Meta's Muse-based scaffold helps medical experts write 10 health-science manuscripts on ~4.95M de-identified patient records, with humans owning cohort and causal choices

Why it matters

【Key points】

On Oct 14 Meta shared a health-science project: medical experts used a human-led research scaffold built on Muse Spark 1.3 and Muse Wonder over de-identified health records of about 4.95 million patients, producing ten manuscripts across clinical fields, posted on Meta AI's research site.

Experts keep the key scientific choices (cohorts, variables, causal assumptions); the AI handles hypotheses, statistics code, literature retrieval and drafting. Jason Wei of Meta's health team said the records were licensed from a vendor.

Two abstracts show the approach. A breast-cancer study found 106 of 156,051 coded patients with sequential orders for two antibody-drug conjugates (SG, T-DXd); the second drug's record span was a median 63 vs 49 days by order, not significant, and the paper calls its analysis descriptive and unadjusted. Another, across 21 US health systems, linked social-needs Z codes to later ED visits and admissions, and a negative-control code showed much of the link reflects being documented at all.

【Verifiable】

Meta on X: https://x.com/AIatMeta/status/2110441716507693450

Paper (breast cancer ADC): https://ai.meta.com/research/publications/human-guided-agentic-ai-for-end-to-end-ehr-research-sequential-antibody-drug-conjugate-pathways-in-breast-cancer/

Paper (Z codes): https://ai.meta.com/research/publications/social-needs-z-codes-and-subsequent-acute-care-across-21-u-s-health-systems/

Jason Wei: https://x.com/_jasonwei/status/2110443686782013909

【Our read】

1. After its Muse Spark math papers, Meta is moving AI-plus-expert research onto real-world data, where the hard part is messy data and causal design. Leaving those choices to humans is a sensible split.

2. Both abstracts state their limits, and one adds a negative control.

3. The manuscripts are self-published, not yet peer-reviewed, and the licensed data can't be replicated outside. The test is whether they pass peer review.