Machine learning & deep learning approaches for respiratory sound-based disease detection

Abstract

Throughout the past decennium, there occurred a lot of scrutiny in the field of spontaneous respiratory sound inspection. The efficacy about decision-making could be improved by the automated systematization about respiratory sounds that could identify anomalies beforehand stages about respiratory dysfunction. It would be advantageous to develop an automatic machine & deep learning-based respiratory sound classification system. The predominant role about the first scientific challenge were to develop algorithms which could distinguish pulmonary audio clips taken originating at both non-medical & medical domains. About 920 recordings in the database were obtained from 126 subjects & were produced by 2 empiricists groups in Portugal & Greece. Recorded were 6898 respiration cycles. Respiratory specialists annotated the cycles as involving Wheezes, Crackles & aggregating both, or absence of abnormal breathing sounds. To facilitate effectiveness of our models, the dataset split up into 3 subcategories: A split of 70% data for training, 15% data for testing, and 15% data for model validation. The importance of having a robust and generalizable model, various data pre-processing techniques were applied. These techniques included Adam Optimizer, Sequential Minimal Optimization and Gradient Descent. We implemented and evaluated several machine learning & deep learning architectures, encompassing Artificial Neural Network (ANN), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Random Forest (RF) & Convolutional Neural Networks (CNN) which are widely used in various fields. Among those, CNN emerged as the most effective model, achieving 99.5% accuracy during training & 98.5% accuracy during testing along with a validation accuracy of 98.5%, thus outperforming the other models. ANN & SVM also demonstrated high performance, with accuracies around 96.6% & 91.5%, while KNN & RF achieved an accuracy of 88.3% & 88.1% respectively. In a clinical setting, the suggested method is very important since it can help doctors with automated illness identification & diagnosis.