Intelligent monitoring for the robust diagnosis of cardiovascular diseases using continuous long term ECG recordings
PI-Technion, Assoc. Prof. Joachim A. Behar (Oxon), PhD, PI-Rambam, Prof. Mahmoud Suleiman, MD, Research Team Dr. Ronit Almog, MD, MPH, Rambam
AI can transform cardiology care by spotting patterns in ECGs, echocardiograms, CT/MRI, labs, and clinical notes that are easy to miss and hard to synthesize quickly. It can help clinicians diagnose earlier, predict risk (like heart failure decompensation or arrhythmias) more accurately, and tailor treatment to a patient’s unique profile. By automating routine tasks such as triage, report drafting, measurement, and follow-up reminders, AI can reduce delays and free clinicians to spend more time with patients.
When combined with remote monitoring from wearables and home devices, it can enable proactive, continuous care instead of reactive visits, provided systems are rigorously validated, privacy-protecting, and used with clinician oversight.
Atrial fibrillation detection and phenotyping using artificial intelligence
Objectives and clinical need
This research aims to enable a robust, low-cost, and remote pathway for diagnosing atrial fibrillation (AF). AF is the most common arrhythmia, affecting about 3% of adults with prevalence increasing with age, and is associated with a markedly higher risk of stroke and mortality as well as substantial healthcare costs. Current diagnostic standards, based on short 12-lead ECGs or typically 24-hour Holter recordings, miss a large proportion of AF cases, leaving many patients undiagnosed and untreated. Evidence shows that substantially longer recordings detect far more AF episodes. The high clinical burden of AF, combined with the limitations of conventional monitoring and the large number of undiagnosed patients, underscores the need for more reliable and scalable diagnostic approaches.
Databases
A dataset of 3,000 (24-hour) Holter ECG coupled with patients EMR data was curated for this research at Rambam Healthcare Campus. Additional external datasets from the USA, Japan and China were used for the research.
Research results
Our research focused on improving the robustness and generalizability of atrial fibrillation (AF) detection across populations, geographies, and signal representations. We developed and validated deep learning models for AF detection using both beat-to-beat interval data and raw single-lead ECG signals, addressing key limitations of prior work that often showed strong performance only on homogeneous datasets.
In a large retrospective study including 4,298 Holter recordings and more than 99,700 hours of continuous data from centers in the USA, Israel, Japan, and China, we demonstrated that deep learning models can reliably detect AF across ethnicities, ages, and sexes. This work was among the first to rigorously evaluate cross-geographical and demographic generalization for AF detection from long-term beat-to-beat intervals. Complementing this, we introduced a novel methodology for evaluating single-lead fibrillatory wave (f-wave) extraction algorithms without requiring ground-truth f-waves. The source code for the AF beat detector is available here and for the f-wave analysis is available here. The software for heart rate variability and ECG waveform analysis is available here.
Publications:
- Biton, S., Aldhafeeri, M., Marcusohn, E., Tsutsui, K., Szwagier, T., Elias, A., Oster, J., Sellal, J. M., Suleiman, M., & Behar, J. A. (2023). Generalizable and robust deep learning algorithm for atrial fibrillation diagnosis across geography, ages and sexes. npj Digital Medicine, 6(1), 44. https://doi.org/10.1038/s41746-023-00791-1
- Ben-Moshe, N., Tsutsui, K., Brimer, S. B., Zvuloni, E., Sornmo, L., & Behar, J. A. (2024). RawECGNet: Deep learning generalization for atrial fibrillation detection from the raw ECG. IEEE Journal of Biomedical and Health Informatics, 28(9), 5180–5188. https://doi.org/10.1109/JBHI.2024.3404877
- Ben-Moshe, N., Biton Brimer, S., Tsutsui, K., Suleiman, M., Sörnmo, L., & Behar, J. A. (2025). Machine learning for ranking f-wave extraction methods in single lead ECGs. Biomedical Signal Processing and Control, 99, 106817. https://doi.org/10.1016/j.bspc.2024.106817