“In an effort to help us advance health equity, can we use AI [artificial intelligence] to get to ‘better’ — faster?” 

It was the question that a group of diverse experts addressed on June 12, 2024 at WBUR CitySpace in Boston. The panel, sponsored by Sanofi and moderated by award-winning Boston Globe reporter Hiawatha Bray, debated the complexities of implementing AI solutions in the healthcare system: ensuring fairness and inclusivity while trying to optimize patient care.

Understanding AI’s impact on health care

The premise of AI in health care is that we can piggyback on large volumes of medical data to understand symptoms and what treatments worked (or not), and use that information to improve future outcomes. 

The technology has become a business imperative at Sanofi, especially given the breathtaking speed with which AI is revolutionizing other industries, says Tanisha Sullivan, head of External Engagement, US Health Equity Strategy, and Massachusetts Government Affairs at the company. 

AI-delivered predictive analytics have been leveraged in health care in some form for a while now, says panelist Collin Stultz, MD, PhD, co-director, Harvard-MIT Health Sciences & Technology, Massachusetts Institute of Technology.

For example, a routine blood draw to calculate cholesterol and determine a 10-year risk of adverse events “has learned what my risk would be from some set of clinical characteristics,” Stultz points out.

But the hot AI topic right now, complex deep learning (like ChatGPT), isn’t exactly right for health care. The medical profession really doesn’t need the kind of large models that generative AI programs run on, with billions of parameters feeding them, Stultz says. Instead, smaller, more purpose-built models are in order. But creating, accessing, and leveraging those are still not small tasks. 

Data and adoption challenges

While the promise of AI in health care to leverage medical data to improve future outcomes is exciting — especially for those historically overlooked by the healthcare system — numerous hurdles lie ahead. 

“Especially when it comes to AI for health equity, there’s a promise … that people who have barriers to access would be able to get better care through the use of AI,” says panelist Irene Chen, assistant professor at UC Berkeley and UCSF, faculty member of Berkeley AI Research (BAIR), and Harvard and MIT alum.

Unfortunately, Chen and other researchers have found that the data AI models feed on is already biased and, if left unchecked, can perpetuate bias in predictive models. “If we naively shove as much data as possible, we’re finding that garbage in [yields] garbage out. Bias in means bias out,” Chen says.

Another challenge: Clinicians and computer scientists often speak past one another, Stultz says. “It’s like a bad marriage,” he jokes. “For [AI] to be successful, there has to be some sort of common language between these two individuals to [address] a common task,” Stultz says.

Existing legislation and healthcare system practices regarding data access might be hindering AI adoption, argues panelist Amar Gupta, a research scientist at MIT. Traditional payment protocols for medical procedures in health care don’t account for AI assistance. As a result, a prototype system for AI-assisted mammogram reading that Gupta’s research team developed found it difficult to gain traction because “there was no reimbursement system for it.” Figuring out how AI-derived efficiencies will fit into current healthcare operating practices will help, Gupta says. 

Simply accessing data isn’t easy either. “One of the biggest problems we have is transferring electronic records from one hospital to another: How do you do it?” Gupta asks, “There is no central place you can go to, to get American medical data.”

Making advanced technologies work for health equity

Data problems and access obstacles aside, panelists agreed that given the potential in delivering better healthcare outcomes, it’s worth continuing to explore AI.

The first step is learning “how we can better start to understand some of this data that might be problematic, or need cleaning, or further introspection,” Chen says. The panelists brought up a couple of ways to begin that “cleaning.”

Chen and her group are evaluating how stigmatizing language that physicians use in clinical notes — patients labeled as “difficult,” for example — can make its way into records and affect healthcare outcomes. 

Stultz suggested that a “confidence level” score for AI-derived predictions, or the ability to understand how accurate the answer from AI could be, might add necessary nuance. “If a patient comes to me and the model says this person is not at high risk of death and it’s wrong, and I know there’s interventions that I could do that would prolong their life, then I missed an opportunity to intervene,” Stultz says, detailing the high stakes involved. 

The future of AI in health care

Panelists agreed that the groundwork toward harnessing digital technologies, including AI, for good in health care continues to gather momentum from its initial spark during COVID.

“Whether we’re talking about helping to ensure that patients and families and communities have access to the treatment and care they need through our global health units, or it is the advancements of telemedicine policy here in the US, or even really helping to advance innovation through the utilization of AI and machine learning; It is our goal every day as we chase the miracles of science, to help ensure that people actually are able to live the lives that they were destined to live,” Sullivan says, citing Sanofi’s commitment to leveraging AI responsibly to help improve health equity. 

In order to live their best lives, people are taking AI data entry into their own hands, aided by organizations, too. For example, Chen is hopeful that the hard work for data integrity in AI is already under way. She is excited about bottom-up approaches, like patients with multiple myeloma who are banding together and voluntarily sharing their data with disease registries. “So people who want to focus on just this one disease can get data on the order of thousands of patients,” Chen says. She adds that the National Institutes of Health’s All of Us is specifically geared toward building its database with data shared by people with diverse backgrounds and communities that haven’t historically been included in health data collection.

Stultz says that digital twin technology, where digital replicas can simulate varying future conditions and predict outcomes, can find success in health care. “Could I give this computer copy of myself three different medications and follow their trajectory and see whether that makes them feel better or not? That’s the Holy Grail in the landscape of medicine,” Stultz says. “Ten or twenty years ago, this was science fiction. Today it really is science.”