A multi-domain graph convolutional network-based prediction model for personalized motor imagery action

Abstract

Motor imagery (MI)-based brain-computer interfaces (BCIs) offer a novel method to decode action imagination. Our previous study demonstrated that actions play a key role in causing individual differences. Cognitive EEG signals showed a positive correlation with MI, reflecting these differences and providing a foundation for predicting suitable MI actions for each individual. This study aimed to propose a multi-domain graph convolutional network (M-GCN) for predicting personalized MI action using cognitive data. The M-GCN extracts time, frequency, and spatial domain features from cognitive tasks to construct multi-domain brain networks using different EEG quantization methods according to the characteristics of the three domains. Subsequently, the M-GCN utilizes spectral GCN to learn the topology relationship between EEG channels by analyzing functional connection strength. Finally, for each action, the M-GCN can accurately map cognitive data to the corresponding MI action and output a personalized action for each subject. A subject-independent decoding paradigm with leave-one-subject-out cross-validation is adopted to validate the model on ten subjects. Compared to baseline and single-domain models, the M-GCN achieves the highest prediction accuracy of 73.60% (p = 7.1 × 10−3), improving by 15.87% (p = 2.0 × 10−4) and by 7.2% (p = 4.0 × 10−4), respectively. This study proves that the M-GCN can precisely predict personalized MI actions, reflecting the efficiency of the multi-domain feature fusion based on cognitive tasks and GCN and offering a novel method for personalized BCI.