Industry Insights into Developing AI-Powered Health Care Solutions

Only a couple of years removed from the unveiling of ChatGPT, advanced artificial intelligence (AI) and large language models (LLMs) have become commonplace in our everyday lives. Search for anything on Google or Bing, and the first thing that shows up is no longer a search result, but an LLM-enabled attempt to summarize the information you’re looking for.

The technology is still, however, far from perfect. While it is adept at picking out relevant information from a vast set of data, it makes plenty of mistakes. These models struggle to recognize nuance and sarcasm, and they can often create completely new (and wrong) information out of thin air; a phenomenon called hallucinations.

While these issues are a minor nuisance to those trying to find a new Salisbury steak recipe or draft a term paper, they could cause a lot of harm in a high-stakes field such as medical research and health care. These challenges, however, won’t stop hospitals and providers from using AI.

They simply can’t afford not to.

“Health care is a five trillion-dollar expenditure in danger of bankrupting the nation in three to four years,” said Rick Shannon, senior vice president and chief medical officer for the Duke University Health System. “We need a sense of urgency in bringing forward these capabilities in health care delivery. But this is not about replacing people and jobs. This is an opportunity, out of the gate, to eliminate defects, errors and waste.”

This was one of the central tenets of the recent Duke Summit on AI for Health Innovation, a three-day event convening experts from health care, engineering and industry to foster a community to push forward AI solutions and capabilities in this space. Held at the JB Duke Hotel from October 9 – 11, the event drew about 100 people eager to share their knowledge and work toward making a positive impact.

“We’re here to explore how to go forward in a collaborative way,” said Jerome Lynch, the Vinik Dean of the Pratt School of Engineering, who provided opening remarks preceding Shannon, the event’s keynote speaker. “And to explore how to go forward while applying an engineering mindset to tackle some of the biggest problems in the AI health care space.”

Shannon’s opening remarks also stressed the importance of engaging frontline workers, reducing care worker burden and keeping equity at the top of mind. After his presentation, the first day featured panel sessions and breakout groups focused on Duke Health’s experiences in developing and evaluating AI-based technologies as well as lessons in product development and design thinking from Duke Engineering’s Christensen Family Center for Innovation.

Two themes often repeated throughout the day were a dearth of AI-based solutions that actually solved any problems for clinicians and a need to pursue the lowest-hanging fruit—using AI technologies to efficiently extract useful information from the overwhelming amounts of data contained in patient electronic records and clinicians’ notes and put it into a standardized format.

The latter half of the second day flipped the viewpoints around and examined some technologies being pursued by Duke researchers that might eventually find their way into the clinic. For example, researchers are using LLMs to design novel biologicals to treat diseases, efficiently model the blood flow in patients to empower predictive medicine, and make sense of all the shorthand and acronyms commonly employed in free-text notes.

Sandwiched between the experiences of clinicians and research faculty was a lively discussion between industry representatives from some of the top companies actively deploying AI-based solutions in the health care domain. Moderated by Jonathan Owens, director of industry engagement for Duke Engineering, the panel included:

• Basia Coulter, a partner for Globant Healthcare & Life Sciences
• Brendan Fowkes, a global industry technology healthcare leader for IBM
• Marissa Halfin, a director of clinical informatics for LabCorp
• Kirsten Lum, a senior data scientist for Johnson & Johnson
• Terry Myerson, CEO and co-founder of Truveta

The conversation kicked off with another reference to the looming opportunity to use LLMs to pull data from patient notes and electronic health records and normalize the information contained in them. Not only would this help individual hospitals run more smoothly, it would open opportunities to run data analytics studies on a population level. Imagine not only being able to prescribe a drug with confidence, but to tailor the dosage to an individual’s physical attributes and medical history.

“The challenge is how to make these records useful for analytics,” said Myerson, whose startup company Truveta has compiled clean electronic health record data for more than 100 million patients representing the full diversity of the U.S. “The core bias in these is that they are coded for reimbursement claims, whereas we need to build AI systems that represent clinical truths in an analytical fashion.”

As several of the participants talked about building and working with their own datasets, the point was raised about how intellectual property is handled across the industry. Qualms over who ultimately owns the data and the AI trained with it can often create barriers to companies, health care providers and researchers working together.

From IBM’s perspective, Fowkes pointed out that his arm of the company doesn’t want to own it, they just want to help take great ideas to market. But these details need to be worked out amongst collaborators in the spirit of partnership before any solutions are developed and deployed.

Lack of internal AI technical resources and concerns regarding external collaborations with other companies and universities shouldn’t be an insurmountable barrier for companies that want to expand in this domain, yet don’t want to try to be something they’re not.

“We have pharmaceutical customers saying they know their mission and they don’t want to become a tech company by hiring whole teams of AI experts and doing this work in-house,” said Coulter. “There are ways to engage companies with decades of tech experience that can complement your own expertise in the health care industry.”

The conversation then moved to the need to keep projects aimed at a specific target set by close collaboration with the people who will actually end up using it. Too often, AI software is created that does a great job at a specific task, but that clinicians don’t actually find useful in aiding decisions or reducing workload.

“How an AI model is actually going to be deployed is really key,” said Lum. “The needs of the people using it on the ground need to be taken into account from the start.” Another key factor is to take into account not just the clinicians who will be using it, but the people who will be affected by the AI technology.

“Pharmaceutical organizations that have treatments for rare diseases on the market look to digital technology and AI for help in reaching those patients because they believe there are more patients out there who are either under or misdiagnosed,” said Coulter. “But rare disease experts point      out that many people don’t want to be diagnosed because of the stigma associated with some diseases. It is a striking revelation that people don’t think about and that shows the power of unintended consequences.”

The panel wrapped with discussions about other areas of digital health that people should be paying attention to, including providing accurate medical information online from trusted health providers, remote patient monitoring and care, precise personalized medicine, and predicting what supplies need to be ordered before anyone realizes they’re needed.

Outside of the panel discussion, Halfin and Fowkes shared more about why their companies, LabCorp and IBM, respectively, decided to sponsor the summit and send them to participate.

IBM wants to help companies like LabCorp build the platforms that will make their data more useful. But they need help from research institutions like Duke to help find new and creative use cases and, eventually, solutions.

LabCorp, on the other side of the equation, is in a unique position where they have an enormous dataset since they run about half of all medical diagnostics ordered across the country. But that data is far more valuable when it is correlated to clinical and genomic data. They’re looking to partners such as Duke to grow ideas and projects into solutions to bring more value to what they think of data as a product.

“LabCorp is navigating the waters of what AI means, and since we’re facing so many opportunities, figuring out if we should build tech, buy tech or partner with someone to develop products that bring more value to our data,” Halfin said. “Partnerships with academic institutions help bring cutting-edge AI solutions to the corporate world while ensuring technologies remain current with the evolving market’s needs.”

“I travel to a number of regional events and weigh in to help people understand what typically works and what doesn’t,” Fowkes said. “And after having a number of conversations with the extended Duke team, this event seems like a layup. Successful adoption/deployment of any AI use case requires proper governance from idea through deployment. The Duke governance program already in place here is far ahead of most others, so collaborations can go faster into action.”