Improving pulmonary disease detection through autoregressive features and K-nearest neighbours classifier

The classification of lung sound signals, concerning autoregressive modelling and machine learning algorithms, is the main objective of this study. Lung sounds recorded from 112 subjects with three filter modes (bell, diaphragm, extended) were then classified into two groups: healthy and pulmonary conditions like asthma, COPD, and pneumonia. The AR modelling extracted five essential features for all modes, combined with a classifier using K-nearest neighbours (KNN). During training, it gives an accuracy of 98.3% for the unhealthy and 95.5% for the healthy cases, whereas during testing, the results were 100% and 92.3%, respectively. The overall accuracy was 98.2%. Dimensionality was reduced with decreased computational load using the method of AR; therefore, the simple model KNN achieved high accuracy. This efficiency makes the approach suitable for hardware implementation in portable, point-of-care diagnostic devices and thus helps in respiratory disease diagnosis in remote and clinical settings.