Cough classification of MEMS microphone signal using time-series and tabular machine learning algorithms

Abstract

Cough is a protective mechanism of the proximal respiratory tract. The frequency and severity of cough provide useful information for the diagnosis of upper respiratory diseases and the evaluation of their treatments. Manual cough classification is subjective, labor-intensive, and time-consuming. In this study, a cough classification technique based on a microelectromechanical system microphone is proposed. For the classification process, various tabular and time-series machine learning algorithms were applied, and their results were compared. With respect to the time-series algorithms, the random interval decision tree, random interval spectral forest, and random convolution kernel transform (ROCKET) methods were used. With respect to the tabular algorithms, a convolution neural network (CNN) with 40 Mel-frequency cepstral coefficients (MFCCs) and a recurrent neural network with 40 MFCCs were used. Voluntary cough and noncough (throat clearing, expiration, speaking, and rest) signals were recorded from 10 healthy subjects. The ROCKET method showed the best accuracy (98.40%). In addition, while its training took the longest time (1628.80 s), this algorithm took a reasonably short time for prediction (0.27 s). The CNN showed the second-best accuracy (97.81%) with short training (454.13 s) and prediction (0.40 s) times. Thus, given its accuracy and prediction time, the ROCKET method is recommended for this type of classification over the CNN algorithm. To validate the application of the proposed methods, two methods were applied to a public Coswara data set. To classify one cough class and three non-cough classes, the ROCKET and CNN showed reasonable accuracies of 90.33% and 89.16%, respectively.