Investigating the potential of MFCC features in classifying respiratory diseases

In the literature so far, classification of respiratory diseases with cough signals has typically involved extracting standard spectral features such as Mel Frequency Cepstral Coefficients (MFCC), and other descriptive features such as Zero-Cross-Rates (ZCR), Entropy, Centroid, etc., from the cough signals, before developing classification models. However, with current trends in audio signal classification gearing towards deep learning, which typically make use of only the spectral features, investigating the potential of MFCCs alone in classifying respiratory diseases becomes quite imperative. MFCCs alone, are in fact theoretically quite powerful in providing all vital information about any audio signal, and therefore using them as the standalone set of features in classifying the respiratory diseases is worth investigating. Furthermore, the classification of respiratory diseases so far has only been limited to no more than two diseases. Hence, in order to make a break in this area, this paper investigates the potential of MFCC features alone in classifying respiratory diseases. This is done through the development of a new classification model that features deep learning model design. This method of investigation is similar to typical feature importance studies that fit models before identifying the contributing features. In this case, however, the features are already filtered, and so the model is optimized only by design to perform the study. Furthermore, in order to substantiate the results of the investigation, the model is made to classify more than just two respiratory diseases. For this we have selected five common respiratory diseases namely Asthma, COPD, ILD, Bronchitis and Pneumonia for the classification. Results show that the MFCC features alone do have the potential of classifying the respiratory diseases. This has been substantiated by achieving training accuracies on the model to fall between 85.86 to 97.83% and test accuracies between 87.02 to 88.50%.