Estimation of Lung Properties From the Forced Expiration Data

Forced expiration is the most commonly applied lung function tests. Despite the problem of spirometry modeling was solved a few decades ago, a relatively small amount of work has been devoted to indirect measurements of lung properties from spirometry data. Just recently, a new method, based on the reduced model for forced expiration and two-stage estimation (global with the feed-forward neural network approximating the inverse mapping (InvNN) and then local with the Levenberg-Marquardt algorithm, starting with the rough estimates yielded by the InvNN) was proposed. The aim of this work was to evaluate the accuracy of the above approach to the indirect measurement of lung properties. To this end, 16,000 synthetic spirometry results were generated, and then used to optimize, train, and validate the InvNN, and to test the entire method. The total estimation errors of model parameters were from 3.7% to 16.6% in relation to their variability ranges. Those original estimates were then recalculated to clinically interpretable airway resistances and compliances, assessed with the relative errors of 7%-35% and 5%-12%, respectively. These outcomes encourage the future use of the method to analyze the results of bronchial challenge or dilation tests.