Pneumonia Detection Using Hjorth Parameters and Spectral Statistical Features of Respiratory Sounds

Abstract

Pneumonia is a serious inflammatory disease that causes pathological changes in the lungs. These changes produce distinctive patterns in respiratory sounds, providing a potential route to pneumonia detection that avoids the patient burden of chest X-ray imaging. This paper proposes a machine learning-based framework for pneumonia detection based on analysis of respiratory sound signals. For this purpose, a balanced set of 38 twenty-second segments, 19 from healthy subjects and 19 from pneumonia patients, was selected from the International Conference on Biomedical and Health Informatics (ICBHI) database. Using MATLAB simulations, three Hjorth parameters (activity, mobility, and complexity) and five spectral statistical features (centroid, spread, skewness, kurtosis, and entropy) were extracted and ranked using Pearson's correlation coefficient. A linear support vector machine (SVM), naïve Bayes (NB), random forest (RF), and decision tree (DT) classification models were trained on the most significant features and tested using 10-fold cross-validation. The results obtained revealed the strength of integrating Hjorth parameters and spectral statistical features compared with either alone. In particular, the SVM classification accuracy, F1 score, and false positive rate reached 94.74%, 0.95, and 0.05, respectively, surpassing the performance of the other classifiers. These findings demonstrate the feasibility of using Hjorth parameters and spectral statistical features with a linear SVM for pneumonia detection and motivate further validation on larger, more diverse datasets to assess the proposed framework as a reliable alternative to chest X-ray-based pneumonia detection.