Feature Extraction of Lung Sounds Based on Bispectrum Analysis

Higher-Order Spectral techniques perform well in non-Gaussian signal processing. In this paper, we propose a novel method for lung sounds feature extraction based on AR model bispectrum estimation. By the bispectral cross correlation analysis, select AR model orders and apply them to estimate the parametric bispectrum of the lung sound signals. Then extract bispectrum features of lung sound signals (normal, pneumonia and asthma) and compare them in bi-frequency domain. To get more information of bispectrum, the method presented divides a cycle into inspiration phase and expiration phase. Peaks of bispectrum, normalized bispectral entropy and parameters of slice spectrum are selected to form the feature vector for lung sounds classification. The results show that bispectrum analysis of lung sounds is applicable and effective. And our work will provide assistant information for early diagnosis of lung-thorax disease.