Epic's Cosmos dataset contains more than 300 million unique patients, with records stitched together across participating health systems.
“Cosmos these days has over 300 million unique patients in it.”
Person
Seth Hain is an Epic Systems executive who spoke on NEJM AI Grand Rounds about Epic's approach to artificial intelligence in healthcare, including its Cosmos patient dataset and its Curiosity prediction models. He said Cosmos holds records for more than 300 million unique patients, and that more than 440 health systems use Epic's generative AI in production.
Hain described Curiosity as generating 20 to 40 possible future timelines for a patient to estimate outcome odds. He said the largest model so far was trained on 118 million patients and has not been deployed in production. He also talked about his own path at Epic, which began with work on its Chronicles database.
1 episodes · 24 quoted claims
Epic's Cosmos dataset contains more than 300 million unique patients, with records stitched together across participating health systems.
“Cosmos these days has over 300 million unique patients in it.”
More than 130 studies in over 80 journals have been published using the Cosmos dataset.
“there's been over 130 studies published in over 80 journals built off of that data set.”
Cosmos is governed by a council whose members are elected from participating health systems.
“This effort is governed by a variety of health systems, and individuals from the health systems that are participating are elected to a board that create the Governing Council for Cosmos.”
Epic's structured medical event model has about a billion parameters and was pre-trained on over 115 billion discrete medical events.
“That training cycle ended up ultimately with a billion-parameter model that used over 115 billion discrete medical events in its pre-training process.”
Epic built 78 evaluations to test its structured medical event model on tasks including diagnosis, prognosis and operational metrics.
“We ended up building out 78 different evaluations for understanding how a structured medical model could impact everything from diagnosis, say, in the near term as you're going in for an office visit,”
The largest Curiosity model to date was trained on 118 million patients.
“The largest model to date we've trained on Curiosity has 118 million patients in it.”
Curiosity was formerly known as Comet in some contexts.
“the current version of Curiosity, or what was formerly known as Comet in some contexts,”
Epic's first Curiosity model was trained only on adults over 18, with pediatric data planned for later training runs.
“The population we trained the first model on was over 18 years old.”
Curiosity generates a patient's predicted future repeatedly, producing 20 to 40 possible timelines to estimate outcome odds.
“you actually generate that multiple times. The result of this is that you have 20, 30, 40 potential timelines that this patient can go down,”
Cosmos is an Epic community effort to bring real-world evidence into the point of care during patient visits.
“Cosmos is an effort across the EPIC community to bring real-world evidence into the point of care”
The Cosmos effort has been underway for almost a decade.
“This is an effort that has been underway for almost a decade, at least at this point,”
Epic has not compared Curiosity against the Epic Sepsis Model, since Curiosity has not been deployed in production.
“we have not done it yet.”
Epic works with researchers, including some at the University of Wisconsin, on evaluation techniques such as LLMs as judges, part of an open-source AI Trust and Assurance Suite.
“which is an open source tool for this.”
Over 440 health systems use Epic's generative AI in production.
“We have over 440 health systems using generated AI in production at this point.”
Epic began rapidly rolling out generative AI in late 2022, starting with outpatient care.
“in late 2022, we started down the path of rapidly rolling out generative AI, first in the context of outpatient care.”
Seth Hain began working across Epic's clinical application suite in January 2020, just as COVID-19 demanded rapid changes such as vaccine administration workflows.
“And so in January of 2020, I started working across the suite of applications”
Hain began at Epic working on scalability of Chronicles, Epic's underlying database foundation, and later worked on virtualization and cloud deployment.
“I started by focusing on very low-end scalability, optimizing all the way down to the chipset, the foundation behind Epic, something we called Chronicles.”
When Seth Hain was a graduate student in Madison, Epic had just signed Kaiser Permanente as a customer.
“they had just signed Kaiser Permanente.”
Epic planned to open the Curiosity models to researchers on Cosmos in the first quarter of the year following the interview.
“which is planned for the first quarter of next year, to see how they do this also.”
Epic's event model begins training with no built-in knowledge of medical concepts such as hypertension or Lipitor; tokens map directly to structured events.
“There is no foundational knowledge of what is hypertension or what is Lipitor.”
Through a capability Hain calls Aura, clinicians can place orders in Epic for specialty diagnostics from other companies, with results flowing back into the medical record.
“something we call Aura, placing orders in Epic for specialty diagnostics that come from a different company”
Epic is developing a tool called diagnosis checker that can generate a differential diagnosis and likely follow-up tests.
“something we're calling diagnosis checker, as an example, which is essentially both a differential diagnosis,”
Epic's software can generate different summaries of medical records for different roles, such as a physician at a physical or a nurse at shift end.
“We have a capability in the software to generate summaries of medical records.”
Hain says Epic's campus features unusual items, including a Fibonacci spiral staircase and a carousel from the closed Ella's Deli.
“you walk by all of this wild stuff but you you sort of it becomes normal”