Classification of Adventitious Respiratory Sound Events: A Stratified Analysis

Abstract

Respiratory diseases are among the deadliest in the world. Adventitious respiratory sounds, such as wheezes and crackles, are commonly present in these pathologies. Automating the analysis of adventitious respiratory sounds can help health professionals monitor patients suffering from respiratory conditions. The ICBHI Respiratory Sound Database, a benchmark dataset in respiratory sound analysis, has large and diverse data available publicly. Given its diversity in data, a stratified analysis by recording equipment, age, sex, body-mass index (BMI), and clinical diagnosis is proposed in this article. Regarding the experiments, three machine learning algorithms (Support Vector Machine - SVM, Random Undersampling Boosting - RUSBoost, and Convolutional Neural Network - CNN) were employed in three tasks: 2-class crackles (crackles vs. others), 2-class wheezes (wheezes vs. others), and 3-class (crackles vs. wheezes vs. others). Overall, the CNNs achieved the best results in almost every category, except when the equipment was Littmann3200 or Meditron, where RUSBoost achieved better results. In terms of stratification categories, we observed significant differences in classification performance, namely in terms of equipment, where the Littmann3200 underperformed the other equipment analysed. In addition, in the 3-class task, the CNNs achieved better results in Male subjects than Female subjects. In terms of BMI, the CNN of the Overweight class in the 2-class wheeze task achieved worse results than the other two BMI classes (Normal and Obese).