Investigation into respiratory sound classification for an imbalanced data set using hybrid LSTM-KAN architectures

Respiratory sounds captured via auscultation contain critical clues for diagnosing pulmonary conditions. Automated classification of these sounds faces the dual challenge of distinguishing subtle acoustic patterns and addressing the severe class imbalance inherent in clinical datasets. This study investigates methods for classifying respiratory sounds into multiple disease categories, with a specific focus on mitigating pronounced class imbalances. In this study, we developed and evaluated a hybrid deep learning model incorporating a Long Short-Term Memory (LSTM) network as a feature sequence encoder, followed by a Kolmogorov–Arnold Network (KAN) for classification. This architecture was combined with a comprehensive feature extraction pipeline and targeted imbalance mitigation techniques. The model was evaluated using a public respiratory sound database comprising six classes with a highly skewed distribution. Strategies such as focal loss, class-specific data augmentation, and Synthetic Minority Over-sampling Technique (SMOTE) are employed to improve minority class recognition. Our results demonstrate that the proposed Hybrid LSTM-KAN model achieves a high overall accuracy of 94.6% and a macro-averaged F 1 -score of 0.703. This performance is notable, given that the dominant class (COPD) constitutes over 86% of the data. While challenges persist for the rarest classes (Bronchiolitis and URTI, with F 1 -scores of approximately 0.45 and 0.44, respectively), the approach shows significant improvement in their detection compared to naive baselines and performs strongly on other minority classes, such as bronchiectasis ( F 1 -score ≈ 0.84). This study contributes to the development of intelligent auscultation tools for the early detection of respiratory diseases, highlighting the potential of combining recurrent neural networks with advanced KAN architectures and focused imbalance handling. • Hybrid LSTM–KAN model proposed for imbalanced respiratory sound classification. • Achieved 94.6% accuracy and 0.703 macro F1-score on the ICBHI 2017 dataset. • Imbalance mitigated effectively using focal loss, SMOTE, and targeted augmentation. • Strong detection of COPD and Bronchiectasis, with challenges in rare classes like URTI and Bronchiolitis. • Provides a clinically relevant step toward automated screening of res- piratory diseases.