Integration of machine learning and signal processing for asthma severity classification using respiratory sounds

Background: Respiratory sounds are a rich source of diagnostic information for diseases such as asthma. However, conventional listening through a stethoscope and feature extraction combined with manual design lack accuracy and are subjective. Objective: In this study, a novel method is presented based on deep learning with a convolutional neural network (CNN), which processes the Mel-spectrograms of wheeze sounds for the categorization of asthma severity into mild, moderate, and severe stages. Methods: The study evaluated a large set of 111 asthma sufferers—the most severity-specific dataset to date—gathered from multiple lung sites and across two respiration phases in two hospitals with ethical permission. Vocal Tract Length Perturbation (VTLP) data augmentation was employed to improve the system’s resistance to noise and variations. The proposed CNN was trained to identify spectral-temporal characteristics from the processed audio. Results: The CNN reached 91.58 % validation accuracy for All dataset, and up to 94 % for data specific to trachea lung positions. Tracheal audios performed better than lower lung bases, and expiratory cycles exhibited more differentiating features than inspiratory ones. Conclusion: To our knowledge, this is the first CNN approach for the recognition of asthma severity, demonstrating that respiratory sound locations and phases reveal distinct characteristics. This work paves the way for non-invasive, objective, and clinically useful decision-making support in asthma therapy.