Automated methods for cough assessments and applications in screening paediatric respiratory diseases

Abstract

Cough has the potential to be developed as a diagnostic tool; however, it has not yet been fully explored. Researchers have attempted to study cough in adults, but none of these studies involved paediatric populations younger than five years of age with respiratory diseases such as pneumonia. The study of automated wet/dry cough classifications in paediatric populations is not yet available either. At the moment, the identification of wet/dry cough is carried out manually by physicians. The results of this process are subjective and depend on the skills and experiences of the observers. Cough is one of the main symptoms of pneumonia, but the World Health Organization (WHO) only uses its existence for screening in pneumonia. An acoustic analysis of the pneumonia cough sounds for diagnosing this disease has not yet been explored. Further, quantitative study of cough analysis is still immature; physicians still have to identify and listen to the cough manually, which is a tedious and time-consuming task. An automated method capable of segmenting cough sound from recordings is urgently required.This thesis proposes the development of innovative cough sound analysis based methods to address the problem of wet/dry cough classification, substituting the bronchodilator test in resource-limited settings and segmenting cough from recordings automatically. In my approach, the cough samples were collected using non-contact sensors at a hospital in a developing country. All subjects included in this thesis are members of the paediatric population suffering from respiratory diseases such as pneumonia and asthma.The supports for my work are the results from preliminary studies and the pathophysiology of respiratory diseases. The infections stimulate the excessive production of mucus in the airways. In pneumonia, the mucus also fills the alveoli and causes lung consolidation. The opening and closing of the collapsed alveoli/airways produce crackle sounds. I hypothesize that the vibration of mucus, the inflammation of airways, the lung consolidation, and the crackle sounds alter the acoustic of pneumonia cough sounds such that they are distinguishable from the coughs of other diseases such as asthma.To capture the cough sound signatures, I extracted features such as non-Gaussianity score, Melfrequency cepstral coefficients, Shannon entropy, formant frequency, and zero crossing rates. These features were used for training classifiers to classify wet/dry coughs automatically, to differentiate pneumonia from asthma, and to segment cough from recordings. My results show that the proposed methods achieve high performance for the designed purposes.This thesis contributes to the development of a pioneering class of technology that addresses fundamental gaps in cough sound analysis. The non-contact technology for cough analysis is perfectly matched for children. In addition, it does not require elaborate sterilization process efforts. The automated wet/dry cough classification method facilitates objective cough assessment, and is useful for long-term wet/dry cough study. My cough based method for separating pneumonia and asthma can revolutionize the diagnosis of these diseases in limited-resource settings. The method can be developed into an affordable system for replacing the bronchodilator test. Further, my automated segmentation method has the potential to be developed as a cough counting device, as well as the front end of cough analysis systems.For future work, my methods should be tested in a larger dataset to develop a robust system. The methods can be developed for smart phone application or deployed in a low cost embedded system.