Blind Separation of Heart Sounds

Abstract

This paper presents a theoretical foundation for the newly developed methodology that enables the prediction of blood pressures based on the heart sounds measured directly on the chest of a patient. The key to this methodology is the separation of heart sounds into first heart sound and second heart sound, from which components attributable to four heart valves, i.e.: mitral; tricuspid; aortic; and pulmonary valve-closure sounds are separated. Since human physiology and anatomy can vary among people and are unknown a priori, such separation is called blind source separation. Moreover, the sources locations, their surroundings and boundary conditions are unspecified. Consequently, it is not possible to obtain an exact separation of signals. To circumvent this difficulty, we extend the point source separation method in this paper to an inhomogeneous fluid medium, and further combine it with iteration schemes to search for approximate source locations and signal propagation speed. Once these are accomplished, the signals emitted from individual sources are separated by deconvoluting mixed signals with respect to the identified sources. Both numerical simulation example and experiment have demonstrated that this approach can provide satisfactory source separation results.