MSBraM: A Multi-scale Self-supervised Brain Foundation Model for Hierarchical EEG Dynamics Learning

2026-07-23

AI Summary

Existing self-supervised EEG models struggle to effectively capture the multi-scale temporal structure of electroencephalographic signals, limiting representation learning and generalization on downstream tasks. To address this, this work proposes MSBraM, the first framework to explicitly model the hierarchical multi-scale dynamics of EEG. MSBraM employs a multi-scale neural tokenizer to discretize raw signals into semantic tokens at varying temporal resolutions and introduces a curriculum-based multi-scale masking strategy that jointly learns fine-grained local patterns and global temporal context during self-supervised pretraining. Evaluated across 12 public datasets and 10 diverse downstream tasks, MSBraM significantly outperforms current state-of-the-art models, demonstrating its superior cross-granularity representational capacity and generalization performance.

Abstract

Self-supervised foundation models have recently shown strong potential for electroencephalogram (EEG)-based analysis. However, existing approaches struggle to capture the inherently multi-scale temporal structure of EEG signals, where local neural patterns and long-range dependencies jointly encode task-relevant information. This limitation hampers cross-scale representation learning and generalization across diverse downstream tasks. To address this challenge, we propose MSBraM, a Multi-Scale self-supervised Brain foundation Model designed to learn hierarchical EEG representations. MSBraM follows a two-stage pretraining framework. First, a multi-scale neural tokenizer discretizes raw EEG signals into semantic codes at different temporal resolutions via vector-quantized reconstruction. Second, the model is pretrained to predict masked codes using a curriculum multi-scale masking strategy, progressively integrating fine-grained local patterns with global temporal context. We pretrain MSBraM on over 2,400 hours of EEG data and evaluate it across 10 downstream tasks on 12 public datasets. Extensive experiments show that MSBraM achieves superior performance on other state-of-the-art pretrained models, demonstrating strong generalization and transferability. These results indicate that explicitly modeling multi-scale temporal dynamics is critical for effective EEG foundation models.

Problem

• multi-scale
• EEG
• self-supervised
• temporal dynamics
• foundation model

Innovation

• multi-scale
• self-supervised
• EEG foundation model
• hierarchical representation
• temporal dynamics