Publication

Abstract : Accurate classification of pulmonary diseases using respiratory sounds remains a challenge due to class imbalance, overlapping acoustic patterns, and limited data for minority conditions. This study proposes a hybrid framework that combines Principal Component Analysis (PCA) for dimensionality reduction with Variational Autoencoder (VAE)-based data augmentation to enhance deep learning model performance. PCA reduces the 32-dimensional acoustic feature space to 25 principal components, retaining critical variance while minimizing redundancy. VAE is trained on these compressed representations to generate label-consistent synthetic samples that improve class balance and intra-class diversity. The proposed pipeline is evaluated across six deep learning architectures—CNN, LSTM, GRU, Bi-GRU, Bi-LSTM, and Conformer. Notably, GRU and CNN—being relatively simpler architectures—exhibited the most significant performance gains, underscoring their compatibility with PCA-VAE-based augmentation. GRU achieved 99.70% accuracy, 98.95% precision, and 98.89% F1-score, including a remarkable +20.45% F1 gain in URTI, a previously underrepresented class. Similarly, CNN showed substantial improvements in detecting Pneumonia and URTI, with F1-score increases of +37.61% and +39.71%, respectively. Both models achieved near-perfect Area Under the Curve (AUC) values post-augmentation, reffecting strong discriminative power and high diagnostic reliability across multiple respiratory conditions. These findings emphasize the complementary roles of PCA and VAE—where PCA enhances feature compactness and removes redundancy, while VAE synthesizes diverse, label-consistent samples in the compressed space. This synergy particularly benefits simpler models like GRU and CNN, enabling them to generalize more effectively and improving their sensitivity to minority classes. Together, PCA and VAE form a robust augmentation pipeline that significantly elevates the performance of respiratory disease classification systems.