PRACTICAL COUGH DETECTION IN PRESENCE OF BACKGROUND NOISE AND PRELIMINARY DIFFERENTIAL DIAGNOSIS FROM COUGH SOUND USING ARTIFICIAL INTELLIGENCE

Abstract

Cough is one of the most common symptoms for many of the diseases. Physicians have been using characteristics of cough for preliminary diagnosis of certain respiratory diseases for ages. But the methods have been subjective and often depend on self-reported history and description of cough by the patients. Recently, with advent of the omnipresent recording devices and advances in machine learning capabilities, many studies have attempted to partially fill the gap. These studies have approached the problem objectively to create devices like cough monitors, cough counters, and partial automatic cough detection using machine learning. There is still a huge gap that exists in detecting and diagnosing the cough in a practical way. This study is an attempt to contribute towards filling this gap. We propose and analyze a machine learning based method to automatically detect cough in presence of background noise. After successful cough detection, we investigate the possibility of preliminary differential diagnosis by distinguishing the cough associated with Asthma, Bronchitis, Bronchiolitis, Pertussis patients and healthy people. As more training data could be collected for cough and non-cough sounds, it allowed us to leverage the potential of powerful deep architecture like ResNet for the cough detection part. For the diagnosis part of the work, not much data was available. In this case the preliminary results show that XGBoost performed better than CNN and ResNet architectures. While the cough detection part of the study offers mature results, lot more cough sound data for the examined diseases is needed before generalizable conclusions can be drawn from the diagnosis results observed in this study.