Respiratory sound analysis in a rabbit tracheomalacia model

Tracheomalacia, which softens the tracheal wall and leads to airway collapse during breathing, can be challenging to detect, especially in children, due to its vague symptoms and the invasive nature of standard diagnostic procedures. In this study, we developed a rabbit model of tracheomalacia by surgically removing a portion of the tracheal cartilage in New Zealand White rabbits. Breath sounds and airway pressure were recorded using a clinical-grade stethoscope setup. From each exhaled breath, 6,373 acoustic features were extracted, and 51 of these, which showed consistent and statistically significant differences (p < 0.01), were selected in at least four out of five animals. We trained and evaluated three machine learning models—LightGBM, logistic regression, and support vector machines—using cross-validation. Among them, LightGBM delivered the highest performance, with area under the curve values > 0.78 for individual breaths and > 0.80 when averaged per subject. The most important features were the low-frequency components of Mel-frequency cepstral coefficients, which likely detected subtle changes in airflow caused by airway collapse. These findings support the potential use of respiratory sound analysis combined with machine learning as a practical and non-invasive tool for identifying tracheomalacia. This model may also be useful for developing pediatric diagnostic methods due to the anatomical similarities between rabbit and neonatal airways.