Representation and classification of breath sounds recorded in an intensive care setting using neural networks

Abstract

Objective. Develop and test methods for representing and classifying breath sounds in an intensive care setting. Methods. Breath sounds were recorded over the bronchial regions of the chest. The breath sounds were represented by their averaged power spectral density, summed into feature vectors across the frequency spectrum from 0 to 800 Hertz. The sounds were segmented by individual breath and each breath was divided into inspiratory and expiratory segments. Sounds were classified as normal or abnormal. Different back-propagation neural network configurations were evaluated. The number of input features, hidden units, and hidden layers were varied. Results. 2127 individual breath sounds from the ICU patients and 321 breaths from training tapes were obtained. Best overall classification rate for the ICU breath sounds was 73% with 62% sensitivity and 85% specificity. Best overall classification rate for the training tapes was 91% with 87% sensitivity and 95% specificity. Conclusions. Long term monitoring of lung sounds is not feasible unless several barriers can be overcome. Several choices in signal representation and neural network design greatly improved the classification rates of breath sounds. The analysis of transmitted sounds from the trachea to the lung is suggested as an area for future study.