麻豆传媒

This article is a collaboration between 麻豆传媒 and:

An app that passively scans heart rate data from the Apple Watch was almost as accurate as a standard electrocardiogram in picking up atrial fibrillation (Afib) but had less impressive results in a cohort closer to real world use, researchers reported.

In a cohort of 51 consecutive patients getting cardioversion for Afib, the Cardiogram app algorithm achieved a C statistic of 0.97 compared with standard 12-lead ECG, with 90% specificity at 98% sensitivity, Gregory Marcus, MD, of the University of California San Francisco, and colleagues reported in JAMA Cardiology.

In another cohort of 1,617 participants in the Health eHeart Study who had an Apple Watch, got the Cardiogram app, and used it while daily taking a single-lead ECG with an AliveCor Kardia device sent to them through the study, the app had a C statistic of 0.72 (95% CI 0.64-0.78) for detecting the 64 individuals (4%) who self-reported having persistent Afib.

The researchers cautioned that "despite the excellent test characteristics observed among sedentary patients undergoing cardioversion, the modest performance in the ambulatory scenario, a context more representative of the ultimate application of this technology, suggests that these data should be primarily interpreted as a proof of concept."

Mintu Turakhia, MD, of Stanford University's Center for Digital Health and the VA Palo Alto Health Care System, both in California, agreed that the more real-world ambulatory cohort data were "indeed humbling," pointing to 68% sensitivity at 68% specificity and a "modest" 8% positive predictive value.

Plenty of prior studies have used a variety of sensors, ranging from photoplethysmography as in this study to a specialized ECG watch band or just a cell phone camera, to measure pulsatile blood flow with fairly high discrimination for Afib, "usually in tightly controlled data," but seldom with prospective real-world validation.

Of note, the U.S. Preventive Services Task Force recently indicated it wouldn't back routine ECG screening for Afib, including that by wearable devices. Key among the reasons for that "I" recommendation (indicating insufficient evidence) was lack of data that finding Afib by such means actually reduces risk of stroke or other adverse events by enough to outweigh potential harms such as anxiety and false positives leading to unnecessary follow-up procedures.

Marcus's Cardiogram study won't change that, as it did not follow patients for clinical outcomes or to determine how Afib detection by the smartwatch app affected clinical care.

"With computational advances and more training data, it is possible that these algorithms may improve," Turakhia wrote. "However, there is also the possibility that they hit a performance ceiling that remains inferior to an accepted gold-standard. What, then, should be the tradeoff that we are willing to accept between high diagnostic accuracy and convenience, ubiquitousness, and continuous monitoring?"

Turakhia pointed to the example of invasive insertable loop recorders, used for managing Afib and detecting the arrhythmia in cryptogenic stroke patients. They have been shown in multiple studies to increase detection, but are by no means perfect, with widely varying positive predictive values (26% to 84%) across these indications for Afib episodes of 2 minutes or longer.

"However, the positive predictive value of insertable loop recorders rises substantially (90%) when restricted to AF episodes of 1 hour or longer owing to a lower likelihood of misclassification of sinus node variability or other ambient arrhythmias. Therefore, the optimal use case for [photoplethysmography] detection may be for longer AF episodes, which are possibly more thrombogenic," the editorialist suggested.

He also noted that the study used a "deep neural network" to derive the algorithm used for Afib detection by the Cardiogram app. This type of artificial intelligence uses empirical data, in this case from 9,750 Health eHeart Study participants with an Apple Watch who downloaded the Cardiogram app and linked it to the study. That makes it hard to "look under the hood" so researchers can learn directly from it what it's seeing and what associations it draws, Turakhia noted.