Support vector machine hidden semi-markov model-based heart sound segmentation

The segmentation of the primary heart sounds within a phonocardiogram (PCG) is an essential step in the classification of pathological cardiac events. Recently, probabilistic models, such as hidden Markov models, have been shown to surpass the segmentation capabilities of previous methods. These models are further improved when a priori information about state duration is incorporated into the model, such as in a hidden semiMarkov model (HSMM). This paper addresses the problem of the accurate segmentation of heart sounds within noisy, real-world PCGs using a HSMM, extended with the use of support vector machines (SVMs) for emission probability estimation. A database of 123 patients with over 20,000 labelled heart sounds were used to train and test the algorithm. Best reported alternatives in the literature were also implemented and tested on the same data. On out-of-sample test data, our method outperforms previously reported methods with sensitivities of 94.9% and 91.0% and positive predictivities of 95.2% and 90.9% for first and second heart sounds respectively.