Generative medical event models trained on 118 million patients in Epic Cosmos (CoMET, renamed Curiosity)
Decoder-only transformers of up to 1 billion parameters, pretrained on 115 billion medical events from 118 million patients in Epic's Cosmos dataset, followed power-law scaling and, by simulating each patient's next events, generally matched or outperformed task-specific supervised models on 78 clinical prediction tasks without task-specific training.
- Model
- Curiosity, first presented as CoMET (Cosmos Medical Event Transformer), up to 1B parameters
- Field
- Medicine
- Date
- 2025-08-16
- Human collaborators
- Shane Waxler, Hoifung Poon, Seth Hain, Rahul Shah
Sources
Original work
Announcement
What was found
Cosmos holds de-identified records of 16.3 billion encounters for more than 300 million patients from 310 health systems that use Epic. The models treat a patient's history as a sequence of discrete events (151 billion tokens in pretraining) and predict the next event autoregressively, so a forecast is made by simulating many possible timelines and counting outcomes. The authors present it as the largest scaling-law study of medical event data, fitting power laws for compute, data and model size and training compute-optimal models from those fits. The 78 evaluation tasks span diagnosis prediction, disease prognosis and healthcare operations, and performance improved with model and data scale. The preprint has since been revised, and later versions call the model family Curiosity.
Novelty check
Europe PMC, titles (transformer OR generative OR foundation model) AND (health records OR medical events OR disease trajectories OR patient timelines), to 15 August 2025, plus the paper's related work: generative models of patient timelines predate it, including BEHRT (Scientific Reports, 2020), Foresight (Lancet Digital Health, 2024) and ETHOS (npj Digital Medicine, 2024), and Delphi-2M followed a month later. What is new is scale, about 118 million patients against hundreds of thousands to a few million, and the scaling-law study.
Caveats and known objections
A preprint, not peer reviewed, announced by the developers, so it enters at claimed. Cosmos is open only to researchers at Epic customer organisations under Epic's terms and the weights are not released, so no outside group can reproduce the results. The evaluation is retrospective, on held-out Cosmos data, and outperforming supervised baselines is a statement about prediction, not about clinical benefit. Autonomy is ai-led on the same reading as other model-output entries: the forecasts are the model's own, while the data, tasks and evaluation were chosen by people.
Nobody outside the lab has checked this yet.
Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.
Or on GitHub: submit a check challenge the grade send a correction or send a pull request
Flag this for triage
Signals order the review queue and nothing else. They are never published, and they never move a grade: that takes a citation.
Entry history (1 event)
- AddedEntered the registry graded Claimed and AI-led.
Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.
Graded claimed for verification and ai-led for autonomy. What these mean.
Cite this entry
whataifound.org. (2025). Generative medical event models trained on 118 million patients in Epic Cosmos (CoMET, renamed Curiosity). whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2025-08-16-epic-cosmos-medical-event-modelsBibTeX
@misc{whataifound-epicsystems-2025-models,
title = {Generative medical event models trained on 118 million patients in Epic Cosmos (CoMET, renamed Curiosity)},
author = {{whataifound.org}},
year = {2025},
howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
note = {Result by Epic Systems / Microsoft Research / Yale School of Medicine. Verification: Claimed. Autonomy: AI-led.},
url = {https://whataifound.org/finding/2025-08-16-epic-cosmos-medical-event-models}
}