Atrial Fibrillation and Atrial Flutter Detection Using Deep Learning

Figure 1. The figure illustrates a method for detecting AFib in single-lead ECG signals using a neural network model.

Figure 2. Visualization of our neural network for AFib detection. It is based on a ConvNextv2 [44] architecture adapted for 1D ECG data.

Figure 3. Effect of merging-threshold on classification metrics. Precision, recall, and F1–score are plotted for a 10 s window and an AFib probability cutoff of 0.75. The F1–optimal threshold lies at 40–50 s, with higher thresholds favoring recall at the expense of precision.

Figure 4. Impact of AFib classification threshold on precision, recall, and F1–score for 10 s windows with a 40 s merging gap. The optimal F1 range occurs at thresholds of 0.7–0.8, where lower cutoffs favor recall and higher cutoffs favor precision.

Figure 5. Guided Grad-CAM visualization for AFib examples. Notably, the highest attribution is observed in the P-waves, as their absence is a key criterion for AFib classification.

Figure 6. Guided Grad-CAM overlays on NSR signals from the SHDB dataset, highlighting cases with incorrect AFib annotations. Despite the mislabels, the model’s strongest attributions consistently fall on the P-wave regions.