HIMSS23: Google proceeds with caution in piloting its LLM in healthcare

CHICAGO—Artificial intelligence tools, including large language models and conversational AI, are generating the biggest buzz at the Healthcare Information Management and Systems Society global conference this year.

Tech giant Google is one company at the forefront of this trend of bringing generative AI into healthcare but is proceeding with caution as it tests out the technology with healthcare partners, Google Health AI and cloud executives told Fierce Healthcare.

The tech giant announced last week that its large language model Med-PaLM 2 will be available to a select group of Google Cloud customers to explore uses and report feedback. Med-PaLM 2 was the first LLM model to reach an “expert” level on the U.S. Medical Licensing Examination-style questions, reaching over 85% accuracy. Google’s new AI-enabled Claims Acceleration Suite was also announced with the goal of streamlining health insurance prior authorizations and claims processing.

"It can be quite overwhelming as a doctor just to get the right information and spend time with your patients," Health AI research lead and physician Alan Karthikesalingam, M.D., Ph.D., told Fierce Healthcare. "Technologies like natural language processing offer this great potential of giving people back the gift of time by summarizing, giving people the right content and putting it nicely in the way you want."

Karthikesalingam sees a future where physicians even ask questions with AI accessing mountains of data, research and patient records to provide an answer. But every step forward with AI in healthcare must be attended to extremely carefully, he said. 

Med-PaLM was birthed by a Google moonshot program and achieved through the training of a LLM on clinical knowledge while previous models had been used on common language. The team brought together six existing medical question-and-answer data sets. The second generation of the model excelled at answering the U.S. and Indian medical licensing exams.  

Karthikesalingam acknowledges that answering licensing exam questions is not a good indicator of whether a model can aid in medical reasoning. In comparison to the answers of practicing physicians, the model was unable to offer medical advice, oftentimes adding extraneous information, he said. 

Through usage with limited customers, the technology can be deployed carefully in order to avoid risking the health of patients, said Aashima Gupta, Google Cloud global director of healthcare strategy and solutions. Early use cases include patient portals being ameliorated with an AI digital assistant. Short and long-form medical summaries are another place the tech is being tested, executives said.

A research article assessing Med-PaLM 1 is currently under peer review. The first iteration of the technology revealed weaknesses that the Google team claims to have made progress with in Med-PaLM 2.

Gupta told Fierce Healthcare that she believes customers will be drawn to the tech giant’s cautious steps.

"When we think about the cloud, and this is what we did even with our FHIR data engine, we become a partner," Gupta said. "That attracts health systems. We want to build an ecosystem first."

The recently announced Claims Acceleration Suite works to help information flow from payers to providers with a special focus on prior authorization.

Google's global leader of health plan strategy and solutions Amy Waldron said the team chose to home in on prior authorization due to the fact that the process requires extensive administrative work. Currently, patients can wait seven to 30 days to get medical care while those data are being processed, according to Waldron.

Waldron is clear that despite the focus on prior authorization, the company is in no way addressing claims denial by functioning as a decision tool. In both announcements, the company seems to be focused on unlocking and presenting data, not making medical or claims decisions.

"We've been able to decrease the administrative process and increase the speed of information getting to a reviewer," Waldron told Fierce Healthcare. "To me, that's critical, because while this process is in play people are waiting for their medicine, they're waiting for treatment, they're waiting for advice."

The suite’s claims data innovator helps solve the ingestion caused by information being locked in unstructured text by using document AI and natural language processing to transform information from unstructured into structured data.

Google’s Healthcare API can then take the information and put it into FHIR. Add in some interoperability, and other parts of the system can access the information to address patient care at all points in the care ecosystem.

Google’s Claims Acceleration Suite has been tested in a few pilot programs that the company believes will further evolve product usage.