Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FoME: A Foundation Model for EEG using Adaptive Temporal-Lateral Attention Scaling

Enze Shi, Kui Zhao, Qilong Yuan, Jiaqi Wang, Huawen Hu, Sigang Yu, Shu Zhang

TL;DR

FoME tackles EEG heterogeneity and data scarcity by training a large, diverse EEG foundation model on over 1.7 TB of data. It introduces time-frequency fusion embedding and adaptive time-lateral attention scaling (ATLAS) to learn robust multi-scale representations across scalp and intracranial recordings. Pretrained with masked signal reconstruction on unlabeled data, FoME achieves state-of-the-art performance across classification, forecasting, and imputation tasks, demonstrating strong transfer and zero-shot potential. This work presents a foundation-model paradigm for EEG with broad implications for BCI, clinical diagnostics, and cognitive neuroscience.

Abstract

Electroencephalography (EEG) is a vital tool to measure and record brain activity in neuroscience and clinical applications, yet its potential is constrained by signal heterogeneity, low signal-to-noise ratios, and limited labeled datasets. In this paper, we propose FoME (Foundation Model for EEG), a novel approach using adaptive temporal-lateral attention scaling to address above-mentioned challenges. FoME is pre-trained on a diverse 1.7TB dataset of scalp and intracranial EEG recordings, comprising 745M parameters trained for 1,096k steps. Our model introduces two key innovations: a time-frequency fusion embedding technique and an adaptive time-lateral attention scaling (ATLAS) mechanism. These components synergistically capture complex temporal and spectral EEG dynamics, enabling FoME to adapt to varying patterns across diverse data streams and facilitate robust multi-channel modeling. Evaluations across four downstream tasks demonstrate FoME's superior performance in classification and forecasting applications, consistently achieving state-of-the-art results. To conclude, FoME establishes a new paradigm for EEG analysis, offering a versatile foundation that advances brain-computer interfaces, clinical diagnostics, and cognitive research across neuroscience and related fields. Our code will be available at https://github.com/1061413241/FoME.

FoME: A Foundation Model for EEG using Adaptive Temporal-Lateral Attention Scaling

TL;DR

FoME tackles EEG heterogeneity and data scarcity by training a large, diverse EEG foundation model on over 1.7 TB of data. It introduces time-frequency fusion embedding and adaptive time-lateral attention scaling (ATLAS) to learn robust multi-scale representations across scalp and intracranial recordings. Pretrained with masked signal reconstruction on unlabeled data, FoME achieves state-of-the-art performance across classification, forecasting, and imputation tasks, demonstrating strong transfer and zero-shot potential. This work presents a foundation-model paradigm for EEG with broad implications for BCI, clinical diagnostics, and cognitive neuroscience.

Abstract

Electroencephalography (EEG) is a vital tool to measure and record brain activity in neuroscience and clinical applications, yet its potential is constrained by signal heterogeneity, low signal-to-noise ratios, and limited labeled datasets. In this paper, we propose FoME (Foundation Model for EEG), a novel approach using adaptive temporal-lateral attention scaling to address above-mentioned challenges. FoME is pre-trained on a diverse 1.7TB dataset of scalp and intracranial EEG recordings, comprising 745M parameters trained for 1,096k steps. Our model introduces two key innovations: a time-frequency fusion embedding technique and an adaptive time-lateral attention scaling (ATLAS) mechanism. These components synergistically capture complex temporal and spectral EEG dynamics, enabling FoME to adapt to varying patterns across diverse data streams and facilitate robust multi-channel modeling. Evaluations across four downstream tasks demonstrate FoME's superior performance in classification and forecasting applications, consistently achieving state-of-the-art results. To conclude, FoME establishes a new paradigm for EEG analysis, offering a versatile foundation that advances brain-computer interfaces, clinical diagnostics, and cognitive research across neuroscience and related fields. Our code will be available at https://github.com/1061413241/FoME.
Paper Structure (36 sections, 8 equations, 7 figures, 5 tables)

This paper contains 36 sections, 8 equations, 7 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Self-supervised Pre-training for EEG Signals
  4. Transfer Learning for Diverse EEG Datasets
  5. Emergence of "Brain Foundation Models"
  6. Method
  7. Overview
  8. Foundation Model Architecture
  9. Patching
  10. Time-Frequency Fusion Encoding
  11. Temporal Encoder
  12. Adaptive Multi-channel Encoder
  13. Pre-training Using Heterogeneous EEG Datasets
  14. Preprocessing
  15. Pre-training Process
  16. ...and 21 more sections

Figures (7)

  • Figure 1: Left: FoME's large-scale heterogeneous pre-training dataset; Right: Parameter scale of existing EEG correlation models.
  • Figure 2: Overview of the FoME architecture and workflow. (A) Raw EEG signal processing and key components of FoME. (B) Self-supervised learning approach using masked signal reconstruction. (C) Fine-tuning process for various downstream EEG analysis tasks.
  • Figure 3: Detailed structural presentation of time-frequency fusion encoding, temporal encoder and adaptive multi-channel encoder.
  • Figure 4: The average performance comparison of FoME and other baselines on all downstream datasets.
  • Figure 5: Visualization of imputation results given by FoME under the 40% mask ratio setting.
  • ...and 2 more figures