An artificial-intelligence algorithm developed at Mayo Clinic could identify left ventricular dysfunction—or a weak heart pump—in most patients based on Apple Watch data, researchers shared at a conference Sunday.
The proof-of-concept study was funded by Rochester, Minnesota-based Mayo Clinic without technical or financial support from Apple.
Left ventricular dysfunction, which affects 2-3% of people globally, might be accompanied by symptoms like shortness of breath, legs swelling or an irregular heartbeat, but it sometimes has no symptoms at all, said Dr. Paul Friedman, chair of Mayo Clinic's cardiovascular medicine department in Rochester and a researcher on the study.
With improved screening for left ventricular dysfunction, providers could prescribe treatments to prevent patients from developing symptoms and lower the likelihood of hospitalization and the risk of death, he added. Diagnosis typically requires a visit to a healthcare facility for an imaging test, such as an echocardiogram, CT scan or MRI.
The AI algorithm developed at Mayo Clinic correctly identified 13 of 16 patients who had a weak heart pump from Apple Watch data during a six-month study period, researchers shared in a presentation at the Heart Rhythm Society conference in San Francisco.
Mayo Clinic's study had two goals, according to Friedman. The first was to assess the accuracy of the AI algorithm, while also figuring out whether researchers could successfully conduct a decentralized trial, in which patients were invited to participate via email and engaged with researchers through digital tools, without an in-person component in Rochester.
Researchers enrolled more than 2,400 Mayo Clinic patients from 46 states and 11 countries, each of whom already had an Apple Watch. The algorithm analyzed data from the Apple Watch's electrocardiogram app, a feature that scored Food and Drug Administration clearance to detect atrial fibrillation in 2018.
While Apple was the first major consumer wearables company to add an FDA-cleared ECG to its smartwatch, since then, competitors like Samsung and Withings have followed suit.
Researchers opted to use the Apple Watch for this study, since Apple has app developer tools that let users share ECG data with researchers.
Study participants downloaded a smartphone app that researchers developed with Mayo Clinic's Center for Digital Health, which uploaded ECGs from the Apple Watch to a data platform at the clinic.
It was up to patients to record an ECG—the data were not automatically recorded in the background—and the app would prompt participants to share ECG readings every two weeks.
Of the 421 participants who had an echocardiogram on file for comparison, 16, or 3.8%, had left ventricular dysfunction. Thirteen of the 16 patients were also identified by the AI algorithm.
The algorithm's sensitivity, or likelihood of accurately identifying a positive result, was 81.2% and specificity, or likelihood of accurately identifying a negative result, was 81.3%.
The study abstract released Sunday is just a first test to show proof of concept, Friedman said. But it suggests "this consumer watch that you buy has the potential to diagnose a potentially asymptomatic and life-threatening disease."
To create the AI tool, researchers adapted an algorithm that had already been developed at Mayo Clinic, which detects a weak heart pump from a standard clinical ECG, which uses 12 electrode leads. The Apple Watch's ECG feature has a single lead, so provides less data.
The 12-lead algorithm was awarded breakthrough device designation by the FDA in 2019 and is licensed to Anumana, a company Mayo Clinic and software startup Nference launched last year to develop and commercialize AI algorithms for early disease detection. The area under the curve—a common measure of accuracy, with the most accurate tests being close to 1—for the 12-lead algorithm was 0.93; the algorithm designed for the Apple Watch was 0.88.
The single-lead algorithm could one day be used to screen heart patients for weak heart pump more easily and outside of a clinical environment, researchers wrote in the study abstract. They plan to continue to test the algorithm in different countries to assess whether it's accurate across populations, what its limitations are and how it could be used in clinical practice.