Gyurin Byun, Huigyu Yang, Hyunseung Choo, Journal of Internet Computing and Services, Vol. 26, No. 5, pp. 75-81, Oct. 2025
10.7472/jksii.2025.26.5.75, Full Text:  HTML
Keywords: Respiratory sound classification, Temporal Convolutional Network (TCN), Multi-spectral fusion, Deep time-series learning

Abstract

Following the pandemic, the demand for remote healthcare and smart monitoring technologies has significantly increased, emphasizing the importance of automated respiratory sound analysis for early diagnosis and long-term monitoring. This study proposes a multi-spectral CNN–TCN model for abnormal respiratory sound classification using publicly available datasets (HF Lung V1 and ICBHI). Three spectral representations—Mel Spectrogram, MFCC, and Chroma—are independently processed and encoded by a VGG16-based CNN pretrained on the ImageNet dataset, followed by individual Temporal Convolutional Network (TCN) modules with dilated convolutions (dilation = 1, 2, 4) to capture short- and long-term temporal dependencies. Experimental results show that the proposed model achieves a mean F1-score improvement of over 4~6% compared with conventional LSTM and GRU models, while maintaining a lightweight architecture with fewer than 0.3 million parameters. These findings demonstrate that combining multi-spectral feature fusion with temporal convolutional learning significantly enhances the accuracy and robustness of abnormal respiratory sound classification.

Statistics
Cite this article