Comparison of Hierarchical and Partitional Clustering in Multi-Source Phonocardiography

Phonocardiography (PCG) has proved a valuable tool over the years to monitor the status of at-risk patients for some cardiovascular diseases. Its multi-source version, consisting of the simultaneous recording of multiple acoustic signals from different points of the patient’s chest, is currently under research as a solution to develop wearable devices based on PCG and bring PCG to the patient’s domicile. When a high number of PCG signals are available, to define the most suitable auscultation area, depending on the clinical question, clustering comes into the picture. In this work, we applied agglomerative hierarchical clustering and k-means to multi-source PCG recordings. A similarity metrics based on the correlation of the signals was used to compare the signals based on their morphological characteristics. The two clustering methods resulted in a Rand Index averagely higher than 0.84, showing a high level of agreement and validating the usage of clustering for the application of interest. Hierarchical clustering allowed for obtaining a better trade-off between the intra-cluster variability and the inter-cluster distance. Adding to its deterministic nature, it should be considered as preferrable with respect to k-means. This work moves one step further to the development a reliable wearable device based on digital auscultation for the monitoring of the patient at its domicile.