METIS: A Language-Guided Multimodal Foundation Model for Brain Signal Analysis

Core Introduction to METIS

METIS is an innovative language-guided multimodal foundation model designed specifically for zero-shot and multi-task brain signal analysis. It combines the semantic understanding capabilities of large language models with neural signal processing, opening up new paths for brain-computer interface and neuroscience research.

Original Author & Source

• Original Author/Maintainer: mingzhi-c
• Source Platform: GitHub
• Original Title: metis-brain-signal-foundation-model
• Original Link: https://github.com/mingzhi-c/metis-brain-signal-foundation-model
• Publish/Update Time: 2026-06-06T09:14:25Z

Brain signal analysis is a core challenge in neuroscience and brain-computer interfaces: traditional methods require large amounts of labeled data for training and face difficulties in cross-task transfer. With the development of large language models and multimodal technologies, the demand for building a universal "brain signal foundation model" has emerged—METIS is an innovative attempt to address this challenge, pioneering a new language-guided paradigm.

METIS consists of three core components:

1. Brain Signal Encoder: Converts raw neural signals (EEG, fMRI, etc.) into high-dimensional semantic representations, capturing spatiotemporal features and aligning with the semantic space.
2. Language Encoder: Processes task descriptions/label definitions based on large language models, providing a semantic understanding framework.
3. Multimodal Fusion Module: Establishes fine-grained correspondence between brain signals and language representations via attention mechanisms, enabling language-guided cross-modal interaction.

Zero-Shot Capability

No labeled data is needed—predictions can be made by describing new tasks in natural language (e.g., recognizing new cognitive states by describing their features), greatly reducing application thresholds.

Multi-Task Capability

A unified architecture handles tasks such as classification (emotion recognition), regression (cognitive load estimation), and generation (signal reconstruction), supporting knowledge transfer and improving versatility.

METIS has broad prospects in multiple fields:

• Brain-Computer Interface: Shortens calibration time, allowing users to adapt quickly by describing their intentions through language.
• Neurological Disease Diagnosis: Assists in analyzing brain signal patterns, facilitating early screening for depression, epilepsy, etc.
• Cognitive Science Research: Verifies correlations between cognitive state hypotheses and brain signals, accelerating scientific discoveries.
• Human-Computer Interaction: Understands users' cognitive states and emotions, enabling more natural intelligent interactions.

Challenges

• Data Heterogeneity: Large differences in brain signals from different devices/paradigms, with insufficient diversity in pre-training data.
• Individual Differences: Strong individual specificity of brain signals, requiring a balance between zero-shot capability and personalized adaptation.
• Causal Mechanisms: Currently only statistical correlations are established; deep causal mechanisms need to be explored.

Future Directions

• Expand multimodal inputs (eye movement, physiological signals);
• Develop efficient fine-tuning strategies for rapid personalization;
• Explore new pre-training methods for brain signals to build "brain signal large models".

METIS represents an important attempt to shift brain signal analysis toward the foundation model paradigm. By combining large language models with brain signal processing, it opens up new paths for universal, flexible, and interpretable brain signal intelligent systems. With technological progress, future human-computer interaction will become more natural and in-depth, and METIS is a key cornerstone of this vision.