Learning feature-wise temporal-spatial gating with Kolmogorov-Arnold networks for EEG seizure detection

This is the authors' abstract. We don't add key points for this paper.

Automated seizure detection from electroencephalogram (EEG) requires modelling how brain signals evolve over time and how different brain regions interact. Most deep learning approaches fuse these streams using fixed, uniform operations that treat all feature dimensions equally, which may not reflect the heterogeneous interplay of temporal and spatial activity during seizures. We propose a framework that learns a separate gating weight per feature dimension through a Kolmogorov-Arnold Network (KAN), applied to embeddings from a Bi-Mamba temporal encoder and a hypergraph convolutional spatial encoder. On the TUSZ benchmark, the model achieves AUROC 0.923, AUPRC 0.743, and F1-score 68.3%, with consistent improvements over unimodal baselines and symmetric fusion strategies across all ablation metrics. Evaluation on CHB-MIT with independent retraining yields AUROC 0.888, demonstrating architectural transferability across recording protocols and electrode configurations. An interpretability analysis shows that ictal activity sharpens temporal attention across electrodes, particularly over right frontal-central regions, increases representational divergence between encoder streams, and is associated with higher gate weighting on a specific subset of latent feature dimensions.