SemSTNet: Medical EEG Semantic Metric Learning With Class Prototypes Generated by Pretrained Language Model

Electroencephalography (EEG) feature learning is crucial for brain-machine interfaces and medical diagnostics. Existing deep learning models for classification often overlook the intrinsic semantic relationships between different EEG classes and rely on overly complex models with a large number of parameters. To address these challenges, we propose SemSTNet, a novel and lightweight framework for EEG analysis. Firstly, we designed an efficient, lightweight convolutional architecture that decouples spatial and temporal feature extraction. Then we propose a framework which introduces a novel semantic metric learning paradigm that uses class prototypes generated by a pretrained language model to better capture inter-class relationships and enhance intra-class compactness. These prototypes are extracted and stored offline, requiring no additional inference from the language model during training or deployment. This design significantly reduces model complexity, resulting in a model with only 23 K parameters-over 100 times fewer than common Transformer-based models. Extensive experiments demonstrate that SemSTNet outperforms state-of-the-art approaches on tasks such as epilepsy classification and sleep staging, highlighting its effectiveness and efficiency. Our work demonstrates that integrating semantic knowledge with a purpose-built lightweight architecture provides a highly effective and efficient solution.