Why St. Jude gives researchers freedom to experiment with AI

Advertisement

Memphis, Tenn.-based St. Jude Children’s Research Hospital has developed an AI governance policy that allows certain researchers to experiment with publicly available AI tools while maintaining stricter oversight for higher-risk applications.

The policy, which is less than four pages long, reflects an effort to give researchers flexibility as AI technology evolves without compromising safeguards around patient information, Renato Umeton, PhD, chief of artificial intelligence and vice president of data sciences at St. Jude, told Becker’s.

“We created an AI policy that basically allows freedom to experiment and does not get in the way of research,” Dr. Umeton said.

One feature of the policy is an exception allowing faculty principal investigators to use publicly available AI tools with data generated in labs they supervise, provided the information does not contain protected patient data.

Researchers who use the exception must accept the potential risks, including the possibility that unpublished research could be exposed or that others could publish similar findings first.

Dr. Umeton said the exception came from St. Jude’s executive team and reflects an acknowledgment that publicly available AI technologies will continue to advance faster than the organization can adopt every emerging tool internally.

The approach allows eligible researchers to explore specialized AI tools, including niche tools that may exist only in their field and may not yet be available through the institution’s internal technology environment.

However, that flexibility does not extend to all AI applications.

St. Jude’s policy uses a risk-based framework that distinguishes among clinical, research and administrative applications. It also considers whether the data involved is protected, confidential or public, creating a three-by-three framework for determining appropriate safeguards.

For example, clinical applications involving protected health information require a higher level of review, while some research applications involving public or confidential data may not need the same degree of additional scrutiny.

“We basically have a risk-aware approach,” Dr. Umeton said.

The policy tiers review based on risk and identifies best practices and the institutional leaders who should be involved, depending on an application’s intended use and the sensitivity of the data. High-risk cases, such as those involving protected health information, require collaboration across the Office of Data Science, Information Services, Legal Services, clinical teams, research teams and other institutional partners.

Developing the framework was the first thing Dr. Umeton worked on after he was hired in February 2025. A draft was ready by April of that year, followed by revisions, and St. Jude finalized several elements in April.

Dr. Umeton said a central priority has been connecting access to AI, technical expertise and governance across the institution through St. Jude’s Office of Data Science, which is led by M. Madan Babu, PhD, chief data scientist, and J. Paul Taylor, MD, PhD, scientific director.

He said that kind of hub is something many organizations lack. In his view, the expertise often exists, but there is no central structure connecting it, and he expects many organizations will want to create one sooner rather than later.

St. Jude also has multidisciplinary councils, including one for clinical work and one for administration and operations, that provide a foundation for evaluating AI applications.

Alongside the policy, St. Jude has expanded access to more than 11,000 AI models, accompanied by educational resources, guardrails and evaluation tools. Dr. Umeton said that effort has supported 170 deployments across the organization, though none is fully clinical yet.

Dr. Umeton said the larger goal is to bring AI closer to the researchers and scientists who understand the problems the technology could help address.

That approach also extends beyond St. Jude’s walls.

The hospital hosted a biohackathon that brought together 250 participants from 62 outside organizations to address unmet research needs identified by St. Jude, according to Dr. Umeton.

Participants were given access to AI models and educational resources, allowing researchers from different organizations to work together on scientific challenges.

Dr. Umeton said the event generated ideas and collaborations that would have been difficult to develop within a single research team.

St. Jude is also looking to technology companies to help address research challenges that academic institutions may not have the resources to tackle independently.

Dr. Umeton pointed to large-scale computing infrastructure as an example of a capability where industry partnerships can provide resources that academic institutions cannot readily replicate.

The hospital invited Google DeepMind, Anthropic, Siemens Healthineers and Hugging Face to its flagship data science conference, with the goal of seeding collaborations between those companies and St. Jude’s research enterprise.

Under the approach Dr. Umeton described, St. Jude’s researchers bring scientific questions, while industry collaborators can contribute technical expertise and access to computing resources. The aim is to accelerate research in pediatric rare diseases.

Dr. Umeton said organizations also need to build expertise in-house. Without investing in workforce training, he said, an organization can have a data science hub, external partners and a policy and still fall behind. He said employees in every role need clear pathways for upskilling.

That focus on durable capability also shapes his advice for health systems weighing AI purchases.

“I would be cautious on long-term commitments on narrowly defined AI products that are difficult to integrate, or difficult to evaluate, or difficult to replace,” Dr. Umeton said.

Because AI technology changes quickly, he said organizations should understand what happens to their infrastructure when a better model becomes available or when needs change, such as when a new state or national policy takes effect.

Before making a substantial investment, Dr. Umeton said leaders should ask what problem they are solving, how they will measure improvement, who will maintain the application and whether the underlying technology can be changed without rebuilding the entire workflow.

“Those are the questions that help distinguish basically a durable investment from an expensive demo,” he said.

Dr. Umeton also pointed to accountability as an unresolved issue. He said it is unfair when AI gets credit for successes while physicians are blamed for failures.

He said the current legal framework is predicated on the direct responsibility of the clinical team, while the FDA’s software as a medical device framework applies to fully autonomous tools.

“But the reality is that few clinical AI builders have the appetite to own both the wins and the shortcomings of their technology,” he said.

As a result, he said, developers generally position their tools as clinical decision support rather than pursuing full autonomy. He said he would be cautious about predicting broad clinical autonomy on a set timeline.

Dr. Umeton said the biggest disconnect between what people expect from AI and what it takes to make it work is the same one the industry faced years ago: the gap between demonstrating a capability and making it a dependable workflow.

“The question is whether a tool can help someone do meaningful work reliably,” Dr. Umeton said. “Remember, we don’t do AI because it’s cool. We do it to advance our mission.”

Advertisement

Next Up in Artificial Intelligence

Advertisement