Getting health technology through the hospital door: what a mathematician and a surgeon look for

The request did not come from the mathematicians. According to Stéphanie Allassonnière, it came from the maternity unit at Necker hospital in Paris, where Yves Ville, the head of department, and two colleagues, Julien Stirnemann and Emmanuel Spaggiari, explained that their prenatal diagnosis centre was overloaded. Some parents only needed reassurance, others needed specialist care, and everyone was seen in the same queue. The mathematicians had to turn that request into a question they could answer, in constant exchange with the clinical team.

That work became Sonio, a company founded in 2020 that makes software to assist fetal ultrasound examinations. Samsung Medison agreed to acquire it in 2024 for around $92 million. Allassonnière, professor of applied mathematics at Université Paris Cité and deputy director of the PR[AI]RIE artificial intelligence institute, is one of its co-founders. In a three-part French-language Tech Anatomy series, co-hosted with Guillaume Laguette, chief revenue officer of the handheld ultrasound start-up echOpen, she and Eric Vibert, liver and transplant surgeon at Paul-Brousse hospital (AP-HP) and founder of BOPA, a chair dedicated to the augmented operating room, discussed why so many health technologies never get past the hospital door.

Three conditions before performance

Vibert sets out three conditions for an innovation to become progress: clinical utility, either because the problem is common or because it is severe; acceptability to the people who will use it; and interoperability with hospital systems.

His own example of failure was a "black box" for the operating theatre. The concept was sound, he says, but staff did not want to be monitored, because "an operating theatre is not an aircraft cockpit" (quotations translated from French). Founders, he adds, underestimate how hard it is to plug into a hospital information system, which will not adapt to them: "It's never Orbis that will change," he says, naming the patient record software used across AP-HP.

Allassonnière's version of the same point comes from Sonio. The screen the sonographer looks at during a pregnancy scan is the one the parents are watching, so nothing could flash an alert. Clinicians who planned to use the tool after the scan found it was too late: the mother could not be called back.

Evidence for a product that ages in two years

Allassonnière describes the mismatch between how long a drug and an algorithm stay relevant. A molecule justifies an 18-month trial. An algorithm, in her estimate, has a useful life of about two years, and "the object itself becomes obsolete during the testing period." Evaluation, she argues, has to become as dynamic as the technology.

Because Sonio targets rare fetal diseases, a prospective trial was not feasible, and she says regulators accepted a retrospective study. The team also ran "challenges": about twenty fetal ultrasound experts worked through scenarios drawn from documented fetal pathology cases, some with the algorithm and some without. Sonio's products have since received several FDA clearances, including one in 2025 for an anomaly detection tool supported by a multicentre reader study, according to trade press.

Her research group applied the same logic to synthetic data. Using pre-anaesthesia questionnaires from the university hospital in Toulouse, it generated artificial patients and, she says, asked three local anaesthetists to separate 50 real patients from 50 generated ones. They made about 50 percent errors and disagreed with one another. The underlying proof-of-concept was published in the Journal of Medical Internet Research in 2025, based on 521 real patients. In the first episode she describes the next step: in a rare disease or narrowly targeted cancer trial, recruiting the full treated arm but only twenty or thirty controls, then completing the control arm with generated patients built on what is already known about placebo or the comparator drug.

Vibert shifts the emphasis. Before testing clinical benefit, he says, test whether people adopt the tool and whether it works with hospital systems. He also wants outcomes measured with patient-reported measures rather than length of stay alone: "it's not only about making the hospital money, it's about treating patients well."

Where funding and regulation pull

Asked whether a start-up should build credibility in the United States first, the guests answer no and argue for Europe, while acknowledging that necessity pulls companies across the Atlantic. The first reason given is regulatory. With CE marking, a file goes to a notified body and comes back as a yes or a no, while the FDA engages in a structured back-and-forth that improves the dossier. Funding is the second factor, as with Nabla, which started in France, grew in the US and is now returning. The discussion then turns to the depth of French talent and to AMI Labs, the Paris-based company founded by Yann LeCun, which raised $1.03 billion in March 2026 and is led by Alexandre LeBrun, formerly chief executive of Nabla.

After the door

Once a tool is inside, both guests stress that the users are often not doctors: sonographers in the United States, theatre nurses, or the transplant coordinator with whom Vibert's team built an AI agent. A new tool also has to remove a task, he says, not add one more layer.

Vibert asks why French teaching hospitals do not issue professional smartphones connected to the patient record and the imaging archive, given how much unofficial use of general chatbots already happens on personal phones. Allassonnière cites Cedars-Sinai in Los Angeles, which, as she was told, blocked public chatbots and offered an internal model on work phones instead. Cedars-Sinai reports more than 6,000 staff using its proprietary GPT platform; we found no published policy matching the ban she described.

Vibert's other answer is the experimentation site. BOpEx, a test bed for digital tools in the operating room run by AP-HP with the Medicen cluster and Université Paris-Saclay, tests reasonably mature products in three or four AP-HP sites for about six months. The aim, he says, is to replace informal one-to-one pilots with a single clinical champion, which often end with the free period. He is explicit that entering BOpEx is not a route into AP-HP purchasing.

Competing readings

One reading of the series is that adoption is mainly a design and integration problem: get the workflow, the screen and the connection to the record right, and evidence follows. The other is that evidence remains the gate, and that what has to change is the evaluation method, through retrospective designs, reader challenges and synthetic control arms. Vibert leans towards the first, Allassonnière towards the second, and both treat them as linked.

What the series does not establish

The episodes reflect the experience of two people with a stake in the solutions they describe, not a review of adoption studies. The Toulouse figures and the two-year lifespan of an algorithm are their own accounts or estimates. Sonio, BOpEx and Vibert's start-up Uncover are described by people involved in them.

What to watch

The indicators that would separate the two readings: whether test beds such as BOpEx publish their six-month reports and how many tested products are later bought, whether regulators accept synthetic or augmented control arms in rare disease trials, and whether French hospitals equip staff with connected professional devices. If integrated tools still struggle without new forms of evidence, the evaluation question will take precedence over the design one.

Listen to the episode

Franchir la porte de l'hôpital - Stéphanie Allassonnière & Eric Vibert (ft. Guillaume Laguette)

Listen · 25 min