Time-frequency based novel deep neural networks for COPD detection from lung sounds on edge device

Chronic obstructive pulmonary disease (COPD) is a global health problem, which requires accurate and efficient diagnostic methods to enable timely and effective treatment. In this study, we propose a novel method for COPD detection that uses time-frequency (TF)-based deep neural networks (DNNs) for lung sound (LS) signal classification. We utilized several TF representation (TFR) methods, including continuous wavelet transform (CWT), Wigner-Ville distribution (WVD), and smoothed pseudo WVD (SPWVD) to represent the LS signals in the TF plane. In the next step, the TFR images are used as input to train pretrained convolutional neural networks (CNNs), namely, EfficientNet-B0, ShuffleNet-V2, MobileNet-V2, and GhostNet for COPD detection. We have also performed the multiscale analysis of the proposed framework by employing the variational mode decomposition (VMD) method. The VMD decomposes LS signals into intrinsic mode functions (IMFs) that facilitate multiscale TFR, which serves as input to the DNN models for COPD detection. The proposed method's performance is evaluated on publicly available ICBHI 2017 and Fraiwan datasets utilizing different performance metrics. In the proposed framework, CWT-based TFR with EfficientNet-B0 achieves superior performance over other models, attaining the highest accuracies of 98.57% and 95.10% in detecting COPD on ICBHI 2017 and Fraiwan's datasets, respectively. The interpretability of the proposed framework is demonstrated using t-distributed stochastic neighbor embedding (t-SNE) plots considering all the TFR methods and CNN models. Furthermore, a statistical analysis of the multiscale deep features is performed using the Kruskal-Wallis statistical test. Finally, the model is deployed as a standalone system on an edge device, enabling real-time inference and classification.