A novel data augmentation technique based on wheezing physiological modeling applied to asthma severity management in respiratory sounds

Wheezing sounds are critical biomarkers for diagnosing respiratory disorders, including asthma. Recent advancements still depend on the availability of high-quality training datasets. The creation of public databases containing a diverse array of wheezing sounds from patient auscultations is vital. Nonetheless, acquiring a sufficiently large and varied collection of wheezing sounds remains challenging because of the significant time required for comprehensive data collection by clinicians, the limited availability of patients exhibiting these acoustic biomarkers, and the infrequency of certain wheezing sounds being made. These factors complicate the generalization of deep learning models and make it particularly difficult to capture sounds in adequate quantities. In this work, we propose a novel data augmentation technique based on wheezing physiology, overcoming the limitations of patient-derived data and sound transformation from data augmentation techniques that are not consistent. The proposed technique mathematically models the temporal and spectral features of wheezing, generating synthetic sounds for use in robust learning, and it can be applied to specific respiratory conditions in which wheezing actively occurs. Finally, the proposed data augmentation method is integrated into an application designed under medical supervision for detecting asthma severity in asthmatic patients. Specifically, the proposed app diagnoses asthma severity by analyzing auscultated sounds by considering both the number of auscultation points where wheezing is detected and the respiratory rate. The experimental results demonstrate that CNN architectures trained with synthetic data achieve competitive performance relative to those trained with real data, with a marginal decrease in accuracy of less than 2.5%. The temporal distributions of energy and vibrato are the most important parameters influencing the similarity between the performance results obtained by considering the synthetic wheezing sounds and the real wheezing sounds. These findings suggest that the proposed data augmentation method is a tool that can be used to generate extensive synthetic data that effectively generalize the main characteristics of most wheezing sounds found in real-world environments. Finally, the developed app for diagnosing asthma severity can serve as a useful tool for expediting medical interventions in cases of high asthma severity.