Axios Pro Exclusive Content

Epic releases first open-source tool to test, monitor AI models

Illustration of a stethoscope tied in knots around binary numbers.

Illustration: Shoshana Gordon/Axios

Health records giant Epic this week released its first open-source tool meant to help health systems test and monitor the performance of EHR-integrated AI models.

Why it matters: As AI-powered health care tools proliferate, a lack of uniform standards could result in a mottled landscape of various unproven systems.

How it works: Epic says its software automates the most time-consuming piece of AI model validation: data collection and mapping.

  • The suite has a reporting dashboard with demographic analyses, a common monitoring template and data schema to make it easier to use the suite for new AI models in the future.
  • The tool lets health care organizations apply demographic-based population filters to see how well a given model is working across various patient cohorts.
  • The Health AI Partnership (HAIP), a collaboration of health systems designing best practices for AI in health care, says two of its sites — Duke Health and University of Wisconsin Health — will use Epic's AI trust and assurance software to conduct a study to locally test and monitor AI tools.

Zoom in: Epic's tool is free and available on GitHub for health systems to download to their health record systems, Corey Miller, Epic's vice president of R&D, told Fierce Healthcare.

  • Epic claims health systems can use the tool to validate any EHR-integrated AI model, whether developed by Epic or another entity.

Caveat: The current version of the tool doesn't validate the performance of generative AI models, though Epic plans to expand it to more AI models in the future.

Between the lines: Adoption of generative AI, particularly for use cases like clinical documentation, remains stifled by lack of interoperability.

What they're saying: "We've been speaking to a lot of founders in the space and one thing we hear is until there's a change in infrastructure — who's using Epic, who's using Cerner — it's really hard to implement the level of accuracy needed to have an impactful inflection point," Montage Ventures VP Nia Patel said at Axios' BFD Expert Voices roundtable this month.

  • "You've got not only these models as moving targets, but their whole stack and other interactions and the ways the tools might integrate," said Brightside Health CEO Brad Kittredge, also speaking at the event.
  • "The market's likely to need some kind of assessment vehicles and some kind of performance benchmarks to get people over that hump, to be ready to take a new tool and integrate it into a stack and add to a sensitive system that's already working," Kittredge added.

Editor's note: This story has been corrected to note that just two of the sites in the Health AI Partnership will be using Epic's software suite, not the entire HAIP.

Go deeper