Among 146 outpatients in South Korea who recorded electrocardiograms on a smartwatch at home, 81% of those with consistently positive results from an artificial intelligence model turned out to have weak heart pumping on a hospital echocardiogram. Only about 6% of those with consistently negative results did. Hak Seung Lee’s team reported the result in Chicago on Nov. 9, during the heart association’s annual Scientific Sessions meeting.
AI, explained every day: The Deep View is the go-to daily newsletter for 600,000+ founders, knowledge workers and professionals who want to stay up to date with artificial intelligence.
We may earn a commission if you sign up.
The condition is left ventricular systolic dysfunction, or LVSD, defined here as a left ventricular ejection fraction of 40% or less. In plain terms, the main pumping chamber pushes out no more than 40% of its blood with each beat. The abstract is preliminary, was not peer reviewed, and comes from a single hospital.
Two groups: 2,211 in clinic, 146 at home
The American Heart Association’s release on the study describes two cohorts. The first had 2,211 adults scheduled for an echocardiogram at Incheon Sejong Hospital, each of whom recorded ECGs on both an Apple Watch and a Samsung Galaxy Watch during the same visit. The AI performed about equally well on the two brands. Even so, 200 participants, or 9%, were positive on one watch and negative on the other, a reminder that a single reading is a noisy measurement.
The second cohort had 146 outpatients who recorded at home, 94% of them on a Galaxy Watch. Of those, 108 recorded on at least five days in the first week, and participants received reminders to keep recording. That prompting may have lifted adherence beyond what ordinary wearers would manage. The Galaxy Watch dominance also means the home result speaks mostly for one brand, even though the clinic cohort tested both. The release does not report sensitivity or specificity for the model as a whole, so the 81% is a share of flagged people who were truly affected, which is a different quantity from how many sick people the system catches.
Repeat positives: from 18 false alarms to 9
The repeat-testing result is the one that carries the 81% figure. Requiring a positive on at least two of three consecutive days cut false positives from 18 to 9 among the 70 people without LVSD. The stricter rule identified 31 of the 38 people who did have LVSD, against 33 under the looser single-positive-day rule, so two cases were lost in exchange for halving the false alarms. Across all 146, 81% of those with consistently positive results had LVSD confirmed by echocardiography, compared with about 6% of those with consistently negative results. Another 49% of participants had mixed results, and 40% of that group had LVSD.
Hak Seung Lee, M.D., the study author, is chief medical officer of Medical AI Co., Ltd., which developed the model, and a cardiologist at the Digital Healthcare Institute of the Sejong Medical Research Institute in Incheon. He said a single smartwatch ECG is a snapshot and that repeated recordings can give different results, adding that larger studies are needed.
A developer evaluating its own model, and the generalization gap
The developer’s role is the first caveat. The association’s page says the company name and location appear inconsistently, as Medical AI Co., Ltd. in one place and Medical AI, Ltd., in Seoul in a photo caption, but the author’s tie to the model’s maker is clear. Funding is not stated in the release; disclosures sit in the abstract itself.
Earlier work from the same group shows why watch data are hard. A European Heart Journal abstract on the single limb lead (Lead 1) of a 12-lead ECG versus a smartwatch, by J. Jang and H. S. Lee, trained a model on 151,065 standard Lead 1 tracings and then tested it on 2,441 smartwatch recordings. Performance trends tracked closely on standard ECGs, at 0.89 or higher, but only 0.54 for Apple Watch and 0.69 for Galaxy Watch, so the authors concluded that models trained on standard ECGs generalize poorly to real-world watches.
The same group has also reported on a dedicated device, an AI-enhanced six-lead portable electrocardiogram evaluated in a prospective single-centre cohort, published in European Heart Journal – Digital Health in March 2025. Beyond the watch study, TCTMD reported in July that two Medical AI models, one flagging an ejection fraction below 40%, have regulatory approval and are used clinically in South Korea.
Open questions before smartwatch screening reaches clinics
Patients were referred for echocardiography, many with known heart failure, so the results may not carry over to healthy people at home, although the authors adjusted for the higher prevalence. Recordings of poor quality, under 3% of the total, were excluded, and no 12-lead ECG comparison was available for the clinic group. The repeat-testing comparison included only the 108 people who recorded on enough days, and the release calls it exploratory.
Mayo Clinic researchers have pursued a related idea, as the maker-community site Hackster reported in describing an AI model applied to Apple Watch traces. The Korean study does not show that using these results leads to earlier treatment, fewer hospitalizations or longer survival, and that is the question any larger trial would have to answer.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
AI, explained every day: Get The Deep View →
We may earn a commission if you sign up.
More from Morning Overview
- A $1 billion gas plant planned for Superior, Wisconsin, has been abandoned
- Shaq spent $1.3 million on three Rolls-Royce Phantoms to prove a point to a salesman
- The Justice Department asked courts to undo approvals for three offshore wind farms, one already making power
- Over 20 trillion gallons of rain fell on the Four Corners states and Nevada in two weeks
