BrainPro: EEG State-Aware Representation Model

• BrainPro is a large-scale EEG representation learning model that captures channel-to-channel and region-to-region dynamics using shared and state-specific encoders.
• It employs a retrieval-based spatial learning block to adapt to variable electrode layouts and a state-decoupling mechanism to enhance representation clarity.
• Pre-training on 2,400 hours of diverse EEG data validated BrainPro's state-of-the-art performance across nine public BCI datasets for varied neurophysiological tasks.

Searching arXiv for the specified BrainPro paper and closely related EEG foundation-model context. arXiv search query: (Ding et al., 26 Sep 2025) BrainPro EEG representation learning BrainPro is a large-scale brain state-aware EEG representation learning model for non-invasive electroencephalography, developed to address two recurrent limitations of prior EEG foundation models: inadequate explicit modeling of channel-to-channel and region-to-region interactions, and the absence of state-aware self-supervised pre-training despite the heterogeneity of EEG corpora across affective, motor, and other brain states. It introduces a retrieval-based spatial learning block for variable electrode layouts and a brain state-decoupling block based on parallel encoders, decoupling losses, and region-aware masked reconstruction, with the stated goal of adapting seamlessly to diverse tasks and hardware settings while achieving state-of-the-art performance and robust generalization across nine public BCI datasets (Ding et al., 26 Sep 2025).

1. Problem formulation and design goals

BrainPro is situated within the recent line of EEG foundation models trained on large-scale datasets for BCI and healthcare, but it is explicitly motivated by shortcomings that remain after simple scaling. Existing models are described as often failing to capture channel-to-channel and region-to-region interactions explicitly, even though such interactions are critical sources of information inherently encoded in EEG signals. Because EEG datasets differ in channel configurations, prior methods either approximate spatial structure with self-attention or restrict training to a limited set of common channels, thereby sacrificing flexibility and effectiveness (Ding et al., 26 Sep 2025).

A second design goal is state awareness. EEG datasets reflect diverse brain states, including emotion and motor processes, yet current models are described as rarely learning state-aware representations during self-supervised pre-training. BrainPro therefore aims not merely to improve transfer in a generic sense, but to organize latent structure so that shared EEG regularities and process-specific regularities can be separated and recombined. This suggests a shift from a single universal encoder toward a structured representation regime in which spatial variability and brain-state variability are treated as first-class modeling constraints rather than nuisances.

2. Overall architecture

At a high level, BrainPro uses parallel encoders: one shared encoder for universal features and multiple brain state-specific encoders for state-dependent features such as affect, motor, and others. During pre-training, a sample updates the shared encoder and only the relevant state-specific encoder; the other state-specific branches are detached. The active shared and state-specific outputs are then fused and passed to a decoder for masked EEG reconstruction (Ding et al., 26 Sep 2025).

Within each encoder, the processing pipeline is hierarchical. A temporal encoder based on multi-scale 1D CNNs first extracts channel-wise temporal dynamics. This is followed by a retrieval-based spatial learner that models channel- and region-level interactions explicitly, after which the features are patchified and linearly embedded into tokens for transformer layers that capture long-range dependencies. The temporal encoder is written as

$$\mathbf{Z}^{(\ell)} = \mathrm{GELU}\!\left(\mathrm{GroupNorm}\!\left(\mathrm{Conv1D}^{(\ell)}(\mathbf{Z}^{(\ell-1)})\right)\right),$$

for $\ell=1,\ldots,L_T$.

This architecture is notable for not relying on a single monolithic transformer to infer all structure implicitly. Instead, BrainPro inserts an explicit spatial stage between temporal feature extraction and transformer token modeling. A plausible implication is that the model treats EEG montage variability as a retrieval and alignment problem before it becomes a sequence-modeling problem.

3. Retrieval-based spatial learning across variable montages

The retrieval-based spatial learning block is the core mechanism by which BrainPro handles heterogeneous electrode layouts. It uses a 60-channel universal template, described as universalized via the SEED dataset, together with 24 functional regions defined according to LGGNet. Two learnable filter banks are introduced:

$$\mathbf{W}_C \in \mathbb{R}^{C_{\mathrm{pre}} \times K_C}, \qquad \mathbf{W}_R \in \mathbb{R}^{N_{\mathrm{region}} \times K_R}.$$

For each input sample, the model identifies which universal channels and regions are present, retrieves the corresponding spatial filters, and computes fine-grained channel features and coarse-grained region features. The fine channel representation is given by

$$\mathbf{H}_{C} = \sigma\left(\mathbf{W}_{C}[I_{\mathrm{ch}}]^\top \mathbf{H}_{\mathrm{temp}}^{\mathrm{reshp}}\right),$$

while region features are formed by averaging over channels within each functional region,

$$\mathbf{M}_{\mathrm{region}}[j] = \frac{1}{|\mathcal{C}_j|}\sum_{c \in \mathcal{C}_j} \mathbf{H}_{\mathrm{temp}}^{\mathrm{reshp}}[c, :],$$

followed by

$$\mathbf{H}_{R} = \sigma\!\left(\mathbf{W}_{R}[\mathcal{R}_{\mathrm{uniq}}]^\top \mathbf{M}_{\mathrm{region}}\right).$$

The final spatial representation concatenates $\mathbf{H}_C$ and $\mathbf{H}_R$ and reshapes the result for downstream tokenization (Ding et al., 26 Sep 2025).

The stated significance of this block is that BrainPro can model channel-to-channel and region-to-region interactions directly regardless of missing electrodes or dataset-specific montages. In contrast to restricting pre-training to a common subset of channels, retrieval-based spatial learning preserves flexibility across datasets and hardware. In contrast to leaving spatial structure to generic self-attention, it imposes a neurophysiologically structured decomposition into fine channel relations and coarse regional relations.

4. Brain state-decoupling and self-supervised objectives

BrainPro’s state-aware component is organized around a shared encoder $\mathcal{E}_{\mathrm{shared}}$ and multiple state-specific encoders $\{\mathcal{E}_k\}_{k=1}^K$. For each sample with brain-state label $\ell=1,\ldots,L_T$0, only the shared encoder and the matching state-specific encoder are updated. This is paired with two losses: a region-aware masked reconstruction loss and a brain-state decoupling loss (Ding et al., 26 Sep 2025).

The reconstruction objective masks a proportion $\ell=1,\ldots,L_T$1 of input patches and weights reconstruction errors according to channel importance vectors specific to brain states. The loss is

$\ell=1,\ldots,L_T$2

where $\ell=1,\ldots,L_T$3 is the set of masked positions and $\ell=1,\ldots,L_T$4 is a brain-state channel prior. The weighting function is

$\ell=1,\ldots,L_T$5

with $\ell=1,\ldots,L_T$6 the sigmoid and $\ell=1,\ldots,L_T$7 increasing with training epoch so that weighting becomes sharper over time.

The decoupling loss is a margin-based cosine penalty that discourages high similarity between the shared representation and the active state-specific representation, and also between the active state-specific representation and inactive state-specific representations:

$\ell=1,\ldots,L_T$8

with margin $\ell=1,\ldots,L_T$9.

The total pre-training loss is

$$\mathbf{W}_C \in \mathbb{R}^{C_{\mathrm{pre}} \times K_C}, \qquad \mathbf{W}_R \in \mathbb{R}^{N_{\mathrm{region}} \times K_R}.$$0

These objectives jointly encode two commitments: masked prediction should be neurophysiology-aware rather than spatially uniform, and latent factors should be explicitly decoupled rather than left to emerge implicitly from a single encoder. This suggests that BrainPro treats pre-training not as generic denoising, but as structured representation disentanglement anchored in known EEG state differences.

5. Pre-training regime and downstream adaptation

BrainPro is pre-trained on approximately 2,400 hours of EEG from diverse datasets spanning affective, motor, and clinical domains. The datasets are grouped into affect, motor, and others to supervise state-specific encoding during pre-training. Channels from each dataset are mapped to the 60-channel universal template, the masking rate is 50%, optimization uses AdamW with a cosine schedule and gradient clipping, training runs for 30 epochs, and batching is organized by dataset for efficiency in spatial filter retrieval (Ding et al., 26 Sep 2025).

Downstream adaptation is described in mixture-of-experts terms. Rather than forcing every task through a single latent space, BrainPro allows arbitrary subsets of encoders to be selected and concatenated:

$$\mathbf{W}_C \in \mathbb{R}^{C_{\mathrm{pre}} \times K_C}, \qquad \mathbf{W}_R \in \mathbb{R}^{N_{\mathrm{region}} \times K_R}.$$1

This design allows a downstream task to use only the shared encoder, the shared encoder plus one state-specific encoder, or the shared encoder plus multiple state-specific encoders, depending on the task structure. The model is therefore presented as adaptable to complex scenarios involving overlapping or mixed brain states. A plausible implication is that BrainPro’s pre-training objective is also a task routing mechanism: the state-specific branches act as reusable specialists while the shared branch stabilizes transfer across domains.

6. Empirical performance, ablation evidence, and interpretability

BrainPro is evaluated on nine public BCI datasets spanning emotion, motor imagery, speech, mental disorder, stress, and attention, using balanced accuracy, Cohen’s Kappa, weighted F1, AUC-PR, and AUROC. The paper reports that BrainPro consistently outperforms prior EEG foundation models and non-foundation baselines, including BIOT, LaBraM, CBraMod, EEGPT, EEGNet, and Conformer, and achieves state-of-the-art results across nearly all tasks and metrics (Ding et al., 26 Sep 2025).

Reported examples include ACC-B $$\mathbf{W}_C \in \mathbb{R}^{C_{\mathrm{pre}} \times K_C}, \qquad \mathbf{W}_R \in \mathbb{R}^{N_{\mathrm{region}} \times K_R}.$$2 on FACED, compared with $$\mathbf{W}_C \in \mathbb{R}^{C_{\mathrm{pre}} \times K_C}, \qquad \mathbf{W}_R \in \mathbb{R}^{N_{\mathrm{region}} \times K_R}.$$3 for the next-best CBraMod; ACC-B $$\mathbf{W}_C \in \mathbb{R}^{C_{\mathrm{pre}} \times K_C}, \qquad \mathbf{W}_R \in \mathbb{R}^{N_{\mathrm{region}} \times K_R}.$$4 on MentalArithmetic, compared with $$\mathbf{W}_C \in \mathbb{R}^{C_{\mathrm{pre}} \times K_C}, \qquad \mathbf{W}_R \in \mathbb{R}^{N_{\mathrm{region}} \times K_R}.$$5 for CBraMod; ACC-B $$\mathbf{W}_C \in \mathbb{R}^{C_{\mathrm{pre}} \times K_C}, \qquad \mathbf{W}_R \in \mathbb{R}^{N_{\mathrm{region}} \times K_R}.$$6 on Imagined Speech, compared with $$\mathbf{W}_C \in \mathbb{R}^{C_{\mathrm{pre}} \times K_C}, \qquad \mathbf{W}_R \in \mathbb{R}^{N_{\mathrm{region}} \times K_R}.$$7 for CBraMod; and ACC-B $$\mathbf{W}_C \in \mathbb{R}^{C_{\mathrm{pre}} \times K_C}, \qquad \mathbf{W}_R \in \mathbb{R}^{N_{\mathrm{region}} \times K_R}.$$8 on an MDD mental-disorder task. The ablation studies are described as confirming that masking, region-aware reconstruction, the decoupling loss, retrieval-based spatial learning, and pre-training all contribute substantially to final performance.

Interpretability is tied to the learned spatial filters rather than to a purely post hoc explanation module. Visualizations of affect and motor filters are reported to highlight neuroscientifically meaningful patterns, with frontal and temporal emphasis for emotion-related processing and central and parietal emphasis for motor processing. The article’s interpretability claim is therefore not only that BrainPro performs well, but that its learned spatial decomposition is aligned with recognizable neurophysiological structure.

7. Scope, related systems, and nomenclature

BrainPro is a representation-learning model for EEG rather than an end-user BCI application, neuroprosthetic controller, or clinical digital twin. This distinction matters because neighboring systems in the broader literature operate at different abstraction levels. BRAVE, for example, is a hybrid EEG- and voice-controlled prosthetic system that combines ensemble learning, human-in-the-loop correction, and automatic speech recognition for real-time upper-limb control, reporting 96% test accuracy and 150 ms latency (Basit et al., 23 May 2025). NeuroSkill is a real-time proactive agentic system for modeling Human State of Mind from BCI and biophysical signals, running fully offline on the edge and structured around foundation EXG models, text embeddings, and a markdown-driven skill layer (Kosmyna et al., 3 Mar 2026). BrainPro, by contrast, is primarily an EEG foundation model intended to supply transferable representations across tasks and montages.

A separate nomenclature issue is that the name “BrainPro” is also used in a later paper for a multi-modal neuro-oncological cognitive digital twin that combines wearable EEG, structural MRI, an Enhanced Vision Transformer, Grad-CAM, and a tumor kinetics engine (Banerjee et al., 30 Sep 2025). That system is unrelated in scope and method to the EEG foundation model discussed here. The overlap in naming can therefore produce bibliographic ambiguity: in the EEG literature, BrainPro denotes a large-scale brain state-aware EEG representation model (Ding et al., 26 Sep 2025), whereas in neuro-oncological digital-twin work the same name denotes an MRI-and-EEG prognostic framework (Banerjee et al., 30 Sep 2025).