Graph convolutional network-based harmonization of EEG for cross-dataset transfer in motor imagery in BCI

OBJECTIVE: Electroencephalogram (EEG) electrode configurations vary across Motor Imagery Brain-Computer Interface (MI-BCI) datasets, limiting transfer learning and system performance due to small dataset sizes. This work proposes a spatial harmonization framework that maps heterogeneous EEG recordings to a common physical electrode montage while preserving task-relevant motor imagery information. APPROACH: Each EEG trial is modeled as a graph, with electrodes as nodes and electrode samples as node embeddings. A two-layer Graph Convolutional Network (GCN) is introduced to capture spatio-temporal relationships between electrodes and EEG samples. The harmonized EEG is evaluated using time-domain, frequency-domain, and spatial-domain analyses, as well as downstream MI classification with EEGNet, FBCNet, and ADFCNN. Performance is assessed under three protocols-within-dataset classification, source-only cross-dataset transfer, and target-domain fine-tuning-across three public MI EEG datasets. MAIN RESULTS: The proposed GCN yields lower harmonization error than spherical spline interpolation while preserving the principal temporal, spectral, and spatial characteristics of motor imagery EEG. In within-dataset classification, combining real and harmonized EEG improved decoding performance, increasing accuracy from 56.57% to 66.20% for EEGNet and from 61.96% to 72.54% for FBCNet on Dataset A. In cross-dataset experiments, the harmonized representation supported both source-only transfer and fine-tuning, with the combined condition generally yielding the highest performance. SIGNIFICANCE: The proposed method addresses electrode-layout incompatibility at the EEG signal level without restricting datasets to a small subset of shared electrodes. By enabling heterogeneous MI EEG datasets to be represented in a common physical montage, the framework provides a practical basis for signal-level harmonization, dataset augmentation, and cross-dataset MI decoding across diverse recording setups.