The project drew on five million de-identified electronic health records, alongside publicly available information. Meta says its internal research scaffold, powered by Muse models, helped produce ten health research manuscripts in four weeks. The work ranged from infectious disease to oral health. Meta calls it an early exploration, not evidence that AI can independently produce reliable medical research. The company outlined the project in a post about Muse-assisted health research.
Researchers used Muse Spark 1.3 and Muse Wonder to explore hypotheses and develop manuscripts. They also used the models to write and run code, drawing on the records and public information. Medical specialists in cardiology and epidemiology took part, alongside colleagues in infectious disease and internal medicine. Meta says some manuscripts grew from an initial prompt in as few as seven days. Participants reported productivity gains of up to tenfold in best-case comparisons with traditional research processes. Muse Spark 1.3 was released on September 2, 2026. Muse Wonder followed on October 12, about 40 days later.
People retained responsibility for the scientific decisions and conclusions. The AI helped with literature synthesis and feasibility checks. Researchers also used it for coding and manuscript development. Muse Wonder is positioned for professional tasks involving medical information. A model tracker lists a 60.5% result for it on HealthBench Professional. A separate review reports a HealthBench Pro score of 62.3 for Wonder Full, with Claude Opus 5.5 scoring higher. Those benchmarks assess answers to medical questions. They do not establish whether a research paper's methods or conclusions hold up.
Fast drafting still left findings to test
The ten manuscripts covered infectious disease and emergency medicine, with other work focused on oncology. The set also included endocrinology and sports medicine, among other fields. Each manuscript received feedback from the medical cohort. A biostatistician reviewed the analyses, and independent third-party experts provided summary reviews. Those reviews also showed what retrospective records cannot establish. A review of clinical AI applications in Healthcare notes that external validation remains limited. Prospective validation across independent populations and healthcare settings is scarce.
One paper tied directly to oral health, "Documented Dental Disease in Patients With Staphylococcus aureus Bacteremia," found dental disease documented about four times more often among patients with Staphylococcus aureus bloodstream infections. Reported rates were highest among people with a history of heart valve infection. Rates were also highest among people with opioid use and injection drug use. The authors said the findings support considering mouth checks during care for these infections. But the study cannot show that dental disease caused the bloodstream infection or determine whether dental treatment would change outcomes. Reviewers warned that differences in how closely patients were examined could also affect the results.
Other analyses show why speed and certainty are different measures. One study found a 73% higher rate of newly diagnosed arrhythmias and conduction disorders among patients with infective endocarditis than among patients with bacteremia without endocarditis. Reviewers said surveillance bias could help explain the association. They described the result as hypothesis-generating, not practice-changing.
Another study examined emergency department patients with behavioral health presentations. Those who left without being seen had higher rates of return visits and admission. They also had higher short-term mortality than patients who completed care. Reviewers called the findings relevant to patient safety and emergency department operations. They also flagged potential confounding and missing information about diagnoses and illness severity. The study did not explain why patients left.
A diabetes analysis found documented remission in 4.3% of previously treated adults. Within five years, remission ended more often through treatment restart than through rising HbA1c. The authors could not establish why treatment was restarted. Reviewers said glycemic relapse should be distinguished from treatment restart.
Researchers kept responsibility for interpretation
Researchers said they used Muse to check whether available records could answer a question. They also used it to refine cohort definitions and challenge analyses. S. Michelle Ogunwole of Johns Hopkins School of Medicine said the system let her explore ideas with little opportunity cost and identify questions that were not worth weeks or months of work. Radiologist Jason Rothschild said Muse Wonder could push back on an idea and suggest another approach instead of simply agreeing.
That assistance did not shift responsibility for interpreting results. Kritika Singh said she read every line and checked each scientific claim before making the final decisions. Epidemiologist R. Felipe Lobelo said he continued to challenge assumptions and refine definitions. He also decided which analyses were scientifically meaningful. Jared C. Durieux cautioned that faster analysis does not necessarily produce more credible evidence. It does not remove the need for epidemiologists and statisticians, or for clinicians and independent validation.
Reviews of a manuscript comparing infective endocarditis outcomes across Staphylococcus species raised a specific concern. Reviewer Storm Dorrough said the paper's mortality claims may be overstated. Its supplement says the analysis "does not establish a mortality difference," and the study did not address existing literature reporting S. lugdunensis endocarditis mortality of roughly 20-42%. Ali Ahmad also questioned the mortality conclusions because results changed in sensitivity analyses. Diagnostic criteria differed across groups as well. Both reviewers called for carefully defined populations and human review of AI-assisted analyses.
Meta says the current work is limited to retrospective analyses. The company describes future research using more comprehensive datasets and prospective trials as a possibility. It also points to supervised AI agents guided by experts in areas where findings cannot be formally verified. More broadly, retrospective or observational findings in medical AI can identify associations and generate hypotheses. Clinical use also depends on validation across independent populations, clinical oversight and ongoing monitoring.
