A neural style transfer-based denoising framework for seismocardiogram signals improved signal-to-noise ratio by 176.7% and yielded an average heart rate estimation error of 0.89 bpm.
Does a neural style transfer-based denoising framework improve signal fidelity and heart rate estimation accuracy in motion-contaminated seismocardiogram signals compared to standard denoising techniques?
A neural style transfer-based denoising framework significantly improves the fidelity of motion-contaminated seismocardiogram signals, enabling highly accurate heart rate estimation during dynamic conditions.
Seismocardiogram (SCG) signals capture the mechanical dynamics of cardiac activity, but their clinical utility is severely limited by motion artifacts during ambulatory monitoring. To overcome this challenge, we propose a neural style transfer (NST)-based denoising framework that converts motion-contaminated SCG recordings into morphology-preserving, rest-like representations. Our method leverages time-frequency spectrograms obtained from continuous wavelet transforms and a pre-trained convolutional neural network (VGG19) to suppress exercise-induced distortions while maintaining physiologically relevant timing and morphology. Across 20 participants, the proposed approach substantially enhanced signal fidelity, improving signal-to-noise ratio and peak signal-to-noise ratio by 176.7% and 152.1%, respectively, and reducing mean squared error and mean absolute error by 95.1% and 83.6%. Structural similarity increased by 70.3%, and correlation with the resting reference more than doubled. Heart rate estimated from denoised signals showed excellent agreement with electrocardiogram measurements, yielding an average error of only 0.89 beats per minute (0.73%). Furthermore, comparative evaluation demonstrated that the proposed restyling and denoising approach consistently outperformed state-of-the-art denoising techniques-including empirical mode decomposition variants, variational mode decomposition, Savitzky-Golay filtering, moving-average filtering, and wavelet-based reconstruction-achieving the lowest heart rate estimation error (0.89 bpm RMSE). Additionally, a controlled simulation confirmed the framework's restyling capability under known ground-truth conditions, yielding a heart rate estimation error of only 0.15 bpm relative to the true reference. These results demonstrate that neural style transfer enables physiology-consistent reconstruction of cardiac mechanical signals in dynamic conditions and represents a highly promising direction toward motion-resilient wearable cardiac monitoring.
Moradi et al. (Thu,) reported a other. Neural style transfer (NST)-based denoising framework vs. State-of-the-art denoising techniques (EMD, VMD, Savitzky-Golay, moving-average, wavelet-based) was evaluated on Heart rate estimation error. A neural style transfer-based denoising framework for seismocardiogram signals improved signal-to-noise ratio by 176.7% and yielded an average heart rate estimation error of 0.89 bpm.