Over the past few years, there has been increased interest in photoplethysmography (PPG) technology, which has revealed that, in addition to heart rate and oxygen saturation, the pulse shape of the PPG signal contains much more valuable information. Lately, the wearable market has shifted towards a multi-wavelength and multichannel approach to increase signal robustness and facilitate the extraction of other intrinsic information from the signal. This transition presents several challenges related to complexity, accuracy, and reliability of algorithms. To address these challenges, anomaly detection stages can be employed to increase the accuracy and reliability of estimated parameters. Powerful algorithms, such as lightweight machine learning (ML) algorithms, can be used for anomaly detection in multi-wavelength PPG (MW-PPG). The main contributions of this paper are (a) proposing a set of features with high information gain for anomaly detection in MW-PPG signals in the classification context, (b) assessing the impact of window size and evaluating various lightweight ML models to achieve highly accurate anomaly detection, and (c) examining the effectiveness of MW-PPG signals in detecting artifacts.
Keywords: PPG; anomaly detection; artifact; machine learning; multi-wavelength PPG; neural networks; photoplethysmography; supervised learning; time series; unsupervised learning.