Generative AI is reshaping healthcare

Over the past six months, artificial intelligence has once again become the hottest topic of tech conversation. New models have caused shockwaves across every industry for their ability to have naturalistic conversations, develop easy-to-read written content, and produce art and images that come close to appearing like authentic human creations.    

The most famous example of this type of technology is ChatGPT, developed by a company called OpenAI. It falls into a class of tools called “generative AI,” which includes any model that can be used to create new content, such as text, images, or videos.   

And healthcare, of course, hasn’t been immune from the hype. Generative AI has huge potential to alter the healthcare delivery landscape – if it is applied correctly to the right use cases.  

“ChatGPT itself is a large language model that is trained against pretty much every piece of content out there on the internet,” explained BJ Moore, Chief Information Officer at Providence, a 51-hospital health system serving the Western United States.   

“Everything went into it, which is great – it gives it lots of depth and lots of personality, but it doesn’t necessarily make it fit for the complexity of healthcare right out of the box. To make it fit for purpose in healthcare, we have to train it on clinical data and tailor it to the healthcare setting so we can start to solve some of our challenges.” 

While the technology is still very much in its infancy, generative AI is already finding niches in healthcare where it can address long-standing concerns about workflow inefficiencies, cognitive burdens for clinicians and staff, and frustrations with the patient experience. 

“The single biggest complaint from physicians is about documentation and the EHR,” Moore stated.  “It’s a fantastic place to start learning about this technology, and it’s exciting to be at the forefront of exploring the possibilities.” 

Applying generative AI to the real-world clinical environment 

At Providence, inbox management is one of the first applications for OpenAI’s GPT model. 

“It’s great that patients can send their providers questions and documents, but our users are just getting overwhelmed with messages,” said Moore. “But if we can unleash generative AI to go through that mailbox on the user’s behalf and triage what’s important, we can make sure that the right information gets in front of that user at the right time so they can do something about it as soon as possible, not when they’re in their pajamas at 10:00 at night, spending their free time catching up on messages right after they’ve already seen the patient that day.” 

OpenAI’s GPT technology is also newly integrated into Nuance’s DAX Express solution, which uses an ambient listening device to capture conversations between a provider and patient, extract relevant information, and automatically generate a clinical note.   

Microsoft has exclusive rights to use OpenAI’s GPT models.  Nuance is now a Microsoft company, which gives them access to the latest iteration of OpenAI’s technology, called GPT-4. 

Originally introduced in 2020, DAX has used a human review process to ensure accuracy and quality before sending the note to the clinician for his or her ultimate approval. The human component resulted in a slight lag of up to four hours before the note appeared in the record. 

With the addition of GPT-4, the generative AI will perform the same work as the human reviewers, eliminating the time lag before clinicians can view the automatically generated note, make any necessary edits or additions, and sign off on the documentation. 

“We’re finding around 95 percent accuracy with the system, which is very good for this type of tool. The majority of edits that clinicians make are usually related to the formatting of the note and the layout of the information, not necessarily to the content,” said Moore. 

Greater reliance on generative AI solutions could be a major step forward for reducing cognitive fatigue and burnout among clinicians, says Dr. Lance Owens, MD, Chief Medical Information Officer of University of Michigan Health-West (UMHW), a longtime Nuance partner. 

“Right now, physicians are the most expensive data clerks in the world,” he said. “Anything we can do to change that is going to be a benefit to the entire healthcare system. UMHW was an early adopter of DAX, and I feel it has almost completely eliminated the documentation burdens I’ve experienced as a clinician.”   

“As a primary care doctor, I’m expected to see patients every fifteen minutes or so while creating meaningful documentation that accurately captures their unique story. Without having to stare at the keyboard and type the whole time, I like my job better. The documentation I’m producing is better.  And I’m a better partner for my patients because I can focus on them and their needs.” 

A bright future for artificial intelligence in healthcare 

Both Moore and Owens believe the industry is just scratching the surface of what generative AI can do for clinicians and for patients. Eventually, AI could support complex clinical decision making on a broader scale – but it will likely be years before the technical, ethical, and legal foundations are in place to make that vision a reality. 

“Bias and trust are major issues in the world of artificial intelligence, and we haven’t answered all those questions yet to the degree where people would be fully comfortable having AI intimately involved in decision-making,” Moore acknowledged. 

While that work is in progress, however, there are plenty of productivity-related use cases available to tackle in the back office, as well as promising opportunities to streamline the clinical process and uncover new insights from existing healthcare data assets. 

“I would love to see implicit ordering and intelligent queuing up of information,” said Owens.  “If I turn to the patient and tell them that we’re going to put them on a certain high blood pressure medication, it would be great to have that order generated for me to review and sign.  Or if a patient is complaining of swollen legs, at some point I expect the system to pull their latest ejection fraction or most recent echocardiogram for me to look at without even asking.” 

For Moore, the large language model behind GPT-4 also has potential applications in data mining and pattern recognition.  

“We have already started to explore what natural language processing can uncover in clinical notes,” he said.  “Several years ago, we did a pilot with Microsoft with the theory there could be items in there, months or years before a diagnosis, that indicate a person is likely to develop cancer. Sure enough, we found some consistent correlations that were suggestive of that being true, which could eventually change the way we conduct screenings or preventive care for people with risk factors identified by AI like that.” 

“I’m hoping that we can use generative AI to pursue these types of ideas further and transform healthcare.  In the last six months, the advances with these models have been greater than any technology advance I’ve seen in my 30-year career. I can’t think of another period that’s been more exciting or held bigger opportunities, but I am very optimistic that we are going to be practicing medicine very differently in the years to come thanks to these tools and what they can do for us.” 

Jennifer Bresnick is a journalist and freelance content creator with a decade of experience in the health IT industry.  Her work has focused on leveraging innovative technology tools to create value, improve health equity, and achieve the promises of the learning health system.  She can be reached at jennifer@inklesscreative.com.