A Neural Network Based Approach for the Evaluation of Spirometric Reference Values

Abstract

It is extremely complex to estimate the reference values of lung parameters obtained by spirometry due to the several potential sources of variability ranging from individual characteristics to regional, ethnical, environmental and technical aspects. Earlier works in estimating reference values provide predictive equations developed using multiple regression analysis on spirometric data of people from a particular region. In this paper, an artificial neural network technique has been applied to obtain the reference values of important lung parameters. This study is confined to South Indian population aged between 15 and 30. In comparison with the regression analysis, neural network model is robust and adaptive, and it does not require knowledge of the underlying relationship between input and output variables. A two layer feedforward neural network with 30 neurons in the hidden layer was trained with age, gender, height, weight, and body mass index of 39 normal subjects as inputs and their lung parameters as outputs. Backpropagation algorithm incorporating Levenburg-Marquardt optimisation technique was used for training the neural network. The trained neural network was verified using 5 sets of test data and observed that this technique provides better proximity to actual values than the conventional predictive equations.