Synthesizing Epileptic Seizures: Gaussian Processes for EEG Generation

Technical Summary: Synthesizing Epileptic Seizures: Gaussian Processes for EEG Generation

Problem Statement
Reliable seizure detection from electroencephalography (EEG) is a critical clinical goal, yet the performance of machine learning methods is frequently constrained by the scarcity, high acquisition costs, and heterogeneity of labeled EEG data. This challenge is compounded by the intrinsic nature of epileptic EEG: it is high-dimensional, non-stationary, and exhibits long-range temporal dependencies. While deep generative models (e.g., GANs, VAEs, diffusion models) have been applied to synthetic EEG generation, they often struggle with long-horizon synthesis, fail to scale to continuous multivariate recordings, and lack interpretability. Furthermore, existing methods typically target short segments or limited channel subsets, failing to preserve the complex spatio-temporal evolution and inter-channel correlations required for clinically realistic seizure modeling.

Methodology: GP-EEG Framework
The authors propose GP-EEG, an interpretable, hierarchical framework for generating patient-specific, long-horizon, multivariate epileptic EEG recordings. The pipeline consists of six distinct stages:

1. Dimensionality Reduction: All recordings for a specific patient are concatenated and decomposed via Singular Value Decomposition (SVD). This separates the data into low-dimensional temporal scores (capturing shared dynamics) and spatial loadings (preserving patient-specific inter-channel structure).
2. Changepoint Detection and GP Regression: The temporal scores are segmented into quasi-stationary (QS) regimes using offline changepoint detection (combining KPSS and ADF tests). Within each regime, a Gaussian Process (GP) regression model is fitted using a quasi-periodic kernel (combining Matérn-3/2 and periodic kernels) to capture local temporal structure with uncertainty quantification.
3. Domain Adaptation: A Convolutional Long Short-Term Memory Variational Autoencoder (Conv-LSTM VAE) is trained to map the raw GP-generated surrogate data to the distribution of real EEG. This step, termed "EEGification," refines spectral content and waveform morphology while preserving the long-range temporal and spatial dependencies established in previous stages.
4. Kernel-State Discretization and Transition Modeling: The continuous GP hyperparameters learned across regimes are discretized into a finite set of kernel states via agglomerative clustering. A discrete Markov chain is then trained to model the transition probabilities between these states.
5. Point-Process Modeling for Changepoints: The timing of regime changes is modeled using an inhomogeneous Poisson process fitted to the observed changepoint locations, providing a principled distribution for sampling synthetic changepoint times.
6. Synthetic Generation and Reconstruction: To generate new data, the model samples a sequence of changepoints from the Poisson process and a sequence of kernel states from the Markov chain. GP samples are drawn for each regime, concatenated to form synthetic temporal scores, and projected back to the full channel space using the patient-specific spatial loadings. Finally, the Conv-LSTM VAE is applied to produce the final synthetic EEG recording.

Key Contributions
The paper outlines four primary contributions:

1. Interpretable Framework: Introduction of GP-EEG, a pipeline that combines low-dimensional compression with explicit regime-based temporal modeling to generate patient-specific, long-horizon multivariate EEG.
2. Preservation of Spatial Structure: By generating temporal scores and reconstructing full-montage recordings via patient-specific spatial loadings, the method maintains realistic inter-channel dependencies without relying on end-to-end learning of high-dimensional spatial correlations.
3. Tractable Long-Horizon Synthesis: The coupling of changepoint-based segmentation with regime-wise GP modeling and an explicit generative model of regime dynamics (Markov chains and Poisson processes) enables scalable synthesis with stage-wise interpretability and control.
4. Open Source: The authors commit to releasing the code following anonymization.

Experimental Results
The method was validated on two open-source datasets: CHB-MIT (pediatric epilepsy) and Siena (adult epilepsy).

• Quantitative Evaluation: GP-EEG outperformed three deep learning baselines (TimeVAE, COSCI-GAN, and ImagenTime) across four metrics: Marginal Distribution Difference (MDD), Autocorrelation Difference (ACD), Skewness Difference (SD), and Kurtosis Difference (KD). Notably, GP-EEG achieved the lowest ACD (0.341 on Siena, 0.304 on CHB-MIT), indicating superior preservation of temporal dynamics.
• Qualitative Evaluation: t-SNE embeddings and Kernel Density Estimates (KDE) showed that GP-EEG generated samples with the closest overlap to real data in both embedding space and amplitude distribution.
• Downstream Utility: In a "Train-on-Synthetic, Test-on-Real" (TSTR) scenario, GP-EEG demonstrated significantly better performance than baselines, which largely failed to learn meaningful seizure features. In data augmentation experiments, adding GP-EEG synthetic data to real training sets improved seizure detection metrics (Accuracy, Precision, Recall, F1) compared to baseline models trained only on real data.

Significance and Claims
The authors claim this is the first study to produce interpretable, multivariate, non-stationary, long-horizon EEG seizure data. They emphasize that their approach addresses the computational and data-scarcity constraints of end-to-end deep learning by leveraging classical mathematical tools (SVD, GPs) combined with modern components (VAEs). The framework is designed to support methodological research in data-scarce clinical settings by enabling the augmentation, stress-testing, and benchmarking of learning algorithms. The authors explicitly state that the intended use is for research and development rather than direct clinical decision-making, and they caution against the over-interpretation of synthetic data as clinically valid evidence without further validation.