How Should Providers Deploy Large Language Models? Experts Weigh In
Now that large language models (LLMs) are the hottest new category of AI to enter the healthcare world, stakeholders are watching closely to see how providers will embed these tools into their workflows and what it will take to do that successfully.
The use of LLMs in healthcare is still quite new, so health systems want to deploy these tools in the least risky way possible. A panel of experts explained how they think health systems can do this during a Wednesday session at MedCity News’ INVEST conference in Chicago.
To integrate LLMs in a responsible way, health systems should start by deploying these AI models in nonclinical settings, said Maia Hightower, UChicago Medicine’s chief digital technology officer.
For example, a health system could implement an LLM to help answer patients’ questions about their bills or assist with appointment scheduling. These nonclinical settings are “safe areas where there’s a lot of opportunity and a lot of administrative burden,” Hightower pointed out.
David McMullin, chief business officer at health AI company Anumanaagreed that providers shouldn’t be rushing to adopt LLMs in clinician-patient interactions.
“When we think about these large language models being implemented to help healthcare, we think about the interaction with the patient. That’s clearly very important, but there are plenty of bottlenecks in the hospital system that have nothing to do with interactions with the patient. There’s plenty of instances where a solution can be deployed with a large language model and also verified so you don’t have the fear of hallucination,” McMullin declared.
The example that comes to the top of his mind is the ability of LLMs to code. He said every health system he has interacted with has had a swamped IT department that often finds itself too overwhelmed to deploy new advances in clinical workflows.
“What if that could be de-bottlenecked through a large language model? The large language model could use code, and that code can be verified — it comes in and you know whether or not it was written correctly. That could have a profound impact on healthcare delivery, even before we’ve gotten to the point where large language models start dialoguing with patients,” McMullin said.
Healthcare certainly has a wide array of inefficiencies and bottlenecks that are ripe for innovation. As health systems begin dipping their toes into the LLM water to solve these problems, Hightower thinks they will be more likely to take this leap with anchor companies than startups.
In her view, it will be a challenge for startups like Hippocratic AI to convince health systems to adopt their AI models. This is because the big vendors that are already a part of hospitals’ ecosystems, like Epic and Amazonare also working hard to deploy LLMs.
“I would imagine a lot of startup folks cried when Epic said that they’re partnering with Microsoft because all of a sudden, their chat bot is like, ‘How am I going to get into Epic if Epic is already in Epic?’” Hightower said. “If I’m a health system, I’m going to double down on my already existing anchor platforms over a high-risk startup.”
The panel acknowledged that healthcare leaders need to come together to erect some guardrails around the use of LLMs in the industry, but they argued that these AI models’ benefits outweigh their risks. Providers should be very excited about the new use cases for LLMs that will be discovered in the next couple of years, they declared.
Photo: venimo, Getty Images