Novel deep learning-based optimization framework for the classification of respiratory diseases using lung sound analysis

Respiratory disease classification using lung sound analysis offers a non-invasive and cost-effective diagnostic alternative to traditional imaging-based methods. This study proposes a deep learning-based optimization framework for automated respiratory disease classification using the publicly available ICBHI 2017 Respiratory Sound Database (Kaggle version), comprising 920 recordings from 126 subjects with annotated respiratory cycles. The collected audio signals are preprocessed using band-pass filtering to suppress noise and irrelevant frequency components. A Denoising Autoencoder (DAE) is employed to extract compact latent representations from the preprocessed signals. The classification stage utilizes an Enhanced Bidirectional Long Short-Term Memory (EBiLSTM) network incorporating residual connections, gate-level normalization, and dropout regularization for improved temporal modeling. Hyperparameters are optimized using the Average and Subtraction-Based Optimizer (ASBO) to maximize classification accuracy. The proposed EBiLSTM-ASBO model achieves statistically significant improvements over representative baselines (FFT-PCR, RFINCA, HST, and RDCNN), attaining an overall accuracy improvement of up to 18.51% with consistent gains in precision, recall, and MCC. Statistical hypothesis testing confirms the robustness of the performance improvements (p < 0.05). The results demonstrate that the proposed framework effectively captures disease-specific acoustic patterns for reliable multi-class respiratory disease classification.