Health care's first AI registry is coming soon

The Coalition for Health AI (CHAI) is building off positive feedback from its applied model cards—or AI "nutrition labels"—by giving them a more sophisticated home.

This morning, CHAI announced its partnership with Avanade, a leading IT consulting firm and Microsoft expert, to develop a public registry for health AI model cards. Brenton Hill, CHAI's head of operations and general counsel, hopes the registry will improve accessibility and inspire continued adoption of model cards.

"If everyone has a model card and they keep it to themselves, it's actually not that valuable," Hill told Newsweek.

Applied model cards give potential users insight into an AI tool's development and any known risks. CHAI's members—including nearly 3,000 health care organizations from the private and public sectors—collaborated on the guidelines.

The new registry will centralize these model cards and leverage CHAI's community of AI developers and health care providers, allowing them to streamline communications while exploring partnerships, according to Hill.

"People are really getting tired of duplication of processes," he said. "A vendor has to send the same thing to 20 different health systems, or health systems have to ask for something 20 different ways. Why not come together and figure out a standard way to get this [information] so we're not all duplicating resources, duplicating efforts and wasting time?"

Many of the organizations that helped develop the model cards have supported the registry. Several leading health systems including Cleveland Clinic, Kaiser Permanente and Stanford Medicine signed today's Model Card Call to Action, along with more than a dozen AI solution providers. The Duke University Health System, based in Durham, North Carolina, was first to be added to the system.

"As the first health system to be onboarded to the registry, we're paving the way for additional members to utilize this important resource that will allow the CHAI community to harness model cards in a meaningful and efficient way," said Michael Pencina, chief data scientist at Duke University and CHAI board member.

Just as anyone can access CHAI's Github page and download their own model card template, anyone will be able to access the registry, including the general public—no membership or health care experience required.

To ensure that the information is both transparent and trustworthy, human reviewers will validate each model card submission. The exact timeline for review will depend on the volume of submissions, according to Hill.

Exact functionalities are still being determined, but he hopes the registry will eventually provide useful insights for vendors and health systems. For example, they may be able to share how many times people searched for ambient scribe models, quantifying interest in a particular tool or capability.

But there are still some kinks to work out, Hill said. The model card concept is gaining steam, but it may need to be retrofitted for generative AI, which requires frequent updates. Plus, the templates will need to be integrated into the registry from their PDF forms, then distilled into a more readable format.

CHAI is looking into ways to integrate the registry with electronic health records, offering accessible information without burdening providers.

"An executive that's making a procurement buying decision is going to care about certain subsections—they probably don't need a whole model card," Hill said. "If we can be flexible and ensure that we're getting people the information that's most pertinent to them, I think that's where this [registry will be] most successful."

The project is launching at the Healthcare Information and Management Systems Society (HIMSS) Conference in Las Vegas from March 3-6. CHAI hopes to engage conference attendees, spreading the news about the registry and building out the sign-in process.