DeeperBrain: A Neuro-Grounded EEG Foundation Model Towards Universal BCI

TL;DR

DeeperBrain targets universal brain-computer interfaces by embedding neurophysiological priors into both model design and learning objectives. It introduces a volume-conduction-aware channel encoding and a neurodynamics-aware temporal encoding, paired with a dual pretraining objective (MER and NSP) to capture both fine-grained waveforms and macroscopic brain states. Pretrained on 14 diverse EEG datasets totaling over $17{,}200$ hours, the model achieves state-of-the-art or competitive results across 10 downstream tasks and maintains superior performance under a rigorous frozen-probing protocol, demonstrating intrinsic universality. This approach narrows the gap between data-driven learning and neuroscientific first principles, offering a robust, transferable EEG encoder for broad BCI deployment and reducing calibration needs in real-world settings.

Abstract

Electroencephalography (EEG) foundation models hold significant promise for universal Brain-Computer Interfaces (BCIs). However, existing approaches often rely on end-to-end fine-tuning and exhibit limited efficacy under frozen-probing protocols, lacking the intrinsic universality required for broad generalization. This limitation stems from adapting general-purpose sequence architectures that overlook the biophysical and dynamical principles of neural activity. To bridge this gap, we propose DeeperBrain, a neuro-grounded foundation model integrating domain-specific inductive biases into its model design and learning objectives. Architecturally, DeeperBrain incorporates a volume conduction-aware channel encoding to model spatial mixing via 3D geometry, and a neurodynamics-aware temporal encoding capturing slow adaptations using oscillatory and exponential bases. For pretraining, we introduce a dual-objective strategy combining Masked EEG Reconstruction (MER) for local fidelity and Neurodynamics Statistics Prediction (NSP). NSP enforces alignment with macroscopic brain states by predicting interpretable order parameters, including spectral power, functional connectivity, cross-frequency coupling, and dynamic complexity. Extensive experiments demonstrate that DeeperBrain achieves state-of-the-art or highly competitive performance under end-to-end fine-tuning. Crucially, it maintains superior efficacy under a rigorous frozen-probing protocol, verifying that embedding neuroscientific first principles endows learned representations with the intrinsic universality essential for universal BCI. The code will be publicly available.

Figures (6)

• Figure 1: DeeperBrain overview.
• Figure 2: Performance comparison (mean $\pm$ std, %) in balanced accuracy on positional encoding. "*" indicates $p<0.05$ and "**" indicates $p<0.01$.
• Figure 3: Performance comparison (mean $\pm$ std, %) in balanced accuracy on pretraining objective. "*" indicates $p<0.05$ and "**" indicates $p<0.01$.
• Figure 4: Comparison of spatial receptive fields on a high-density 10-5 montage. Unlike the baseline (left) which exhibits a discrete impulse response, DeeperBrain (right) explicitly models volume conduction via a continuous spatial decay kernel (visualized here with $\tau=8$ cm), aligning with the physics of scalp potentials.
• Figure 5: Comparison of temporal positional encodings. (Left) Standard sinusoidal PE relies on generic, symmetric spectral priors. (Middle) Learnable absolute PE starts as unstructured noise, lacking explicit physical constraints. (Right) DeeperBrain PE incorporates strong neurophysiological inductive biases: Slow Oscillations for quasi-periodic state maintenance and Adaptive Decays for modeling the dissipative "arrow of time".