Discrete State-Space Embedded Diffusion Model for Epileptic EEG Data Augmentation

Epileptic seizure detection is a challenging task due to the severe imbalance between background activity and seizure EEG data. To address this problem, we propose the DSDM-EEG framework, a Discrete State-space embedded Diffusion Model that is able to generate high-quality and diverse EEG samples. First, a variational autoencoder (VAE) is used to learn the continuous latent representations of background activity in EEG signals, capturing the underlying features of normal brain activity. Next, a vector quantized variational autoencoder (VQ-VAE) module constructs a discrete latent space for seizure signals, enabling efficient representation of abrupt, non-continuous seizure waveforms via a finite set of codebook entries. Then, the extracted continuous and discrete features are embedded into a denoising module through a cross-attention mechanism. Finally, a dynamic denoising augmentation strategy diversifies single-sample generation by adjusting the number of diffusion steps. Experiments on the CHB-MIT dataset demonstrate that DSDM-EEG outperforms mainstream methods in sensitivity, specificity, and accuracy.