MRMR-based feature selection for automatic asthma wheezes recognition

In this paper application of the mRMR (minimum Redundancy Maximum Relevance) algorithm to reduction of the number of lung sounds features used for asthma wheezes recognition is proposed. The paper presents the reduction of following features: Tonal Index (TI), Kurtosis (K), Energy Ratio (ER), correlation feature (CF1), Difference to Mean ratio (D2M), Eigen Value Decomposition feature (EVD), Linear Prediction feature (LP),Spectral Flatness (SF), Spectral Peaks Entropy (SPE), and two features that has not been presented yet in wheezes detection: Audio Spectral Envelope (ASE) taken from ISO/IEC MPEG-7 standard and Vector Comparison (VC). As a classifier the SVM algorithm was used.