Clustering Seismocardiographic Events using Unsupervised Machine Learning

Abstract

Seismocardiographic (SCG) signal morphology is known to be affected by cardio-pulmonary interactions, which introduce variability in the SCG signal. Hence, grouping of SCG signals according to their respiratory phase can reduce their morphological dissimilarity. In addition, correlating SCG with pulmonary phases may provide more insights into the nature of cardio-pulmonary interactions. This study uses unsupervised machine learning to cluster SCG events based on their morphology. Here, K-means clustering was employed using the time domain amplitude as the feature vector. The method is applied on measured SCG data from 5 male subjects (Age: 30 ± 5.8 years). The mean Silhouette values for different number of clusters suggested that optimal clustering was reached when SCG waveforms were divided into two groups. Using respiratory flow information, SCG waves were labeled as inspiratory vs. expiratory or high vs. low lung volume. The SCG clusters were then compared with these labels and purity values were calculated. The distributions of clustered SCG events in relation to respiratory flowrate and lung volume phases showed consistent trends in all subjects. Results suggested that grouping SCG based on lung volume phases would yield more homogeneous groups and, hence, would keep SCG variability (within each group) to a minimum. The demonstrated utility of the proposed machine learning approach in identifying respiratory phases from SCG waveforms may obviate the need for simultaneous respiratory measurements.