SAN FRANCISCO — Real-world data and artificial intelligence are front and center among Mayo Clinic’s plans for the future.
That was the takeaway from the system's late Monday J.P. Morgan Healthcare Conference presentation, where the Minnesota nonprofit unveiled a new partnership with AI startup Cerebras Systems.
The pair will develop a foundation AI model (an AI model trained so that it can be used in a range of use cases) upon Mayo Clinic’s various structured and unstructured data, including “clinical notes, imaging reports and lab values,” said Matthew Callstrom, M.D., Mayo Clinic’s medical director for strategy and chair of its radiology department.
The arrangement, reported to involve multiple millions of dollars, is a way for AI to capture the medical care health systems like Mayo Clinic deliver that exceed what’s etched out in clinical guidelines, Callstrom said.
“Our first effort will be to use over 100,000 patients that we have complete genomic sequencing [for] to build a language model around genomics data,” he said during the conference. “We’ll be able to ask relatively simple questions, for instance, how well patients respond to treatment. And one of the first use cases will be to find out whether or not patients with rheumatoid arthritis might respond to methotrexate. With them, we might be able to more quickly pivot patients to the right treatment.”
But the new deal with Cerebras was just one component of Mayo Clinic’s bet on novel AI technologies. Callstrom described to attendees two other ongoing AI projects intended to help improve patient outcomes.
The first focuses on pancreatic cancer. The disease has a 39% five-year survival rate when diagnosed early versus just a 3% five-year survival rate when spotted late—an issue that’s compounded when radiologists “miss about half” of cases when reviewing CT scans, Callstrom said.
To address this, Mayo Clinic is currently conducting a clinical trial to see if AI can reduce imaging noise and make it easier to spot signs of pancreatic cancer when it can still be surgically removed.
“The sensitivity and specificity for this model trained on early-stage cancer is .97,” he explained. “That means that you can nearly pick up all cancers, makes the radiologists look better—we [radiologists] feel good about that—and we can actually help patients dramatically. We may move the needle on early-stage pancreatic cancer.”
The other project he described was also focused on early detection, this time for atrial fibrillation. By training a model on pre-diagnosis ECGs from patients who ultimately developed atrial fibrillation, “they found that there’s [a] signal in the ECG” that predicts the condition months in advance.
“This [model] has been deployed within our practice and over 400,000 patients are screened for this finding, and we have a dashboard integrated within Epic so that we can actually look at this,” he said. “These patients then go into monitoring, so you’re actually changing care for patients with this model.”
Callstrom said the care-focused AI projects are running alongside other early efforts that stand to greatly reduce administrative burden, an oft-cited target for healthcare AI among providers. These potential implementations are often brought forward by Mayo Clinic’s own staff; for instance, an RFA for generative AI skunkworks project yielded 250 proposals.
“It’s a little overwhelming,” he said. “We were able to fund about 15 and we’re making great progress, and many of them are reducing administrative burden.”
All the while, the $17-billion-dollar system is still pushing forward with its Mayo Clinic Platform. A body of deidentified real-world patient data bolstered by dozens of provider and application developer partners, Callstrom described it as a contrast to the more traditional pipeline approach that is speedier, more broadly accessible to partners and can help remove bias.
Mayo Clinic is also pairing the work with a push for responsible use of AI within healthcare. The organization was a co-founder of the Coalition for Health AI, an industry group that created a blueprint for “trustworthy” healthcare AI and has signed onto similar efforts from the White House.
Callstrom said these commitments and an emphasis on transparency are necessary for patients to accept the new technology—though he noted that, at least anecdotally, more people are starting to come on board.
As for the other side of the fence, executives said that current healthcare workers at Mayo Clinic and other provider employers have little to fear from the new tools.
“[Employees] know that most of these areas, there’s workforce shortages. This is actually covering the gap. I don’t see that, even if we were super successful, that we’d be able to reduce our headcount. Would it slow down the growth of our headcount? I think so, eventually, as we have best practices [and] build things to be more efficient.”
“That’s absolutely right,” Mayo Clinic Chief Financial Officer Dennis Dahlen added. “You can’t shelter yourself from change. … We can soften that with principal leadership and thoughtful approaches, but the real opportunity is just to bend the curve of staff growth in the future.”