BDAN: Mitigating Temporal Difference Across Electrodes in Cross-Subject Motor Imagery Classification via Generative Bridging Domain

TL;DR

BDAN tackles cross-subject MI classification by addressing temporal-electrode data distribution shifts across sessions and electrodes. It combines a three-block spatial feature extractor with a generative bridging domain and two-stage adaptation, governed by four bridging losses and a compact matrix-arithmetic framework, to align source and target distributions in a bridging space. Evaluated on BCIC-III-IVa and BCIC-IV-2a, BDAN outperforms deep learning and global-domain adaptation baselines, with ablations confirming the necessity of both adaptation stages and the Gaussian bridging perturbation. The approach yields robust temporal and electrode-domain alignment, enabling more reliable unsupervised target-domain classification and offering a path toward electrode-agnostic cross-subject MI decoding.

Abstract

Because of "the non-repeatability of the experiment settings and conditions" and "the variability of brain patterns among subjects", the data distributions across sessions and electrodes are different in cross-subject motor imagery (MI) studies, eventually reducing the performance of the classification model. Systematically summarised based on the existing studies, a novel temporal-electrode data distribution problem is investigated under both intra-subject and inter-subject scenarios in this paper. Based on the presented issue, a novel bridging domain adaptation network (BDAN) is proposed, aiming to minimise the data distribution difference across sessions in the aspect of the electrode, thus improving and enhancing model performance. In the proposed BDAN, deep features of all the EEG data are extracted via a specially designed spatial feature extractor. With the obtained spatio-temporal features, a special generative bridging domain is established, bridging the data from all the subjects across sessions. The difference across sessions and electrodes is then minimized using the customized bridging loss functions, and the known knowledge is automatically transferred through the constructed bridging domain. To show the effectiveness of the proposed BDAN, comparison experiments and ablation studies are conducted on a public EEG dataset. The overall comparison results demonstrate the superior performance of the proposed BDAN compared with the other advanced deep learning and domain adaptation methods.

Figures (7)

• Figure 1: Schematic of the proposed temporal-electrode data distribution difference problem. A, Factors of the temporal-electrode data distribution difference problem, where the two main factors are illustrated diagrammatically. B, Electrode data distribution difference across sessions, where the proposed temporal-electrode data distribution difference is diagrammatically presented.
• Figure 2: Visualization of temporal-electrode data distribution problem and solutions. A, Electrode data distribution across sessions from different subjects, where each electrode ball represents the data distribution of one electrode, the colour of the electrode ball indicates the corresponding subject and electrode, and the arrows denote the data distribution differences across sessions. B, Computation of the generative bridging domain, where the bridging domain is computed via all the features from the source and target subjects.
• Figure 3: Bridging domain adaptation network (BDAN), where the blue and orange arrows represent the data flows from the source and target domains respectively, and the purple arrow represents the data flow of the generative bridging domain. A, Training stage of the BDAN, where bridging domain adaptation is adopted between source and target data, minimising the temporal-electrode data distribution difference and transferring the known spatio-temporal knowledge from the source domain to the target domain. B, Testing stage of the BDAN, where unsupervised classification is adopted on the target domain via the trained model, automatically reducing the time-related electrode data distribution difference and improving the classification performance.
• Figure 4: Transformation operations of deep features. A, "Permute" and "Reshape" transformation operations, where the 4D-tensor with shape $n_{s} \times f_{s} \times e_{s} \times p_{s}$ is transformed into 3D-tensor with shape $e_{s} \times n_{s} \times t_{s}$ via "Permute" and "Reshape" operations. B, "Expand" transformation operation, where the 3D-tensor with shape $e_{s} \times n_{s} \times t_{s}$ is expanded into 4D-tensor with shape $1 \times e_{s} \times n_{s} \times t_{s}$.
• Figure 5: Performance and results of the proposed BDAN and the other algorithms. A-B, comparison experiments. C-D, ablation studies.