EEG Decoding for Datasets with Heterogenous Electrode Configurations using Transfer Learning Graph Neural Networks

TL;DR

The paper tackles the problem of cross-dataset heterogeneity in EEG-based motor imagery decoding, where differing electrode layouts and channel counts hinder generalization. It introduces an end-to-end framework that combines dataset-specific encoders, graph neural networks for spatial modeling, and a shared latent alignment module driven by an MDD loss to perform supervised domain adaptation across multiple small datasets. Empirical results on three MI EEG datasets show superior accuracy and stability compared with baselines, demonstrating effective integration of heterogeneous data without discarding non-common channels. This approach enhances generalization for BMI applications and offers a pathway to apply similar GNN+transfer-learning strategies to other biosignals with variable sensor configurations.

Abstract

Brain-Machine Interfacing (BMI) has greatly benefited from adopting machine learning methods for feature learning that require extensive data for training, which are often unavailable from a single dataset. Yet, it is difficult to combine data across labs or even data within the same lab collected over the years due to the variation in recording equipment and electrode layouts resulting in shifts in data distribution, changes in data dimensionality, and altered identity of data dimensions. Our objective is to overcome this limitation and learn from many different and diverse datasets across labs with different experimental protocols. To tackle the domain adaptation problem, we developed a novel machine learning framework combining graph neural networks (GNNs) and transfer learning methodologies for non-invasive Motor Imagery (MI) EEG decoding, as an example of BMI. Empirically, we focus on the challenges of learning from EEG data with different electrode layouts and varying numbers of electrodes. We utilise three MI EEG databases collected using very different numbers of EEG sensors (from 22 channels to 64) and layouts (from custom layouts to 10-20). Our model achieved the highest accuracy with lower standard deviations on the testing datasets. This indicates that the GNN-based transfer learning framework can effectively aggregate knowledge from multiple datasets with different electrode layouts, leading to improved generalization in subject-independent MI EEG classification. The findings of this study have important implications for Brain-Computer-Interface (BCI) research, as they highlight a promising method for overcoming the limitations posed by non-unified experimental setups. By enabling the integration of diverse datasets with varying electrode layouts, our proposed approach can help advance the development and application of BMI technologies.

Figures (4)

• Figure 1: The 3 motor imagery EEG datasets showing different numbers of electrodes and their arrangements, visualised in the EEG 10-20 system: From left to right: BCIC 2a tangermann2012review, PhysioNet MI schalk2004bci2000, OpenBMI lee2019eeg (from left). A naive strategy to combine and utilise these datasets by only using the data recorded from these common sensors (yellow circles) discards a lot of data recorded from the other sensors (white circles).
• Figure 2: (a) Our proposed frameworks for training with multiple motor imagery EEG datasets. The framework starts with Temporal CNN blocks with similar architecture to EEGNet lawhern2018eegnet without spatial filters. After each temporal convolution layer, we apply batch normalization, PRelu activation layer, average pooling and dropout layers to introduce non-linearity and reduce the risk of overfitting. After each GCN block, the number of channels is reduced by approximately a factor of 2. At the final stage of the GCN block, we apply the global mean pooling operator, which averages the remaining node features and produces a graph feature vector. The feature vector passes through an MLP latent alignment block and finally, a Linear classifier was used to perform the classification. (b) Illustration of the adjacency matrix and the local aggregation process in each GCN block.(c) Illustration of the SAGPool process in each GCN block.
• Figure 3: (a) Illustration of the PhysioNet MI sensor grid representation on the EEG 10-20 system and (b) its adjacency matrix $\mathbf{A}$ with shape $64 \times 64$, constructed using the neighbourhood method. We draw here the full graph with sensors and their immediate neighbours connected by a black line. As an illustration of the neighbourhood concept, $C_z$ (red node) is connected to its immediate neighbour nodes (green nodes). Additional global connections are shown in blue. Note, that the colour coding is binary in the adjacency matrix (right), as we only consider connections that exist (yellow) and that do not exist (dark blue). (c) The adjacency matrix constructed using the correlation method. (b) The resulting adjacency matrix when only the top 5 connections for each node are kept. The numerical index of each sensor is shown on the right.
• Figure 4: (a) Average inter-subject cross-validation results. Single CNN 1: A single CNN model trained with the training samples from a single dataset, Single CNN 2: A single CNN model trained with samples from all datasets by selecting the common channels. tSNE plots of the latent features of test data from three EEG datasets without latent alignment:(b) Colours represent different datasets. (c) Colours represent different subjects, Circle and Square represent left and right motor imagery classes respectively. tSNE plots of the latent features of test data from three EEG datasets with latent alignment:(d) Colours represent different datasets. (e) Colours represent different subjects, Circle and Square represent left and right motor imagery classes respectively.