EEG Foundation Models

EEG foundation models are pretrained on large collections of brain recordings to create representations that can be reused for tasks such as sleep staging, seizure detection, clinical classification, and brain-computer interface decoding. Common approaches include masked-signal reconstruction, contrastive learning, prediction of latent representations, and alignment with clinical text; Transformer and state-space models aim to capture temporal, spectral, and spatial structure while accommodating different electrode layouts. Models are adapted through fine-tuning, lightweight adapters, or frozen-feature classifiers, with growing interest in long recordings and low-label or cross-subject settings. A central tension is whether broad pretraining reliably beats well-tuned task-specific models: results vary across tasks and benchmarks, and identity, artifacts, and recording-site differences can masquerade as transferable brain information.