MIT researchers press FDA on unreviewed clinical AI tools

An MIT team has put numbers on something the clinical AI industry has been quietly aware of for years: a large share of the AI tools nudging treatment decisions in American hospitals were never reviewed by the FDA. As laid out in reporting on the MIT viewpoint by Medical Xpress, this is a roughly $4 billion market of clinical decision support tools operating largely beyond public accountability, and a 2023 survey of more than 2,400 hospitals found 65% reported using AI or predictive models, primarily to identify high-risk patients.

The piece, published in The Lancet Digital Health00146-3/fulltext) and co-authored by MIT CSAIL professor Regina Barzilay with Maëlle-Marie Corso and Kendall Square Policy Strategies' Paul Kim, traces how this happened. The 2016 21st Century Cures Act carved certain clinical decision support out of device regulation. The FDA's 2019 draft guidance interpreted that carveout relatively permissively, and by the time the 2022 final guidance significantly tightened it, the tools were already embedded in clinical practice. The authors call this regulatory whiplash.

What that looks like in deployment is the part worth attention. The viewpoint contrasts tools that went through FDA review (Sepsis ImmunoScore via the De Novo pathway in 2024, PeraTrend, eCART) with widely used ones that did not (the Epic Sepsis Model, the Epic Deterioration Index, the Tyrer-Cuzick breast cancer risk model). Their framing question, as captured in the MIT CSAIL announcement, is direct: why should an identical tool be exempt when developed in-house by a large interstate health system but cleared as a device when sold by a vendor? Their answer is three recommendations centered on transparency: a public registry that discloses which CDS tools hospitals use and how they were validated, structured dialogue between the FDA, developers and clinicians, and updated FDA guidance reflecting current technology.

The honest caveat is that this is a viewpoint paper proposing a direction, not a finalized policy package. The reporting does not quantify patient harm from the tools currently outside FDA review, nor does it spell out who would fund or enforce a national registry. The FDA issued revised guidance in January 2026 with expanded enforcement discretion, and how that plays alongside the MIT call for transparency is still open.

For anyone building or procuring clinical AI, the takeaway is to assume disclosure expectations are going up, not down. Vendors that already hold FDA clearance gain a credibility moat, and health systems willing to publish validation methods now will be better positioned when registries arrive.