A Two-Level Classifier for Prediction of Healthy and Unhealthy Lung Sounds Using Machine Learning and Convolutional Neural Network

The ability to detect lung disorders with the aid of lung sound auscultation has become very common these days. The early diagnose of these diseases can cause a reduced risk to loss of life. Auscultation has a disadvantage that it requires a skilled practitioner to listen and remember the lung sound from different parts of the posterior chest and predict the kind of lung disease. In this paper, a comparative study of one-dimension convolutional neural network (1D CNN), and machine learning methods is demonstrated on the lung sound database obtained from Kaggle. SMOTE (Synthetic Minority Over-Sampling Technique) is applied to treat the imbalance in the unhealthy recordings. The sound features like Mel spectrogram, MFCC (Mel Frequency Cepstral Coefficient), MFCC Delta, MFCC Delta-Delta, pitch, spectral centroid, spectral entropy, spectral flux, spectral kurtosis, spectral skewness and spectral spread was obtained and tested on different machine learning models using 10-fold, 5-fold cross validation and 33% hold out. The results obtained under normal and abnormal classification for 10-fold, 5-fold and 33% hold out showed an accuracy of 78.6%, 77.5% and 76.4% for MFCC Delta-Delta features respectively for fine KNN machine learning model. Under Asthma and COPD binary classification, proposed study showed an accuracy of 91.6%, 90.6%, 89.1% respectively for 10-fold, 5-fold and 33% hold out for subspace KNN and fine KNN models. The 1D CNN neural network implements the Normal-Abnormal and Asthma-COPD two level binary classification showed an accuracy of 94.67% and 75.92%.