Meta shares ten health studies doctors ran with its Muse AI models

What did Meta publish on 14 October?

On 14 October 2026 Meta’s research team published Accelerating Scientific Research in Health Data Science With Muse Models, a post that shares ten medical research manuscripts. Each was produced by outside physician-researchers working with an internal coding agent powered by Muse Spark 1.3 and Muse Wonder, the frontier model Meta launched two days earlier.

The researchers came from cardiology, epidemiology, infectious disease, internal medicine, oncology and emergency medicine, with affiliations including Penn State, Baylor, Emory, Stanford, the University of Washington and Johns Hopkins. Meta says the group did original research over four weeks. Alongside the papers it posted summary reviews from contracted outside experts, short testimonials from participants, and a seven-page report in which the team describes what the model did well and where it failed.

Meta frames the project as an initial exploration of AI for health science. It is not a product launch. The agent is internal, the health database is licensed, and Meta has not said whether others will get access.

How did the Muse research setup work?

The system had two kinds of data. Most studies used a licensed, de-identified, longitudinal US database of electronic health records and claims covering about 4.9 million patients; individual papers give the figure as 4.95 million. Others used public sources: the National Electronic Injury Surveillance System, which samples US emergency department injury visits, combined with Google Trends, Wikipedia pageviews and the GDELT news database.

According to Meta, researchers used the agent to brainstorm hypotheses, write and execute all of the analysis code, and draft the manuscripts with them. The team built shared skills to steer it, including an ideation skill one participant credits with letting her test many ideas cheaply. They also wrote what the report calls a data contract: one shared definition of what the records contain, how they link and how studies may access them, with checks that enforce those rules, so each project did not have to rediscover the database.

Every study went through at least two rounds of internal peer review plus a biostatistical and computer-science review of all code and analyses. The report says the most valuable early gain was not faster writing but faster stopping: the model flagged questions already answered in the literature and mismatches between a question and the available data, so weak projects ended early.

Meta says the experts went from the first prompt to a written manuscript in under two weeks, with some papers taking as few as seven days, and that participants reported up to a tenfold productivity gain in the best cases. Those are self-reported figures from a project Meta ran and supported.

What did the ten studies find?

The ten studies are observational: they describe patterns in existing records rather than test treatments. The table below lists them with the main result Meta’s summary gives for each. Several revisit known clinical questions with a larger cohort than earlier work; a few respond to very recent events.

The fungal infection study is the clearest example of speed. After the CDC’s MMWR reported on 2 October 2026 that Cladophialophora bantiana identifications had risen at one reference laboratory, Meta-affiliated authors used the system to search the records and describe 21 patients within days. Paddy Ssentongo of Penn State is first author on three of the papers, and Christos Colovos, listed with a Meta Platforms affiliation on that fungal paper, is a co-author on all ten.

The breast cancer paper doubles as a methods paper. Its abstract says the workflow found 106 patients who received both antibody-drug conjugates and that the order made no clear difference to how long the second drug was used, while stressing that the analysis was descriptive, unadjusted and based on orders rather than confirmed treatment.

What did outside reviewers say?

Meta published the outside reviews under each paper, and they are mixed in a useful way. A Penn State cardiologist called the endocarditis rhythm study methodologically exemplary for observational research, but said a signal in its negative-control analysis and differences in heart monitoring between groups left surveillance bias as a live explanation, and judged it hypothesis-generating rather than practice-changing.

Reviewers of the Staphylococcus lugdunensis endocarditis paper were sharper. One wrote that the mortality claims are overstated, noting the supplement itself says the analysis does not establish a mortality difference and that the paper did not address literature reporting much higher mortality. Another was less convinced by the mortality conclusions because they changed in sensitivity analyses.

Other reviewers pointed to missing confounder adjustment in the emergency department study, no clinician review of cases in the e-scooter study, and a case count too small for conclusions in the fungal series. Several still praised the scale of the cohorts and the candour about limits. These are short summary reviews by experts Meta contracted, not journal peer review.

Where did the AI go wrong?

The researchers’ own report is unusually direct about failure. It describes the system as a high-throughput, unevenly reliable research assistant: strong at drafting, data extraction and code, including methods beyond a given investigator’s expertise, and unreliable at judging whether a source existed, whether a recorded field measured what it was meant to, and whether its own output was correct.

It says the model was, on documented occasions, assertively incorrect, with no uncertainty signal a reviewer could use. More than once it told the team that something was not in the dataset when it was, and a clinician who knew the data caught the error. The investigators also say the model’s first choices of cutoffs, variable definitions and statistical methods often had to be challenged, and that its fluent first drafts were heavily revised.

The report places the approach deliberately between fully autonomous AI scientist systems, which it says risk unverifiable citations and opaque provenance, and narrow tools that automate one step. That echoes other recent results, such as TheoPhysEval, where agents scored very well on calculations but choosing and checking problems stayed human work.

What does this change for medical research, and what does it not show?

The practical change is time. The report says work that once took months of data cleaning, or weeks to answer a reviewer, can now advance in days, and argues this could help smaller research groups that cannot afford data engineers and statisticians. One participant, a retired NIH deputy branch chief, wrote that he no longer had to wait for busy analysts to pull and analyse data.

The limits are just as clear. All ten are retrospective studies on one licensed dataset and public injury data, so they can show associations, not causes. The productivity figures are self-reported by participants in a Meta-run project. Faster studies also mean more studies, and the reviews show that speed does not remove confounding or overstated conclusions. Clinical AI evaluations such as Google’s AMIE primary-care feasibility study have moved toward prospective testing; Meta says its own next step is richer data to generate hypotheses that can be tested in prospective trials.

For researchers, the most transferable lessons are procedural: write a shared data contract, make the model justify every cutoff and definition, keep an expert who knows the data in the loop, and treat a confident answer from the model as something to check rather than something to trust.