Abstract / Summary
One of the chronic respiratory conditions is asthma, which causes airway inflammation and affects millions of individuals worldwide. Early and accurate asthma detection is essential for timely clinical intervention; however, existing respiratory-audio-based detection mechanisms suffer from high training complexity, limited generalization capability, inadequate feature representation, and inconsistent classification performance. To address these challenges, this research develops a Cultural Guidance Optimized Patch-Mix Contrastive Learning enabled Convolutional Neural Network Light Gradient Boosting Machine (CGO-PC2BM) framework for automatic asthma detection and classification. The novelty of the proposed framework lies in the synergistic integration of Spectrogram Statistical Audio Features (S2AF), Patch-Mix Contrastive Learning, CNN-LightGBM ensemble classification, and Cultural Guidance Optimization Algorithm (CGOA)-based adaptive hyperparameter optimization within a unified respiratory-audio analysis framework. The proposed S2AF mechanism combines VGGish embeddings, hybrid CQT-STFT spectrogram representations, and statistical audio descriptors to capture complementary semantic, spectral, temporal, and statistical characteristics of respiratory sounds. Furthermore, Patch-Mix Contrastive Learning enhances discriminative representation learning and improves generalization toward unseen respiratory audio samples, while LightGBM accelerates model training and improves classification accuracy. In addition, CGOA optimizes model hyperparameters and improves convergence towards the optimal solution. Experimental results demonstrate that the proposed framework achieves specificity, accuracy, F1-score, precision, and sensitivity values of 96.08%, 96.61%, 96.39%, 95.76%, and 97.03%, respectively, on the Asthma Detection Dataset Version 2, demonstrating its effectiveness for respiratory-audio-based asthma detection and classification.