Leveraging Spectral Signatures and Attention Enhanced BiLSTM Network for Comprehensive Respiratory Sound Analysis

Respiratory conditions, including pneumonia, bronchitis, asthma, and chronic illnesses like Chronic Obstructive Pulmonary Disease (COPD) significantly impact human health, and early detection is crucial for effective treatment. Conventional diagnostic techniques like X-rays and CT scans frequently encounter challenges, such as high costs, lengthy procedures, and limited availability in underserved regions. Although existing approaches commonly use convolutional and recurrent neural networks for automated diagnosis, this study proposes a novel method using a bidirectional long- and short-term memory (BiLSTM) network paired with an attention mechanism to detect and classify respiratory sounds into eight distinct disease categories.The attention mechanism improves classification accuracy by emphasizing critical temporal regions within the respiratory sound signals, ensuring the model focuses on diagnostically relevant features.To address class imbalance and limited dataset size, advanced data augmentation techniques, including the GriffinLim algorithm, were employed to generate high-fidelity, diverse audio samples, thereby balancing the representation of disease classes. The proposed BiLSTM-Attention model achieved an impressive classification accuracy of $96 \%$ and an F1 score of $96 \%$ across eight classes, highlighting its effectiveness as a reliable and accessible tool for the early detection of various lung diseases.