Learning Periodic Patterns in ECG Signals Using TimesNet for Automated Cardiac Classification

Background/Objectives: Although deep learning methods have achieved promising performance in recent years, comparatively less attention has been given to explicitly modeling periodic and multi-scale temporal dynamics for ECG-specific representation learning within TimesNet-based frameworks. In this work, we propose an ECG-specific TimesNet-based framework for multi-label classification of multi-lead ECG recordings that incorporates periodicity-aware temporal modeling. Methods: The proposed framework utilizes Fast Fourier Transform (FFT)-guided temporal decomposition to identify dominant frequency components and reshapes ECG sequences into period-aligned representations to better capture intra-period morphological patterns and inter-period rhythm dependencies. Multi-scale convolutional TimesBlocks are further employed to learn rhythm-aware and morphology-aware temporal representations. Results: The proposed framework was evaluated on the PTB-XL dataset using two experimental settings: Three-Class classification (NORM, AFIB, PVC) and Five-Class classification (NORM, AFIB, MI, PVC, STTC). Experiments were conducted using a one-vs-rest multi-label learning strategy with independent class probability estimation. The framework achieved mean one-vs-rest test AUC values of 0.956 and 0.913 for the Three-Class and Five-Class settings, respectively. Experimental results indicated that the reduced classification complexity in the Three-Class setting was associated with improved feature separability, more stable decision boundaries, and enhanced discriminative representation learning. Latent-space visualization using UMAP and PCA demonstrated clearer clustering in the Three-Class configuration, while gradient-based interpretability analysis highlighted physiologically relevant ECG waveform regions contributing to model predictions. In addition, computational profiling demonstrated practical feasibility with approximately 1.957 million trainable parameters, 13.14 GFLOPs computational complexity, 5.230 ms average inference latency per ECG recording, and a throughput of approximately 191 ECG recordings per second on GPU hardware. Conclusions: These findings suggest that periodicity-aware temporal modeling can improve ECGF representation learning while demonstrating practical potential for computationally efficient and interpretable automated ECG analysis applications.