Smartwatch AI Misses Heart Dysfunction and Gives Conflicting Results

Smartwatch screening for weakened heart pumping could help identify heart failure earlier, when treatment can make a meaningful difference. But a Korean study reported at the American Heart Association meeting suggests that promising overall performance does not necessarily translate into dependable results across watches or repeated measurements at home.

Researchers applied an artificial intelligence model to watch electrocardiograms, which record the heart's electrical activity. More than 2,000 patients referred for heart ultrasound testing used both Apple and Samsung Galaxy watches. Overall discrimination was strong, but sensitivity was 69.2% for Apple and 53.4% for Galaxy. Specificity was 88.9% and 93.4%, respectively, indicating that both missed cases and false alarms remained concerns.

At the same predefined cutoff, 9% of participants received conflicting classifications between watches: 158 tested positive only with Apple and 42 only with Galaxy. Apple scores tended to run higher. In a smaller home monitoring group, mostly using Galaxy watches, 49.3% recorded both positive and negative results within one week using the same device.

These inconsistencies matter because a screening tool must balance finding disease against unnecessary additional testing. Averaging recordings across a week offered limited improvement. Stricter interpretation reduced false positives but increased missed cases. The findings suggest that different devices may need separately validated cutoffs, rather than assuming one threshold works equally well for every watch and every recording.

For users, a single negative recording should not provide reassurance on its own, and mixed results warrant clinical evaluation rather than self-diagnosis. These experimental predictions do not replace heart ultrasound or professional assessment. Important caveats include recruitment from patients already referred for imaging, limited details about the smaller home cohort, and reporting from a conference presentation rather than a complete published paper. The supplied article is also truncated, so full methods and independent validation should guide any clinical adoption.