An app researchers designed for the Apple Watch allows patients to transmit electrocardiograms to a central processing center, where analysis driven by artificial intelligence (AI) can detect left ventricular (LV) dysfunction with a promising degree of accuracy.
The technology is envisioned as an easy, inexpensive way to extend routine LV functional monitoring, without clinic visits or potentially costly imaging, to vastly more people with high-risk conditions than would otherwise be practical.
The approach could be especially useful, for example, in the frail elderly, patients with hypertension or diabetes, or patients who are being treated with cardiotoxic chemotherapy agents, observed Zachi I. Attia, MSEE, PhD, an AI engineer at the Mayo Clinic, Rochester, Minnesota.
“This is a great way to keep monitoring and testing patients who are at higher risk,” Attia told theheart.org | Medscape Cardiology. “We can imagine them getting an ECG every day by the watch, which will allow us to monitor them much more efficiently.”
The processing system for Apple Watch single-lead tracings, obtained in sinus rhythm, had been adapted from a neural network trained to look for signs of LV dysfunction in more discerning 12-lead ECGs.
Applied to 421 patients with recent, clinically indicated echocardiograms for comparison, the app-based technique identified 13 of the 16 with ejection fractions of 40% or lower, for a predictive accuracy of almost 88% based on area-under-the curve (AUC) assessment. Its sensitivity and specificity were both about 81%.
The cohort was a subgroup of 2454 patients — more than half of them women — from 46 US states and 11 countries that had downloaded the ECG-transmitting app and, in turn, uploaded thousands of single-lead tracings to the Mayo researchers over 5 months.
They had been invited to join the study mostly because they were already in the Mayo system and owned an Apple Watch with an ECG sensor, said Attia who reported the findings May 1 at the Heart Rhythm Society (HRS) 2022 Scientific Sessions. The meeting was held virtually and live in San Francisco.
The system’s predictive power “I think outperforms some of the other screening paradigms that we’ve heard about using AI,” observed James Freeman, MD, MPH, who isn’t connected to the study. “That’s important, because what we can’t have is tools that have us then doing a tremendous amount of downstream testing on people, and raising a lot of anxiety.”
“The next step would be to look at very enriched populations, and look then at real hard outcomes,” Freeman, from Yale University School of Medicine, New Haven, Connecticut, told theheart.org | Medscape Cardiology.
“It’s very promising and very impressive” that the AI-based system could pick up LV dysfunction from a single-lead ECG, added Adam C. Lee, MBBS, University of California, San Francisco, also not part of the study. But it has yet to be shown that it can improve patient outcomes without increasing the burden of tests. The technique still “needs to be looked at in terms of downstream effects,” he said.
The system’s predictive accuracy was “very similar to what you would get from a 12-lead, clinically done ECG,” Attia said. But “it’s important to say this is a proof-of-concept study” for the app-based technique.
All of the AI-driven screening and monitoring applications researched at his center, regardless of their predictive power, he said, are primarily aimed at improving patient care “in a way that doesn’t add burden or unnecessary tests.”
Indeed, a similar neural network, when previously applied to conventional 12-lead ECGs obtained in practice in a prospective randomized study, allowed clinicians to identify about 30% more cases of LV dysfunction than were detected by conventional reading. And it did so without requiring additional echocardiograms, Attia observed.
The investigators “successfully developed an infrastructure to recruit patients remotely, transmit data securely, and maintain patient engagement. And the results were quite impressive,” Sana Al-Khatib, MD, MHS, said as invited discussant after Attia’s formal presentation of the study.
Still, she noted, the results are based on only a small proportion of a select group of patients who had been invited to participate based simply on their owning an ECG-equipped Apple Watch, and who had analyzable data.
“To me this raises questions about the generalizability of their findings. It will be important to see these findings validated and replicated by other groups,” said Al-Khatib, from Duke University Medical Center, Durham, North Carolina.
“While this technology is innovative and quite promising, it’s not ready for application in clinical practice.”
But it did appear to be appealing enough for those who participated in the study. A user-friendly system that would encourage a high degree of patient engagement was a primary goal of the development process, Attia said.
Most of the study’s patients interacted with the app about eight times, a frequency that corresponded to push notifications send out every 2 weeks as reminders to check their rhythm. Although participants were allowed to ignore or block the push notifications, he noted, most didn’t and chose to continue uploading their ECGs to the processing center.
There was no evidence to support the almost cliché contention that proficiency with and enthusiasm for smartphone-based apps and related technology is inversely related to a person’s age, Attia said.
The mean age was 53 years, and “we had patients from 22 up to 92,” he observed. Surprisingly, “patients who were 60 and older actually used the app more often than the younger patients.” Possibly, he speculated, older people were more likely to be concerned about their health and welcome the app as a helpful medical device.
The study presentation stated that “Attia and other Mayo staff invented the low ejection fraction algorithm, and they and Mayo Clinic may benefit from its commercialization. This study received no financial or technical support from Apple.” Attia discloses holding stock options for AliveCor, Eko Devices, and Anumana, which holds the license for the algorithm tested in the study. Lee discloses owning stock in AtriCure and Abbott. Freeman discloses receiving honoraria or fees for speaking or consulting from Medtronic, Biosense-Webster, Boston Scientific, and Janssen Pharmaceuticals. Al-Khatib discloses receiving research grants from Medtronic, Abbott, and Boston Scientific.
Heart Rhythm Society (HRS) 2022 Scientific Sessions: Abstract LB-736. Presented May 1, 2022.