RunDAE achieved significantly higher and more stable signal-to-noise ratio improvements compared to conventional DAE and CDAE models in PPG signal denoising.
Running Denoising Autoencoder (RunDAE) models effectively suppress noise in PPG signals without introducing the segment boundary discontinuities seen in classical autoencoders.
Photoplethysmography (PPG) is vital for monitoring cardiovascular dynamics in clinical and wearable systems. However, PPG signals are highly susceptible to noise, especially during movement. The three primary sources of interference are: baseline wander, muscle artifacts, and Gaussian white noise. The vulnerability of PPG to these artifacts complicates its reliable use in real-world conditions. To remove these noises and ensure PPG signal fidelity, this study leverages sample-by-sample denoising using two Running Denoising Autoencoder architectures: a fully connected model (RunDAE) and a convolutional (RunCDAE). These models are evaluated on real PPG recordings with varying segment lengths, corrupted by additive mixtures of noise at multiple input signal-to-noise ratios. Performance metric, i.e., SNR improvement, demonstrates that the RunDAE significantly enhances signal quality while preserving morphological features of PPG waveforms.
Samann et al. (Mon,) conducted a other in Photoplethysmography (PPG) noise suppression (n=50). Running Denoising Autoencoder (RunDAE) vs. Classical Denoising Autoencoder (DAE) and Convolutional Denoising Autoencoder (CDAE) was evaluated on Signal-to-noise ratio (SNR) improvement. RunDAE achieved significantly higher and more stable signal-to-noise ratio improvements compared to conventional DAE and CDAE models in PPG signal denoising.