Electrode Domain Adaptation Network: Minimizing the Difference Across Electrodes in Single-Source to Single-Target Motor Imagery Classification

Because of electrode positioning error and brain nonlinear dynamics, the data distributions across electrodes are different in motor imagery (MI) study, eventually reducing the MI classification performance. In this paper, a novel inter-electrode data distribution problem is systematically illustrated and summarised for both intra-subject and inter-subject scenarios. To deal with the electrode data distribution difference problem, a novel electrode domain adaptation network (EDAN) is proposed, aiming to improve classification accuracy and enhance model robustness. Specifically, deep features by electrode from raw MI data are extracted by a specially designed spatial-temporal convolutional neural network (CNN). Then, with the customized intra-subject and inter-subject electrode loss functions, the electrode domain adaptation is conducted on the deep features to reduce the data distribution difference across electrodes. The comparison experiments and ablation studies of STS MI classification are conducted on a public EEG dataset to show the effectiveness of the proposed EDAN. The visualization of the deep features intuitively demonstrates the effectiveness of the electrode domain adaptation compared with global and class domain adaptations. The overall comparison results demonstrate the proposed EDAN can handle the inter-electrode difference and improve the classification accuracy compared with the other advanced deep learning and domain adaptation methods.