Nature Medicine published a mother-and-child clinical AI agent on 4 September; the Crossref deposit is stamped 17:15:13 UTC. It will be reported as an LLM agent that predicts maternal and infant outcomes. The abstract makes clear that the predictions belong to something else.
Where the numbers come from
"MoChiAgent's core predictive engine, MoChiFormer, was developed and internally evaluated using 4,401,599 longitudinal clinical visits and externally validated using independent maternal and infant cohorts consisting of 263,452 and 23,192 visits." And: "MoChiFormer accurately identified key gestational conditions, achieving AUROCs of 0.89 for placental abruption, 0.89 for premature rupture of membranes, and 0.91 for preterm labour." MoChiFormer is a supervised transformer over EHR laboratory sequences. It also "reconstructs missing laboratory values, reduces batch effects" and supports trajectory modelling — none of that is language modelling.
What the LLM layer is measured on
Not discrimination. The paper's source data lists "individual physician evaluation scores for each case and each system (MoChiAgent, ChatGPT, Gemini and OpenEvidence) across diagnostic accuracy, evidence traceability, completeness of the diagnostic and therapeutic plan, and clinical safety". That is a physician preference exercise, which is a legitimate measurement of a very different thing. The LLM orchestrates tools and retrieves guideline text; the forecast is already made by the time it does.
What the received framing gets wrong
"LLM agent predicts preterm labour at 0.91 AUROC" would be false. A supervised transformer does; an LLM writes it up and cites guidelines. This is now the standard shape of clinical-agent papers, and it matters because the two components generalise differently — a transformer trained on one health system's lab sequences fails in recognisable ways, while an LLM wrapper's failures are fluent. Crediting the wrapper with the classifier's discrimination hides which half needs revalidating at a new site.
Two caveats the paper carries itself
The external validation is small relative to development — 263,452 and 23,192 visits against 4.4 million, roughly 6% and 0.5% — and the infant cohort is counted in visits, not infants. The cluster findings (elevated risks of neonatal jaundice at HR 2.81, 95% CI 2.60-3.03, and haematological disease at HR 2.83, 95% CI 2.62-3.05) are associations within the cohort. And this is an Accelerated Article Preview, not the final version of record, so the figures can still move.