Triple-Factorized attention framework with dynamic adaptive windowing for premature contraction classification

<title>Abstract</title> <p>Premature atrial contraction (PAC) and premature ventricular contraction (PVC) are prevalent arrhythmias with significant clinical implications. However, their transient nature and subtle morphological variations challenge reliable detection by conventional electrocardiogram (ECG) analysis. To address this, we propose historical recognition with dynamic adaptive window network (HRDAWNet), a novel attention framework integrating Triple-Factorized Attention and Dynamic Adaptive Windowing for dynamic PAC and PVC recognition. Triple-Factorized Attention mechanism synergistically encodes channel-spatio-temporal dependencies to resolve the limitation of isolated feature processing in existing methods. Dynamic Adaptive Windowing also exploits learnable relevance scoring to selectively aggregate contextual embeddings from informative historical states to overcome the fixed-length context constraints. Validated on the China Physiological Signal Challenge 2018 (CPSC2018) and MIT-BIH arrhythmia (MITBA) databases, HRDAWNet achieves stable improvement with 99.3% accuracy (CPSC2018) and 99.0% accuracy (MITBA) compared to the existing methods. Ablation studies confirm the model's robustness to physiological variability, rapid convergence and clinical interpretability via gradient-weighted class activation mapping, which localizes the critical features such as inverted P-waves in PAC and wide QRS complexes in PVC. These advancements establish a new paradigm for adaptive ECG analysis and demonstrate significant potential for real-time arrhythmia monitoring in clinical and wearable settings.</p>