The math is unforgiving. For every 15 minutes a stroke goes untreated, a patient loses roughly a month of disability-free life. Which makes what happens in the back of an ambulance—before the CT scan, before the hospital, sometimes before anyone is certain what's wrong—perhaps the most consequential window in stroke care.
A French startup thinks it has found a way to shrink the guesswork. AI-Stroke, based in Montpellier, closed a $4.6 million seed round on November 18 to commercialize software that transforms any tablet or smartphone into what it calls an "AI neurologist." The tool is designed for the chaotic minutes when emergency responders are trying to decide whether a patient needs a stroke center capable of advanced intervention or can go to a closer facility without specialized capabilities.
The funding, led by Heka—Newfund's investment vehicle focused on brain-tech ventures—also drew backing from Bpifrance and a handful of undisclosed angel investors. AI-Stroke will use the capital to navigate FDA regulatory approval and launch multi-site clinical trials across leading U.S. stroke centers, according to CEO Cédric Javault.
When Every Minute Costs a Month
Stroke detection in the field has long been a weak link in emergency medicine. Paramedics rely on manual assessments—acronyms like FAST (Face, Arms, Speech, Time) or CPSS (Cincinnati Prehospital Stroke Scale)—that are easy to administer but surprisingly prone to error. A study of New York City Fire Department responses found that emergency personnel missed more than a third of strokes, detecting them with just 62.4% sensitivity. In South Carolina, an analysis of over 100,000 cases showed similar results: sensitivity hovered around 59%.
Those missed cases matter. Stroke patients routed to the wrong hospital lose time. Some never recover what those delays cost them.
AI-Stroke's software guides users—EMS personnel, or even bystanders in a pinch—through a 30-second video assessment. The phone's camera captures three classic stroke indicators: whether one side of the face droops, whether an arm drifts downward, whether speech slurs or garbles. The system outputs a stroke probability and likely type within seconds, aiming to inform triage decisions before the ambulance pulls away from the curb.
In trials involving more than 2,000 French first responders (conducted with SDIS, the French fire and rescue service), the AI demonstrated 82% sensitivity compared to 68% for human responders. Specificity was 80% versus 79%. Accuracy: 84% against 74%. The algorithms were trained on over 20,000 videos and 6 million images, collected in partnership with CHU Nîmes, a regional hospital center in southern France.
A Polytechnique Pedigree Meets a Clinical Problem
Javault, who studied at École Polytechnique and the Corps des Mines before completing a master's in AI in 2021, founded AI-Stroke in 2022. He has assembled a clinical advisory board heavy with stroke care luminaries—including Dr. Gregory W. Albers, who directs the Stanford Stroke Center and co-founded RapidAI; Prof. Gary Ford of Oxford University Hospitals, a co-developer of the FAST protocol itself; and Dr. Sean I. Savitz, who leads the Institute for Stroke and Cerebrovascular Disease at UTHealth Houston. Dr. Pankajavalli Ramakrishnan, a neurointervention specialist at Westchester Medical Center Health Network, also advises.
The startup is threading a narrow path between pre-hospital diagnostics and the hospital-based AI tools that have attracted hundreds of millions in venture capital. Companies like RapidAI and Viz.ai analyze CT and CTA scans after patients arrive—valuable, but too late to fix routing mistakes. AI-Stroke is betting on a different moment: the few minutes when decisions get made in the field, before imaging is even an option.
Forest Devices, which secured FDA Breakthrough Device designation for its AlphaStroke LVO detection system, operates in adjacent terrain. So does CVAid Medical, which raised $4 million in 2022 for a telestroke platform. But the market for pre-hospital stroke tools remains relatively uncrowded, at least compared to the feeding frenzy around in-hospital AI diagnostics.
The Regulatory Road Ahead
Before the seed round, AI-Stroke had already pulled in €1.5 million in non-dilutive funding from Bpifrance under France's Deep Tech Plan in March 2025. The company has collected a string of government grants—French Tech Emergence, French Tech Seed, i-Lab—that have kept it moving without heavy equity dilution. In January, it joined a cohort of French startups at J.P. Morgan Week in San Francisco, part of a Business France delegation.
The new funds will support an FDA submission, though Javault has not yet disclosed whether the company will pursue a De Novo pathway for novel devices or attempt to clear via 510(k) by showing equivalence to an existing predicate. The choice matters—De Novo offers more certainty but takes longer; 510(k) is faster but riskier if the FDA disagrees about your predicate.
Clinical trials across multiple U.S. sites will run in parallel. According to a Bpifrance profile, AI-Stroke is targeting commercialization by the end of 2026, an ambitious timeline given regulatory uncertainties and the typically glacial pace of hospital adoption cycles.
Whether the company hits that mark may depend less on the technology than on the usual impediments: reimbursement codes, liability concerns, resistance from EMS systems wary of changing protocols. Then again, the current protocols miss a lot of strokes. That's the opening AI-Stroke is betting on—one 30-second video at a time.
