Conditional GAN based augmentation for predictive modeling of respiratory signals.

Area of Science:

• Medical imaging and diagnostics
• Artificial intelligence in healthcare
• Signal processing for respiratory analysis

Background:

• Respiratory illnesses pose a significant global health burden, leading to mortality and morbidity.
• Current diagnostic methods for respiratory disorders often lack optimal accuracy and vary between experts.
• Deep learning models require extensive datasets, which are limited for respiratory data.

Purpose of the Study:

• To address the limited dataset challenge in respiratory signal analysis using data augmentation.
• To evaluate the efficacy of conditional generative adversarial networks (cGAN) for generating synthetic respiratory signals.
• To improve the accuracy of respiratory disorder classification using deep learning models.

Main Methods:

• Utilized conditional generative adversarial networks (cGAN) for data augmentation (DA) to create synthetic respiratory signals.
• Employed publicly available datasets (ICBHI 2017, RALE, Think Labs Lung Sounds Library) for training and validation.
• Assessed signal similarity between original and augmented data, and classified signals using scalogram representations fed into pre-trained deep learning models (Alexnet, GoogLeNet, ResNet-50).

Main Results:

• Conditional generative adversarial networks (cGAN) effectively generated synthetic respiratory signals.
• Similarity measures confirmed the efficacy of the data augmentation approach.
• Classification using ResNet-50 with cGAN-augmented data achieved high accuracy rates of 92.50% and 92.68% on two datasets.

Conclusions:

• The proposed cGAN-based data augmentation method significantly enhances the performance of deep learning models for respiratory signal classification.
• This approach offers a promising solution for improving the accuracy and reliability of diagnosing respiratory disorders, especially with limited data.
• The study demonstrates the potential of advanced AI techniques to overcome data scarcity in medical diagnostics.