Generative AI with WGAN-GP for boosting seizure detection accuracy.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Biomedical Signal Processing

Background:

• Imbalanced electroencephalogram (EEG) datasets present significant challenges for developing accurate automated epileptic seizure detection systems.
• Generative Artificial Intelligence (AI) offers a potential solution by augmenting minority class data to improve model performance.

Purpose of the Study:

• To investigate the effectiveness of various Generative Adversarial Network (GAN) variants for data augmentation in seizure detection.
• To evaluate the performance of a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) integrated with a bidirectional Long Short-Term Memory (LSTM) architecture.

Main Methods:

• Exploration of multiple GANs (WGAN-GP, Vanilla GAN, Conditional GAN, Cramer GAN) for data augmentation with Random Forest classifiers.
• Integration of the best-performing GAN variant (WGAN-GP) with a bidirectional LSTM network.
• Comparison of the proposed generative AI approach against traditional and synthetic oversampling methods.

Main Results:

• WGAN-GP demonstrated superior performance as a data augmentation technique compared to other GAN variants when used with Random Forest classifiers.
• The WGAN-GP and bidirectional LSTM combination achieved a 91.73% accuracy on augmented imbalanced EEG data, outperforming the 86% accuracy on non-augmented data.
• The generative AI approach significantly improved performance metrics crucial for reliable seizure detection.

Conclusions:

• The WGAN-GP generative AI technique, when combined with bidirectional LSTM, effectively enhances seizure detection accuracy for imbalanced EEG datasets.
• This advanced approach surpasses traditional oversampling and class weight adjustment methods, offering a promising tool for epilepsy monitoring and management.
• The study highlights the potential of generative AI in improving automated seizure detection systems.