Custom-Built Deep Convolutional Neural Network for Breathing Sound Classification to Detect Respiratory Diseases

Abstract

In any living being, the respiratory system plays a vital role and is responsible for taking oxygen required for body organs and blood. It has a substantial impact on global health. Detecting and diagnosing the respiratory diseases is a challenging task for the medical practitioners. When it comes to addressing COVID-19 in the current situations, it becomes substantially more dangerous and leading to death since the virus directly effecting the human respiratory system. In COVID times, wearing masks for longer times is also leading to respiratory diseases. The traditional method used for observing the respiratory disorders is auscultation. It is well-known for being less costly, non-invasive, and safe, requiring less diagnosis time. However, the accuracy of diagnosis with auscultation is dependent on the physician’s expertise and understanding and therefore necessitates substantial training. This paper suggests a solution that depends on a deep CNN for diagnosing respiratory diseases. A customized deep convolutional neural network was developed by considering the stacked LSTM model to classify the breathing sounds to detect respiratory diseases. The developed model was built to categorize the six different types of breathing cycles included in the “ICBHF17 scientific challenge respiratory sound database”, and it performs well with 98.6% accuracy. We compared the developed model’s efficiency against state-of-the-art models.