WMFNet: A wavelet-guided multi-scale feature-preserving network for atrial fibrillation detection

Atrial fibrillation (AF) is a major global public health concern due to its high prevalence and association with severe complications. While deep learning has demonstrated remarkable success in automated AF detection from electrocardiograms (ECGs), prevailing models suffer from the inevitable loss of fine-grained features caused by inherent architectural operations like feature compression and selection. This often results in the loss of subtle yet diagnostically critical waveform anomalies. To address this limitation, we propose a wavelet-guided multi-scale feature preserving network (WMFNet). Our model preserves intricate ECG details throughout the feature hierarchy by integrating a feature pyramid with Haar wavelet-based downsampling, thereby mitigating critical information loss. A dual-attention cross-fusion mechanism is further employed to dynamically select and integrate the most discriminative features across scales. Extensive evaluation on the CinC2017 dataset yielded competitive performance, with an accuracy of 89.44% and an F1-score of 88.80%, achieving competitive performance against several established benchmarks. External validation indicated reasonable cross-dataset transferability, achieving 95.43% accuracy on CPSC2018 and 93.61% on ChapmanECG. Critically, Grad-CAM visualizations provided visual evidence that the model generally aligns with AF-relevant ECG regions. These attributes collectively position WMFNet as a highly promising clinical diagnostic assistance tool. Its accurate, generalizable, and interpretable analysis will enhance diagnostic efficiency in cardiac monitoring, offering a practical tool to mitigate the growing burden of AF on healthcare systems.