SCD-FuseNet: spectral common-difference decoupling and lightweight cross-attention fusion for multi-lead ECG arrhythmia recognition

Multi-lead electrocardiogram analysis plays a crucial role in arrhythmia diagnosis, yet conventional interpretation methods remain limited by expert dependency and poor scalability. Deep learning methods offer improved automation but remain challenged by redundant inter-lead information, insufficient common-difference modeling, and inadequate representation of multi-scale temporal morphology. To address these limitations, we propose SCD-FuseNet, a spectral common-difference decoupling and lightweight cross-attention fusion network for multi-lead ECG classification. The model first estimates a cross-lead common spectrum in the complex frequency domain and obtains lead-wise differential spectra by spectral subtraction, so that diagnostic inter-lead variations are enhanced before temporal feature extraction. A hierarchical multi-scale residual module captures both short-duration QRS-related patterns and longer ST/T morphological trends, while a lightweight cross-attention fusion module uses the common branch as a stable query to selectively recalibrate the differential branch. We evaluate the method on three public ECG datasets: CPSC2018, Georgia, and PTB-XL. SCD-FuseNet achieves macro-AUCs of 95.19%, 90.45%, and 92.48%, respectively, using a unified 5-fold cross-validation protocol, with only 0.356M parameters and 179.16 MFLOPs. The current evaluation follows a recording-level fold assignment with window-level assessment, while patient-wise validation will be investigated in future work. These results indicate that SCD-FuseNet provides a promising and resource-aware framework for automated multi-lead ECG classification.