Statistical and Machine Learning Analysis of Respiratory Sounds for Multilevel Asthma Severity Classification

Respiratory sounds originating from the lung parenchyma carry essential diagnostic information about different lung diseases. Nevertheless, most of the current databases do not provide detailed anatomical information about the site of recording, respiratory phase, and asthma severity, which may result in incorrect clinical interpretation. We present here an anatomically precise dataset that was checked through a completely new integrated framework for the diagnosis and severity assessment of asthmatic patients with high accuracy. Wheeze segments were extracted from recordings taken from mild, moderate, and severe asthma patient groups. Our analytical methods were comprehensive, based on both univariate and multivariate statistical analyses of frequency-based features. Based on these statistical properties, we introduce a new two-stage classifications framework that aims to improve diagnostic accuracy by a large margin. We tested the capabilities of Support Vector Machine (SVM), k-Nearest Neighbors (KNN), and Ensemble classifiers in two experiment setups. Results show that Framework II scores much higher than Frame-work I, reaching a maximum sensitivity of 97%, a specificity of 99%, and an overall accuracy of 98%. It is worthy of note that data taken from the tracheal area gave the best classification result after comparison with other anatomical sites. The paper presents a strong, scalable diagnosis option with a high degree of clinical usefulness for health care around the world, especially in settings where diagnostic tools are scarce.