Wav2Vec 2.0 detects pediatric asthma from stethoscope recordings with 84% accuracy
AI listens for asthma: Wav2Vec 2.0 identifies pediatric cases from breath sounds alone
Using pretrained self-supervised speech representation models (HuBERT, WavLM, Wav2Vec 2.0) for feature extraction from 30-second digital stethoscope recordings of 31 pediatric patients (10 asthmatic, 21 non-asthmatic) in the emergency department, Wav2Vec 2.0 combined with histogram-based gradient boosting achieved 0.84 accuracy, 0.80 sensitivity, 0.86 specificity, and 0.76 F1-score under both patient-level stratified group 5‑fold cross-validation and leave-one-patient-out validation. This non-invasive approach could provide objective asthma assessment in emergency settings where pulmonary function testing is limited.
- Wav2Vec 2.0 combined with histogram-based gradient boosting achieved 84% accuracy on pediatric asthma detection
- 30-second breath sound recordings from six chest locations were collected from 31 emergency department patients
- Model showed consistent performance across two validation strategies, suggesting strong generalization to unseen patients
Why It Matters
Enables rapid, non-invasive pediatric asthma diagnosis in emergency settings where objective pulmonary testing is often unavailable.