CGM-EEG: Cross-Gated Mamba for Spatio-Temporal Eeg Representation Learning

Abstract

Accurate decoding of electroencephalography (EEG) signals is essential for advancing clinical diagnostics and brain-computer interface (BCI) applications. Although sequence-based models have shown promise in capturing long-range dependencies, they often struggle to represent the intricate spatio-temporal structure of EEG data. To address this limitation, we propose CGM-EEG (Cross-Gated Mamba for EEG Decoding), a fully Mamba-based dual-branch architecture that processes spatial and temporal features in parallel. To enhance feature interaction, our framework introduces a Cross-Gate Module (CGM) that enables each branch to integrate information from the other through a lightweight crossgated mechanism. Extensive experiments on three public clinical EEG benchmarks demonstrate that CGM-EEG achieves up to a 6.6 % higher balanced accuracy and a 7.1 % lower inference latency than recent transformer-based models. These results highlight the potential of our approach as an efficient and scalable foundation for real-time EEG decoding in biomedical and neuroimaging applications.

Related publications

• MARBLE: Lightweight EEG-to-fMRI Translation via Mamba-Attention and ROI-Conditioned DecodingLecture notes in computer science · 2026 · Same author: Emil Kim
• Multi-Scale EEG Feature Extraction with Attention Mechanism for Seizure DetectionJournal of KIISE · 2026 · Same author: Emil Kim
• HSST-EEG: A Hybrid State-Space and Transformer Architecture for EEG DecodingIEEE Access · 2026 · Same author: Emil Kim
• Compact convolutional transformer for subject-independent motor imagery EEG-based BCIsScientific Reports · 2024 · Citations: 30 · Shared topic: EEG and Brain-Computer Interfaces