Automatic classification of normal and sick patients with crackles using wavelet packet decomposition and support vector machine

• Published in: Biomedical signal processing and control Vol. 67; p. 102521
• Main Authors: Stasiakiewicz, Paweł, Dobrowolski, Andrzej P., Targowski, Tomasz, Gałązka-Świderek, Natalia, Sadura-Sieklucka, Teresa, Majka, Katarzyna, Skoczylas, Agnieszka, Lejkowski, Wojciech, Olszewski, Robert
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.05.2021
• Subjects: Crackles; Support vector machine; Wavelet packets; Crackles; Wavelet packets; Support vector machine
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2021.102521

•Method of wavelet packets properties analysis was proposed and non-intuitive exchange of frequency sub-bands was observed.•Feature generation was conducted using wavelet packets and respiratory phase detection algorithm.•Higher order wavelets such as db20 are better suited to generate the distinctive features of lung sounds with crackles.•System generalisation capability was checked using cross-validation method with subject crossing.•Classifiers ensemble is characterized by 95 % sensitivity and 91 % specificity using 10-fold cross-validation method. Auscultation of the respiratory system – a key system in a human body – is a complicated procedure and it requires a doctor to have good perception skills and profound experience. During auscultation, specific sounds are identified by the doctor who then associates the acoustic phenomena heard with pathological processes. This article is an attempt at developing a classification system, using wavelet packets, a genetic algorithm, and a Support Vector Machine (SVM), which distinguishes between healthy patients and patients with crackles caused by pneumonia, pulmonary fibrosis, Heart Failure (HF) or Chronic Obstructive Pulmonary Disease (COPD). The system is elaborated and tested over a dataset consisting of 62 healthy (166 recordings) and 58 sick patients (187 recordings). A reliable system is described, consisting of 5 wavelet classifiers, featuring approx. 95 % sensitivity and 91 % specificity, applying 10-fold cross-validation.