Scalogram based prediction model for respiratory disorders using optimized convolutional neural networks.

Area of Science:

• Medical Diagnostics
• Signal Processing
• Artificial Intelligence

Background:

• Lung auscultation is crucial for diagnosing respiratory conditions like COPD.
• Previous analyses of lung sounds relied on wavelet transforms or spectrograms.
• Accurate prediction models for respiratory disorders remain a challenge.

Purpose of the Study:

• To develop an optimized Convolutional Neural Network (CNN) architecture for accurate respiratory disorder prediction.
• To model respiratory sound signals using empirical mode decomposition (EMD) and scalogram representations.
• To evaluate the performance of the proposed CNN model against baseline methods.

Main Methods:

• Respiratory sound signals were decomposed into intrinsic mode functions (IMFs) using EMD.
• Bump and Morse scalograms were generated from IMFs to represent signal energy.
• A pre-trained, optimized Alexnet CNN was utilized for classification of lung sound datasets.
• Stochastic gradient descent with momentum (SGDM) and adaptive data momentum (ADAM) optimizers were employed.

Main Results:

• The proposed CNN model achieved a validation accuracy of 83.78%.
• This represents a significant improvement over standard Bump and Morse wavelet transform methods (79.04% and 81.27%).
• The scalogram representation of EMD-derived IMFs enhanced prediction accuracy.

Conclusions:

• The optimized Alexnet CNN with EMD-based scalograms offers a highly effective approach for respiratory disorder prediction.
• This method demonstrates superior performance compared to existing state-of-the-art techniques.
• The findings pave the way for more accurate and reliable non-invasive respiratory diagnostics.