PPG denoising using maximum-mean discrepancy based variational autoencoder with data from multiple datasets

In this study, we implemented a maximum-mean discrepancy based variational autoencoder (MMD-VAE) for the denoising of photoplethysmogram (PPG) signals, using data from multiple datasets. We applied random masking to generate noisy counterparts for clean 10-second segments. We report evaluation results on PPG-DaLiA and WESAD. Using only PPG data, our approach outperforms existing methods on WESAD, and achieves performance similar to the state-of-the-art on PPG-DaLiA. The results highlight the importance of leveraging multiple datasets for effective model training. Overall, the findings validate the suitability of the MMD-VAE for PPG denoising.