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

Area of Science:

• Medical Diagnostics
• Artificial Intelligence in Healthcare
• Signal Processing

Background:

• Chronic obstructive pulmonary disease (COPD) presents a significant global health challenge, necessitating advanced diagnostic tools.
• Accurate and efficient detection of COPD is crucial for timely and effective patient treatment.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework for COPD detection using lung sound (LS) signal classification.
• To explore the efficacy of various time-frequency representation (TFR) methods and deep neural network (DNN) architectures for this task.

Main Methods:

• Lung sound signals were transformed into time-frequency representations (TFRs) using methods like Continuous Wavelet Transform (CWT).
• Pre-trained Convolutional Neural Networks (CNNs) including EfficientNet-B0, ShuffleNet-V2, MobileNet-V2, and GhostNet were trained on TFR images.
• Variational Mode Decomposition (VMD) was employed for multiscale analysis, decomposing LS signals into intrinsic mode functions (IMFs).

Main Results:

• The CWT-based TFR combined with EfficientNet-B0 demonstrated superior performance, achieving 98.57% accuracy on the ICBHI 2017 dataset and 95.10% on the Fraiwan dataset.
• Model interpretability was assessed using t-distributed stochastic neighbor embedding (t-SNE) plots.
• Statistical analysis using the Kruskal-Wallis test was performed on multiscale deep features.

Conclusions:

• The proposed TF-based DNN framework offers a highly accurate and efficient method for COPD detection from lung sounds.
• The model's deployment on edge devices enables real-time inference, facilitating practical clinical application.
• This approach holds promise for improving the diagnosis and management of chronic obstructive pulmonary disease.