Real-time airway monitoring system using binary classification model based on respiratory sounds of rabbits with a tracheostomy tube

Tracheostomy is a medical procedure used to ensure airway integrity. As patients with tracheostomies produce excess secretions obstructing the airway, proper airway management is required. Medical staff primarily assess airway status through respiratory sounds, but this assessment heavily depends on their experience and expertise. Therefore, a continuous and standardized airway assessment system is needed, and it would be even more beneficial if it could operate in real time. Due to challenges in obtaining controlled respiratory sound data from humans, respiratory sounds from rabbits with tracheostomy tubes were utilized. Airway obstruction was induced using artificial sputum. Collected respiratory sound samples were converted into spectrograms and analyzed via deep learning. A total of 1,443 respiratory cycles, representing 402 samples of 4-second respiratory sound segments, were recorded from 29 New Zealand rabbits. The trained convolutional neural network (CNN) binary classification model, evaluated on the validation dataset, achieved an accuracy of 0.9375 and an area under the receiver operating characteristic (ROC) curve of 0.9900 in classifying normal and obstructive respiratory sound samples. Furthermore, in testing experiments simulating a medical scenario, the developed Internet-of-Things-based device enabled real-time remote data transmission. As a result, 42 respiratory sound samples from two rabbits, collected using the developed device, were used as the testing dataset for the CNN classification model, which achieved an accuracy of 0.9524 and an area under the ROC curve of 0.9953. This is the first study using deep learning to assess the airway condition of rabbits with tracheostomy tubes, suggesting potential applications in human airway monitoring.