Style Transfer–Assisted Deep Learning Method for Photoplethysmogram Denoising

Wearable devices like photoplethysmogram (PPG) sensors are prone to motion artifacts, affecting the quality of cardiovascular data. Traditional denoising methods often degrade the signal, while AI-driven solutions like deep learning (DL) struggle with random and systematic distortions, requiring large datasets for training. To overcome these limitations, this study proposes a style transfer-assisted cycle-consistent generative adversarial network (stccGAN) for denoising 3-channel PPG signals (red, green, and infrared) acquired by the patented Soundi chest sensor. Two identical devices were used: one to collect the chest PPG signal (to be denoised) and another to obtain a synchronized finger PPG signal (reference signal). The proposed stccGAN uses style transfer with dual generators featuring U-Net, GRU, and LSTM layers to improve chest PPG quality. Validation with data from 30 subjects (20 for training, 10 for testing) showed an average 70% correlation with the reference signal, with a 15% improvement over raw chest PPG. This demonstrates effective signal restoration, enabling accurate cardiac assessment and blood pressure estimation from chest PPG signals.