Personalized detection of explosive cough events in patients with pulmonary disease

Abstract

We present a new method for the discrimination of explosive cough events based on a combination of spectral and pitch-related features. The method was tested on 16 distinct partitions of a database with 9 patients. After a pre-processing stage where non-relevant segments were discarded, we have extracted eight features from each of the other segments and have fed them to the classifiers. Four types of algorithms were implemented to classify the events, with Bayesian classifiers achieving the best performance. Preliminary results showed that performance increased when the analysis was performed on individual subjects and when specific sensor locations were chosen. These results demonstrate that personalizing the analysis is a promising approach and shed some light on where to put sensors when automatic analysis is performed in the future.