Photoplethysmography Motion Artifacts Removal Based on Signal-Noise Interaction Modeling Utilizing Envelope Filtering and Time-Delay Neural Network

Physiological features extraction from Motion Artifacts (MA) corrupted Photoplethysmography (PPG) signals are of great difficulty due to the non-linear interaction between clean PPG and noise. To accurately extract features such as heart rate, blood pressure, and Oxygen Saturation in Blood, a robust PPG de-noising and waveform recovering method is required. However, traditional adaptive filtering or signal decomposition methods fail to achieve this goal. In this paper, a Signal-Noise Interaction modeling based algorithm for motion artifacts removal (SniMA) under strenuous physical exercise utilizing Envelope Filtering (EF) and Time-Delay Neural Network (TDNN) is proposed. Envelope Filtering normalizes raw data and eliminates imbalance training problem for neural networks, while TDNN is a hidden-state included method aiming at modeling PPG signal-noise interaction. An accelerometer as well as a gyroscope functions as the noise reference. Results comparing 8 performance indexes extracted from 4 types of waveform related features (heart rate, waveform correlation, pulse width, and pulse area) on an open access dataset we built show SniMA's effectiveness for PPG MA removal. The proposed algorithm provides a significant improvement in PPG waveform recovery and PPG feature estimation accuracy. This is the first time that a TDNN has been introduced for PPG de-noising. This study also indicates the potential of other hidden-state involved signal-noise interaction modeling methods for PPG de-noising.