Self-Supervised Audio Representation Learning for Pediatric Asthma Detection in Emergency Care Using Digital Stethoscope Recordings

Abstract

Accurate diagnosis of pediatric asthma in emergency departments remains challenging due to overlapping respiratory symptoms, time constraints, and the limited feasibility of pulmonary function testing in young children. This study investigates the feasibility of pediatric asthma detection in the emergency department using breath sound recordings and machine learning. Thirty-second breath sounds were collected from six chest locations in 31 pediatric patients (10 asthmatic, 21 non-asthmatic) and analyzed using pretrained self-supervised speech representation models (HuBERT, WavLM, and Wav2Vec 2.0) for feature extraction, with patient age and sex incorporated into the feature representations. Conventional machine learning classifiers were trained and evaluated using patient-level stratified group 5-fold cross-validation and leave-one-patient-out validation to ensure the generalizability of the findings. Among the evaluated approaches, Wav2Vec 2.0 combined with histogram-based gradient boosting achieved the strongest and most consistent performance, yielding an accuracy of 0.84, sensitivity of 0.80, specificity of 0.86, and F1-score of 0.76 under both evaluation protocols. The consistency of performance across validation strategies suggests promising generalization to unseen patients. These findings suggest that pretrained self-supervised audio representations offer a promising, non-invasive approach for pediatric asthma detection in real-world emergency department settings, where objective respiratory assessment is often limited.