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

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning shows promise in ECG analysis, but explicit modeling of temporal dynamics remains underexplored.
• Existing TimesNet frameworks lack ECG-specific periodicity and multi-scale temporal feature learning.

Purpose of the Study:

• To propose an ECG-specific TimesNet framework incorporating periodicity-aware temporal modeling for multi-label classification.
• To enhance ECG representation learning for improved diagnostic accuracy and interpretability.

Main Methods:

• Utilized Fast Fourier Transform (FFT)-guided temporal decomposition for frequency component identification.
• Reshaped ECG sequences into period-aligned representations to capture intra-period and inter-period dynamics.
• Employed multi-scale convolutional TimesBlocks for rhythm-aware and morphology-aware feature extraction.

Main Results:

• Achieved mean one-vs-rest test AUC values of 0.956 (Three-Class) and 0.913 (Five-Class) on the PTB-XL dataset.
• Demonstrated improved feature separability and clearer latent-space clustering in the Three-Class setting.
• Showcased practical feasibility with efficient computational complexity and low inference latency.

Conclusions:

• Periodicity-aware temporal modeling significantly enhances ECG representation learning.
• The proposed framework offers a computationally efficient and interpretable solution for automated ECG analysis.
• Findings suggest potential for improved clinical diagnostic tools in cardiology.