Classification of Asthma, COPD and Healthy Lung Sounds Using Fourier Bessel Series Expansion in Machine Learning and Deep Learning Paradigm

Lung diseases are very common now a day and a low cost early detection and diagnoses of these diseases are very important to increase the life expectancy. The objectives of this study are: (i) to identify and investigate the effectiveness of Fourier Bessel Series Expansion (FBSE) in representing non stationary lung sound signals; (ii) to explore and compare the various machine learning and deep learning models viz. support vector machines (SVM), k nearest neighbors (kNN), ensemble bagged tree classifier, one-dimension convolutional neural network (1D-CNN) etc., in classifying the lung sounds into healthy and unhealthy signals. Fourier Bessel Series Expansion (FBSE) utilizes Bessel function which have inherent non stationary and damping characteristics thereby suitable for effective representation of lung sounds. The robustness and efficacy of FBSE was justified with the results obtained under two class i.e. normal vs abnormal and asthma vs COPD classification wherein ensemble subspace k-NN achieves the highest accuracy of 94.1%, 91.46% and 92.9% under normal-abnormal classification and ensemble bagged tree achieved 96.6%, 96.6% and 92.1% accuracy under asthma - COPD classification corresponding to 10-fold, 5-fold cross validation and 33% hold out respectively. The proposed 1D-CNN network deployed using FBSE expansion showed an accuracy of 96.4% and 81.1% for normal vs abnormal and asthma vs COPD classification respectively.