Multichannel lung sound analysis to detect severity of lung disease in cystic fibrosis

Objective: Respiratory disease in Cystic fibrosis (CF) patients is one of the main causes of the reduction in pulmonary function and death. The primary goals of CF treatment include maintaining or improving pulmonary function and reducing the rate of pulmonary function decline. Therefore, the severity of lung disease should be monitored in CF patients. The objective of this study is to examine multichannel lung sound analysis in detecting the severity of lung disease in CF patients. Methods: 209 multichannel lung sound samples were recorded from thirty seven CF patients using a thirty channel acquisition system. Then, expiration to inspiration lung sound power ratio features in different frequency bands (E/I F) were extracted from large airway, upper airway and peripheral airway channels. These features were compared between the groups with different severity levels of the lung disease using Support Vector Machine, Artificial Neural Network, Decision tree and Naive Baysian classifiers by 'leave-one-sample-out' method. Results: It was shown that features of upper airways and peripheral airways were more effective in discriminating normal from mild (91.1%) and moderate from severe (92.8%) respiratory sound samples, respectively. The best result for discriminating between all groups of severity was related to neural network classifier which performs 89.05% average accuracy. Also, 'leave-one-subject-out' method confirmed the results. Conclusion: The proposed multichannel lung sound analysis method was successful in discriminating different severity levels of CF lung disease. Moreover, analysis of different lung region signals in consecutive levels of lung disease was consistent with regional damage of lung in CF.

Automated summary

Multichannel lung sound analysis was successful in discriminating different severity levels of CF lung disease. Features from upper airways and peripheral airways were more effective in distinguishing normal from mild (91.1%) and moderate from severe (92.8%) respiratory sound samples, respectively. The neural network classifier showed the best performance in discriminating among all severity groups with an average accuracy of 89.05%.