Authors:
Kaat Vandecasteele
1
;
2
;
Jesús Lázaro
2
;
1
;
Evy Cleeren
3
;
Kasper Claes
4
;
Wim Van Paesschen
3
;
Sabine Van Huffel
1
;
2
and
Borbála Hunyadi
2
;
1
Affiliations:
1
KU Leuven, Belgium
;
2
imec, Belgium
;
3
KU Leuven, University Hospital, Belgium
;
4
UCB, Belgium
Keyword(s):
Artifact Detection, Photoplethysmography, Wrist, Feature Selection.
Abstract:
There is a growing interest in monitoring of vital signs through wearable devices, such as heart rate (HR). A
comfortable and non-invasive technique to measure the HR is pulse photoplethysmography (PPG) with the
use of a smartwatch. This watch records also triaxial accelerometry (ACM). However, it is well known that
motion and noise artifacts (MNA) are present. A MNA detection method, which classifies into a clean or MNA
segment, is trained and tested on a dataset of 17 patients, each with a recording duration of 24 hours. PPG-and
ACM-derived features are extracted and classified with a LS-SVM classifier. A sensitivity and specificity
of respectively 85.50 % and 92.36 % are obtained. For this dataset, the ACM features do not improve the
performance, suggesting that ACM recording could be avoided from the point of view for detecting MNA in
PPG signals during daily life.