Abstract:
Stethoscope-based auscultation is the most efficient, non-invasive and inexpensive technique for
assessing lung conditions on the base of lung sounds analysis. However, it provides a subjective
perception of lung sounds. Additionally, the stethoscope recording is highly vulnerable to
different noises which could mask the important features of lung sounds and may lead to
misdiagnosis. Moreover, given the non-stationary nature of lung sounds, the exclusive time or
spectral domain analysis is not effective for analysis. Hence, the time-frequency analysis of lung
sounds is paramount. In this research, a method for efficient analysis of lung sounds used for
further classification of lung diseases has been proposed. An electronic stethoscope has been
constructed for signal acquisition. The lung sound signals used in the study were normalized at a
uniform sampling frequency of 44.1KHz, and16 bit-depth. We employed wavelet multiresolution
analysis technique to analyze the lung sound signals. A six-level DWT was performed for better
analysis of lung sound signals by decomposing them into details and approximation. Wavelet
denoising technique was used for the pre-processing task. Four wavelet functions (Db4, Db10,
Sym5, and Sym13) with soft and hard thresholding methods and four different threshold selection
rules were used to analyze the required performance of denoising the lung sound signals. Sym13
Wavelet function with soft thresholding method outperforms all other wavelet functions in
denoising lung sound signals. In the feature extraction task, a total of 16 features have been
extracted by following a DWT-based feature extraction procedure. Moreover, one-way ANOVA
was applied for the feature selection task to select the most relevant features. After feature
selection, a total of 13 features were used for final classification of data. A comparison among
the classification accuracies of the different machine learning classifiers was performed and Fine
Gaussian SVM has been selected due to its higher classification accuracy. At last, we optimized
the selected model using Bayesian optimization technique and an accuracy, specificity, and
sensitivity of 99%, 99.2%, and 99.04%, respectively, has been achieved on the unseen data. The
proposed method delivered a considerable improved result for classification of lung diseases as
asthma, pneumonia, COPD, URTI, LRTI, bronchiectasis, bronchiolitis or healthy. Furthermore,
as a future work, the proposed method can be further improved using deep learning techniques
especially for larger lung sound data
