Audio analysis with convolutional neural networks and boosting algorithms tuned by metaheuristics for respiratory condition classification

Abstract

In contemporary medical research, respiratory disorders have become a primary focus. Improving patient outcomes for any medical condition largely depends on early identification and prompt treatment. Traditionally, medical professionals diagnose respiratory diseases by auscultating a patient’s breathing. However, this method has inherent limitations, as it may not enable physicians to accurately identify every respiratory condition. This study explores the potential of using convolutional neural networks (CNNs) in conjunction with audio analysis for the identification of respiratory problems. This work proses a novel two-tier framework that integrates CNNs with extreme gradient boosting (XGBoost) and adaptive boosting (AdaBoost) models to classify respiratory conditions. Additionally, modern optimization techniques are employed to enhance classification efficiency, recognizing the significant impact that appropriate hyperparameter tuning has on machine learning (ML) and deep learning (DL) performance. This research introduces a modified version of particle swarm optimization (PSO) tailored to meet the specific needs of ML and DL tuning. The proposed approach is validated using a real-world clinical dataset. Two studies, both based on mel spectrograms of patient breathing patterns, were conducted: the first aimed at determining whether patients have respiratory conditions (binary classification), while the second employed the same data structure for multi-class classification. In both scenarios, advanced optimizers were utilized to optimize model architecture and training settings. Under identical testing conditions, the proposed PSO metaheuristic achieved an accuracy of 98.14$\text{\%}$ for respiratory condition detection in binary classification and a slightly lower accuracy of 81.25$\text{\%}$ for specific condition identification in multi-class classification.