Skip to content
DSML.KZ
Publications

Article · IEEE Access · January 1, 2026

HSST-EEG: A Hybrid State-Space and Transformer Architecture for EEG Decoding

Emil Kim, Jin Kyu Gahm

Emil Kim
Open via DOI

Abstract

Accurate interpretation of electroencephalography (EEG) signals requires effective modeling of high-dimensional, non-stationary, and long-range spatio-temporal dynamics, which remains a central challenge in both clinical and BCI-oriented decoding tasks. Recent approaches that combine convolutional neural networks (CNNs) or explicit spectral feature extractors with downstream sequence models have improved EEG decoding performance. However, such pipelines remain limited in their ability to jointly capture long-range temporal dependencies and dynamic cross-channel interactions while maintaining computational efficiency. Transformer-based architectures capture pairwise temporal and spatial features via self-attention, enabling flexible long-range dependency modeling but incurring quadratic computational complexity. In contrast, Mamba-based state-space architectures achieve favorable linear-time complexity by modeling temporal dependencies implicitly through shared state evolution, which may limit their capacity to represent complex spatio-temporal EEG structures compared to similarly sized Transformer models. To address these limitations, we propose HSST-EEG (Hybrid State-Space Transformer for EEG Decoding), a hybrid architecture that integrates Mamba-based state-space modeling with Transformer-based attention to balance computational efficiency and representational capacity. Experiments conducted on two widely used clinical EEG benchmarks, the CHB-MIT Scalp EEG Database and the TUH EEG Events Corpus, demonstrate that HSST-EEG achieves up to a 5% improvement in classification accuracy compared to recent EEG decoding baselines.

EEG and Brain-Computer InterfacesEmotion and Mood RecognitionECG Monitoring and AnalysisEEG decodingMambastate-space model

Related publications