Deep learning-based parameter estimation for pulmonary hysteresis modeling

Abstract

Respiratory hysteresis, the difference in the pressure-volume curve between inspiration and expiration in each respiratory cycle, is a phenomenon that becomes significant in certain diseases. In order to study such effect and analyze ventilatory control strategies, it is useful to model this aspect of respiratory physiology. This work explores the use of a general analytical hysteresis model, which has emerged in the literature to represent the relationship between applied forces and corresponding displacements in various structures and materials. A model expression was obtained that can be matched to the output of a multilayer perceptron neural network. Then, using data from real patients, the model parameters were adjusted using a deep learning-based parameter estimation method, with errors less than 8% as well as training hyperparameters to obtain better fits. This parameter estimation method represents a significant contribution to the modeling of respiratory hysteresis and could be applied to other fields.