Computerized Classification of Normal and Abnormal Lung Sounds by Multivariate Linear Autoregressive Model

Abstract

This work proposes multichannel acquisition of lung sounds by a microphone array, feature extraction by a multivariate AR (MAR) model, dimensionality reduction of the feature vectors (FV) by SVD and PCA and, their classification by a supervised neural network. A microphone array of 25 sensors was attached on the thoracic surface of the subjects, who were breathing at 1.5 L/sec. The supervised neural network used the backpropagation learning algorithm based on the Levenberg-Marquardt rule. Figures of merit for the classification task by the neural net include the percentage of correct classification during training, testing and validation phases as well as sensitivity, specificity and performance. MAR in combination with PCA provided the best average percentage of correct classification with acoustic information not seen by the neural network during the training phase (87.68%). The results state the advantages of a microphone array for the classification of normal and abnormal acoustic pulmonary information in diffuse interstitial pneumonia and for this goal, the authors assume that not only the crackles and their number indicates the severity of the disease, but the basal respiratory signal could be also affected.