Classification of pulmonary diseases based on impulse oscillometric measurements of lung function using neural...

Central and peripheral airflow obstructions frequently occur in patients with chronic obstructive lung disease or asthma and may have different pathophysiological mechanisms of obstruction and require different therapeutic interventions. Impulse oscillometry (IOS) is a patient-friendly method for studying respiratory function in health and disease. The enormous variety of patterns and the high degree of variability in the measured lung function parameters has made the automated diagnosis of pulmonary diseases very desirable by pulmonary physiologists and clinicians. Computer aided diagnosis can serve as a second but quantitative opinion to diagnosis and screening. Presently, there are no robust algorithms to classify the IOS patterns into particular disease groups. In this work, an artificial neural network (ANN) was used to recognize and classify the disease of the central and peripheral airways. Using IOS measurements in 131 patients, a training set was created and used in a feedforward ANN that was trained by backpropagation algorithm to classify pulmonary diseases as either central or peripheral. After supervised training, the ANN was presented with the same data and produced a 98.47% correct classification rate. When a new set of unseen data was used, the classification accuracy was reduced to 61.47%. Having produced a promising classification rate in the first case, the accuracy of this classifier could be further improved. Inclusion of more training samples combined with fuzzy logic decision rules could facilitate the development of a software tool that assists pulmonary specialists with their diagnosis of lung function using the patient-friendly IOS system.