Comparison of classical and deep learning models for respiratory sound classification

This study compares classical machine learning and modern deep learning approaches for the automatic classification of respiratory sound signals on the ICBHI 2017 dataset. After preprocessing, Mel-spectrograms and spectral features were extracted. Classical algorithms (SVM, KNN, Random Forest, XGBoost) were trained on handcrafted features, while convolutional architectures (the GhostNet family, ResNet-50, EfficientNet-B0, MobileNetV3) were trained on Mel-spectrograms. A patient-independent (subject-wise) split was adopted, and class-balancing data augmentation was applied only to the training folds to avoid validation/test leakage. The results show that GhostNetV4 achieved the best performance with 89.9% accuracy and macro-F1 = 0.898, outperforming both deep and classical baselines. These findings indicate that computationally efficient deep learning models are strong candidates for integration into clinical decision support systems for respiratory sound analysis. The study also contributes a transparent evaluation protocol that improves comparability across works by coupling a patient-independent split with a clearly defined augmentation policy.