Deep recurrent neural network-based autoencoder for photoplethysmogram artifacts filtering

Recently, the need for fast, cost-effective, convenient, and non-invasive cardiovascular analysis techniques has been the primary and most attractive reason to use photoplethysmogram (PPG). Most wearable devices on the market today can collect PPG data and enable the measurement of important features such as heart rate, respiration rate, and blood pressure, in addition to detecting irregular pulses and cardiovascular diseases. One major drawback of PPG data is their high sensitivity to motion, resulting in distorted and meaningless signals. This paper proposes a neural network-based filtering method to remove corrupted windows from the collected PPG data in an unsupervised manner. It also proposes a PPG data summarization and augmentation strategy which optimizes the network performances. Experimental results show that the proposed approach was capable of achieving 90% precision and 95% recall when processing PPG data collected from a Shimmer3 GSR+ sensor.