智能手表检测左心室收缩功能障碍?研究结果不尽人意
Smartwatch Detection of Reduced LVEF? Not So Fast, Study Says.
一项韩国研究显示,尽管人工智能心电图模型在区分左心室收缩功能障碍(LVSD)方面表现出良好的鉴别能力,但智能手表在实际应用中存在一致性问题。在2000多名患者的临床心超检查中,两种智能手表技术的AUC分别为0.907和0.895,但9%的患者在两种手表上检测结果不一致。
— Disappointingly, good discrimination of an AI-ECG model did not translate into consistent results
by Nicole Lou, Senior Staff Writer, MedPage Today
November 9, 2026 • 4 min read
The enthusiasm for smartwatch detection of left ventricular systolic dysfunction (LVSD) was tempered by a Korean study indicating difficulties with the technology in its current state.
Using an artificial intelligence (AI)-ECG model trained to detect LVSD, researchers set a threshold for a result of LVSD versus no LVSD when applied to Apple Watches and Samsung Galaxy Watches.
Over 2,000 patients undergoing clinically-indicated echocardiography were provided both watches for ECG pairing, and the area under the curve came out to 0.907 and 0.895, respectively, between the two smartwatch technologies, with similar sensitivities of 69.2% and 53.4%, and specificities also comparable at 88.9% and 93.4%.
At the prespecified shared model cutoff, however, 9% of the cohort had discordant results: 158 people were positive for LVSD only on Apple Watch and 42 only on Galaxy Watch. Apple Watch scores systemically trended higher, reported Hak Seung Lee, MD, of Sejong Medical Research Institute in Bucheon and Medical AI Co. in Seoul, South Korea, in his presentation at the American Heart Association (AHA) annual meeting held in Chicago.
Problems also arose regarding within-person inconsistencies. In a smaller cohort of patients referred for echocardiography who recorded at least two ECGs a day at home, 49.3% reported both positive and negative recordings in the same week at the same cutoff using the same device (a Galaxy in the vast majority of cases), according to Lee.
Attempts to tinker with these results for a clearer finding, including a strategy of averaging over a week, produced limited success: A stricter rule dropped false positives but was offset by more missed LVSD cases.
"So, a single negative home recording is not reassuring on its own, and mixed results should prompt further evaluation," Lee said during a press conference. "Smartwatch AI is making progress, but we still need more research and more evidence before we can apply it in [a] real-world setting ... We need [a] more specific device, specific cutoff for fine-tuning."
"Finding the right balance is an important next step," he concluded.
Smartwatch detection of LVSD has appeal for making it easier and more accessible to detect LVSD in people with heart failure or pre-heart failure. This seemed like a promising approach given a 2022 study reporting good accuracy of an AI model, using the single-lead ECG of a smartwatch, for detecting LVSD.
"We have incredibly effective treatments for heart failure, yet implementation is poor ... There is an urgent and unmet need that patients with or at risk for heart failure need to be identified early to promote treatment," said Jessica Golbus, MD, MS, of the University of Michigan in Ann Arbor, a panelist at the AHA press conference.
Regarding smartwatch LVSD detection, the question is "how, as a health system, we're going to deal with the false positive burden while balancing that against the urgent and unmet need to identify individuals with asymptomatic and symptomatic LVSD," Golbus said. "We must develop strategies that can identify LVSD without overwhelming healthcare resources."
She also underscored that "appropriate strategies are in place" when LVSD is detected, so that "patients can be connected with their health systems, and the treatments that we know are life-saving can be instituted. And so, an important next step is going to be looking at how these results translate into clinical outcomes."
The present study covered two prospective cohorts at Incheon Sejong Hospital in South Korea.
Lee said the AI-ECG model used for the study, AiTiALVSD-1L, is a single lead model pre-trained on nearly 3 million ECGs without labels and fine-tuned for LVSD on about 360,000 clinical lead I ECGs. A cutoff of 23.5 was set in a separate Galaxy Watch tuning set at 98% specificity; this cutoff was applied to Apple Watches, too.
The study defined LVSD as an ejection fraction of 40% or less on ECG within 14 days of echocardiography.
The analysis of Apple Watch versus Galaxy Watch included 2,211 people with a clinical indication for echocardiography. There was known LVSD in 6.6% of this cohort, and 31% had known heart failure.
Separately, 146 people took at-home ECG readings using one smartwatch. Of this group, 33.6% had LVSD and 55.5% had known heart failure.
Chief among the study's limitations was its single-center design. Investigators also did not assess smartwatch screening performance in asymptomatic people without heart failure.
"We also need to understand the performance of these devices in different patient populations, such as across age, race, ethnicity, and ejection fraction," added Golbus.
来源:MedPage Today · medpagetoday.com