Heart failure affects 11.8% of adults aged 65 and older, reducing quality of life and longevity, according to a study published on arXiv. Researchers at the Technion and Leumit Health Services have developed a deep learning model called DeepHHF that uses 24-hour single-lead electrocardiogram (ECG) data to predict the risk of heart failure within five years. The model achieved an area under the receiver operating characteristic curve (AUC) of 0.80, outperforming models based on 30-second ECG segments and clinical scores.
Deep Learning on Day-Long ECG Data
The study leveraged the Technion-Leumit Holter ECG (TLHE) dataset, which includes 69,663 recordings from 47,729 patients collected over 20 years. DeepHHF processes the full 24-hour waveform, capturing paroxysmal events that shorter recordings miss. The authors reported that high-risk individuals identified by DeepHHF had a twofold chance of hospitalization or death compared to those not flagged as high-risk.
Explainability and Clinical Relevance
Explainability analysis of DeepHHF showed that the model focused on arrhythmias and heart abnormalities when making predictions. The authors emphasize that artificial intelligence applied to single-lead Holter ECG is non-invasive, inexpensive, and widely accessible, making it a promising tool for heart failure risk prediction in clinical settings.
Model Performance Comparison
The table below summarizes the performance of DeepHHF compared to alternative approaches, as reported in the study:
| Model / Approach | AUC |
|---|---|
| DeepHHF (24-hour ECG) | 0.80 |
| Model using 30-second segments | Lower (not specified) |
| Clinical score | Lower (not specified) |
The authors note that DeepHHF's ability to model continuous 24-hour ECG data enables it to detect transient cardiac events that are critical for reliable risk assessment.
Implications for Preventive Cardiology
By identifying individuals at high risk of developing heart failure years in advance, DeepHHF could enable earlier interventions to reduce morbidity and mortality. The study, published on arXiv under reference 2601.00014, demonstrates the feasibility of deep learning for modeling long-duration physiological signals. The authors conclude that this approach could be integrated into routine Holter monitoring workflows, providing a scalable screening tool for older adults.