Estimation of Lung Properties Using ANN-Based Inverse Modeling of Spirometric Data

Abstract

Spirometry is the most commonly used test of lung function because the forced expiratory flow-volume (FV) curve is effort-independent and simultaneously sensitive to pathological processes in the lungs. Despite this, a method for the estimation of respiratory system parameters, based on this association, has not been yet proposed. The aim of this work was to explore a feedforward neural network (FFNN) approximating the inverse mapping between the FV curve and respiratory parameters. To this end, the sensitivity analysis of the reduced model for forced expiration has been carried out, showing its local identifiability and the importance of particular parameters. This forward model was then applied to simulate spirometric data (8000 elements), used for training, validating, optimizing and testing the FFNN. The suboptimal FFNN structure had 52 input neurons (for spirometric data), two hidden nonlinear layers with 30 and 20 neurons respectively, and 10 output neurons (for parameter estimates). The total relative error of estimation of individual parameters was between 11 and 28%. Parameter estimates yielded by this inverse FFNN will be used as starting points for a more precise local estimation algorithm.