Hypercube-S4GNN: A Multi-Edge Graph Approach Using State Space Models on Multivariate EEG for Seizure Detection

Electroencephalogram (EEG) signals contain long contextual information and complex spatial connections across its multivariates, making its spatiotemporal interdependence challenging to explain. Deep learning algorithms have demonstrated the ability to accurately predict seizures from EEG data. We propose Hypercube-S4GNN, which combines Structured State Spaces for Sequence Modelling (S4) with Graph Neural Networks (GNN) to effectively capture both the intra-series (temporal) and inter-series (spatial) dependencies from EEG. Our study combines temporal and spatial modeling using diverse graph connectivity measures, including distance, correlation, Directed Transfer Function (DTF), and coherence. The study examines each graph connectivity and their combinations as a hypercube of edge features, emphasizing the synergistic benefits that can be achieved by integrating distinct graph connectivity with S4 embeddings for improved seizure detection. Our proposed framework outperforms baseline models substantially through a comprehensive set of experiments conducted on a benchmark dataset in seizure analysis. By employing State Space Models (SSMs) and graph convolution, the model is capable of discerning spatial and temporal dependencies in EEG data, presenting a valuable contribution to the evolving landscape of multivariate time series (MTS) forecasting.