Publication

Abstract : Automated respiratory sound classification faces the dual challenge of extracting discriminative acoustic features and handling severe class imbalance inherent in clinical datasets. This study presents a systematic integration of complementary imbalance-handling techniques within a single-stage ensemble classifier combining XGBoost, LightGBM, and Neural Networks with soft voting for multi-class respiratory disease diagnosis. The proposed framework integrates Label-Distribution-Aware Margin (LDAM) loss combined with focal weighting, Effective Number class weighting, SMOTE-based synthetic oversampling, multi-strategy feature selection prioritizing minority class discriminators, and per-class threshold tuning for F1 optimization. To avoid optimistic bias from threshold optimization, we employ nested cross-validation where thresholds are tuned on inner folds. The ensemble architecture avoids error cascades inherent in hierarchical approaches by directly classifying all six disease categories. Evaluated on the ICBHI respiratory sound database using 5-fold stratified group cross-validation with patient-level grouping, the proposed method achieves macro F1-score of 0.46  ±  0.04 with strong majority class performance (COPD F1=0.96) and improved minority class recognition compared to single-model baselines. Comprehensive analysis including ablation studies, baseline comparisons with CNN and Transformer architectures, statistical significance testing with bootstrap confidence intervals, ROC curves, precision-recall curves, and calibration diagrams demonstrates the effectiveness of the systematic combination. While individual techniques (LDAM, SMOTE, gradient boosting) are established, this work contributes their rigorous integration and evaluation for imbalanced respiratory sound classification, providing a practical, deployable solution for computer-aided diagnosis with complete open-source implementation.