Automatic cough segmentation from non-contact sound recordings in pediatric wards

Cough is a common symptom of almost all childhood respiratory diseases. In a typical consultation session, physicians may seek for qualitative information (e.g., wetness) and quantitative information (e.g., cough frequency) either by listening to voluntary coughs or by interviewing the patients/carers. This information is useful in the differential diagnosis and in assessing the treatment outcome of the disease. The manual cough assessment is tedious, subjective, and not suitable for long-term recording. Researchers have attempted to develop automated systems for cough assessment but none of the existing systems have specifically targeted the pediatric population. In this paper we address these issues and develop a method to automatically identify cough segments from the pediatric sound recordings. Our method is based on extracting mathematical features such as non-Gaussianity, Shannon entropy, and cepstral coefficients to describe cough characteristics. These features were then used to train an artificial neural network to detect coughs segment in the sound recordings. Working on a prospective data set of 14 subjects (sound recording length 840 min), proposed method achieved sensitivity, specificity, and Cohen's Kappa of 93%, 98%, and 0.65, respectively. These results indicate that the proposed method has the potential to be developed as an automated pediatric cough counting device as well as the front-end of a cough analysis system.