A Machine Learning Framework for Multi-Class Lung Disease Classification Using Lung Sounds

Abstract

Diagnostic ambiguity between chronic pulmonary disease like asthma and Chronic Obstructive Pulmonary (COPD) presents significant diagnostic challenges due to their overlapping clinical symptoms. This overlap often leads to misdiagnosis and increased mortality, highlighting the urgent need for effective diagnostic methods. To address this issue, this study proposes a computerized lung sound-based method. The methodology is evaluated using the publicly available ICBHI and KAUH lung sound datasets. Lung sound denoising is performed using a Butterworth bandpass filter. A combination of acoustic features such as Zero-Crossing Rate (ZCR), Mel spectrogram, and Mel-Frequency Cepstral Coefficients (MFCC) are extracted, followed by Z-score normalization to ensure zero mean and unit variance. Four machine learning classifiers, namely Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbors (kNN), and Multi-Layer Perceptron (MLP), are compared. Additionally, the system implements hyperparameter tuning strategies, such as Grid Search (GS), to distinguish among healthy, asthma, COPD, and other lung disease cases. After hyperparameter tuning, the RF classifier achieved the highest validation accuracy of 96.72% and a test accuracy of 96.59%, while kNN achieved the largest performance gain, improving accuracy by 4.1% compared to its baseline. The high accuracy of the proposed framework, which employs a computationally efficient feature set, highlights its potential as a reliable decision-support tool for lung disease classification, offering considerable utility in resource-constrained clinical settings.