Automated lung sound detection via Bi-GRU-modified SqueezeNet architecture with new stock well feature set.

Lung sound analysis is critical for diagnosing respiratory diseases such as asthma, bronchiectasis, bronchiolitis, COPD, LRTI, pneumonia, and URTI. Traditional diagnostic methods rely heavily on physicians' expertise, making them time-consuming and subjective. To address these limitations, this study introduces a novel deep learning-based model, Bidirectional-Gated Recurrent Unit-Modified SqueezeNet (BGRMSNet), for automated lung sound detection and classification. The proposed approach consists of four key phases: preprocessing, feature extraction, data augmentation, and detection. In the preprocessing stage, a Threshold-based Wiener Filtering (T-WF) technique effectively removes impulse noise and outliers. The feature extraction phase captures comprehensive frequency-domain characteristics using permutation entropy, Modified Stockwell Transform (MST), Short-Time Fourier Transform (STFT), spectral centroid, and spectral rolloff. These features are further enhanced through random sampling-based data augmentation to improve model robustness.The detection phase employs the BGRMSNet architecture, which integrates Bidirectional Gated Recurrent Units (Bi-GRU) for modeling temporal dependencies and a Modified SqueezeNet (MSNet) for efficient feature extraction. MSNet incorporates enhancements including Improved Batch Normalization (IBN), multi-head attention, dropout, dense layers, and an improved exponential Softmax activation function. The combined architecture allows BGRMSNet to capture both temporal and spatial features effectively. Comprehensive evaluations, including ablation studies, statistical analysis, and k-fold cross-validation, demonstrate the model's high performance. The BGRMSNet model achieved an accuracy of 0.970, specificity of 0.987, and negative predictive value (NPV) of 0.972, outperforming conventional diagnostic approaches. These results highlight the potential of BGRMSNet as a robust and accurate tool for automated lung disease detection, supporting enhanced diagnostic decision-making in clinical environments.