Convolutional Neural Network for Respiratory Mechanics Estimation during Pressure Support Ventilation

Abstract

Abstract In mechanically ventilated patients, some lung injuries can be reduced or avoided with therapy individualization, while the lung function is evaluated continuously, breath by breath. However, obtaining information on respiratory mechanics (respiratory system resistance and compliance) in the presence of respiratory effort is challenging, even if using invasive and complex procedures. The contribution of this work is to predict both respiratory system resistance and compliance over time using a convolutional neural network (CNN) and estimate the respiratory effort profile using the respiratory dynamics. Therefore, the approach used in this work was to generate a large amount of simulated data to feed a CNN so it could learn how to predict the correct values of the respiratory system resistance and compliance. Then, the respiratory effort was estimated by solving a first-order linear model. The main results showed a normalized mean squared error of 5.7% for the respiratory system resistance and 11.56% for compliance from Bland-Altman plots derived from the computational simulator. Finally, the method was validated using real data from an active lung simulator within which respiratory mechanics varied, and some ventilator settings were adjusted to mimic actual patient situations. The active lung simulator effort profile was obtained with a normalized mean squared error of 8.31% considering the use of an active lung simulator. The results have shown that the simulated data were valuable for the CNN training, while the performance over the real data suggested that the network was generalized accordingly for estimating respiratory parameters and effort profile.