Meta shares ten health studies that researchers wrote with its Muse models

Meta built an internal coding scaffold, a tool that writes and runs analysis code, on Muse Spark 1.3 and Muse Wonder, the model it introduced on 12 October. Medical researchers in fields such as cardiology and infectious disease used it for four weeks to brainstorm hypotheses, write and run all the code and draft manuscripts, keeping every scientific decision for themselves. It could reach public data and a licensed database of five million de-identified electronic health records.

The ten manuscripts range from sports medicine to oncology. One, a case series on the fungus Cladophialophora bantiana, found 21 patients with it in brain or lung samples. Each paper had feedback from peers, a biostatistician’s review and summaries from independent experts Meta contracted. Participants reported up to ten times their usual productivity in the best cases.

What the researchers found

In a report Meta hosts, the researchers call the system fast but unevenly reliable. At times it was confidently wrong with nothing to warn a reviewer, and more than once it said data was missing when it was in the dataset. It put mortality from Staphylococcus aureus bacteremia, a bloodstream infection, at 1%; an infectious disease doctor knew published studies report closer to 30%, which led to an audit of how deaths were recorded.

The studies are retrospective, using existing records only. Meta says future work on larger datasets could produce hypotheses to test in prospective trials. It hasn’t said whether researchers outside the project will get the tool.