Reconstruction of Corrupted Photoplethysmography Signals Using Recursive Generative Adversarial Networks

This article explores how motion artifacts (MAs) affect photoplethysmography (PPG) signals measured from the radial artery of the wrist through our wearable system called WrisTee. We propose a recursive generative adversarial network (GAN) model that reconstructs corrupted PPG signals across a wide range of signal-to-noise ratios (SNRs) from -33 to 25 dB. To train and evaluate the model's performance, we constructed a dataset of PPG signals obtained from 15 subjects and developed an algorithm for generating synthetic noisy data. The proposed GAN model enables the measurement of heart rate (HR) from synthetic data with a mean absolute error (MAE) of 1.7 bpm. Finally, our model successfully processed 77% of PPG segments from real noisy data. This method provides a foundation for developing generative models aimed at reconstructing noisy PPG data affected by MAs, therefore enhancing the accuracy of personal health monitoring devices.