In a South Korean cardiology clinic, a machine learning model was asked to listen, in a sense, to the thirty-second electrical rhythm of a wristwatch and to hear what the body often hides: a heart whose main chamber has begun to pump less than it should. The condition, left ventricular systolic dysfunction, can sit silently for years, and the question here is whether a device already on so many wrists might offer an early signal. The answer, so far, is promising but unsteady. The model worked about equally well on Apple and Samsung watches, yet the same cutoff disagreed between brands for a me
AI Reads Smartwatch ECGs to Flag Hidden Weak Heart Pumping, But Results Vary by Brand
A single smartwatch ECG is a snapshot, and repeated recordings can differ
Mark
So what did the researchers actually set out to do?
Mimi
Test whether AI reading a 30-second smartwatch ECG could flag people with a weak left ventricle, someone who might need a proper echocardiogram. LVSD can be silent, which is the whole worry.
Mark
And did it work?
Mimi
On discrimination, about equally well on Apple and Samsung watches in 2,211 clinic patients. But with one fixed cutoff, 200 people, 9%, got opposite answers on the two watches.
Luke
That cutoff was set using Galaxy Watch data, so some mismatch may reflect calibration rather than the watches being fundamentally different. The report doesn't say which.
Mark
What about the home recordings?
Mimi
Requiring two positives out of three consecutive days cut false positives from 18 to nine among 70 people without LVSD. But it caught 31 of 38 true cases instead of 33.
Luke
Keep the denominator in view. That's 108 people, and the authors call it exploratory. Halving false positives sounds big, but it's nine people.
Mark
Why does the 81% figure matter?
Mimi
Of those consistently positive, 81% had confirmed LVSD, versus about 6% of consistent negatives. Consistency seemed to carry information.
Luke
But those patients were referred for echocardiograms, many with known heart failure. That inflates how often a positive is right. The authors adjusted to lower prevalence, but I can't tell you those adjusted numbers from this account.
Mark
And whether it helps patients?
Luke
Unknown. No evidence yet on earlier treatment, fewer hospitalizations or survival. It's one hospital, and the model's developer is a study author.
O Pulso
- A thirty-second wristwatch ECG, read by AI, can separate people with weak heart pumping from those without, on both Apple and Samsung devices.
- The same fixed cutoff labeled about 9% of clinic participants differently depending on which watch recorded them, unsettling the idea of one universal threshold.
- Home users' single readings wavered day to day, and nearly half of home participants showed mixed results that could not be cleanly sorted.
- Requiring two positives in three days halved false alarms among people without the condition, but it also missed a few real cases, forcing a tradeoff between alarm and omission.
- The findings remain exploratory, drawn from one hospital and referred patients, and the study has not yet shown that early flags lead to better outcomes.
In a South Korean cardiology clinic, a machine learning model was asked to listen, in a sense, to the thirty-second electrical rhythm of a wristwatch and to hear what the body often hides: a heart whose main chamber has begun to pump less than it should. The condition, left ventricular systolic dysfunction, can sit silently for years, and the question here is whether a device already on so many wrists might offer an early signal. The answer, so far, is promising but unsteady. The model worked about equally well on Apple and Samsung watches, yet the same cutoff disagreed between brands for a meaningful minority of people, and repeated readings on different days sometimes told different stories, which reminds us that a single heartbeat-trace is only a moment in a longer pattern.
At an outpatient cardiology clinic at Incheon Sejong Hospital, 2,211 adults scheduled for an echocardiogram sat quietly while staff recorded their heart rhythm twice, first on an Apple Watch and then immediately on a Samsung Galaxy Watch. Each recording was a single lead lasting thirty seconds. The purpose was to find out whether an AI model could read these brief strips and detect a weakened heart that the patient might never feel.
The condition under study, left ventricular systolic dysfunction, means the heart's main pumping chamber no longer contracts well. The researchers defined it as an ejection fraction of 40% or less. It often causes no symptoms, yet left untreated it can progress to heart failure. The model, developed by Medical AI Co., was led by cardiologist Hak Seung Lee, who said many people quietly lose pumping strength and that the team hoped smartwatch ECGs could identify those who deserve further screening.
In the clinic, the model distinguished affected patients from unaffected ones about equally well on both watches. The difficulty appeared with a fixed cutoff, set in advance from Galaxy Watch data and then applied to both devices. Under that single threshold, 9% of participants, 200 people, tested positive on one watch and negative on the other.
A second group of 146 outpatients recorded at home, most with Galaxy Watches, logging a median of 2.4 ECGs daily. Among the 108 who recorded on at least five days in the first week, requiring positives on two of three consecutive days cut false positives among unaffected people from 18 to nine. The cost was two additional missed cases among the 38 with the condition. Across all home participants, 81% of those consistently positive had confirmed disease, compared with about 6% of those consistently negative, while the 49% with mixed results fell in between.
Lee summed up the lesson: one smartwatch reading is a snapshot, and repeated readings can disagree. The researchers stressed that the findings come from a single hospital, referred patients, and a small exploratory comparison. Whether flagging these individuals leads to earlier treatment, fewer hospitalizations, or longer survival remains unknown, and larger studies will need to answer that.
At an outpatient cardiology clinic at Incheon Sejong Hospital in South Korea, 2,211 adults who were scheduled for an echocardiogram sat while staff recorded their heart rhythm twice: once on an Apple Watch, then right away on a Samsung Galaxy Watch. Each tracing was a single lead lasting 30 seconds. The point was to see whether an AI model could read those short strips and spot a weak heart that the patient might not feel.
The condition is left ventricular systolic dysfunction, or LVSD. The left ventricle, the heart's main pumping chamber, fails to contract effectively. The study defined it as an ejection fraction of 40% or less, meaning the chamber expels no more than 40% of its blood with each beat. It can exist without symptoms, and left untreated it can progress to symptomatic heart failure. The findings were presented as preliminary, independent late-breaking work at the American Heart Association's Scientific Sessions 2026, held Nov. 6-9 in Chicago.
The AI model was developed by Medical AI Co., Ltd. Its chief medical officer, Hak Seung Lee, is the study's author and also a cardiologist at the Digital Healthcare Institute, Sejong Medical Research Institute in Incheon. Lee said some people have reduced pumping function without noticeable symptoms, and the team wanted to know whether smartwatch ECGs could identify those who might benefit from further screening.
In the clinic group, the model separated patients with and without LVSD about equally well on both brands of watch. The trouble came with a fixed threshold. The researchers set the AI score cutoff in advance using separate Galaxy Watch data, then applied it to both watches. Under that single cutoff, 200 participants, 9%, were positive on one watch and negative on the other.
A second, separate group tested a different problem: what happens when people record at home, repeatedly. It included 146 outpatients who used their own smartwatches, 94% of them Galaxy Watches, with staff reminders to keep recording. They logged a median of 2.4 ECGs per day, and the first week of recordings was analyzed. Of these, 108 recorded on at least five days that week.
Among those 108, the researchers tried requiring a positive AI result on at least two of three consecutive recording days. Of the 70 participants without LVSD, false positives fell from 18 to nine, a 50% reduction. The cost was sensitivity: the rule identified 31 of 38 participants with LVSD, versus 33 when any single positive day counted. That is two more missed cases. This was one of several repeat-testing rules examined.
Across all 146 home participants, 81% of those with consistently positive results had LVSD confirmed by echocardiography, compared with about 6% of those with consistently negative results. Nearly half, 49%, had mixed results, and 40% of that group had confirmed LVSD. Because the participants were referred for echocardiograms, and many already had known heart failure, LVSD was far more common than in the general population. The researchers adjusted the analysis and estimated results at lower prevalence.
Lee put the lesson plainly: a single smartwatch ECG is a snapshot, and repeated recordings can give different results. Requiring repeated positives cut false alarms but also missed cases that should have been flagged, he said, and larger studies are needed to decide how to use the results to guide further evaluation.
The limits are considerable. The work came from one hospital in South Korea, in patients referred for echocardiography, so it may not apply to the general public. The repeat-testing comparison rested on only 108 people and is exploratory. Recordings of poor quality, under 3%, were excluded, and home participants were reminded to record. ECGs were taken within 14 days of the confirming echocardiogram. No 12-lead ECGs were available for comparison in the clinic group, and the study did not test whether using the AI results leads to earlier treatment, fewer hospitalizations or better survival. Those questions are what larger studies would need to answer.
Citações Notáveis
A single smartwatch ECG is a snapshot, and repeated recordings can give different results.— Hak Seung Lee, study author, Medical AI Co., Ltd.
Quer a matéria completa?
Leia o original em Medical Xpress ↗
Technology