“We learn more from failure than success,” said Dr John Halamka about Epic’s sepsis prediction model.
Mayo Clinic digital health leaders have put AI to the test and learned that some parts of it are more useful than others.
Dr John Halamka, president of Mayo Clinic Platform, and Paul Cerrato, Mayo’s senior research analyst, recently published an article outlining the poor results of Epic’s AI-based algorithm that predicts the onset of sepsis.
They said that the digital health marketplace was crowded with attention-getting tools but that “the evidential support for many recently approved medical devices varies widely”.
The Epic Sepsis Model (ESM) is part of the Epic EHR system and is widely used by Epic customers in the US. It was designed to predict the risk of sepsis – a life-threatening complication that is often acquired in healthcare settings and is one of the most frequent adverse events during care delivery. The ESM is a proprietary tool and used a set of approximately 80 data elements including vital signs, laboratory results, comorbidities, and demographic factors.
Epic dominates the US electronic health record market and, since 2018, Epic has been used in all 90 Mayo Clinic hospitals and clinics, and by 52,000 staff. The global electronic health record is now also used by ACT Health, Melbourne’s Parkville Precinct and will soon become NSW Health’s sole provider.
Dr Halamka said that because ESM was a proprietary algorithm, there was a “paucity of information available on the software’s inner workings or its long-term performance”. However, it’s recent performance had been found to fall very short of expectations, clinical or otherwise.
Dr Halamka cited research in JAMA Internal Medicine from 2021. The study reviewed 28,000 patients in the University of Michigan health system that used Epic’s AI-based algorithm to assess sepsis risk every 15 minutes from arrival in the emergency department.
Roughly 2500 of all patients developed sepsis but Epic’s ESM failed to identify two-thirds of them. Epic’s ESM performed worse in centres with more cases of sepsis, more patients with multiple health conditions and more oncology patients.
Dr Halamka said that these kind of findings “amplify the reservations many clinicians have about trusting AI-based clinical decision support tools”. He said that the cliché, “we learn more from failures than successes” is relevant when looking at the outcomes of AI in healthcare.
“Two types of failures are worth closer scrutiny: algorithms that claim to improve diagnosis or treatment but fall short for lack of evidence or fairness; and failure to convince clinicians in community practice that evidence-based algorithms are worth using,” he said.
However, Dr Halamka is not a harbinger of AI-doom. On the contrary, he says that these results should not be used to dismiss the benefits of AI carte blanche.
“Reports like [JAMA Internal Medicine study] unfortunately tend to make clinicians not just skeptical but cynical about all AI-based tools, which is a missed opportunity to improve patient care,” he said.
In a recent podcast with health thought leader Dr Eric Topel, Dr Halamka said that alongside the need for guidelines and guardrails to AI in healthcare, on the whole, “people are starting to recognise that…this could be very good”.
He said that the US currently had “a perfect storm” with the technology, the policy and the cultural change, that would make the next couple of years really productive.
Mayo Clinic has been exploring AI use for a number of years and is now exploring how generative AI can assist healthcare using Google’s Enterprise Search in Generative AI App Builder.
Following the podcast with Dr Halamka, Dr Topel subsequently summarised the interview into an infographic that charted the maturity of AI across six, broad use cases.
Using AI to monitor patients got top billing, with hospital-at-home being identified as reaching full maturity in unconventional healthcare settings.
Patient’s seeking diagnosis via Google or generative AI, like ChatGPT, ranked least mature.
Dr Halamka is speaking at the Wild Health Summit on 11 September about the integration of AI, and specifically ChatGPT, in medicine. He is also speaking about the implications of using large EHR’s, like Epic, for Australian health systems.
Do you have a story tip for us, or a topic you would like to see us cover? Contact the editor at editor@healthservicesdaily.com.au.