An automated and unobtrusive system for cough detection

Abstract

Cough monitoring is useful for people suffering from chronic obstructive pulmonary disease (COPD) since cough is associated with an increased risk of frequent exacerbations and hospitalizations. Differently from what exists in the literature, this paper presents an automated cough detector that can be used for long term and remote monitoring. A dataset of sound traces collected in 7 COPD patients' home was used to test the performance of different machine learning approaches. Audio sounds were recorded for 90 days and they contain cough events or environmental noises with the last being the larger proportion. This suggests us to consider supervised methods that can deal with class imbalance learning. The data allow also to investigate the possibility to distinguish between patient coughs and coughs coming from other people in the same environment. The results, presented using a stratified leave-one-subject out cross validation, are promising since the area under the Receiver Operating Characteristic (ROC) curve gets as high as 0.916 ± 0.027.