Motion Artifact Detection and Classification for Unobtrusive Cardiorespiratory Signals Using Machine Learning

Onno Linschmann, Carl Revander, Steffen Leonhardt, Markus Lueken
RWTH Aachen


For personal health care applications, more and more unobtrusive sensors, such as reflective photoplethysmography (rPPG), capacitive electrocardiography (cECG) or ballistocardiography (BCG), are used. While these sensors provide more comfort for the user, they exhibit a lower signal-to-noise ratio and especially suffer from motion artifacts. Therefore, methods for reliable detection and in case of e.g. sleep analysis also classification, are researched.

In this paper, Support Vector machines (SVM) are investigateds for detection, i.e. binary classification, of motion artifacts. Furthermore, two methods for also classifying these artifacts with regard to eight classes of movements (torso, head, arms, standing up) are presented. First, a direct multi-class classification, and second, a multi-class classification after perfect detection. Eight waveform-related features which are calculated for overlapping windows were selected and used for classification. For the evaluation, the UnoVis dataset which provides nine recordings of six channels of unobtrusive sensors (rPPG, cECG, BCG (2), high-frequency impedance sensors (2)) with annotated motion is used. Different combinations of channels and features are tested. For that, the dataset is divided into a test set and a validation set. Parameter tuning is performed using a leave-one-out cross-validation. After parameter tuning a validation is performed on the validation set.

For the binary classification, an accuracy, sensitivity and specificity of 91%, 71%, 97% (test set) and 92%, 73% and 98% (validation set) are achieved respectively. Best performance is achieved for using a reduced feature set and only the BCG channels. The window size does not influence the results significantly. For the direct multi-class classification, the SVM's performance is rather poor with mean accuracies, sensitivities and specificities of 77%, 21% and 93% (test set) and 78%, 28% and 93% (validation set) respectively. Similar results were achieved for perfect prediction meaning that while motion artifacts are detected well, their different sources are often confused.