A pretrained foundation model for headache disorders based on magnetoencephalography

Area of Science:

• Medical Artificial Intelligence
• Functional Neuroimaging
• Neurology

Background:

• Foundation models show promise in AI but are underutilized in magnetoencephalography (MEG).
• High-dimensional MEG data and limited labeled datasets pose challenges for clinical research.
• Headache disorders require advanced diagnostic tools for better patient outcomes.

Purpose of the Study:

• To develop a domain-specific, self-supervised foundation model for MEG data tailored to headache disorders.
• To address challenges of high-dimensional data and limited labeled datasets in clinical MEG research.
• To improve the diagnostic accuracy for neurological conditions like migraine using AI.

Main Methods:

• Developed a transformer-based foundation model pretrained on multi-state MEG recordings from 416 participants.
• Employed a self-supervised masked-signal reconstruction strategy to learn spatiotemporal neural representations.
• Evaluated model performance via signal reconstruction, attention weight visualization, and downstream migraine classification tasks.

Main Results:

• The foundation model successfully reconstructed continuous MEG signals and stimulus-evoked responses.
• Attention weight visualization aligned with sensory brain regions, indicating neurophysiological interpretability.
• Classifiers using model-derived features significantly outperformed those using original MEG signals for migraine diagnosis.

Conclusions:

• Introduced a scalable, data-efficient framework for clinical MEG analysis, reducing reliance on manual feature extraction and labeled data.
• Demonstrated the efficacy of foundation models in decoding complex neural dynamics from MEG data.
• The model shows potential for understanding neuropathology and advancing precision diagnostics in neurology.