Respiratory Disease Classification Using NMF-Enhanced Log-Mel Spectrograms and Convolutional Recurrent Neural...

Area of Science:

• Medical acoustics
• Artificial intelligence in healthcare
• Signal processing for biomedical applications

Background:

• Respiratory disease classification from lung sounds is difficult due to signal noise and overlapping acoustic patterns.
• Existing methods struggle with the heterogeneity of audio data and diverse disease presentations.

Purpose of the Study:

• To develop and evaluate a novel framework for multi-class respiratory disease classification using enhanced lung sound analysis.
• To improve the accuracy and robustness of automated detection for conditions including Asthma, COPD, Pneumonia, and others.

Main Methods:

• Lung sound recordings from public datasets were harmonized into a unified seven-class label space.
• A Non-negative Matrix Factorization (NMF) enhancement was applied to log-mel spectrograms for improved feature salience.
• A Convolutional Recurrent Neural Network (CRNN) with attention was developed and compared against other deep learning architectures.

Main Results:

• The proposed CRNN model achieved high performance, with 96.14% accuracy and 94.05% Macro-F1 score on the seven-class dataset.
• The CRNN demonstrated superior class separation and more balanced recognition across different respiratory diseases compared to RDLINet, ResNet, and YOLO-style models.
• Class-wise analysis and confusion matrix evaluation confirmed the CRNN's effectiveness.

Conclusions:

• NMF-enhanced spectro-temporal modeling combined with CRNNs offers a powerful approach for automated multi-class respiratory disease classification.
• This framework shows significant potential for clinical applications in diagnosing respiratory conditions from lung sound recordings.
• The study highlights the benefits of advanced signal processing and deep learning for overcoming challenges in respiratory sound analysis.