HST-SN: A Hierarchical Framework Learning Spatiotemporal Semantic Features From EEG Signals

Electroencephalography (EEG) has become increasingly pivotal in clinical diagnostics, cognitive psychology, brain computer interfaces and other cognitive tasks. Due to the data variability across different cognitive tasks, how to design the general-purpose EEG feature extractor which is effective for different tasks remains challenging. Building on neuropsychological foundations that all cognitive tasks depend on timed interactions between specialized brain regions, a framework Hierarchical Spatio-Temporal Semantic feature extraction Network (HST-SN) is proposed to extract the spatio-temporal features within and among the semantic units, which refers to temporally or spatially discrete segments of EEG signals that carry semantic information - for example, a brain region with specific function or a brief temporal segment in cognitive process. The proposed method could not be limited by specific application scenarios while still maintaining high decoding performance. Evaluated on four distinct cognitive tasks, including attention, fatigue, preference and motor imaginary tasks, the new model outperforms other state-of-the-art (SOTA) general-purpose methods, demonstrating strong generalization capability and robustness. In addition, a set of comparative experiments from both temporal and spatial perspectives were conducted to validate its high efficiency. By enhancing the practicality of EEG measurement and decoding, this work facilitates the deployment of brain-computer interfaces across diverse real-world scenarios.