Determining optimal size of HMM-GMM models to classify bio-acoustic signals

Abstract

This paper demonstrates the analysis and proposed HMM-GMM models architecture to classify heart and lung sounds (HS and LS) signals to emphasize the model size optimization. Respiratory and cardiovascular diseases continue to represent one of the major worldwide healthcare problems associated with a high mortality rate, which can be reduced by an early and effective diagnosis; in this context, the use of digital tools utilizing signal pattern recognition allows efficient screening for abnormalities and their quantitative assessment. In particular, the HMM-GMM models demonstrated their efficiency in normal and traditionally noisy environments in light of very low intensities of these auscultation signals used as diagnostic indicators. Furthermore, applied MFCC and Quantiles feature extractors improve overall classification. While characterization with silhouettes, dendrograms and algorithms such as BIC was inconclusive when GMM was applied, however, they were used as a starting point in the determination of a size of the model as it allowed a reduction in the number of iterations considering different model size. In addition, one can note that application of MFCC or Quantiles allowed differentiating the characteristics of normal HS and LS from those associated with pathological conditions. Furthermore, it was observed that a large amount of data leads to more robust and adapted models, but does not limit the calculation demand. Overall, this approach may enhance efficiency and precision of the diagnostic screening for abnormal auscultation indicators.