The excitement over generative AI—and AI in general—has reached the multi-trillion-dollar healthcare industry, driven by the news of ChatGPT passing the United States Medical Licensing Exam (USMLE) and the rapid introduction of new healthcare-related AI applications. Bill Gates, for example, is recommending the use of generative AI tools for primary diagnoses of patients. While acknowledging that AI will inevitably misdiagnose patients, Gates argues that the upside is worth it.
Preventing misdiagnoses that can impact, at the very least, a patient’s quality of life, depends a lot on the quality and availability of the health data that is fed into the AI model. The current excitement notwithstanding, the development of AI healthcare solutions has been severely constrained by the dearth of comprehensive and representative real-world health data.
“So much of healthcare today holds data very close, it’s very challenging,” says Dr. John Halamka, President of Mayo Clinic Platform. In a paper published last year, Halamka and other researchers discussed the importance of high-quality data for the successful implementation of healthcare AI: “…even robustly built models using state-of-the-art algorithms may fail once tested in realistic environments due to unpredictability of real-world conditions, out-of-dataset scenarios, characteristics of deployment infrastructure, and lack of added value to clinical workflows relative to cost and potential clinical risks.”
Mayo Clinic’s solution to the health data challenge and to testing AI solutions in real-world healthcare setting was to establish three years ago a data platform, based initially on the de-identified records of 10 million patients. For good measure, the Mayo Clinic Platform also includes 644 million clinical notes from 5.3 million patients, collected over 40 years; 3 million echocardiograms, showing the heart muscle in action; 111 million electrocardiograms, showing electrical activity of the heart; 1.2 billion lab test results; 9 billion pathology reports; and 595 million diagnoses and 771 million procedures.
“In creating this entire corpus of multimodal, de-identified data, we’ve enabled over 150 organizations to come in to develop and validate AI models,” says Dr. Halamka.
One of these organizations is Atropos Health, offering physicians a detailed and accurate consultation and possible answers to even the most obscure medical questions. This “digital consultant” may even discover and suggest emergent treatments. It started as a research project at Stanford University, pursuing the idea that “you should be able to push a button and have the patient in front of you analyzed against every patient of the past and be told, well, you better consider this, that, or the other disease,” says Dr. Halamka.
Today, running its proprietary search and analysis programs, Atropos Health’s AI tool generates an automated report, what it calls a Prognostogram, in response to a physician inquiry, assisting in medical decision-making. While it typically takes weeks to determine the best course of treatment for an individual with a challenging case, a Prognostogram suggests clinicians’ actions, with evidence, in a matter of days.
When Mayo Clinic and Atropos Health announced their collaboration last November, Brigham Hyde, CEO of Atropos Health, explained its expected benefits: “The wealth of real-world evidence at Mayo Clinic would make the Prognostogram more efficient at advising physicians as they deliver care. In turn, this could empower Mayo Clinic to make new discoveries, develop new treatments, and improve patient health all over the globe.”
A study of the responses by the Atropos’ technology to the first 100 requests from its service concluded that “on-demand evidence generation to inform clinical decision-making is an achievable goal,” and that “as large patient data repositories are created, the potential to benefit from such a service is immense.”
The Mayo Clinic data repository not only provides a wealth of high-quality, comprehensive, and longitudinal clinical data but it is incrementally refreshed as more data is added to patients’ medical records, giving users access to continually expanding data with near real-time information.
The data, however, was initially limited to patients’ experience with the Mayo Clinic. Last July, the Mayo Clinic Platform extended its reach by collaborating with Mercy, one of the 25 largest U.S. health systems. The combination of the different populations and geographic locations of Mercy (with its 500 million de-identified patient encounters) and Mayo, is expected to improve accuracy, reduce model bias and create more diverse, and therefore stronger, treatment recommendations for patients. (See here for more on Mayo Clinic Platform and on Truveta, another health data platform, from Forbes contributor Seth Joseph).
“This is what we call a distributed data network of bringing organizations around the world together so that privacy-protected model development and validation can happen at multiple institutions. This gets us to a better state as opposed to locking our data in silos, as has been so much of the past,” says Dr. Halamka.
Dr. Halamka got his medical degree 30 years ago. “A few things have changed since then…” he wryly notes, observing that “50% of what I learned in medical school is wrong. I just don’t know which 50%.” This is probably true—or possibly even more so—for physicians that got their degrees more recently. “So what AI can do for me is it can digest the corpus of the literature as well as the multimodal patient data and say, John, consider these five diagnoses,” says Halamka.
But he warns the generative AI enthusiasts that turning the decision making entirely over to a statistical model is fraught with dangerous possibilities. “I don’t think AI means ‘artificial intelligence,’” says Dr. Halamka. “AI means augmented intelligence.”