Temporal Phenotyping of Paroxysmal Atrial Fibrillation Reveals Prognostic Circadian Subtypes

Joachim Behar
3 days ago
3 min read

Updated: 2 days ago

This work represents a close collaboration between the following authors: Shany Brimer Biton, Jonathan Sobel, Anat Reiner Benaim, Eran Zvuloni, Ronit Almog, Julien Oster, Izhar Laufer, Ilan Green, Mahmoud Suleiman, Kenta Tsutsui, Leif Sörnmo, and Joachim A. Behar.

URL: [] GitHub:

The collaboration brings together expertise from the Faculty of Biomedical Engineering at the Technion (Israel); Geneva University Hospitals and the University of Geneva (Switzerland); Ben-Gurion University of the Negev (Israel); Rambam Health Care Campus (Israel); Inserm and Université de Lorraine (France); Leumit Health Services (Israel); Teikyo University School of Medicine (Japan); and Lund University (Sweden).

Why Timing Matters in Atrial Fibrillation

Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia, affecting millions worldwide. Clinically, AF is usually classified into broad categories such as paroxysmal or persistent, and risk is assessed using summary metrics like overall AF burden. But this binary or aggregate view hides something fundamental: AF is a dynamic condition that unfolds over time.

Patients often report that their AF episodes cluster at specific times of day such as during sleep, early morning hours, or daytime activity. While previous studies hinted at circadian patterns, results were inconsistent, limited by small cohorts, predefined time windows, or simplistic definitions of AF onset.

Our latest work takes a different approach: letting the data speak for itself.

From Raw ECGs to Circadian Phenotypes

In this study, we analyzed nearly 59,000 24-hour Holter ECG recordings from over 40,000 patients, collected across 20 primary care centers over more than a decade.

The analysis pipeline combined two powerful machine-learning components:

Deep learning for AF detection: AF episodes were identified using ArNet2, a state-of-the-art deep neural network trained on over 50,000 hours of ECG data. Rather than relying on noisy rule-based methods, this model detects AF with high robustness across ages, sexes, and recording conditions.
Unsupervised learning for temporal phenotypingInstead of focusing on when AF starts, we calculated short-term AF burden in consecutive one-hour windows across the day. These 24-hour AF burden profiles were then clustered using hierarchical unsupervised learning that is without assuming any predefined circadian structure.

The result: distinct, data-driven circadian “chronophenotypes” of atrial fibrillation.

Three Circadian AF Chronophenotypes

Among patients with paroxysmal AF, three clear temporal patterns emerged:

Nocturnal-to-Morning AFAF activity peaks during the night and early morning hours.
Evening-to-Early Morning AFAF burden rises in the late afternoon and persists into the night.
Daytime AFAF predominates during daytime hours.

Importantly, these patterns were not explained by differences in age, sex, comorbidities, or medications. They reflect how AF distributes itself across the 24-hour day, not who the patients are on paper.

Figure: Chronophenotyping AF using 24-h Holter recordings. Following preprocessing and AF detection, the short-term AFB is computed in 1-h segments and clustered into chronophenotypes using unsupervised learning. Each cluster is characterized by demographics and survival outcomes. (Figure created with BioRender.com)

Same AF, Different Outcomes

One of the most striking findings was that timing mattered as much as quantity.

The Evening-to-Early Morning chronophenotype had the highest AF burden, yet was not associated with increased mortality or heart failure risk.
In contrast, the Nocturnal-to-Morning and Daytime chronophenotypes showed a significantly higher risk of death, even after adjusting for age and sex.
Daytime AF was also associated with a higher risk of heart failure.

These results suggest that AF burden alone does not tell the full story. When AF occurs may reflect different underlying mechanisms such as autonomic nervous system balance or circadian regulation of cardiac electrophysiology with distinct prognostic implications.

Why This Matters Clinically

This work introduces a new way of thinking about AF:

Beyond binary diagnosis: AF is not just present or absent, it has structure in time.
Beyond average burden: Two patients with the same AF burden may have very different risks depending on when their AF occurs.
Toward personalized chronotherapy: Understanding a patient’s AF chronophenotype could guide:
- Optimal timing of ECG monitoring
- More refined risk stratification
- Future time-optimized treatments aligned with circadian biology

In short, AF has a clock — and that clock matters.

A Scalable Framework for Temporal Phenotyping

While this study focused on atrial fibrillation, the methodology is broadly applicable. Combining deep learning with unsupervised temporal clustering opens the door to chronophenotyping across many physiological signals, from arrhythmias to sleep disorders and beyond.

As long-term wearable and ambulatory monitoring becomes routine, such approaches may play a key role in the next generation of personalized, time-aware medicine.

Looking Ahead

Future work will explore longer monitoring periods, external validation across populations, and integration with sleep, activity, and autonomic data. But even now, these findings highlight a simple yet powerful insight: