Atrial fibrillation (AF) is typically classified into paroxysmal (PAF) and persistent (PeAF) forms. Accurately identifying the AF subtype is crucial for optimizing patient management. However, distinguishing between these subtypes using short ECG recordings remains a clinical challenge.

This study aims to develop and evaluate deep learning (DL) models capable of differentiating PAF from PeAF using short-duration ECG segments.

We leveraged the IRIDIA-AF database, a curated collection of 988 Holter recordings, comprising annotated segments labeled as PAF. An equivalent number of ECG segments in AF were randomly extracted from 200 Holter recordings of patients with PeAF. Several ML architectures were trained and validated on these datasets. The models’ performance was assessed using standard classification metrics.

The best-performing model achieved an AUROC of 0.84 on the external test set ( Table 1 ). Attention map analysis suggested that the network captured both temporal and morphological components of the ECG waveform, potentially identifying subtle features beyond classical rhythm irregularities. In parallel, machine learning-based models indicated that most of the discriminative signal resided in RR interval dynamics, even in the absence of sinus rhythm, highlighting the informative value of ventricular response patterns in AF.

Our findings demonstrate the feasibility of using DL models to distinguish paroxysmal from persistent AF based on short ECG recordings. Trained on the IRIDIA-AF dataset, the proposed model shows promise for enhancing clinical decision-making in AF management.