A GRAPH CONVOLUTION BASED DUAL-BRANCH NETWORK FOR DECODING OF UPPER-LIMB MOTOR IMAGERY EEG SIGNALS

Analysing motor imagery EEG signals using conventional Euclidean feature computation is challenging due to low spatial resolution, non-stationarity, and complex inter-channel dependencies, which limit robust subject-independent generalization in brain-computer interface systems.

This study proposes a graph-based convolutional framework for subject-independent decoding of upper-limb motor imagery. The model integrates graph convolution with skip connections and handcrafted time-frequency features, together with parallel spatial and temporal convolutional pathways. Graph convolution captures functional connectivity between electrodes, while convolutional modules enhance spatial-temporal representation learning. Subject-independent evaluation is ensured using Stratified Group K-Fold (SGKF) and Leave-One-Subject-Out (LOSO) validation protocols.

The framework was evaluated on the PhysioNet, BCI IV-2a, and BCI IV-2b datasets using multiple electrode configurations. For the PhysioNet dataset, three upper-limb classes were considered, achieving SGKF accuracies of 86.42%, 91.01%, and 93.91% using 64, 21, and 5 electrodes, respectively, and LOSO accuracies of 92.25%, 95.72%, and 96.01%. On the BCI IV-2a dataset, SGKF accuracies of 86.13%, 89.01%, and 92.33% were obtained, with LOSO accuracies reaching 97.12% using five electrodes. For the BCI IV-2b dataset, accuracies of 88.60% (SGKF) and 93.35% (LOSO) were achieved using only three electrodes.

The proposed graph-based framework demonstrates competitive performance to representative graph-based methods while requiring fewer trainable parameters and EEG electrodes.

These findings demonstrate that graph-based deep learning enables accurate, subject-independent motor imagery decoding with reduced model complexity and minimal electrode usage, supporting practical and resource-efficient brain-computer interface systems.