Source adaptivity for cardiac sound separation

Physiological sounds exhibit a variety of waveforms and hence, their probability density functions (pdf) show a wide variation in their shapes ranging from subGaussian to superGaussian to multimodal. In blind source separation (BSS) applications, where extraction of cardiac sounds from their mixtures is the main objective, assuming a particular density profile for the unknown source signals is almost impossible. In this paper we show that separation of cardiac sounds, from their mixtures, could be improved if online density estimation algorithms are used to estimate the unknown source density near the global optimum of BSS algorithms. Experimental results are included.