AI heart-screening study finds smartwatch choice changes results for 9% of patients

An AI heart-screening result can change with the watch recording it. In preliminary research presented November 9 at the American Heart Association’s Scientific Sessions 2026, Apple and Samsung watches produced conflicting classifications for 9% of clinic participants. Repeated checks at home reduced false alarms, but also missed more people with weakened heart pumping.

The study explored whether a familiar wrist device could help identify people who need a heart ultrasound, called an echocardiogram. Its target was left ventricular systolic dysfunction, or LVSD: weakness in the heart’s main pumping chamber. The condition can exist without noticeable symptoms and, untreated, can lead to symptomatic heart failure.

Medical AI Co., Ltd. developed the model, which analyzed 30-second smartwatch electrocardiograms, or ECGs—recordings of the heart’s electrical signals. Study author Hak Seung Lee is the company’s chief medical officer and a cardiologist. The American Heart Association cautioned that the conference abstract has not been peer-reviewed; its findings remain preliminary.

Two watches, one clinic visit

Researchers first enrolled 2,211 adults scheduled for echocardiograms at Incheon Sejong Hospital in South Korea, between July 2023 and May 2024. Staff supervised back-to-back ECG recordings on an Apple Watch and a Samsung Galaxy Watch during the same clinic visit. Heart ultrasounds provided the reference for checking whether participants had LVSD.

The model distinguished people with and without the condition about equally well across the two brands. But that overall finding did not mean each person received the same result. Applying one score cutoff, set beforehand using separate Galaxy Watch data, classified 200 participants as positive on one watch and negative on the other.

At home, repeated checks change the tradeoff

A second group of 146 outpatients enrolled between July 2024 and December 2025 recorded ECGs at home on their own watches. Most—94%—used a Galaxy Watch. Participants received reminders to keep recording, and researchers analyzed the first week. The comparison of repeat-testing rules covered 108 people who recorded on at least five days.

One rule required positive AI results on at least two of three consecutive recording days, rather than a single positive day. That halved false-positive classifications, meaning alerts in people whose ultrasounds showed no LVSD. The stricter rule also missed two additional people who did have the condition. It was one of several repeat-testing approaches examined.

In the full home-recording group, nearly half had a mixture of positive and negative results. Ultrasounds confirmed LVSD in 40% of that mixed-result group. Even consistently negative readings were not a guarantee: about 6% of participants with that pattern had confirmed LVSD.

The next test is broader than detection

Both groups came from one hospital and consisted of patients referred for echocardiography, many with known heart failure. LVSD was therefore more common than in the general population. Researchers adjusted their analysis to estimate results at lower prevalence, but the study’s patient selection limits how widely its findings can apply.

The home repeat-testing comparison was exploratory, and poor-quality ECG recordings were excluded from the study. More fundamentally, the research did not determine whether acting on AI results produces earlier treatment, fewer hospitalizations or longer survival. Lee said larger studies are needed to establish how these readings should guide further evaluation—not replace traditional screening exams.