A Pre-trained EEG-to-MEG Generative Framework for Enhancing BCI Decoding

TL;DR

This work tackles the data bottleneck in MEG-based BCIs by learning a cross-modal EEG-to-MEG mapping that yields MEG-quality signals from EEG. It uses a pre-trained EEG backbone to extract general neural representations and adds a spatial focus mapper, a latent diffusion generator, and a broadband spectral calibrator to synthesize MEG that matches real MEG in time-frequency characteristics and source-space activations. The model improves BCI decoding when combining EEG with synthesized MEG on paired datasets and also enhances EEG-only decoding on independent data, indicating robust cross-task generalization. This approach offers a scalable route to augment MEG observations, reducing hardware costs and expanding the applicability of MEG-informed BCIs.

Abstract

Electroencephalography (EEG) and magnetoencephalography (MEG) play important and complementary roles in non-invasive brain-computer interface (BCI) decoding. However, compared to the low cost and portability of EEG, MEG is more expensive and less portable, which severely limits the practical application of MEG in BCI systems. To overcome this limitation, this study proposes the first cross-modal generation framework based on EEG-MEG spatiotemporal coupled representations to synthesize MEG signals cost-effectively. The framework first extracts general neural activity representations through a pre-trained EEG model. Building upon these representations, the framework effectively learns the lower spatial dispersion and higher high-frequency sensitivity of MEG via the spatial focus mapping module and the broadband spectral calibration module. Experimental results demonstrate that the synthesized MEG signals show high consistency with the real MEG in both time-frequency characteristics and source space activation patterns. More importantly, downstream BCI decoding experiments demonstrate that using synthesized MEG leads to performance enhancements not only on paired EEG-MEG datasets but also on independent EEG-only datasets. Overall, this framework opens a new avenue for overcoming data bottlenecks in BCI.

Figures (9)

• Figure 1: Comparative analysis of the strengths and limitations of EEG and MEG.
• Figure 2: Flowchart of the pre-trained EEG-to-MEG generative framework. The framework comprises four core components: a EEG representation encoder based on Pre-trained model to extract general neural activity representations. A spatial focus mapping module for EEG-MEG feature alignment. A latent conditional diffusion generator for latent space MEG feature generation. A broadband spectral calibration module for MEG signal recovery. The model is jointly optimized end-to-end via a multi-dimensional loss function.
• Figure 3: Comparison of topographic maps from original EEG, real MEG, and synthesized MEG under task-activation states. The first row corresponds to the Somatomotor dataset, and the second row corresponds to the NOD-MEG dataset. The three columns display the input EEG, real MEG, and synthesized MEG topographies, respectively. The synthesized MEG topographies exhibit high spatial consistency with the real MEG, accurately capturing the task-specific activation patterns for both datasets.
• Figure 4: Quantitative quality assessment of synthesized MEG across different frequency bands. This figure shows the comparison of four key performance metrics between synthesized and real MEG across five standard frequency bands (Delta, Theta, Alpha, Beta, Gamma). Red dots represent the Somatomotor dataset, while blue dots represent the NOD-MEG dataset. Data points indicate the mean values on the test set, and error bars denote the standard deviations. Quantitative analysis confirms that the synthesized MEG maintain high fidelity to real MEG across the full frequency spectrum, exhibiting consistent temporal alignment and spatial structural similarity.
• Figure 5: Comparative assessment of BCI decoding performance. The left and right panels illustrate classification performance on the Somatomotor dataset and the NOD-MEG dataset, respectively. Three input modalities are compared: EEG only (red), EEG combined with synthesized MEG (light blue), and EEG combined with real MEG (dark blue). Box plots display the statistical distributions of accuracy, precision, recall, and F1 scores derived from ten runs of 10-fold cross-validation. The central horizontal line represents the median, and the upper and lower box edges indicate the quartiles. The results demonstrate that combining synthesized MEG and EEG significantly enhances decoding performance across both datasets compared to the only EEG baseline.