An Adaptive Wavelet–Attention Framework with Dependency-Aware Feature Modeling for Robust ECG Arrhythmia Classification

Accurate identification of cardiac arrhythmias from Electrocardiogram (ECG) signals is an important element of cardiac care, especially as wearable and portable monitoring devices become more widespread. Despite advances in deep learning, the reliability of automated arrhythmia classification often declines when ECG recordings are affected by baseline drift, motion artifacts, or power-line interference. This work introduces a two-stage framework that combines adaptive ECG denoising with dependency-aware deep feature learning. The first stage, the Adaptive Threshold–Driven Time–Frequency Framework (AT-TFF), converts raw ECG signals into refined time–frequency representations using adaptive wavelet thresholding, retaining diagnostically important waveform structures while removing narrow-band and motion-induced noise. The denoised scalograms are then processed by the proposed Feature Dependency Modeling and Extraction Network (FDMENet), which combines dynamic convolutional feature extraction with a Hierarchical Shifted Dependency Attention Block to capture both local morphological detail and long-range temporal relations. The model was trained and evaluated on two public datasets, MIT-BIH Arrhythmia and PTB-XL, reaching accuracies of 96.8% and 93.5%, respectively, which exceed those of CNN, recurrent, and transformer baselines on Macro-F1, AUROC, and PR-AUC. The results show that combining adaptive signal cleaning with context-driven feature learning improves robustness and diagnostic reliability in both clinical and ambulatory ECG monitoring. Ablation and noise-robustness experiments confirm the contribution of each component.