Abstract:
Purpose: Lung diseases are the third leading cause of death worldwide. Stethoscope-based auscultation is the most commonly used,
non-invasive, inexpensive, and primary diagnostic approach for assessing lung conditions. However, the manual auscultation-based
diagnosis procedure is prone to error, and its accuracy is dependent on the physician’s experience and hearing capacity. Moreover, the
stethoscope recording is vulnerable to different noises that can mask the important features of lung sounds which may lead to
misdiagnosis. In this paper, a method for the acquisition of lung sound signals and classification of the top 7 lung diseases has been
proposed for improving the efficacy of auscultation diagnosis of pulmonary disease.
Methods: An electronic stethoscope has been constructed for signal acquisition. Lung sound signals were then collected from people
with COPD, upper respiratory tract infections (URTI), lower respiratory tract infections (LRTI), pneumonia, bronchiectasis, bronch iolitis, asthma, and healthy people. Lung sounds were analyzed using a wavelet multiresolution analysis. To choose the most relevant
features, feature selection using one-way ANOVA was performed. The classification accuracy of various machine learning classifiers
was compared, and the Fine Gaussian SVM was chosen for final classification due to its superior performance. Model optimization
was accomplished through the application of Bayesian optimization techniques.
Results: A test classification accuracy of 99%, specificity of 99.2%, and sensitivity of 99.04%, have been achieved for the 7 lung
diseases using the optimized Fine Gaussian SVM classifier.
Conclusion: Our experimental results demonstrate that the proposed method has the potential to be used as a decision support system
for the classification of lung diseases, especially in those areas where the expertise and the means are limited.
Keywords: auscultation, classification, denoising, discrete wavelet transform, feature extraction, lung diseases, lung sounds
