EEG Foundation Models: Progresses, Benchmarking, and Open Problems

TL;DR

This work delivers a comprehensive taxonomy and fair benchmark for EEG foundation models, revealing that pre-trained encoders often require full-end adaptation rather than fixed feature extraction to generalize across tasks. By evaluating 12 open-source foundation models and 7 specialist baselines on 13 EEG datasets spanning 9 BCI paradigms under LOSO and within-subject few-shot settings, it shows that larger models do not automatically yield better generalization and that specialist models trained from scratch can remain highly competitive. The findings emphasize the need for improved pre-training objectives, noise-robustness, and evaluation protocols that reflect real-world deployment, ultimately aiming to standardize progress and accelerate practical BCI development. The study also provides a formal problem definition, data collection and preprocessing guidelines, and a taxonomy of pre-training objectives to guide future research in transferability and robustness across heterogeneous EEG data.

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.

Figures (11)

• Figure 1: Overview of BCI foundation models. Models are pre-trained on large scale heterogeneous EEG data collected from devices with diverse electrode configurations across various paradigms. Through self-supervised pre-training, the learned representations may generalize to a wide range of downstream tasks.
• Figure 2: EEG foundation model pre-training pipeline. Raw EEG trials are first standardized through channel selection or unification, followed by dataset dependent preprocessing and normalization/alignment. The standardized signal is then used for self-supervised pre-training with representative objectives: (a) Masked reconstruction of raw EEG signals in the time domain; (b) Masked reconstruction of embedded tokens after tokenization; (c) Frequency domain reconstruction, where the target can be the spectrogram, spectral amplitude, or phase related representation; (d) Codebook based reconstruction, where a tokenizer maps the signal to discrete codebook indices or codebook embeddings and the model learns to predict the corresponding discrete units; and, (e) Autoregressive or causal reconstruction using causal masking, implemented with causal Transformer blocks or large language models.
• Figure 3: Overview of 50 existing EEG foundation models.
• Figure 4: Dataset usage statistics across existing EEG foundation models. (a) Frequency ranking of datasets used during pre-training; (b) Frequency ranking of downstream datasets used for generalization evaluation; and, (c) Frequency ranking of datasets used in pre-training or downstream evaluation.
• Figure 5: Overview of datasets and evaluation scenarios used in the benchmark. (a) The 13 downstream datasets spanning 9 representative BCI paradigms, including motor imagery, P300, SSVEP, clinical detection, emotion recognition, visual decoding, fatigue detection, sleep stage analysis, and workload detection; (b) Illustration of the two evaluation scenarios: the leave-one-subject-out (LOSO) scenario, which aggregates labeled data from multiple subjects for fine-tuning and evaluates on a held-out subject, and the within-subject few-shot scenario, which uses only a small amount of labeled data from the target subject for adaptation.

Theorems & Definitions (2)

• Definition 2.1: Foundation Model
• Definition 2.2: BCI Foundation Model