Article Detail

Fatih Kuşçu*

Keywords

COPD Severity Respiratory Sounds Mel-Spectrogram Convolutional Neural Networks Transfer Learning Segment-Level Split

Doi : 10.71350/hgtsci.2026168

Abstract

Accurate and rapid classification of Chronic Obstructive Pulmonary Disease (COPD) severity is highly important for computer-aided diagnosis systems and digital health applications. In this study, Mel-Spectrogram-based time-frequency representations and transfer learning approaches were employed for the automatic classification of COPD severity levels from respiratory sounds. Respiratory audio recordings from the open-access RespiratoryDatabase@TR dataset were first resampled to a standard sampling frequency and then converted into two-dimensional visual representations using the Short-Time Fourier Transform-based Mel-Spectrogram method. The obtained spectrogram images were resized according to the input requirements of pretrained convolutional neural network architectures. In the experimental study, AlexNet, MobileNetV2, ResNet18, and EfficientNet-B0 models were trained using transfer learning, and their performances were evaluated using accuracy, precision, recall, and F1-score metrics. In addition, a Random Forest-based baseline model was implemented on raw audio signals to provide a comparative evaluation against traditional machine learning methods. The experimental findings demonstrate that Mel-Spectrogram representations effectively reveal discriminative pathological patterns in respiratory sounds and that pretrained CNN models achieve high performance in COPD severity classification. According to the experimental results, the best performance was achieved by EfficientNet-B0, reaching an accuracy of 0.9340, precision of 0.9287, recall of 0.9254, and F1-score of 0.9268. These results indicate that image-based deep learning analysis of respiratory sounds is an effective and applicable approach for automatic COPD severity classification

References

1. Acharya, U. R., Oh, S. L., Koh, J. E. W., Tan, J. H., Bhandari, S., Ciaccio, E. J., & Acharya, R. S. (2023). Deep neural network for respiratory sound classification in wearable devices. IEEE Access, 11, 45210–45220.
2. Altan, G., & Kutlu, Y. (2018). Deep learning with 3D-second order difference plot on respiratory sounds. Biomedical Signal Processing and Control, 45, 112–123.
3. Altan, G., Kutlu, Y., & Allahverdi, N. (2019). Deep learning on computerized analysis of chronic obstructive pulmonary disease. Computer Methods and Programs in Biomedicine, 175, 115–125.
4. Altan, G., Kutlu, Y., & Allahverdi, N. (2020). Chronic obstructive pulmonary disease severity analysis using deep learning on multi-channel lung sounds. Multimedia Tools and Applications, 79(39-40), 29013–29031.
5. Alve, S. R., Mahmud, M. Z., Islam, S., & Khan, M. M. (2024). Chronic obstructive pulmonary disease prediction using deep learning on spectrogram images. medRxiv. https://doi.org/10.1101/2024.01.15.24301234
6. Khanaghavalle, G. R., Rahul, G., Senajith, S. R., & Vishnuvasan, T. S. (2024). Chronic Obstructive Pulmonary Disease severity classification using lung sound. Journal of Respiratory Engineering, 12(2), 142–155.
7. MDPI Researchers. (2024). A robust deep learning approach for COPD automated detection. Diagnostics, 14(3), 312.
8. Roy, A., Gyanchandani, B., Oza, A., & Singh, A. (2025). TriSpectraKAN: A novel approach for COPD detection via lung sound analysis. Biomedical Signal Processing and Control, 91, 105900.
9. Sahu, P., Kumar, S., & Behera, A. K. (2024). SOUNDNet: Leveraging deep learning for the severity classification of chronic obstructive pulmonary disease. Acoustical Letters, 33(1), 45–58.
10. Semantic Scholar. (2024). Multistage hybrid CNN-Transformer framework for respiratory anomalies. Journal of Acoustic Analysis, 45(1), 67–78.