Classification of Pulmonary Diseases Using Cough Sounds Using Deep Learning

Abstract

Cough sound analysis has attracted interest as a potential low-cost diagnostic tool for low- resource settings, where the burden of pulmonary disease is quite high. However, published results on cough sound analysis are generally limited to specific pulmonary diseases (e.g. detection of Whooping cough - Pertussis) and the study sizes are small. In this paper, we present a general framework for cough sound analysis, which includes automatic cough segmentation, feature extraction and a general classification design that can be applied to a wide range of pulmonary diseases. For our analysis, three evidence-based features were selected (variance, kurtosis, and zero crossing Irregularity) as well as an additional feature that we developed (rate of decay). Our cough sound analysis framework was tested using voluntary cough data collected from 54 patients presenting a combination of pulmonary conditions (COPD, asthma, and allergic rhinitis) equally sampled from all patients arriving at a pulmonary clinic, as well as 33 healthy individuals. All study subjects were examined with a stethoscope auscultation, clinical questionnaire, and peak flow meter, and were given a full pulmonary function test (spirometer, body plethysmograph), which was the gold standard used to determine each patient's diagnosis.When the classifiers were trained using cough sounds alone, the accuracy (as determined by the AUC of the ROC curve) was 74% for Healthy vs Unhealthy, 80% for Obstructive vs non-Obstructive, and 81% for Asthma vs COPD. We also compared the performance of our cough sound analysis against other low- cost diagnostic tools and observed that cough sounds surprisingly had better performance than lungs oundauscultation alone, but had significantly lower performance compared to our clinical questionnaire or peak flow meter test. From these data, we conclude that cough sounds have value as a rapid and simple screening tool, but are of less diagnostic value compared to a clinical questionnaire or peak flow meter.