A lot of hype surrounds AI in healthcare and the rush of digital health companies seeking to cash in.
But experts are unsure how generative AI applications like ChatGPT and GPT-4 will influence clinical diagnosis and decision-making. Most say the first wave of adoption will take place in areas where administrative redundancies exist.
Related: Microsoft, Nuance introduce ChatGPT successor to healthcare
“Obviously, there’s a lot of energy and a lot of concern,” said Dr. Greg Ator, chief medical informatics officer at the University of Kansas Health System. “People just get way out in front of their skis on some of these technologies.”
Instead, early adoption of generative AI in healthcare is taking place in the less flashy area of clinical note taking. Ator is part of the team implementing generative AI technology at the academic health system to aid clinician note taking. The system is working with Abridge, a medical AI company, to summarize clinical conversations from recorded audio during patient visits.
Abridge’s generative AI technology is similar to Nuance Communications, a clinical documentation software company owned by Microsoft. Last Monday, Nuance said it is adding OpenAI’s ChaptGPT successor GPT-4 to its latest application, which will be used in electronic health record systems.
In both cases, users must describe what they’re seeing for the software to work properly. For example, if a patient presents with a sore throat, specific commentary of what the clinician is seeing must be verbally shared for the program to enter the information.
In addition to inputting relevant information to the EHR, both applications remove conversations not applicable to the care plan.
“They’re power tools,” said Abridge’s co-founder and CEO Shiv Rao. “[Generative AI is] a powerful tool in the context of a much bigger set of technologies that, orchestrated together, amounts to a solution that can create value in the workflow.”
Medical records are a logical place to begin because clinicians can quickly identify where AI-produced results were derived, Ator said. Clinicians can easily listen to a visit recording again if the AI misses valuable information.
“What you build beside, underneath and above these foundation models like GPT-4 is the secret sauce,” Rao said. “There’s a certain layer of technologies that are now available to all of us, but how we integrate those tools into larger solutions is going to be the difference between really magical experiences for doctors and their patients and solutions that feel like off-the-shelf toys."
Investor interest remains strong
Investments in healthcare AI totaled $4.4 billion in 2022, according to data from Rock Health, a research and digital health venture firm. While last year’s total was down more than 50% from 2021, it was in line with 2020.
The same data revealed 2021 set a high watermark with 224 deals for companies using AI technology. While 2022 was not as fruitful, it was higher than 2020. Though experts say the levels from 2021 won’t be returning any time soon, the space remains of interest.
While AI investments have remained strong, not as many of those technologies are ready for widespread adoption, experts say.
“I think that for some time forward, we're going to continue to need to have humans in the loop because the AI is far from perfect,” said Erik Brynjolfsson, director of the digital economy lab at Stanford University’s Institute for Human Centered AI. “It can't do a lot of things."
Brynjolfsson said trained medical professionals are able to quickly dismiss abnormalities on a scan or medical image whereas AI may make a wrong diagnosis. While there is potential to eventually replace some roles of clinicians, experts say human input remains critical.
Generative AI is also time consuming to install and even in promising areas isn't quite ready for primetime. Nuance is rolling out its GPT-4 feature in the summer.
Ator said University of Kansas Health System is implementing the technology over the few months. He is optimistic it will be completed by the end of the year, but did not want to provide a specific timeline. This is largely due to the time it takes to train clinicians and the integration required with the provider’s EHR platform Epic.
"Anytime you're working with a complex system, like Epic, which is our base medical system, we have to interact with them. Some of [the implementation] is driven by their timetable," Ator said.
Another potential barrier to adoption could be patient acceptance. A Pew Research Center survey conducted in December found 60% of adult US patients would feel uncomfortable if their healthcare provider relied on AI for their medical care. Less than a third felt the quality of their care would increase as AI was implemented.
While the study did not specifically ask respondents about analyzing audio recordings of their visits, the report’s authors found “concern over the pace of AI adoption” was broadly shared in medicine.
Brynjolfsson said dictation and medical imaging are areas where providers could improve processes. But he said the future of healthcare will continue to require clinician and human oversight.
Others, though, are more bullish about future adoption.
“What we're seeing today is just a sign of what's going to come,” said Dr. Robert Pearl, the former CEO of Oakland, CA-based Kaiser Permanente and a current professor at Stanford University. “Everyone's focusing on the mistakes of the day or the shortcomings of today. They're irrelevant."
This story first appeared in Digital Health Business & Technology.