Wheeze and Crackle Discrimination Algorithm in Pneumonia Respiratory Signals

A new pneumonia detection method is proposed to provide both pneumonia detection in respiratory sound signals and wheeze and crackle discrimination when pneumonia episodes are detected. In the proposed method, two-step hierarchy, classifying pneumonia in the first step and discriminating wheezing and crackling in the second step, is considered; the conventional pneumonia detection method is modified to improve pneumonia detection performance, while wheezing and crackling discrimination functionality is added to facilitate the application of appropriate remedies for each case. We used resampling techniques to address the imbalance in the ICBHI pneumonia dataset. The random forest algorithm is used to classify pneumonia from healthy respiratory data, as well as to distinguish between wheeze and crackle from pneumonia data. Against the ICBHI respiratory dataset, the proposed random forest-based hierarchy pneumonia detection method provides 85.40% accuracy in detecting pneumonia and 82.70% accuracy in discriminating wheeze from crackling, respectively. Keywords—signal processing, machine learning, audio signals, pneumonia, hierarchical classification.