Heart sound segmentation using switching linear dynamical models

Localization of exact positions of the fundamental heart sounds (FHS) is an essential step towards automatic analysis of heart sound phonocardiogram (PCG) recordings, the automatic segmentation allows for data-driven classification of heart pathological events. Current approach using probabilistic models such as hidden Markov models (HMMs) has improved accuracy of heart sound segmentation. In this paper, we propose a switching linear dynamic system (SLDS) of piece-wise stationary autoregressive (AR) processes for segmenting the heart sounds into four fundamental components with distinct second order structure (auto-correlation). The SLDS is able to capture simultaneously both the continuous state-space in the hidden dynamics in PCG, and the regime switching in the dynamics using a discrete Markov chain. This overcomes limitation of HMMs which is based on a single-layer of discrete states. Compared to AR processes, the Gaussian mixture densities in HMM do not account for the temporal autorrelation structure in PCG which has one-to-one correspondence to frequency content a distinctive feature of HS components. We introduce three schemes for model estimation: (1) switching Kalman filter (SKF) model. (2) refinement by switching Kalman filter (SKS), and (3) fusion of SKF and the duration-dependent Viterbi algorithm (SKF-Viterbi). Results on a large PCG dateset of Physionet/Challenge 2016 shows SKF-Viterbi significantly outperforms SKF by improvement of segmentation accuracy from 71% to 84.2%.