Hidden Markov model in spectro-temporal tracking of asthmatic wheezing in respiratory sounds

Abstract

Interval of respiratory cycle occupied by asthmatic wheezing can be an indicator of severity of asthmatic attack. This information can be estimated from respiratory sounds spectrogram, by performing spectro-temporal tracking of duration of frequency lines of asthmatic wheezing. In this paper we model wheezing using hidden Markov model, considering its instantaneous frequency as hidden state. We estimate the hidden state using Forward-backward and Viterbi algorithm from a series of observations drawn from STFT. We present a simplified model focusing on tracking of a single frequency line (monophonic wheezing). In comparison to a referent wheeze tracking algorithm, average results show 10%increase in tracking accuracy, and a significant gain in robustness (same tracking error at 10 dB lower SNR). Execution speed is analysed in order to evaluate suitability of the method for m-health application of asthmatic patients monitoring. Real-time operation was verified for Forward-backward algorithm on an Android smartphone.