一种针对手臂精细动作想象的细粒度图神经分类模型OA

Existing methods face challenges in fully capturing the complex spatiotemporal coupling features of motor imagery EEG signals from fine arm movements,and there remains room for improvement in their decoding precision and accuracy.Fo-cusing on these challenges,this study completed the construction of an EEG database for arm motor imagery,and proposed a fine-grained graph convolutional classification model for motor imagery EEG signals.The model embedded a self-attention mechanism to focus on core electrode channels,and accurately captured deep dependencies between nodes to suppress redun-dant interference.This study designed a fine-grained graph neural module.Through time window segmentation and multi-scale subgraph construction,the module alleviated the limitation of graph neural networks that tend to prioritize spatial information over temporal characteristics.The adopted convolutional classification layer efficiently completed the integration and discrimina-tion of spatiotemporal features.This study conducted comparative experiments based on the self-built arm movement imagery dataset and public datasets.Experimental results show that the model achieves favorable classification performance compared with multiple mainstream baseline classification models,with a maximum within-subject accuracy of 76％and a maximum cross-subject accuracy of 59％on the arm movement dataset.These results confirm that the proposed model architecture is ef-fective,provide a new perspective for the spatiotemporal joint modeling of motor imagery in complex EEG signals,and offer technical support for the development of high-performance brain-computer interface systems.