Generative AI helps Providence make specialty referrals

Renton, Wash.-based Providence is using ChatGPT-like technology in a new clinical decision support tool for providers.

The 51-hospital system in integrating generative artificial intelligence into MedPearl, its clinician education and referral platform that is built into its EHR. The program gives primary care providers advice on whether — and where — to send patients for specialty care.

"Primary care is on the frontlines, they're seeing patients, they're trying to get through their day. They're not going to remember every little step that needs to happen in that care transition process," Eve Cunningham, MD, chief of virtual care and digital health at Providence, told Becker's. "They might not have all of that in their heads. And so we need to empower those frontline clinicians with good information, good knowledge."

Since Providence started offering MedPearl in January, about 2,700 of its 4,000 primary and urgent care providers have employed it, searching the application more than 90,000 times. MedPearl has its own content creation feature that is reviewed and validated by clinicians.

The health system set out to develop the platform 2 ½ years ago — before ChatGPT even existed publicly — to help solve a problem: mistaken specialty referrals. Dr. Cunningham said about half the time one of three things happens: Patients are sent to the wrong specialist, not enough workup is done, and (most commonly) they don't need to see a specialist at all.

Because while physicians spend a lot of time in EHRs — and often get burnt out by them — those platforms typically only have information on patients.

"Medical knowledge does not live in the electronic health record," Dr. Cunningham said. "I'll say that again: Medical knowledge does not live in the electronic health record."

In a pilot last year involving more than 200 Providence providers, 72 percent said MedPearl improved their workups, 20 percent said it changed which specialists they referred to, and another 20 percent said they were able to care for the patient themselves without the need of a specialist, Dr. Cunningham said. MedPearl's net provider score is 61, between good and excellent.

Providence has since integrated generative AI to enhance MedPearl's search function by giving suggestions and helping interpret medical acronyms.

The health system also plans to institute semantic search, where AI assists by learning to understand the searcher's intent, and explore how generative AI can quickly update MedPearl when new medical guidelines come out.

Providence uses a closed generative AI platform from Microsoft and OpenAI, the maker of ChatGPT.

Dr. Cunningham said generative AI is going to have to improve before it gets more involved in clinical decision-making.

"Any large language model that's going to be used in healthcare for this type of use case needs to be safely created, clinically validated, constantly updated with good information," she said. "If I go into ChatGPT — and I play with it almost every day — and I ask similar questions to what we have in MedPearl, it lies, it hallucinates, it gives me bad information. It doesn't have anything after 2021. You can't trust it, right?"

She said MedPearl and its future versions can also help address other pressing healthcare issues.

"We have a shortage of doctors, we have burnout of our providers, and we have a massive amount of pressure coming from all of these Medicare-age patients aging into the system," Dr. Cunningham said. "So solving this problem has a tremendous amount of value. We're making every visit count in a meaningful way, and we're getting the patient transitioned to the right level of care at the right time with the right type of specialist."

‍