Jii Kwon, Youmin Shin
Neural signals such as EEG, ECoG, and intracortical recordings offer a valuable window into brain dynamics but remain difficult to decode due to high dimensionality, nonstationarity, and substantial interindividual variability. Traditional machine learning and deep learning models often show limited generalizability and insufficient interpretability in these settings. Foundation models (FMs)-large-scale architectures pretrained on diverse datasets-have recently emerged as a promising paradigm for building robust, transferable, and physiologically grounded neural representations. Among these modalities, EEG currently serves as the most practical and representative platform for FM development due to its large-scale open datasets, standardized protocols, and broad clinical applicability, while the same conceptual framework remains generalizable to other neural recording types. This review synthesizes emerging FM approaches for neural decoding and critically examines representative EEG-based architectures. We highlight three essential design principles: physiology-aware representation learning that captures oscillatory and dynamic structure, structure-aware architectures that incorporate spatial and anatomical priors, and interpretability mechanisms that ensure neuroscientific and clinical validity. Although models such as the Patched Brain Transformer, CBraMod, and BrainGPT demonstrate encouraging adaptability, many still inherit objectives from non-neural domains and underutilize spatial priors such as electrode topology or functional connectivity. While this review focuses on EEG as the most data-rich and scalable testbed, the same framework can extend to ECoG and intracortical recordings to support unified neural representations across spatial and temporal scales. Fully realizing the potential of neural FMs will require biologically informed objectives, structure-aware architectures, interpretable representations, and standardized data ecosystems.