Comparison Of Two Different Methods For Clustering SCG Signals

• Published in: Journal of cardiac failure Vol. 29; no. 4; p. 597
• Main Authors: Sandler, Richard H., Ahdy, Sherif, Azad, Md Khushidul, Raval, Nirav, Mentz, Robert, Mansy, Hansen
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.04.2023
• ISSN: 1071-9164; 1532-8414
• DOI: 10.1016/j.cardfail.2022.10.127

Seismocardiography (SCG) refers to the measured vibrations at the chest wall surface that are induced by cardiac activity. Grouping similar SCG events into “clusters” may help reduce SCG signal variability and thereby increase utility for monitoring and diagnosis of cardiac diseases including HF. Clustering involves optimizing “distance” both within a “cluster” grouping and between cluster groups. This is done by using various signal processing techniques such as “cross correlation” and “dynamic time warping” (DTW). The effect of the different distance measures on SCG clustering has not been fully investigated. Investigate the effect of two distance measures on clustering SCG events and compare waveform variability before and after clustering to determine which distance measure would be more computationally efficient plus which would provide optimal clustering separation. 15 healthy males (19-31 y) were studied w/ simultaneous measurement of SCG at 36 chest surface locations, along with electrocardiography (ECG) and spirometry. The lung volume changes were determined by integrating the respiratory flow rate. SCG events were segmented using the ECG-R wave, and clustered into two groups with similar waveform morphologies using the “k-medoid” algorithm based on prior research. Two distance measures (cross correlation and DTW) were used in clustering. The computational time, intra-cluster variabilities before and after clustering and post-clustering inter-cluster variability were calculated for each distance measure. Figures 1(a) and (b) show the percent drop in the intra-cluster variability with clustering and the inter-cluster variability after clustering for each subject and distance measure. The values shown are averaged over all SCG measurement locations. The two distance measures tended to have comparable percent drop in intra-cluster variability with no clear trend. However, better inter-cluster separation was achieved with the cross-correlation method in all subjects. In addition, the computational time for the cross-correlation approach was about one third of DTW. The cross-correlation distance measure method performed better than DTW since it resulted in more cluster separation and required less computational time. Therefore, the cross-correlation method may be preferable for future studies, especially when dealing with large datasets.