Epilepsy is a common neurological disorder affecting over 50 million people worldwide, characterised by recurrent seizures accompanied by abnormal neuronal electrical activity. Electroencephalogram (EEG) is a technique for recording brain electrical signals, widely employed for epileptic seizure (ES) prediction due to its high temporal resolution, portability, and cost-effectiveness. However, reliable ES prediction based on EEG remains challenging, primarily owing to the limited duration of recorded pre-ictal states in publicly available datasets and the typically low signal-to-noise ratio (SNR) in non-invasive recordings. To mitigate these issues, we propose a Conditional Deep Convolutional Generative Adversarial Network (CDCGAN), which combines the representational power of Deep Convolutional Generative Adversarial Network (DCGAN) with the categorical conditioning mechanism of Conditional Generative Adversarial Network (CGAN) to generate class-specific EEG samples. By synthesising target samples, CDCGAN aims to alleviate class imbalance and enhance the quality of low-resolution spectral representations. To evaluate the practical utility of generated data, we trained a Convolutional Neural Network (CNN) on the augmented dataset and compared its performance against prior studies. Under the Leave-One-Seizure-Out cross-validation (LOSO-CV) protocol, our method achieved an average AUC of 0.876 at a 60% augmentation rate with 50 training epochs. The AUC improvement relative to corresponding control settings demonstrates that GAN-based data augmentation provides additional effective training samples for ES prediction while preserving task-relevant and discriminative pre-ictal EEG features.
Huang et al. (Thu,) studied this question.