Discrimination among different types of heart sounds has a significant impact in designing pHealth systems based upon this bio-signal, since (i) it enables the optimal selection and tuning of the analysis algorithms and (ii) it may be applied as a first level strategy for heart dysfunction diagnosis. In this paper we introduce an algorithm for heart sound type discrimination into three classes: healthy heart sounds, heart sounds with murmur produced by native heart valves and heart sounds produced by prosthetic mechanical heart valves. The algorithm is based on a nonlinear dynamical model of phase space reconstruction for various frequency bands. For each frequency sub-band the chaotic nature and the complexity of the signal is assessed using the largest Lyapunov exponents (LLE) and the correlation dimension (CD). The effectiveness of the method has been tested with heart sounds of 45 subjects (15 subjects of each class). It was concluded that LLEs and the CDs exhibit complementary significance in the discrimination among different classes of heart sounds.