Quartiles and Mel Frequency Cepstral Coefficients vectors in Hidden Markov-Gaussian Mixture Models classification of merged heart sounds and lung sounds signals

Abstract

This paper presents integrated Hidden Markov and Gaussian Mixture Models (HMM-GMM) to classify lung sounds (LS) and heart sounds (HS) characteristics. In order to optimize the models' size, several methodologies encompassing dendrograms, silhouettes and the Bayesian Information Criterion (BIC) were applied. The experiments were carried out extracting features from the LS and HS with MFCC (Mel-Frequency Cepstral Coefficients) vectors and Quantile vectors, specifically Quartiles. The merged HMM-GMM architecture for the signals using Quartiles, overall offered consistent classification results. In both types of vectors, a high degree of classification efficiency was obtained reaching up to 96% for the studied sets of signals. For MFCC the classification results were not conclusive. An assessment of the number of clusters using dendrograms, silhouettes, and BIC linked with the models' size. Consequently this allows to enhance efficiency of merged HMM-GMM models in diagnostic classification of cardiopulmonary acoustic signals.