Medtech startup AI-Stroke, founded in 2022 by a team of engineers, clinicians, and former investors, has raised a $4.6 million seed round led by Heka, the BrainTech-focused vehicle of Newfund VC, with participation from Bpifrance and angel investors, to bring artificial intelligence to the earliest, most time-critical moments of stroke care: the minutes before a patient even reaches a CT scanner. The company’s tool lets ambulance crews and emergency department teams screen for stroke using nothing more than a smartphone, analyzing facial symmetry, arm movement, and speech in a roughly 30-second video to flag signs of stroke within seconds.
Why Seconds Matter So Much in Stroke Care
Stroke treatment operates on one of medicine’s most unforgiving clocks. Every minute a stroke goes untreated, the brain loses roughly 1.9 million neurons, and the standard clot-dissolving drugs used to treat ischemic strokes lose effectiveness the longer they’re delayed. Paramedics traditionally rely on manual screening checklists—like asking a patient to smile or raise both arms—that are useful but subjective and depend heavily on individual training and experience under pressure.
What the Study Found
In a study involving roughly 2,000 paramedic staff, AI-Stroke’s system detected twice as many true stroke cases compared to standard manual screening protocols. That’s a significant claim: it suggests the AI isn’t just faster than a human assessment, it’s catching cases that trained paramedics using conventional checklists were missing entirely, potentially because subtle facial asymmetry or speech changes are easier for a computer vision model to quantify consistently than for a human eye under stress and time pressure. The model itself was built on what the company describes as the world’s largest clinically annotated stroke-video dataset, comprising roughly 20,000 videos and 6 million images, and the fresh funding is earmarked specifically for expanding clinical validation across multi-center studies in Europe and the United States as AI-Stroke pursues an FDA approval pathway.
A Parallel Effort From Dispatch Centers
Ambulance-based screening is only part of the picture. Separate research has trained AI models directly on emergency call audio, using a dataset of nearly 3,000 calls from 2019 and 2022 that combined transcribed audio logs with structured clinical data. That dispatch-center model achieved a sensitivity of 81 percent and specificity of about 79.6 percent for identifying stroke calls—meaningful accuracy for a tool meant to help 911 dispatchers decide how urgently to route an ambulance, even before paramedics arrive on scene.
How This Fits Into the Bigger Prehospital AI Push
Stroke detection joins a growing list of AI tools designed to compress the “time to treatment” window in emergency medicine generally, alongside AI-assisted ECG devices already used in some ambulances to remotely detect ST-elevation heart attacks before a patient reaches the hospital. The common thread is pushing diagnostic intelligence earlier in the chain of care—out of the hospital and into the ambulance or even the 911 call center—rather than waiting for a patient to arrive at an emergency department equipped with a CT scanner.
The Skeptical View
Clinicians caution that a smartphone screening tool is not a replacement for imaging. Even a highly accurate AI facial/speech assessment can’t distinguish between an ischemic stroke (treatable with clot-dissolving drugs) and a hemorrhagic stroke (where those same drugs would be dangerous)—that distinction still requires a CT scan. The realistic value of tools like AI-Stroke is triage: helping paramedics decide faster whether to divert to a stroke-capable hospital and alert the stroke team in advance, shaving minutes off the process rather than replacing the diagnostic pathway itself.
What’s Next
With fresh funding in hand, AI-Stroke is expected to expand its paramedic pilot programs and pursue regulatory clearance pathways similar to other prehospital AI tools already in ambulances. Longer term, researchers are exploring whether lightweight versions of these algorithms could run on portable CT scanners or edge devices carried directly on ambulances, potentially collapsing the diagnostic gap between “suspected stroke” and “confirmed, treatable stroke” even further before a patient ever reaches a hospital bay. AI-Stroke has said it intends to use part of the new funding to pursue multi-center clinical studies specifically designed to satisfy regulators on both sides of the Atlantic, since a tool used to help triage which hospital an ambulance diverts to carries real consequences if its false-negative rate isn’t rigorously characterized across diverse patient populations, skin tones, and speech patterns rather than just the population represented in its initial training dataset.
Photo: OpenDataInstitute / BY-SA via flickr
