Clustering of SCG Events Using Unsupervised Machine Learning

Abstract

Seismocardiography (SCG) is the measurement of chest surface vibrations induced by cardiac activity. SCG beats are typically averaged to reduce noise and determine average SCG waveforms and features. Variability in SCG morphology impedes precise determination of average waveforms. Hence, it is desirable to group SCG beats into clusters with minimal intra-cluster heterogeneity. Cardio-pulmonary interactions are known to contribute to SCG variability. Therefore, grouping SCG signals by their respiratory phase may be helpful. SCG signals and respiratory flowrate were simultaneously measured in seventeen subjects (Age: 23 ± 3.5 years, 7 female). Unsupervised machine learning was implemented to cluster SCG beats according to their morphology. The time domain amplitudes of the SCG beats were used as the feature vector. K-medoids clustering was employed with dynamic time warping (DTW) distance as the heterogeneity measure. The quality of the clustering was measured using mean silhouette values and the elbow method for varying clusters numbers. Optimal clustering was achieved when SCG beats were split into two groups. Using respiratory flow information, SCG beats were labeled as inspiratory vs. expiratory, and as high vs. low lung volumes. The SCG groups determined by machine learning were compared with these labels. Grouping SCG based on lung volume phases yielded more homogeneous clusters than grouping by inspiration vs. expiration (p < 0.01). Unsupervised clustering reduced the intra-cluster variability by an average of 15% across subjects. Grouping by lung volume and inspiration vs. expiration reduced variability by 6% and 3%, respectively. The variability reduction may help more precise determination of average SCG waveforms and features, thereby improving SCG diagnostic utility.