Classification between normal and abnormal respiratory sounds based on stochastic approach

In this paper, we propose a novel classification procedure for distinguishing between normal and abnormal respiratory sounds on the basis of stochastic approach. The main characteristic of our procedure is that two stochastic models are used to detect abnormal respiratory sounds precisely: (1) hidden Markov models (HMMs) for acoustic spectral features and (2) bigram models for the occurrence of acoustic segments in each inspiratory/expiratory period. The classification procedure comprises a training process and a test process. In the training process, acoustic models for normal and abnormal respiratory sounds are trained using a transcribed database. In the test process, the classification procedure detects the segment sequence with the highest total likelihood and yields the classification results. Our procedure achieved a classification rate of 84.2% between normal and abnormal respiratory sounds. Experimental results revealed that for the classification, use of the segment bigram led to a 4.8% reduction of error rate in comparison with the classification rate of a conventional method that uses deterministic rules to describe segment sequences instead of the segment bigram.