Multilabel and Multiclass Approaches Comparison for Respiratory Sounds Classification

Abstract

Respiratory diseases are one of the leading causes of death worldwide according to ten World Health Organization (WHO) due to fatal issues and produce a decreasing of the life quality for people who suffer it. Therefore, there is a necessity to generate tools that allow agile and reliable diagnostic support systems for management of these diseases. Recently, different approaches based on artificial intelligence (AI), mostly at employing artificial neural networks (NN) have been validated to be a successful alternative in respiratory diseases diagnosis using images and signals as information sources. The present proposal uses AI algorithms used on auscultation signals from the respiratory system, identifying respiratory sounds associated to pulmonary diseases (crackles and wheezes). The records used were extracted from the Respiratory Sound Database of the ICBHI 2017 Challenge. Different works have used this database to apply a multiclass classification with satisfactory performance results. However, the ICBHI holds the labels in a multilabel format. Due to this, the present work explores the use of the multilabel target for the classification of these respiratory sounds. Statistics from time and frequency features were used to train five classic machine learning (ML) models for a comparison between multilabel and multiclass classification. A k-fold cross-validation was employed to evaluate the performance of the models with similar results compared to the classical multiclass classification, but with the advantages of the multilabel employment objective such as better represents the problem, make it a better alternative.