AINEWS
A Wake Forest University School of Medicine team trains an AI on a single lead ECG (the kind of single sensor heart tracing many smartwatches already capture) and flags three forms of heart failure.
Heart failure affects more than 6 million Americans and is a leading cause of hospitalization and death. The hardest form to catch in routine exams is heart failure with preserved ejection fraction (HFpEF), and it often stays invisible until it has already done damage. A Wake Forest University School of Medicine team has now trained an AI to read a single-lead ECG, the kind of single-sensor heart tracing many smartwatches can already capture, and flag all three of the major heart-failure phenotypes, including the one clinicians most often miss.
The study, led by corresponding author Oguz Akbilgic, was published in the Journal of the American Heart Association. A preprint version is available at SSRN abstract 5024367. The model classifies three cardiac phenotypes from a standard 12-lead ECG: reduced ejection fraction (rEF), mildly reduced ejection fraction (mEF), and HFpEF. It then performed the same classification using a single lead, the same configuration captured by the simplified ECG sensors built into many consumer wearables.
HFpEF is the load-bearing arm of the result. The condition is defined by a heart that pumps out a roughly normal percentage of blood with each beat but fails to fill and relax properly. Standard checks can look unremarkable until the disease is well advanced, which is why HFpEF tends to be diagnosed late. Reduced-EF screening from ECG already has prior literature, so adding HFpEF to a single-lead model is what makes the paper new.
The current reference standard for confirming any of the three phenotypes is echocardiography, an ultrasound of the heart that is not always available in every care setting. Akbilgic and colleagues position the AI as a triage flag for further evaluation, not a standalone diagnosis. "This is a major step forward," Akbilgic said in the release. "By identifying electrical patterns humans can't easily see, we hope to catch heart failure earlier, especially HFpEF, which is notoriously difficult to detect in routine exams."
The wearable beat is more tentative. The release frames the single-lead result as a step toward consumer-device deployment, and the same lead on a smartwatch could, in principle, run a similar model. The study did not test data collected from actual wearables, and Akbilgic and the Wake Forest team describe that adaptation as a "could eventually" rather than a delivered result. Any framing that puts this work as "AI diagnoses heart failure on your wrist" is reading ahead of the paper.
A separate peer-reviewed paper, published in 2025 in the Journal of Cardiovascular Development and Disease (doi 10.3390/jcdd13070340), describes an independent AI-ECG model aimed at HFpEF detection. The mechanism is related, but the cohort, features, and validation are distinct, so it functions here as a comparator rather than a replication.
What the study did not show is also worth naming. There is no clinical-outcome or mortality data attached to the model's flag, no wearable-collected data in the validation set, and no multi-site replication visible in the available excerpt. It is a peer-reviewed screening tool, not yet a deployed one, and it has not been tested on the rhythms a consumer device would actually record outside a clinic.
What it does offer is a useful framing: a model trained and validated on clinical single-lead ECG data can flag reduced, mildly reduced, and preserved ejection-fraction phenotypes, including the one that most often hides until late. The smartwatch version is a plausible next step, not a finished one. The gating questions for the next decade are familiar: replication across cohorts, clinical-outcome evidence, and a validation path on data the consumer devices actually collect.