Application of Clustering Techniques for Lung Sounds to Improve Interpretability and Detection of Crackles

Abstract

Due to the subjectivity involved currently in pulmonary auscultation process and its diagnostic to evaluate the condition of respiratory airways, this work pretends to evaluate the performance of clustering algorithms such as k-means and DBSCAN to perform a computational analysis of lung sounds aiming to visualize a representation of such sounds that highlights the presence of crackles and the energy associated with them. In order to achieve that goal, Wavelet analysis techniques were used in contrast to traditional frequency analysis given the similarity between the typical waveform for a crackle and the wavelet sym4. Once the lung sound signal with isolated crackles is obtained, the clustering process groups crackles in regions of high density and provides visualization that might be useful for the diagnostic made by an expert. Evaluation suggests that k-means groups crackle more effective than DBSCAN in terms of generated clusters.

Keywords

• Lung Sounds
• Crackles
• Vesicular Sounds
• Adventitious Sounds
• Transformed Wavelet
• Decomposition Wavelet
• Symlet
• Clustering
• k-means
• DBSCAN
• Log-energy