EEG Foundation Models:
Beyond Hand-Crafted Features.

Research note · October 2026

EEG foundation models are moving the field from task-specific feature engineering toward reusable neural representations learned from heterogeneous recordings. A 2026 benchmarking study reviewed 50 representative models and evaluated 12 open-source foundation models across 13 EEG datasets and nine BCI paradigms, highlighting the importance of consistent preprocessing, cross-subject evaluation, and transfer protocols.

The shift

Traditional EEG pipelines often begin with band power, connectivity, entropy, or other engineered features. Foundation-model approaches attempt to learn representations directly from large collections of neural recordings. That can make downstream adaptation easier, but scale alone does not guarantee better generalization.

What current research suggests

• Benchmarking protocols can materially change conclusions about transferability.
• Leave-one-subject-out and few-shot calibration are useful stress tests for real deployment.
• Linear probing can be insufficient for some EEG foundation models.
• Specialist models trained for a task can remain competitive with larger pretrained models.

Why this matters for Neumage

Neumage treats EEG as an executable research object: acquisition → QC → preprocessing → representation → model → subject-disjoint evaluation → evidence. This makes it possible to compare foundation models against conventional baselines without changing the scientific protocol silently.

Source: Liu et al., EEG Foundation Models: Progresses, Benchmarking, and Open Problems, 2026.