Multiresolution for the detection of xiphoidal sounds in noisy medical audio signals

We present in this paper a method we have developed in order to detect particular events (xiphoidal sounds) in swallowing audio signals. The difficulty of the problem comes from the noisy nature of our medical audio signals, due to biological and environmental reasons. Besides, xiphoidal sounds have a highly non-stationary behaviour that makes conventional analysis methods such as Fourier Transform or parametric model based methods non suitable. Thus, the detection method we have developed is based on a wavelet multiresolution analysis of the signal which is a powerful concept particularly adapted to signals dominated by transient behaviour or discontinuities. The decomposition levels of signals that comprise significant information are automatically selected and a multicriteria decision function is used to detect xiphoidal sounds among selected peaks.