Automated Diagnosis of Pulmonary Diseases Using Lung Sound Signals

Pulmonary diseases are the third biggest cause of deaths worldwide. A prominent method to detect these diseases is the observation of lung sound signals. There is an increasing need for an efficient technique that can automatically diagnose such diseases with high accuracy. In this paper, two popular public datasets are considered, and every lung sound signal is decomposed into 8 frequency bands using rectangular zero-phase filters. Features are extracted from every band, including energy, kurtosis, mean absolute deviation and Lp norm. The extracted features are utilized for classification using machine learning schemes. The proposed method achieves 99.9% accuracy for multi-class classification on the combined dataset, and for binary classification, we have evaluated normal signal versus pathogenic signal, which is found to be 100% accurate for most of the diseases. High accuracy is obtained for the individual datasets as well. Top 20 features selected using minimum redundancy maximum relevance algorithm also yield 98.6% accuracy. Therefore, the proposed method can be easily deployed in real-time systems.