Multimodal acoustic models HMM applied to heart and lung sounds Abstract

Abstract

This paper demonstrates classification processes of normal and abnormal bioacoustics signals recorded over a human thorax which encompasses heart and lung sounds. The specific aim is to design a signal classification system based on acoustical modeling techniques employing particularly HMM models to detect events' sequences, and GMM to model clusters corresponding to the data events. The modalities for extracting data characteristic are the MFCC and Octile vectors. These approaches have a potential of enhancing the classification accuracy of these auscultatory diagnostic indicators as the initial studies demonstrated that the HMM based models are less sensitive to the noise. Preliminary results demonstrate over 95\% accuracy in classification of the evaluated sound signals. This is particularly critical taking into account environmental interference in a variety of medical care settings. As the heart sounds frequency components parallel those of the lungs sounds, but with a different periodicity, they can be modeled with the same recording. The preliminary experimental results are supportive of this approach and demonstrate the feasibility of a development of an automated early diagnostic assessment of patients' auscultatory diagnostic indicators utilizing low-cost technologies.