The main challenge in utilizing deep learning (DL) for clinical diagnostic support is its lack of explainability and interpretability. Recent approaches aim to explain DL decisions from electrocardiogram (ECG) analysis by tracing model explanations back to beat segments. Fibrillatory (F) waves, as a main characteristic of atrial fibrillation (AF), are irregularly distributed over the signal and have not yet been considered.
Using 477 publicly available AF ECGs, we systematically investigated the relationship between F waves and reliable model explanations. F waves were detected using peak detection after removing beat-aligned QT templates. We employed a convolutional neural network, derived from an explainable ECG architecture (xECGArch), which uses self-learned morphology features for AF detection.
Analysis of variance revealed an increased mean relative relevance (rR) of the F waves compared to the rR of the full waveform (+13.5 %, p < .001). When limiting F wave detection to areas without overlap with other morphologic features, the rR increased by 13.1 % and exceeded the rR of the full waveform by 28.3 % (p < .001).
The rR peaking in the flank facing the QRS complex indicates the importance of the distance to QRS complexes for differentiation from P waves. For the first time, we attributed DL explanations to F waves, improving DL interpretability, which is essential for clinical use.