Classification of pulmonary diseases using artificial neural networks
Abstract
Pulmonary diseases can be divided into central and peripheral conditions, depending on the area affected. Making this distinction by looking at patterns composed by measurements of respiratory system impedance is a difficult task for the untrained eye. But using Artificial Neural Networks (ANN) can help to classify and characterize the features of major respiratory diseases such as asthma by analyzing physical parameters of the lung's mechanics. Impulse Oscillometry (IOS) provides values of resistance and reactance of airways as a function of frequency, from the Pressure/Flow ratio. IOS has brought a sense of fresh air to clinicians because it's a highly reproducible, non-invasive and sensitive test. This study utilizes IOS measurements to create an input set for training a back-propagation ANN seeking to classify pulmonary diseases. NN1 used a data set containing 131 patterns belonging to three different classes ( central, peripheral, and other) and produced 98.47% classification accuracy during validation and 61.54% during generalization. NN2 used a data set containing 361 patterns belonging to two different classes (asthmatic and non-asthmatic) and produced a satisfactory 95.01% classification accuracy during validation and a remarkable 94.44% during generalization. Further investigation of this last result produced an impressive generalization result of 98.61% classification accuracy. NN1 yielded promising classification rates and its accuracy can be improved by including more training patterns combined with fuzzy logic decision rules. Although NN2 proved to be quite accurate, these ANN classifiers are envisioned as an aid to the knowledgeable physician, not a stand-alone entity.
Subject Area
Electrical engineering|Biomedical engineering
Recommended Citation
Barua, Miroslava, "Classification of pulmonary diseases using artificial neural networks" (2004). ETD Collection for University of Texas, El Paso. AAI1423707.
https://scholarworks.utep.edu/dissertations/AAI1423707