Recent studies have shown that deep learning systems can outperform conventional algorithms for the automated identification of atrial fibrillation (AF). However, understanding how deep learning algorithms make their decisions is notoriously hard, particularly in the context of Electrocardiogram (ECG) classification. This stands as a barrier to the widespread clinical adoption of deep learning systems. To address this, we propose an ante-hoc explanation framework for deep learning systems detecting AF using ECG data. Our approach introduces BeatVision, an invertible two-dimensional representation of ECG signals designed to capture relative heartbeat information concurrently. Deep learning models trained on this representation predict cardiac conditions, while an attribution method highlights the most relevant regions of the BeatVision matrix as a class activation map (CAM). To explain model decisions, an inverse mapping algorithm projects the CAM back onto an attention wave that allows identification of the most influential R-R intervals contributing to the model’s decision. Rather than relying on post-hoc interpretability techniques, BeatVision integrates interpretability directly into the model architecture and data representation. The generated explanations have been validated through heart rate variability (HRV) analysis and expert evaluation.