Anomaly Detection of Motion Artifact in Photoplethysmography (PPG) Sensors using Unsupervised Learning

Photoplethysmography (PPG) sensors, which are optical sensors collecting human pulse signals, are known to be susceptible to the user's motion. Thus, for wide usage of PPG sensors in various applications, detecting motion artifacts (MAs) is urgently needed. This study proposes an anomaly detection algorithm for MAs using a long short-term memory (LSTM) autoencoder, loss functions customized to PPG sensor types, and a double-check algorithm supporting the loss function. The tested PPG sensors were first classified into two types. In sensors Type A, the fluctuation amplitude of the normal signal is larger than that of the MA signal. In sensors Type B, the fluctuation amplitude of the normal signal is smaller than that of the MA signal. These observations highlight the need for customized features tailored to each PPG type. Subsequently, the collected PPG raw data were processed using a moving average filter, and MA occurrences are identified by comparing the filtered data in the same sliding window with the customized loss function. Experimental results demonstrate that our algorithm successfully detects MAs regardless of the PPG sensor type. However, in very specific situations, the used loss function may not correctly detect the MA signals. To address this, a supporting double-check algorithm was additively proposed. As a result, this unsupervised learning-based algorithm comprehensively solves the commonly observed MA problems by using only normal signals as training datasets. The proposed algorithm requires low time complexity.