HMM-GMM model's size selection methodology for bioacoustics-based diagnostic classification

This paper presents a methodology for optimized utilization of merged Hidden Markov Models and Mixed Gaussian Model to classify lung sounds (LS) and heart sounds (HS) as a part of cardiopulmonary diagnostic assessment. Specifically this method was used as a criterion to determine most advantageous number of clusters for the HMMGMM model calculation. For this purpose, the LS and HS characteristics were evaluated in terms of MFCC (Melfrequency cepstral coefficients) and Quantile vectors. The analysis for the number of clusters was based on utilization of dendrograms, silhouettes, and the Bayesian Information Criterion (BIC). The merged HMM-GMM models for LS signals with Quartiles offered excellent results, while for HS signals, the best results were obtained with MFCC vectors. In both groups of LS and HS signals, a high degree classification efficiency was obtained reaching 100% for studied sets of signals. In particular, the results demonstrate that utilizing BIC or dendrograms as a part of optimized criterion enhances efficiency of merged HMM-GMM models in diagnostic classification of cardiopulmonary acoustic signals.