An efficient lung sound classification technique based on MFCC and HDMR

Abstract

In this work, an efficient feature extraction scheme is developed for classifying the pulmonary diseases. The proposed method is hybrid which combines two important techniques that are Mel Frequency Cepstral Coefficients (MFCC) and High-Dimensional Model Representation (HDMR). MFCC is capable of imitating the human ear; therefore, it is capable of characterizing the lung sounds acquired by a stethoscope. On the other hand, HDMR performs decorrelation and denoising to the high-dimensional data. The MFCC entries establish a two-dimensional feature matrix, which is decomposed in terms of less dimensional entities by the application of HDMR. These entities are considered feature vectors that are then fed to the relevant machine learning classification algorithms and then the overall accuracies are calculated. According to the results, the proposed algorithm achieves 97.2% classification accuracy which is competitive with other existing state-of-the-art methods in the literature. HDMR also improves significantly the classification efficiency of the proposed technique. The results emphasize that HDMR can be employed as an efficient method in recognizing pulmonary disease tasks.