At Stanford Health Care, for example, its Data Science Executive Committee (DSEC) oversees governance for AI models that includes providing approvals and education on change management. The committee also oversees new tech funding, and Lindmark has one chance per year to ask for operating money for the health system’s budget.
“Any new AI agent that's going to come into the organization that will affect the clinical outcomes in any way has to be presented to the DSEC, which is chaired by our CEO and our CMIO,” Lindmark says. “They review all those requests and sign off to ensure that they agree with the assessment that's been done, and it’s responsible and safe for deployment.”
It also helps when a solution isn’t reliant on one model. “The models are changing so fast, it’s best to ensure that whatever you're building can be model-agnostic as newer versions come, or if you want to use a lower-cost model, have the ability to go between models fairly easily,” Lindmark says.
As to what’s next, Revri plans to create more agentic workflows at Stanford Health Care to build on the tumor board’s work.
“We are starting to move more toward a service-line focus, which is cancer treatment,” he says. “What is needed for cancer treatment? We look at it holistically to see if a series of agents can help.”
Similar to how oncologists can choose among a set of agents to help with their job, cardiologists could as well, Revri says. In addition, Revri and his team are evaluating the ROI of low-cost models versus frontier AI models.
Stanford Health Care will also fine-tune the agents to improve their efficiency, Lindmark adds.
READ MORE: Here's what infrastructure can and cannot do for your data.
Allina Health Develops AI Agents to Assist Clinical Workflows
Minneapolis, Minn.-based Allina Health is developing a next-generation data architecture with Databricks to connect all of its data pieces together so that AI agents can operate on top. “Getting the data structured, having it well defined and then continuing to manage and govern it is critical,” says Chief Digital and Information Officer Dr. Dave Ingham.
An interface layer is a key aspect of the agentic AI data foundation, he notes, describing it as a computing structure that acts like a translator to communicate in the language of various systems to request data.
“They don't talk to enough things, and that's why we have the robust interface engine to go get the data and perform the tasks,” Ingham says. “That's what really unlocks automated complex workflows that until now have always required humans.”
Allina Health has developed an AI agent called Alli, which acts as both an operational and clinical AI assistant. The idea is to take pressure off contact center human agents and help with patient triaging to find out why they are calling, Ingham says.
Meanwhile, another AI agent reviews the charts of incoming psychiatry patients for violent history, Ingham says. It uses thresholds with a color code to determine which patients have been violent. “It helps keep our patients safe, keep our care team members safe or at least more prepared for potential outbursts,” Ingham says.
