His keynote address kicked off a conference and expo that included more than 900 sessions, 300 exhibits, and an array of technical workshops. Amid all that action, one panel discussion took a closer look at “The Role of Generative AI in Modern Medicine.” Participants included Eric Topol, professor and EVP at Scripps Research, Katherine Wood, CEO and CIO at ARK Investment Management, and Peter Lee, president of Microsoft Research.
Moderated by Kimberly Powell, VP of NVIDIA Healthcare and Life Sciences, the panel explored how advancements in AI are reshaping everything from the delivery of care to the discovery of new medicines—and considered how AI could accelerate medical discoveries and improve patient outcomes moving forward.
Looking Back to Consider Tech’s Future Potential
To set the stage for their forward-thinking discussion, the panel first reflected on some of the ways technology has attempted to revolutionize healthcare in the past, highlighting some of the notable breakthroughs that have made a significant impact and some of the disappointments that have fallen short of initial hopes or expectations.
ARK Investment Management’s Katherine Wood recalled that the first time she heard the phrase personalized medicine was back when she was investing during the tech and telecom bubble. At the time, she said there was talk of therapies that would emerge within just a few years or so to address specific diseases.
“There was too much capital chasing too few opportunities too soon,” she said, “so it ended badly in the tech and telecom bust.” Ever since, and particularly following the 2008-09 financial crisis, she has sensed a certain amount of fear swirling around the innovation taking place in sectors like healthcare and biotech.
Wood also highlighted the dramatic shift in sequencing technologies that has taken place over time. Sequencing the first human genome took over a decade and billions of dollars to accomplish. Although this was a major milestone, she pointed out that “there was no way that was ready for prime time.”
This same feat currently takes just a few hours and a few hundred dollars, give or take. And now that the cost of sequencing has dropped, it is becoming increasingly possible to extract meaningful data and identify early signs of disease.
She said it has been fascinating to watch the financial markets along the way, including how the “multi-omic revolution” has lagged during a bull market. Even so, she believes the “convergence between sequencing and artificial intelligence” will ultimately lead to profound results that could transform human, animal, and plant life.
As the conversation turned to others, Eric Topol of Scripps Research highlighted electronic health records (EHRs) as a notable digital health disappointment, calling this “a tremendous bust.” He also pointed out that “genomics hasn’t really delivered yet.” Although there may be a variety of drugs labeled for genomics interactions, for example, hardly any are currently used clinically.
At least part of the issue has been the inability to deal effectively with big data. “It was basically a collection of data with nowhere to go, all dressed up,” he said. But Topol believes we’re entering a new phase and this scenario is about to change. “We can start to handle many layers of any individual’s data, which we couldn’t before.”
Peter Lee of Microsoft Research said he believes “the technology industry has made the well-intentioned mistake of jumping first to diagnostics.” In the realm of omics and in areas like protein design or prediction, “there are some amazing things happening that I have no doubt will lead to advances,” he continued.
However, he believes much of the “early meaningful applications” have centered on healthcare professionals’ productivity and daily working satisfaction. In recent years Microsoft acquired Nuance, for instance, which is focused on “reducing the notetaking burden for doctors and nurses” and involves a “considerable amount of AI.”
Lee also circled back to the mention of EHRs, pointing out that there could be opportunity moving forward despite some of the disappointment to date. Now that virtually all health records have become digitalized, one remaining question is, “What good can we make of all that?” He feels optimistic that AI can play a role.
How AI is Altering Care Delivery, Drug Discovery, and More
As the panel looked ahead, they spoke of some of the ways in which AI could help improve relationships in healthcare and enhance drug discovery efforts, among other possibilities.
According to Topol, there has been a growing “erosion of the patient/doctor relationship” in recent decades. Although the switch from paper to electronic health records can be viewed as progress, he said this has in some ways turned clinicians into data clerks. There has been an ongoing need to free doctors from these types of functions and restore this relationship, he pointed out, and generative AI could help.
Even though machines have no idea what empathy is, they do have the potential to promote greater levels of it among healthcare professionals, and his hunch is that someday down the road, “every clinician will have to go through coaching by a large language model to be maximally empathetic and communicative.”
Microsoft has been involved with the recent integration of GPT-4 into EHRs, and Lee noted that one surprising aspect to emerge is what he referred to as “reverse prompting.” Essentially, the large language model can pick out an interesting tidbit from a conversation and provide it to the clinician.
Perhaps the patient is about to become a grandparent, for example, or about to travel to an exciting destination. When brought to a doctor’s attention, this can offer an opportunity to pause, reflect on that patient’s life, and decide whether or not to accept the suggestion in an effort to connect on a more personal level.
Companies such as Abridge are using AI to turn clinical conversations into structured notes, Kimberly Powell of NVIDIA pointed out, and Hippocratic AI is developing generative AI healthcare “agents” that can take on tasks like calling a patient following surgery or checking on medication compliance.
She also highlighted the intersection of generative AI and drug discovery efforts. It is becoming increasingly possible to represent the world of drugs using computers, Powell said, which means there are growing opportunities to start building sophisticated models that can predict interactions at multiscale.
In AI research, models can learn from nature, Lee explained, and one prime example pertains to proteins and molecules. Microsoft Research as well as top research universities and labs across the globe are all pursuing this, and one reason it works is that the physics equations governing the molecules and atoms are known.
One crucial question is whether we can “move up the biological stack” into omics areas. “There, we don’t necessarily have the physics equations for this,” Lee said. “We have some idea of the mechanisms, but most of the data that we would really need to train these models and move up the stack comes from direct observation.” These observations are often “destructive on the self,” however.
Another notable challenge is that today’s bioscience, biopharma, and biotech industries are protective of their data. He foresees industry becoming even more protective in the years ahead considering generative AI offers the possibility of extracting value from this data.
This hurdle will need to be overcome to reach a point where we can begin to “pull adequate amounts of data,” he added. But if we do get there, “the sky’s the limit” and “we can really start to emulate a huge amount of human biology.”
As the conversation circled back to Topol, he pointed out that using AI to discover new drugs or new applications for existing medications is already happening. During the pandemic, for example, BenevolentAI helped identify an arthritis drug that could also be helpful for the treatment of COVID-19. “That’s the first example,” he said, but there are plenty of other examples that are now in phase one or two clinical trials.
When he first arrived at Scripps Research over a decade ago, “it would take two or three years to develop the crystal structure of a protein,” Topol added, whereas that now takes just minutes or even seconds. “So this is zooming.”
Elephant in the Room: Regulating AI in Health and Medicine
The panel also touched on the potential regulatory issues related to the use of generative AI in health and medicine. Powell posed a handful of questions to the group, such as: What sorts of conversations are taking place with regulatory bodies? What standards are being developed? And how can companies be proactive?
“The medical community itself should assertively take ownership of the questions of whether, when, and how this technology should or shouldn’t be used in the practice of medicine and in the advancement of medical science,” Lee said.
He noted that the American Medical Association has issued guidelines and anticipates that the National Academy of Medicine will issue guidelines later this year. “A lot is happening, and every major leader in healthcare and medicine, I think, is really just confronting this head on.”
Lee also suggested that regulators need to be permitted to act, and from his perspective, the White House Executive Order and recently passed EU AI Act are “signals to regulators that they’re allowed to do something.” He called generative AI “the technological centerpiece of national competitiveness” and emphasized that regulation is needed “to provide clarity to companies like ours.”
Yet another element of regulatory considerations involves history and how expectations can change over time, he said, along with the need to understand where things are headed down the road.
“People reacted badly to the idea that doctors should wash their hands before laying hands on you,” Lee said. Another example: “You would never expect your obstetrician to just put an ear to a pregnant woman’s belly and then make a pronouncement on the health of that baby,” he continued. The expectation, rather, is that ultrasound technology be put to good use.
“In the very near future,” he added, “it may be just as foolhardy to think that the doctor might practice medicine on you without the assistance of AI.”
Paul Nicolaus is a freelance writer specializing in science, nature, and health. Learn more at www.nicolauswriting.com.