Automated detection of air trapping from mechanical ventilation waveform through interpretable dual-channel 1D convolutional neural network

Abstract Objective: Air trapping is a major symptom of respiratory diseases like chronic obstructive pulmonary disease (COPD) and asthma, and has always been a significant problem in treating patients using mechanical ventilation. If not handled timely, it can pose risk of severe respiratory dysfunction and potential life-threatening complications. Currently, the assessment of air trapping for ventilated patients largely relies on clinical experience of medical staffs. 
Approach: We introduced an interpretable dual-channel one-dimensional convolutional neural network (DC-1DCNN) with a simple structure, which enables fast inference. This model is designed to classify whether a respiratory waveform is indicative of air trapping. A global average pooling (GAP) layer was integrated into the DC-1DCNN model to highlight the segments of the respiratory waveform that the model focused on when making a classification. An air trapping index (ATI) was introduced to quantify the condition of air trapping in the ventilated patients and to evaluate the effectiveness of bronchodilator nebulized treatments. 
Main results: The results demonstrate a satisfactory accuracy of 96.2% in identifying air trapping breath cycles, with highlighted critical sections in breath cycles that match the understanding of clinical experts for air trapping. The efficacy of bronchodilators can be well assessed by the ATI. 
Significance: The findings suggest that the proposed DC-1DCNN can help detect air trapping in real-time, and help the clinicians better monitor the airway condition of the ventilated patients.