Recognition of Retained Secretions in Central-Airway for Adult Patients Receiving Mechanical Ventilation

In critical care settings, mechanical ventilation is a common practice employed to support patients. Retained secretions commonly affect the effectiveness of mechanical ventilation, and there is currently no autonomous recognition method available. This study proposes an autonomous recognition method based on deep learning to analyze respiratory sounds for effective recognition of retained secretions. Initially, binary classification was performed to detect the presence or absence of secretions, achieving a notable accuracy of 91.5% and a precision of 89.5%. Based on these promising results, the study progressed to a more detailed ternary classification to provide further analysis of suction requirements, categorizing requirements as 'no suction required', 'suction monitoring required', and 'suction required'. Experiments conducted on 1,512 seconds of respiratory sound data demonstrated the efficacy of the binary classification, but also highlighted the challenges in the ternary classification. This study demonstrates the potential of deep learning techniques in augmenting critical care, promising significant improvements in patient management and care outcomes.