Augmented reality paradigm and EA-DGCNN decoding algorithm for motor imagery in stroke patients

Stroke represents one of the leading causes of long-term motor dysfunction on a global scale. Existing motor imagery (MI) brain-computer interface (BCI) rehabilitation training systems face critical challenges of suboptimal patient engagement and limited electroencephalography (EEG) feature decoding capacity. To overcome these challenges, this study presents two primary original contributions: an independently designed wearable augmented reality (AR) MI training paradigm and a novel Euclidean Alignment-based Dynamic Graph Convolutional Neural Network (EA-DGCNN) end-to-end decoding architecture.

The immersive AR-MI paradigm heightens patient engagement and enhances MI performance. The EA-DGCNN model incorporates Euclidean alignment to minimize inter-subject variability in EEG distributions. The model extracts multi-dimensional spatiotemporal features via convolutional block attention, dynamic graph attention, and multi-scale convolutions, incorporating a Broad Learning System to achieve precise classification results.

Experimental results demonstrate that the model achieves an accuracy of 85.13 ± 10.53% on the public BCI Competition IV 2a dataset. On a custom clinical dataset of 20 stroke patients, the AR paradigm yields a classification accuracy of 71.9 ± 8.9%, outperforming the 68.9 ± 8.0% of the traditional 2D paradigm. Furthermore, evaluations on the 64-channel PhysioNet benchmark demonstrate robust cross-subject transferability with a superior grand-average accuracy of 83.67 ± 8.52%. Neurophysiological feature analysis indicates that the AR paradigm elicits more pronounced event-related desynchronisation and event-related synchronisation responses. The t-SNE visualisation demonstrates that the AR paradigm, coupled with the proposed model, produces highly discriminative feature clustering patterns.

This study delivers a more effective paradigm and a robust decoding solution for clinical BCI rehabilitation.