Meta said it developed an AI-assisted system for health research and is sharing 10 medical research manuscripts produced with the system. Medical researchers used an internal coding scaffold powered by Muse Spark 1.3 and Muse Wonder to work with public data and a licensed retrospective database containing five million de-identified electronic health records. Domain experts retained responsibility for research decisions and conclusions.

Researchers in cardiology, epidemiology, infectious disease and internal medicine used the scaffold over four weeks to develop hypotheses, write and run code, and collaborate with the models on manuscripts. Meta says the researchers moved from an initial prompt to a written manuscript in less than two weeks; some papers took as few as seven days. Participants unanimously reported productivity gains, reaching up to tenfold in best-case scenarios compared with traditional research processes.

The 10 manuscripts cover infectious disease, sports medicine, emergency medicine, oncology, endocrinology and oral health, among other areas. Each received feedback from peers in the medical cohort, analysis review by a biostatistician, and summary reviews from contracted independent experts.

One paper examined arrhythmias and conduction disorders after infective endocarditis, a serious infection of the heart’s inner lining and valves. It reported a 73% higher rate of new arrhythmias and conduction disorders than among patients with bacteremia without infective endocarditis; the one-year risks were 7% and 4%, respectively. Reviewers described the work as hypothesis-generating and said differences in cardiac monitoring could explain some of the association, so it does not establish causation or change standard care.

Meta described the project as an initial exploration based on retrospective analyses. It said future work using more comprehensive datasets could generate hypotheses for testing in prospective trials, and characterized its intended direction as supervised auto-research agents guided by domain experts, especially where findings cannot be formally verified. Medical experts also reported using the scaffold to synthesize existing literature, check whether available data could answer a hypothesis, and critique collaborators’ work.

Meta said it developed an AI-assisted system for health research and is sharing 10 medical research manuscripts produced with the system. Medical researchers used an internal coding scaffold powered by Muse Spark 1.3 and Muse Wonder to work with public data and a licensed retrospective database containing five million de-identified electronic health records. Domain experts retained responsibility for research decisions and conclusions.

Researchers in cardiology, epidemiology, infectious disease and internal medicine used the scaffold over four weeks to develop hypotheses, write and run code, and collaborate with the models on manuscripts. Meta says the researchers moved from an initial prompt to a written manuscript in less than two weeks; some papers took as few as seven days. Participants unanimously reported productivity gains, reaching up to tenfold in best-case scenarios compared with traditional research processes.

The 10 manuscripts cover infectious disease, sports medicine, emergency medicine, oncology, endocrinology and oral health, among other areas. Each received feedback from peers in the medical cohort, analysis review by a biostatistician, and summary reviews from contracted independent experts.

One paper examined arrhythmias and conduction disorders after infective endocarditis, a serious infection of the heart’s inner lining and valves. It reported a 73% higher rate of new arrhythmias and conduction disorders than among patients with bacteremia without infective endocarditis; the one-year risks were 7% and 4%, respectively. Reviewers described the work as hypothesis-generating and said differences in cardiac monitoring could explain some of the association, so it does not establish causation or change standard care.

Meta described the project as an initial exploration based on retrospective analyses. It said future work using more comprehensive datasets could generate hypotheses for testing in prospective trials, and characterized its intended direction as supervised auto-research agents guided by domain experts, especially where findings cannot be formally verified. Medical experts also reported using the scaffold to synthesize existing literature, check whether available data could answer a hypothesis, and critique collaborators’ work.