Respiratory Sounds Classification employing a Multi-label Approach

摘要

Respiratory diseases are one of the leading causes of death worldwide, which also reduces the quality of life of the people who suffer from them.. Therefore, there is the necessity to generate tools that allow a rapid and reliable diagnosis support to make an appropriate management of these diseases. Different approaches based on artificial intelligence (AI) have been contributed to these problems, which has shown to be useful in assisting the diagnosis of such diseases. The present proposal holds the use of AI algorithms to identify respiratory sounds that are associated with respiratory diseases (crackles and wheezes), for this, the database Respiratory Sound Database from the ICBHI 2017 Challenge was employed. The proposed models use statistics of time and frequency features, and a multi-label approach for the classification, which is a different approach from that used in related work, where a multi-target approach is employed. As a result, three machine learning algorithms were trained for both a multi-label and a multi-class classification, obtaining comparable results between them. For the case of the multi-label models we obtained at most an average output accuracy of 81.9%.