Deep Learning-based Translation of Compressed Seismocardiogram into Electrocardiogram and Breathing Signals

Wireless wearable biosignal monitoring systems face two major challenges: 1) the large volume of data generated during long-term monitoring, which heavily burdens storage and transmission, and 2) the need to use many electrodes or sensors, which is cumbersome and impractical for daily use. To simultaneously address these challenges, we propose a deep learning (DL) architecture that translates compressed seismocardiograms (SCG) into electrocardiograms (ECG) and breathing signals. Since both SCG and ECG are biosignals generated by heart vibrations, and breathing influences SCG, we successfully disentangled ECG and breathing signals from the compressed SCG data. For signal reconstruction, we developed a modified residual attention U-Net architecture consisting of a linear block, a U-Net backbone, and additional convolutional layers. Extensive experiments were conducted to evaluate the proposed DL architecture using the Combined Measurement of ECG, Breathing, and Seismocardiogram (CEBS) database across various compression ratios (CR) and noise levels. Performance was assessed in terms of Root Mean Square Error (RMSE), Percentage Root Mean Square Difference (PRD), and Pearson Correlation Coefficient (CC). At a CR of 50% and a noise level of 10 dB, the mean RMSE, PRD, and CC values were 0.051, 11.640, and 0.946 for ECG signals, and 0.126, 35.617, and 0.882 for breathing signals, respectively. Additionally, clinical analysis of the reconstructed ECG demonstrated an R peak detection with the matching rate of 98.54% and an averaged R-R interval error of 1.81 ms.