A Novel AI-Inspired Method and System Implementation for Detecting and Classifying Pediatric Respiratory Sound Events

Physicians utilize sound records for diagnosing respiratory diseases, but human error in judgment can lead to misdiagnosis and delay treatment. We propose an AI-inspired system separated into event detection and event classification. For event detection, we use voice activity detection to detect the presence of sound events. For event classification, we use the trained DenseNet model to classify various sound events associated with respiratory conditions. Ultimately, the process correlates sound events to different respiratory diseases based on the identified sound events. The F-score and Error Rate are 0.3296 and 1.36 for the overall evaluation metrics with the SPRSound dataset.

I. Introduction

Respiratory diseases are common pediatric conditions with high mortality rates. By observing respiratory sound spectra, different respiratory conditions can be identifiedby their distinct sound patterns. Using this fact, we use an extensive respiratory sound dataset to develop a sound event detection algorithm to assist doctors in diagnosing respiratory diseases.