Modeling Day-Long ECG Signals to Predict Heart Failure Risk with Explainable AI

A deep learning model trained on 24-hour electrocardiogram recordings can predict heart failure risk within five years, achieving an area under the receiver operating characteristic curve of 0.80, according to a study posted on arXiv [1]. The model, called DeepHHF, was developed using the Technion-Leumit Holter ECG dataset, which contains 69,663 recordings from 47,729 patients collected over 20 years [1]. Heart failure affects 11.8% of adults aged 65 and older, reducing quality of life and longevity [1]. The study's authors hypothesized that artificial intelligence applied to day-long single-lead ECG data could capture paroxysmal events and circadian variations that shorter recordings miss [4]. DeepHHF was trained in two stages. First, an encoder extracted features from independent 30-second windows of the Holter recordings. Those features were then fed into a sequential component that accounted for the full 24-hour recording, producing a score for the probability of developing heart failure within five years [4]. The model outperformed both a version trained only on 30-second segments, which achieved an AUROC of 0.77, and a clinical score [4]. Individuals flagged as high-risk by DeepHHF had a two-fold chance of hospitalization or death incidents [1]. An explainability analysis showed the model focused on arrhythmias and heart abnormalities, with key attention concentrated between 8 AM and 3 PM [3]. The American Heart Association defines arrhythmia as a heartbeat that is not in a regular pattern, and notes that abnormalities of heart rate sometimes indicate heart disease [6]. The study excluded 6,042 Holter examinations with a prior documented heart failure diagnosis from the cohort [4]. The researchers defined the heart failure endpoint by extracting the first documented diagnosis per patient, with recordings within five years prior classified as positive cases [4]. The approach uses single-lead Holter ECG, which the authors describe as non-invasive, inexpensive, and widely accessible [1]. The paper was submitted by Eran Zvuloni and revised in June 2026 [1]. The work appears on arXiv under the Electrical Engineering and Systems Science section, with code and data links available through Hugging Face [5].