Respiratory Disease Classification Using IPMix and Custom Deep Learning Architectures

This paper introduces a comprehensive framework for respiratory disease classification, addressing challenges such as class imbalance and feature variability in lung sound patterns. The proposed approach employs advanced data augmentation techniques, including IPMix and other audio-specific methods, to enhance data diversity and ensure balanced class representation. A multi-input CNN architecture is designed, utilizing Logmel Spectrogram, MFCC, and CQT Spectrogram as input features to capture diverse acoustic characteristics critical for disease detection. In addition, a hybrid ResNet-LSTM model is proposed, effectively capturing both spatial and temporal dependencies in respiratory sounds. To further optimize classification performance, an ensemble method is applied, combining predictions from the custom architectures to leverage their complementary strengths. Evaluated on the ICBHI 2017 Respiratory Sound Database, the framework achieves notable results in both binary (Healthy vs. Unhealthy) and multi-class (eight respiratory conditions) classification tasks. This study underscores the potential of combining advanced data augmentation techniques with tailored deep learning architectures to enhance respiratory disease diagnosis and support clinical applications.