Publication

Abstract : Conventional auscultation of respiratory sounds by physicians is often limited by subjective interpretation and inconsistent clinical outcomes. This research presents an efficient, interpretable, and computationally lightweight machine learning framework for classifying respiratory sounds into five diagnostic categories: Bronchiolitis, COPD, Healthy, Pneumonia, and URTI. A comprehensive set of acoustic features—including Mel- Frequency Cepstral Coefficients (MFCCs), formants, and spectral descriptors—was extracted and characterized using statistical descriptors to capture relevant temporal and spectral patterns. Several feature selection techniques, including InfoGain, ReliefF, Wrapper-based selection with Best First Search and Principal Component Analysis (PCA) was applied to reduce the original feature space from 260 to approximately 50 near-optimal features. Three traditional classifiers—Random Forest, Support Vector Machine (SVM), and k-nearest Neighbors (k-NN)—were systematically evaluated. The k-NN model combined with Wrapper-based feature selection achieved the highest classification accuracy of 98.51%, outperforming other traditional and common deep-learning-based methods. The proposed approach offers high diagnostic performance while maintaining low computational complexity, making it suitable for real-time deployment in resource-constrained clinical environments.