Feature Selection of Voluntary Cough Patterns for Detecting Lung Diseases

Cough is a classic symptom of respiratory disease. Airflow patterns produced during a cough represent a portion of the maximum expiratory flow-volume curve which has often been used to diagnose lung disorders. We have previously described a system for detecting lung disease that was based on both the airflow and the acoustic properties of a voluntary cough. The system used 26 representative features of the cough airflow measurements and 111 of the cough sound pressure wave. Redundancy within the feature set was eliminated using principle component analysis (PCA). A classifier was developed based on the projections of the principle components. The objective of this study was to determine the effect of eliminating irrelevant features of the cough prior to the PCA classifier to maintain, or even improve, overall system accuracy. Four types of feature selection methods were examined. They included forward sequential selection (SFS), backward sequential selection (SBS), sequential plusl-take away r (SLR), and genetic algorithm (GA) techniques. Three coughs from 112 individual with and without lung disease were classified using this system, and the results were compared with the diagnosis of pulmonary physicians. The overall classification accuracy was 94% when no attempt was made to optimize the feature set. This can be compared with the results of the genetic algorithm which used only 59 out of 137 features and increased the average classifier accuracy to 97.6%. The accuracy (number of features) using the above-mentioned algorithms was; 97.32% (35) for the SFS; 96.71% (111) for the SBS; 97.08 % (42) for the LRS; and 97.62% (59) for the GA. In conclusion, all feature selection methods improved the classification accuracy while simultaneously reducing the number of features.