Reliable pneumonia detection and wheeze-crackle classification via data balancing and augmentation

초록

Pneumonia diagnosis based on X-ray imaging can be costly and inaccessible in resource-limited settings. We propose a machine learning framework that classifies respiratory sounds to detect pneumonia and further distinguish wheeze and crackle, which reflect different pathologies. To improve reliability under class imbalance, we apply a hybrid resampling strategy that combines ADASYN and Tomek Link, and incorporate acoustically plausible augmentations through noise injection and volume control to enhance robustness while preserving signal characteristics. Experiments on the ICBHI 2017 dataset show that XGBoost consistently outperforms Random Forest, achieving higher predictive performance and more confident classifications. These results suggest that the proposed approach enables reliable sound-based respiratory disease screening, even with small and imbalanced datasets.

키워드