STF-CSNet: A spatio-temporal fusion deep compressive sensing method for ECG signals

Electrocardiogram (ECG) signals are crucial for medical diagnosis and health monitoring, yet their processing and transmission often face challenges due to high data volume and the need for real-time analysis. Traditional methods for ECG signal compression and reconstruction struggle with balancing efficiency and signal quality. In this paper, we propose STF-CSNet, a novel spatio-temporal fusion deep compressive sensing method designed to address these challenges. The key idea of STF-CSNet is to leverage both spatial and temporal information in ECG signals to achieve efficient compression while preserving signal integrity. Specifically, we introduce a channel attention mechanism to prioritize key features across different ECG leads, and a multi-head mixed-attention mechanism to capture long-term temporal dependencies. A spatio-temporal fusion approach is then used to enhance the representation of ECG signals by combining both time and space domain features. Our extensive experiments on the PTB diagnostic ECG database demonstrate that STF-CSNet achieves a PRD of 3.10% and an SNR of 30.74 dB at 10% sensing rate, outperforming existing compressive sensing methods for ECG signals. Additionally, our method shows strong generalization ability across multiple datasets, including the St. Petersburg INCART and STAFF III databases, achieving competitive PRD and SNR values. These results demonstrate that STF-CSNet provides a more efficient and robust solution for ECG signal compression, reducing data transmission and storage costs without sacrificing signal quality, offering significant improvements over current state-of-the-art methods.