Construction of Effective HMMs for Classification between Normal and Abnormal Respiration

In many situations, abnormal sounds termed adventitious sounds are included as the lung sound of a subject suffering from a pulmonary disease. Thus, we aimed to detect abnormal sounds from auscultatory sound automatically. For this purpose, we expressed the acoustic features of normal lung sound for healthy subjects and abnormal lung sound for patients by using HMMs (Hidden Markov Models) and distinguished between normal and abnormal lung sounds. Furthermore, we detected abnormal sounds under a noisy environment including heart sounds by using a heart sound model. However, the duration time and the property for segments of respiratory, heart, and adventitious sounds varied. In our previous method, we constructed the HMMs with the same number of states and mixtures (topology) for all kinds of segments. Since we did not consider an appropriate topology, the classification rate between normal and abnormal respiration was low (88.96 %). In this paper, we proposed to construct the appropriate HMMs for each segment. By selecting a suitable topology for each segment, the classification rate was increased (91.35 %). The result showed the effectiveness of the proposed method considering the topology of HMMs.