STHMA: Decoupling Spatio-Temporal Dynamics in EEG via Hybrid State Space Modeling.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Decoding affective states from Electroencephalography (EEG) is crucial for Brain-Computer Interfaces.
• Non-stationary physiological signals and complex spatio-temporal dynamics in EEG hinder accurate emotion recognition.
• Existing Transformer models struggle with neural signal dynamics and computational complexity.

Purpose of the Study:

• To propose a novel framework, Spatio-Temporal Hybrid Mamba-Attention (STHMA), for enhanced EEG-based emotion recognition.
• To explicitly disentangle and model the complex dynamics of EEG signals using linear-complexity State Space Models.
• To improve the accuracy and scalability of affective state decoding from neural signals.

Main Methods:

• Introduced a Dual-Domain Physics-Aware Embedding module fusing temporal convolutions and spectral features for neurophysiological fidelity.
• Developed a Decoupled Spatial-Temporal Scanning strategy to separate functional connectivity learning from emotional state evolution tracking.
• Utilized linear-complexity State Space Models to efficiently handle high-dimensional EEG data.

Main Results:

• The STHMA framework achieved state-of-the-art performance on the FACED and SEED-V datasets.
• Significantly outperformed random chance baselines, demonstrating robust emotion recognition capabilities.
• Validated the effectiveness of the proposed embedding and scanning strategies in capturing EEG dynamics.

Conclusions:

• Combining Physics-Aware Embeddings with decoupled state-space modeling provides a scalable and effective paradigm for EEG emotion recognition.
• The STHMA framework offers a promising advancement for non-invasive Brain-Computer Interfaces.
• The proposed methods successfully address the challenges of non-stationarity and spatio-temporal entanglement in EEG signals.