Photoplethysmogram Signal Quality Assessment Using Support Vector Machine and Multi-Feature Fusion

Abstract

Background: Noise is unavoidable in the physiological signal measurement system. Poor quality signals can affect the results of analysis and disable the following clinical diagnosis. Thus, it is necessary to perform signal quality assessment before we interpreting the signal. Objective: In this work, we describe a method combing support vector machine (SVM) and multi-feature fusion for assessing the signal quality of pulsatile waveforms, concentrating on the photoplethysmogram (PPG). Methods: PPG signals from 53 healthy volunteers were recorded. Each had a 5 min length. Signal quality in each heart beat was manual annotated by clinical expert, and then the signal quality in 5 s episode was automatically calculated according to the results from each beat segments, resulting in a total of 13,294 5-s PPG segments. Then a SVM was trained to classify clean/noisy PPG recordings by inputting a set of twelve signal quality features. Further experiments were carried out to verify the proposed SVM based signal quality classifier method. Results: An average accuracy of 87.90%, a sensitivity of 88.10% and a specificity of 87.66% were found on the 10-fold cross validation. Conclusions: The signal quality of PPGs can be accurately classified by using the proposed method.