Autonomous Recognition of Retained Secretions in Central‐Airway Based on Deep Learning for Adult Patients Receiving Invasive Mechanical Ventilation

Prolonged invasive mechanical ventilation can lead to secretion retention in airways, which increases the risk of airway obstruction and lung infections. Clearing secretions in time is essential to guarantee ventilation effectiveness and patients’ safety. However, manual and scheduled auscultations, widely adopted in clinical practices, can hardly continuously recognize the extent of secretion retentions to determine an optimal timing of secretion suction, which also place a burden on clinicians. Herein, an autonomous recognition model of secretion retention is proposed based on a ResNet‐34 convolutional neural network to perform triple classifications, which represent three levels of secretion retention, namely no secretions, secretions appear, and secretions retained. Respiratory sounds in central airways are continuously recorded and fed into the model in the form of Mel spectrograms. Experimental results demonstrate that the model can continuously monitor the breathing sounds and autonomously recognize different extent of secretion retentions in real time with overall accuracy of 89.08%. An edge computing module to integrate the recognition model into ventilators with limited computing power is also developed and approved, which indicates its potential clinical applications. The proposed autonomous recognition model facilitates clinicians to optimize secretion‐suction schedules to guarantee the ventilation effectiveness and patients’ safety.