EMBC Special Issue: Neural Style Transfer-Based Denoising of Seismocardiogram Signals Under Dynamic Conditions

Area of Science:

• Biomedical Engineering
• Signal Processing
• Cardiology

Background:

• Seismocardiogram (SCG) signals offer insights into cardiac mechanics but are prone to motion artifacts during ambulatory monitoring.
• Existing denoising methods struggle to preserve signal morphology and physiological timing.

Purpose of the Study:

• To develop and validate a novel neural style transfer (NST)-based framework for denoising SCG signals contaminated by motion artifacts.
• To convert motion-corrupted SCG recordings into morphology-preserving, rest-like representations.

Main Methods:

• Utilized time-frequency spectrograms from continuous wavelet transforms and a pre-trained VGG19 convolutional neural network.
• Applied NST to suppress exercise-induced distortions while maintaining physiologically relevant SCG morphology and timing.

Main Results:

• Significantly improved SCG signal fidelity, with SNR and PSNR increases of 176.7% and 152.1%, respectively.
• Reduced MSE and MAE by 95.1% and 83.6%, while increasing structural similarity by 70.3%.
• Achieved highly accurate heart rate estimation (0.89 bpm RMSE), outperforming state-of-the-art denoising techniques.

Conclusions:

• NST enables physiology-consistent reconstruction of cardiac mechanical signals in dynamic conditions.
• The proposed framework offers a promising solution for motion-resilient wearable cardiac monitoring.