EEG Foundation Models: Progresses, Benchmarking, and Open Problems
Dingkun Liu, Yuheng Chen, Zhu Chen, Zhenyao Cui, Yaozhi Wen, Jiayu An, Jingwei Luo, Dongrui Wu
2026-01-30
Summary
This paper investigates the new trend of using 'foundation models' for brain-computer interfaces (BCIs) that rely on electroencephalography (EEG) data, which measures brain activity. These models aim to learn general patterns from lots of different EEG recordings so they can be used for various BCI tasks.
What's the problem?
While these EEG foundation models are showing promise, there hasn't been a good, fair way to compare them. Different research groups use different methods for preparing the data, building the models, and testing how well they work. This makes it hard to know which models are truly the best and why. It's like trying to compare athletes in different sports with different rules!
What's the solution?
The researchers tackled this problem by thoroughly reviewing 50 different EEG foundation models, categorizing their key design choices like how they handle data and the type of model architecture used. Then, they took 12 publicly available models and tested them on 13 different EEG datasets covering nine different BCI tasks. They tested how well the models worked when applied to new people (cross-subject generalization) and when quickly adapted to a single person (few-shot learning). They also compared fully retraining the models versus just using the pre-trained parts, and looked at whether bigger models always perform better.
Why it matters?
This research is important because it provides a standardized comparison of these new EEG foundation models. The findings show that simply using the pre-trained model without further training isn't always enough, and that models specifically designed for a task can still be very competitive. Surprisingly, making the models larger doesn't automatically guarantee better performance with the current data and training methods. This helps researchers focus their efforts on the most promising approaches for building better brain-computer interfaces.
Abstract
Electroencephalography (EEG) foundation models have recently emerged as a promising paradigm for brain-computer interfaces (BCIs), aiming to learn transferable neural representations from large-scale heterogeneous recordings. Despite rapid progresses, there lacks fair and comprehensive comparisons of existing EEG foundation models, due to inconsistent pre-training objectives, preprocessing choices, and downstream evaluation protocols. This paper fills this gap. We first review 50 representative models and organize their design choices into a unified taxonomic framework including data standardization, model architectures, and self-supervised pre-training strategies. We then evaluate 12 open-source foundation models and competitive specialist baselines across 13 EEG datasets spanning nine BCI paradigms. Emphasizing real-world deployments, we consider both cross-subject generalization under a leave-one-subject-out protocol and rapid calibration under a within-subject few-shot setting. We further compare full-parameter fine-tuning with linear probing to assess the transferability of pre-trained representations, and examine the relationship between model scale and downstream performance. Our results indicate that: 1) linear probing is frequently insufficient; 2) specialist models trained from scratch remain competitive across many tasks; and, 3) larger foundation models do not necessarily yield better generalization performance under current data regimes and training practices.