All news

The first autonomous AI clearance makes clinical trials the moat

A cobalt tower on a pale stone island surrounded by a round moat of still dark water

US regulators let IDx-DR diagnose eye disease with no doctor reading the image. Its edge is a trial in primary care, not the best score on archived images.

For the first time, American regulators have let software make a medical diagnosis on its own. The FDA authorised IDx-DR on 11 April: the system reads retinal photographs taken in a family doctor's office and decides, with no specialist looking at the image, whether a patient with diabetes has more than mild diabetic retinopathy and needs an eye doctor.

The decision sets the price of entry for every company selling AI diagnostics, and the price is evidence. IDx did not win with the most accurate algorithm on paper. It won with a prospective trial run where the product will actually be used, and that trial, not the model, is now the asset investors pay for.

The numbers make the point. The pivotal study enrolled 900 patients with diabetes at 10 primary care sites; the software found the disease 87.4% of the time and correctly cleared patients without it 89.5% of the time. Google researchers published higher figures in December 2016, with sensitivity up to 97.5%, after training on 128,175 archived images graded by panels of ophthalmologists. Their scores came from archives. IDx's came from patients walking into family practices, photographed with one specified camera against endpoints fixed in advance, and that is the kind of number a regulator accepts.

The commercial logic follows from who now buys. More than 30 million Americans live with diabetes, and about half of them skip their yearly eye examination. A result that needs no specialist turns the primary care practice into the customer, able to offer a screening it used to refer out, and takes the specialist out of the buying decision. The route also creates a new device category. Rivals with the same intended use can now seek clearance through the shorter path that cites an existing device as precedent, so the first company pays for the road and its followers drive on it. IDx was founded in 2010 and needed eight years to reach this point. Fanaura looks at AI-first companies in regulated markets with that clock in mind, and founders should budget the trial as the largest capital cost in the plan, not as a final formality.

Clearance is not adoption, and that is the strongest argument against the trial-as-moat view. The authorisation covers a single camera model that practices must buy, and it settles nothing about who pays for an examination performed by software. A clever rival with better retrospective scores and a cheaper camera could follow through the door IDx opened. The test comes over the next year: how many practices install the system, and whether insurers pay for the exam at a rate that covers the hardware. Those numbers will show whether being first buys a lasting position, a question that sits at the centre of the patient capital our thesis describes.

In medical AI, the algorithm is the cheap part. The trial that proves it in a real clinic is what the company is worth.

For founders

Building AI diagnostics with a prospective trial planned in the setting where the product will be used? Send us the project.

Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.

Apply in five minutes
Subscribe

Get the news by email

One short note each weekday. No marketing, and you can unsubscribe at any time.