶. In Kawasaki, a simple scale reached an area under the curve of 0.768 for blockage (abstract), and the study compared no machine-learning model⁷. A Madrid study of ambulance records reported 91.67% recall and 64.71% precision for a blockage model, outperforming the reference scale on all metrics except specificity (abstract)⁸. In 88 patients in a Sydney emergency department, a model using pulse-sensor signals reached a median test area under the curve of 0.77 (95% CI 0.71 to 0.82); the authors name the small sample as the main limitation, and the setting was a hospital, not an ambulance (arXiv version)⁹.

These figures mean that machine learning has beaten bedside scales in retrospective data, by a margin the authors call modest, and that none of it has yet been tried live in an ambulance.

Taking patients straight to a thrombectomy centre. The RACECAT trial in Catalonia randomised 1,401 patients with suspected blockage to a thrombectomy-capable centre (688) or the closest local stroke centre (713), between March 2017 and June 2020⁴. The adjusted common odds ratio for 90-day disability score was 1.03 (95% CI 0.82 to 1.29), not statistically significant⁴. Thrombectomy was given to 48.8% against 39.4% (OR 1.46, 95% CI 1.13 to 1.89), and clot-dissolving drugs to 47.5% against 60.4% (OR 0.59, 95% CI 0.45 to 0.76), in direct-transport versus local-centre order as the paper lists them⁴. Enrolment was halted for futility, and allocation could not be blinded⁴. This result is more than three years old.

Bringing the scanner to the patient. A mobile stroke unit is an ambulance with a CT scanner. A Berlin trial presented in February 2020 reported that availability of mobile stroke unit transport reduced disability and death at three months (adjusted OR 0.72, 95% CI 0.60 to 0.87), with thrombolysis in 60% against 48% (conference abstract, as reported by Medscape)¹⁰. A Cleveland Clinic cost model, based on 400 patients in 2014 and 2015 and summarised by the clinic (operator-reported), put annual cost at $856,482 for the mobile unit against $785,869 for standard transport, and found the unit cheaper only above thresholds such as 391 stroke patients a year¹¹.

What does not exist. No trial was found in which an algorithm's output decided which hospital a stroke patient went to. A study of machine-learning blockage detection on mobile-unit scans was identified but could not be opened, so nothing about it is used here.

The possibility

One way this could work: an ambulance crew enters routine findings into a scoring model, or a mobile unit scans the patient and software reads the scan, and the result feeds a destination rule agreed with the receiving hospitals. If the probability of a blockage is high, the patient goes to a thrombectomy centre; otherwise to the nearest stroke centre.

The AI part is narrow. It would be the prediction or detection step: a model that scores the likelihood of a blockage from patient data, or software that flags a blockage on a scan. Everything else would be ordinary systems, logistics and policy: ambulance dispatch, radio or data links, rules for bypassing hospitals, capacity at thrombectomy centres, transfers, payment and the clinical decision itself. If simple scales already sort most patients correctly, AI would be a small part of the solution, and the harder work would sit in the destination rule and the hospital network.

What would have to be true

1. The prediction would have to be clearly better than a simple scale, not marginally. The authors of the Dutch study call the gains limited⁶, and a simple scale in Kawasaki reached 0.768⁷.
2. The prediction would have to hold up live. The studies found are retrospective analyses of past records, and the Dutch authors call for prospective implementation studies⁶.
3. Bypassing the nearest hospital would have to help patients where it is applied. In RACECAT, 90-day disability did not differ⁴, and a later analysis of patients with confirmed blockage found a common odds ratio of 1.05 (95% CI 0.80 to 1.40) for the local-centre route⁵.
4. The cost of lost thrombolysis would have to be small or avoidable. Clot-dissolving drugs were given less often in the direct-transport arm⁴.
5. Thrombectomy centres and transfer services would have to absorb extra arrivals. No capacity figures were found.
6. The cost would have to be acceptable. In the Cleveland model, the mobile unit cost $70,613 more per year in the base case, and the model did not count any improvement in clinical outcomes¹¹.

What could stop it

What the evidence shows

In RACECAT, the primary outcome showed no difference, enrolment was halted for futility and thrombolysis was less frequent in the direct-transport arm⁴. A secondary analysis was post hoc, was not designed to compare hospital characteristics, and its authors say generalising to other health systems is uncertain; the study was supported by an unrestricted Medtronic grant to a foundation⁵. In the same analysis, median door-to-needle time was 47 minutes at telestroke-supported local centres and 30 minutes at primary stroke centres⁵.

The machine-learning studies are small, retrospective or both. The Sydney study had 88 patients⁹, and the Madrid result is abstract-only⁸. The Cleveland model is a hospital-perspective model from 2014 and 2015 data¹¹, and the Berlin result reached this paper through press coverage of a conference abstract¹⁰.

Reasoning

The challenge question is what would make this less useful or more harmful than it first appears. A missed blockage sends a patient who needs thrombectomy to the nearest hospital, which is where they are sent today. A false alarm could send a patient with a haemorrhage or a mimic on a longer journey, and if thrombolysis is delayed, the patient who bears that risk is not the patient the model was built to help. The extra miles also have a price paid by ambulance services and receiving centres, who may not be the ones who gain.

If RACECAT's null result held outside Catalonia, a more accurate model would only reach the right centre faster for patients who may do no better there. Simpler options, such as better scales, telestroke support at local centres and faster transfer, might achieve similar results at lower cost. If bypass helps mainly where distances are short and centres have spare capacity, the model might matter least where the evidence is thinnest.

Signs to watch

• Prospective implementation studies of machine-learning triage in ambulances, which the authors of the Dutch study say are needed⁶.
• Replication of RACECAT-type bypass trials in other health systems, which its authors call for⁴.
• Any trial in which an algorithm's output sets the destination hospital, with 90-day disability as the outcome.
• Cost studies of mobile stroke units that include patient outcomes, which the Cleveland model did not¹¹.
• Studies that compare machine learning against simple scales in the same patients, which the Kawasaki study did not⁷.