MEG-GPT: A transformer-based foundation model for magnetoencephalography data

Area of Science:

• Computational Neuroscience
• Deep Learning
• Neuroimaging

Background:

• Traditional methods struggle with complex spatio-temporal patterns in magnetoencephalography (MEG).
• Foundation models have shown success in other AI domains, offering potential for neuroscience applications.

Purpose of the Study:

• Introduce MEG-GPT, a transformer-based foundation model for MEG data analysis.
• Develop a novel data-driven tokeniser to preserve high temporal resolution in MEG signals.
• Evaluate MEG-GPT's ability to generate realistic brain data and improve downstream decoding tasks.

Main Methods:

• Developed MEG-GPT, a transformer model utilizing time-attention and next time-point prediction.
• Introduced a novel data-driven tokeniser for continuous MEG data.
• Trained MEG-GPT on tokenised brain region time courses from a large-scale MEG dataset (N=612).

Main Results:

• MEG-GPT generates data with realistic spatio-spectral properties, including transient events and population variability.
• The model improves zero-shot generalization in downstream decoding tasks across sessions and subjects compared to PCA.
• Fine-tuning MEG-GPT on smaller labeled datasets boosts performance in cross-subject decoding.

Conclusions:

• MEG-GPT establishes a powerful foundation model for electrophysiological data analysis.
• This work paves the way for advanced applications in computational neuroscience and neural decoding.
• The novel tokeniser preserves crucial temporal resolution in MEG data.