MSBraM: A Multi-scale Self-supervised Brain Foundation Model for Hierarchical EEG Dynamics Learning
Electroencephalogram (EEG) signals are complex, containing neural information that fluctuates across different timeframes—from rapid, millisecond-long spikes to slow-wave oscillations that last for seconds. While existing AI models for EEG analysis have made progress, they often struggle to capture these multi-scale temporal dynamics, leading to representations that are too rigid for diverse medical and research tasks. MSBraM is a new foundation model designed to solve this by explicitly learning hierarchical EEG representations, allowing it to better generalize across a wide range of brain-related applications.
A Two-Stage Learning Framework
MSBraM uses a two-stage pretraining process to build its understanding of brain signals. In the first stage, the model employs a "multi-scale neural tokenizer." This component breaks down raw EEG data into semantic codes at various resolutions, effectively mapping both fine-grained details and broader patterns into a structured format. In the second stage, the model undergoes "masked EEG modeling." By hiding parts of the signal and forcing the model to predict the missing information, MSBraM learns to reconstruct the underlying structure of the brain activity.
Capturing Multi-scale Dynamics
To ensure the model understands the full scope of EEG data, the researchers implemented a specialized architecture. Each stage of the model includes a downsampling module to create coarser views of the data, a multi-branch Transformer to process these different scales in parallel, and a fusion module (BiFPN) that intelligently combines information from all scales. This allows the model to maintain a "fine-to-coarse" perspective, ensuring that local neural patterns and long-range dependencies are integrated into a single, coherent representation.
Curriculum Masking
A key innovation in MSBraM is the "curriculum multi-scale masking strategy." Instead of using a static approach to hide data during training, the model starts with a low masking ratio and gradually increases it. This curriculum approach encourages the model to first master local, fine-grained features before moving on to more complex, global temporal contexts. By ensuring that masks are consistent across all scales, the model is prevented from "cheating" by using easy-to-guess information from one scale to fill in the gaps of another.
Performance and Generalization
The researchers pretrained MSBraM on over 2,400 hours of EEG data. When tested across 10 different downstream tasks and 12 public datasets—covering areas such as disease diagnosis, sleep staging, and mental state assessment—the model achieved state-of-the-art performance. These results suggest that by explicitly modeling the multi-scale nature of neural signals, MSBraM provides a more robust and transferable foundation for EEG analysis than previous models that treated these signals as uniform data.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!