BrainPro: Process-conditioned Pretraining with Topographic Retrieval for Large EEG Models

Yi Ding; Muyun Jiang; Weibang Jiang; Shuailei Zhang; Xinliang Zhou; Chenyu Liu; Shanglin Li; Yong Li

cs.LGSep 26, 2025

BrainPro: Process-conditioned Pretraining with Topographic Retrieval for Large EEG Models

Authors: Yi Ding, Muyun Jiang, Weibang Jiang, Shuailei Zhang, Xinliang Zhou, Chenyu Liu, Shanglin Li, Yong Li,

Organizations: Nanyang Technological University · Shanghai Jiao Tong University · Advanced Telecommunications Research Institute International · Southeast University

Abstract

EEG foundation models are pretrained on heterogeneous corpora, but two challenges remain underexplored: (1) experimental protocols provide mental-process cues that are rarely used in masked pretraining, and (2) electrode montages vary across datasets, making spatial representations montage-dependent. Existing methods typically learn a single general-purpose representation and model spatial structure implicitly, limiting their ability to exploit process-related information under montage heterogeneity. We introduce BrainPro, a process-conditioned self-supervised pretraining framework that couples topographic retrieval with shared and process-associated representation learning. BrainPro maps dataset-specific montages to a universal channel-region template and retrieves channel- and region-level spatial filters to form a topographically aligned spatial basis. Over this basis, BrainPro learns a shared encoder for general EEG structure and additional affective, motor-related, and auxiliary residual branches for process-associated variation. Protocol-derived mental-process cues condition branch activation and region-weighted masked reconstruction, incorporating topographic spatial priors and process information. Across nine public BCI benchmarks, BrainPro achieves strong performance among evaluated baselines. Ablations, channel-drop analysis, encoder-configuration studies, and spatial-filter visualizations suggest that topographic retrieval and process-conditioned representation learning jointly improve EEG decoding. Our code is available at https://github.com/yi-ding-cs/BrainPro.

Figures & tables

Figure 1: BrainPro framework. BrainPro consists of a shared encoder $\mathcal{E}_{S}$ , affective and motor process-associated encoders ( $\mathcal{E}_{A}$ and $\mathcal{E}_{M}$ ), and an auxiliary residual encoder $\mathcal{E}_{O}$ for heterogeneous non-affect/non-motor data. Pretraining combines process-guided region-weighted reconstruction and representation-level branch separation to learn shared and process-associated representations over a topographically aligned spatial basis.

Figure 2: EEG encoder of BrainPro. Each encoder consists of three stages: (1) a temporal learning encoder based on temporal CNNs; (2) a topographic retrieval encoder that performs channel- and region-based spatial retrieval; and (3) patchification and token embedding, followed by Transformer encoders for spatiotemporal modeling

Methods	FACED (9-Class)			SEED-V (5-Class)			BCI-IV-2a (4-Class)
Methods	ACC-B	Kappa	F1-W	ACC-B	Kappa	F1-W	ACC-B	Kappa	F1-W
EEGNet	0.3692 ±0.0103	0.2880 ±0.0113	0.3693 ±0.0111	0.2408 ±0.0031	0.0536 ±0.0042	0.1908 ±0.0088	0.5521 ±0.0183	0.4028 ±0.0244	0.5346 ±0.0228
Conformer	0.4566 ±0.0108	0.3849 ±0.0121	0.4555 ±0.0098	0.3087 ±0.0095	0.1294 ±0.0134	0.2837 ±0.0229	0.4879 ±0.0183	0.3171 ±0.0244	0.4561 ±0.0206
BIOT	0.2992 ±0.0119	0.2105 ±0.0147	0.2954 ±0.0149	0.3575 ±0.0052	0.1971 ±0.0081	0.3636 ±0.0078	0.2392 ±0.0280	-0.0144 ±0.0373	0.1063 ±0.0060
EEGPT	0.2809 ±0.0116	0.1912 ±0.0129	0.2810 ±0.0120	0.2421 ±0.0048	0.0557 ±0.0077	0.2438 ±0.0055	0.3849 ±0.0226	0.1799 ±0.0301	0.3272 ±0.0336
LaBraM	0.5224 ±0.0116	0.4610 ±0.0126	0.5259 ±0.0108	0.3986 ±0.0200	0.2491 ±0.0256	0.4040 ±0.0197	0.5255 ±0.0329	0.3674 ±0.0439	0.5095 ±0.0405

Table 1: Comparison results of different methods on downstream tasks.

Panel of Figure 3 — Figure 3: Visualization of the learned spatial filters for the Affect, Motor, and Shared encoders after pretraining. Each subplot shows a 2D scalp topography in which warmer colors indicate larger filter weights and cooler colors indicate smaller weights, reflecting the relative contribution of each scalp region to the encoder’s representation. Emotion- and motor-related regions are outlined.

Methods	FACED (9-Class)			BCI-IV-2A (4-Class)
Methods	ACC-B	Kappa	F1-W	ACC-B	Kappa	F1-W
w/o reconstruction	0.5391 ± 0.0151	0.3854 ± 0.0202	0.5401 ± 0.0177	0.4354 ± 0.0329	0.2472 ± 0.0439	0.3758 ± 0.0548
w/o separation	0.5638 ± 0.0138	0.4184 ± 0.0185	0.5603 ± 0.0107	0.4870 ± 0.0184	0.3160 ± 0.0246	0.4725 ± 0.0219
w/ random retrieval	0.5747 ± 0.0098	0.4329 ± 0.0130	0.5674 ± 0.0068	0.5125 ± 0.0308	0.3500 ± 0.0410	0.5052 ± 0.0379
w/o spatial prior	0.5634 ± 0.0271	0.5060 ± 0.0319	0.5672 ± 0.0330	0.5496 ± 0.0143	0.3994 ± 0.0191	0.5495 ± 0.0141
w/ random process cue	0.5269 ± 0.0288	0.4682 ± 0.0341	0.5330 ± 0.0329	0.5232 ± 0.0065	0.3642 ± 0.0087	0.5104 ± 0.0127

Table 2: Ablation studies of the main components on affective and motor-related downstream tasks.

Panel of Figure 4 — Figure 4: Performance of BrainPro, CBraMod, and LaBraM under different levels of channel-drop rate.

	FACED (9-Class)			BCI-IV-2A (4-Class)
Method	ACC-B	Kappa	F1-W	ACC-B	Kappa	F1-W
S	0.5635 ± 0.0031	0.5088 ± 0.0050	0.5710 ± 0.0049	0.5527 ± 0.0218	0.4036 ± 0.0290	0.5504 ± 0.0236
S + A	0.5937 ± 0.0087	0.5418 ± 0.0092	0.6023 ± 0.0061	0.5499 ± 0.0044	0.3999 ± 0.0059	0.5469 ± 0.0060
S + M	0.5828 ± 0.0142	0.5296 ± 0.0143	0.5904 ± 0.0144	0.5674 ± 0.0148	0.4232 ± 0.0198	0.5653 ± 0.0169
S + O	0.5821 ± 0.0144	0.5292 ± 0.0150	0.5881 ± 0.0133	0.5525 ± 0.0173	0.4033 ± 0.0231	0.5491 ± 0.0192
S + A + M	0.5988 ± 0.0057	0.5477 ± 0.0058	0.6098 ± 0.0043	0.5797 ± 0.0229	0.4396 ± 0.0305	0.5764 ± 0.0245

Table 3: Effects of different encoder configurations on downstream performance.

Appendix figures & tables28 assets

Supplementary material from the paper’s appendix.

Appendix

BCI Task	Dataset	$f_{s}$	# Ch.	T	# Samples	Label
I.Emotion Recognition	FACED	250 Hz	32	10 s	10,332	9-class
	SEED-V	1000 Hz	62	1 s	117,744	5-class
	SEED-VII	1000 Hz	62	1 s	281,679	7-class
II.Motor Imagery	BCI-IV-2A	250 Hz	22	4 s	5,088	4-class
	SHU-MI	250 Hz	32	4 s	11,988	2-class
III.Imagined Speech	BCIC2020-3	256 Hz	64	4 s	5,250	5-class

Appendix

Table 4: Overview of downstream BCI tasks and datasets.

Figure 5: The brain region definition used in spatial learning block.

Module	Input	Operation	Process cue	Purpose
Topographic mapping	Dataset montage	Map supported channels to the 60-channel template	$\times$	Establish a common spatial basis
Channel and region retrieval	Temporal features and template indices	Retrieve spatial filters from each encoder using montage-defined indices	$\times$	Associate spatial parameters with scalp locations
Branch routing	EEG segment and process label	Route gradients to the shared and assigned process encoders	$\checkmark$	Learn shared and process-associated representations
Separation loss	Branch representations	Apply margin-based cosine regularization with detached inactive outputs	$\checkmark$	Discourage excessive similarity between branches
Reconstruction	Shared and active branch outputs	Decode the unmasked EEG target in template space using the active decoder	$\checkmark$	Provide a masked reconstruction objective
Region weighting	Reconstruction errors, process priors, and epoch	Apply epoch-dependent channel weights, with uniform weights for the auxiliary group	$\checkmark$	Emphasize coarse process-relevant regions

Appendix

Table 5: Overview of the main modules in BrainPro.

Brain Process Category	Datasets
Affect	Emobrain, SEED, SEED-IV, SEED-GER, SEED-FRA, and SHJT EEG Data
Motor	Grasp and Lift EEG Challenge, EEG Motor Movement/Imagery, KU, and HGD
Others	Raw EEG Data, Resting State EEG Data, SPIS Resting State, TUEP, TUSZ, and TUSL

Appendix

Table 6: Categories of the pretraining datasets.

Type	Factor	Value
Temporal Encoder	Input channels	{1, 32, 32}
	Output channels	{32, 32, 32}
	Kernel size	{15, 3, 3}
	Stride	{1, 1, 1}
	Padding	{7, 1, 1}
Spatial Encoder	Channel-wise Filter Number	32

Appendix

Table 7: Hyperparameters for pretraining.

Dataset	Learning Rate	Hidden Factor	Dropout	Token Selection
FACED	5e-4	8	0.1	All
SEED-V	3e-4	1	0.1	All
SEED-VII	3e-4	8	0.1	All
SHU	1e-4	1	0.1	Aggr
BCI-IV-2A	1e-3	1	0.2	Aggr
Mumtaz2016	1e-4	2	0.1	Aggr

Appendix

Table 8: Tuned hyper-parameters for downstream tasks.

Hyper-parameters	Settings
Epochs	50 (Others) / 30 (ATTEN)
Batch size	64
Optimizer	AdamW
Adam $\beta$	(0.9, 0.999)
Adam $\epsilon$	1e-8
Weight decay	5e-2 / 1e-1 (BCIC2020-3)

Appendix

Table 9: Fixed hyper-parameters for fine-tuning.

Methods	ACC-B	Kappa	F1-W
EEGNet	0.1892 ± 0.0056	0.0562 ± 0.0071	0.1472 ± 0.0123
Conformer	0.2719 ± 0.0032	0.1566 ± 0.0039	0.2712 ± 0.0033
BIOT	0.3005 ± 0.0054	0.1859 ± 0.0070	0.3055 ± 0.0069
EEGPT	0.2213 ± 0.0060	0.0954 ± 0.0071	0.2147 ± 0.0120
LaBraM	0.3346 ± 0.0122	0.2256 ± 0.0146	0.3391 ± 0.0121
CBraMod	0.3318 ± 0.0056	0.2236 ± 0.0065	0.3384 ± 0.0059

Appendix

Table 10: Results on the SEED-VII dataset.

Methods	ACC-B	AUC-PR	AUROC
EEGNet	0.9113 ± 0.0104	0.9632 ± 0.0045	0.9512 ± 0.0096
Conformer	0.8422 ± 0.0344	0.9442 ± 0.0502	0.9613 ± 0.0064
BIOT	0.8789 ± 0.0190	0.9744 ± 0.0083	0.9664 ± 0.0136
EEGPT	0.8475 ± 0.0933	0.9695 ± 0.0076	0.9669 ± 0.0069
LaBraM	0.8986 ± 0.0018	0.9791 ± 0.0041	0.9754 ± 0.0050
CBraMod	0.8909 ± 0.0037	0.9769 ± 0.0047	0.9726 ± 0.0062

Appendix

Table 11: Results on the Major Depressive Disorder dataset.

Methods	ACC-B	Kappa	F1-W
EEGNet	0.2755 ± 0.0183	0.0943 ± 0.0228	0.2737 ± 0.0180
Conformer	0.3339 ± 0.0104	0.1673 ± 0.0130	0.3269 ± 0.0153
BIOT	0.3016 ± 0.0206	0.1270 ± 0.0257	0.2999 ± 0.0217
EEGPT	0.2787 ± 0.0061	0.0983 ± 0.0076	0.2758 ± 0.0060
LaBraM	0.4880 ± 0.0161	0.3600 ± 0.0201	0.4871 ± 0.0165
CBraMod	0.5061 ± 0.0211	0.3827 ± 0.0263	0.5050 ± 0.0212

Appendix

Table 12: Results on the Speech dataset.

Drop	BrainPro		CBraMod		LaBraM
	Mean $\pm$ std	$\Delta$	Mean $\pm$ std	$\Delta$	Mean $\pm$ std	$\Delta$
ACC-B $\uparrow$
0%	$\mathbf{0.5937}\pm 0.0087$	$0$	$0.5669\pm 0.0094$	$0$	$0.5224\pm 0.0116$	$0$
10%	$\mathbf{0.5747}\pm 0.0014$	$-0.0190$	$0.5218\pm 0.0042$	$-0.0451$	$0.4972\pm 0.0036$	$-0.0252$
20%	$\mathbf{0.5373}\pm 0.0015$	$-0.0564$	$0.4890\pm 0.0054$	$-0.0779$	$0.4725\pm 0.0086$	$-0.0499$
30%	$\mathbf{0.4995}\pm 0.0056$	$-0.0942$	$0.4473\pm 0.0035$	$-0.1196$	$0.4230\pm 0.0085$	$-0.0994$

Appendix

Table 13: Robustness under random channel dropping on FACED. Values are mean

\pm

std;

\Delta

denotes the absolute score change relative to no channel dropping. Bold indicates the highest mean in each row.

Method	FACED (9-Class)			SHU-MI (2-Class)
Method	ACC-B	Kappa	F1-W	ACC-B	AUC-PR	AUROC
BIOT (Fixed)	0.2775 $\pm$ 0.0318	0.1839 $\pm$ 0.0512	0.2599 $\pm$ 0.0271	0.5123 $\pm$ 0.0206	0.5215 $\pm$ 0.0197	0.5146 $\pm$ 0.0264
LaBraM (Fixed)	0.3004 $\pm$ 0.0458	0.2377 $\pm$ 0.0617	0.2943 $\pm$ 0.0375	0.5233 $\pm$ 0.0272	0.5329 $\pm$ 0.0284	0.5218 $\pm$ 0.0275
CBraMod (Fixed)	0.3146 $\pm$ 0.0346	0.2579 $\pm$ 0.0542	0.3077 $\pm$ 0.0298	0.5217 $\pm$ 0.0247	0.5304 $\pm$ 0.0253	0.5238 $\pm$ 0.0317
BrainPro (Fixed)	0.3581 $\pm$ 0.0052	0.2749 $\pm$ 0.0060	0.3564 $\pm$ 0.0055	0.5992 $\pm$ 0.0104	0.6289 $\pm$ 0.0018	0.6483 $\pm$ 0.0023

Appendix

Table 14: Linear probing results on representative affective and motor-related downstream tasks.

Figure 6: Effects of learning rate on BrainPro fine-tuning.

Encoder No.	FACED				BCI-IV-2A
Encoder No.	ACC-B	Kappa	F1-W	MLP Ratio	ACC-B	Kappa	F1-W	MLP Ratio
1	0.5635 ± 0.0031	0.5088 ± 0.0050	0.5710 ± 0.0049	0.51	0.5527 ± 0.0218	0.4036 ± 0.0290	0.5504 ± 0.0236	0.59
1	0.5808 ± 0.0048	0.5255 ± 0.0068	0.5840 ± 0.0079	1.02	0.5403 ± 0.0206	0.3871 ± 0.0275	0.5368 ± 0.0245	1.16
2	0.5937 ± 0.0087	0.5418 ± 0.0092	0.6023 ± 0.0061	1.00	0.5674 ± 0.0148	0.4232 ± 0.0198	0.5653 ± 0.0169	1.00
3	0.5938 ± 0.0069	0.5427 ± 0.0069	0.6039 ± 0.0042	0.93	0.5889 ± 0.0165	0.4519 ± 0.0220	0.5868 ± 0.0157	1.07
3	0.5988 ± 0.0057	0.5477 ± 0.0058	0.6098 ± 0.0043	1.49	0.5797 ± 0.0229	0.4396 ± 0.0305	0.5764 ± 0.0245	1.41

Appendix

Table 15: Effect of encoder count and relative MLP size on performance. MLP Ratio is computed w.r.t. BrainPro.

Figure 7: Effects of noisy samples on BrainPro pretraining.

Setting	BCI-IV-2A			Mental Arithmetic
Setting	ACC-B	Kappa	F1-W	ACC-B	AUC-PR	AUROC
w/o reset	0.5222 ± 0.0406	0.3630 ± 0.0541	0.5109 ± 0.0490	0.8017 ± 0.0122	0.8955 ± 0.0113	0.8471 ± 0.0184
reset	0.5674 ± 0.0148	0.4232 ± 0.0198	0.5653 ± 0.0169	0.8083 ± 0.0156	0.8980 ± 0.0052	0.8512 ± 0.0083

Appendix

Table 16: Effect of resetting temporal position embedding on BCI-IV-2A and Mental Arithmetic datasets.

Token Merge	BCI-VI-2A			Mental Arithmetic
Token Merge	ACC-B	Kappa	F1-W	ACC-B	AUC-PR	AUROC
Aggr	0.5674 $\pm$ 0.0148	0.4232 $\pm$ 0.0198	0.5653 $\pm$ 0.0169	0.7367 $\pm$ 0.0315	0.8353 $\pm$ 0.0175	0.7822 $\pm$ 0.0141
Mean	0.4090 $\pm$ 0.0470	0.2120 $\pm$ 0.0627	0.3664 $\pm$ 0.0703	0.8083 $\pm$ 0.0156	0.8980 $\pm$ 0.0052	0.8512 $\pm$ 0.0083
All	0.5373 $\pm$ 0.0315	0.3831 $\pm$ 0.0421	0.5295 $\pm$ 0.0335	0.5883 $\pm$ 0.0347	0.6753 $\pm$ 0.0131	0.6293 $\pm$ 0.0160

Appendix

Table 17: Comparison of different token selection strategies on BCI-VI-2A and Mental Arithmetic datasets.

Method	Parameters	FLOPs
EEGNet	0.005M	10.03M
Conformer	0.17M	50.09M
BIOT	3.28M	229.35M
EEGPT	51.66M	9.31G
LaBraM	9.43M	392.68M
CBraMod	8.45M	350.79M

Appendix

Table 18: Comparison of model complexity.

Panel of Figure 8 — Figure 8: Visualization of learned filters for three groups of encoders after pretraining: Affect (top row), Motor (middle row), and Shared (bottom row).

Figure 9: SEED-VII analysis. (a) Performance of three branch configurations. (b) Class recall (7 classes) calculated as

n_{ii}/\sum_{j}n_{ij}

from single-run confusion matrices.

Figure 10: Effect of pretraining data size on FACED and BCI-IV-2A. Each panel uses a restricted vertical range to show performance changes.

Figure 11: Effect of the separation margin on FACED and BCI-IV-2A. The three panels show ACC-B, Kappa, and F1-W.

Figure 12: Comparison of regional weighting schedules on FACED and BCI-IV-2A. The settings use uniform weighting, fixed sharpness of 15, and epoch-dependent sharpness.

Figure 13: Comparison of spatial retrieval configurations on FACED and BCI-IV-2A. Region only uses regional retrieval. Channel only uses channel retrieval. Both combines the two spatial scales.

Figure 14: Comparison of feature fusion methods on FACED and BCI-IV-2A. The alternatives are addition, learned weighting, and concatenation.

Figure 15: Linear probes on frozen representations. (a) Dataset identity prediction within each process family. The matched process branches are A, M, and O. Panels (b), (c), and (d) show process prediction with one dataset excluded from probe training. Dataset identity comparisons are made within each family. The dashed line indicates chance balanced accuracy of

1/3

Figure 16: Branch exposure after data cleaning. Bars show each process family’s proportion of retained pretraining windows. Labels give the corresponding window counts.

Figure 17: Montage coverage for (a) 19 pretraining configurations and (b) nine downstream datasets. Skyblue indicates retained channels. Light gray indicates discarded channels. Labels give retained and input channel counts. Retained template indices are unique in all evaluated mappings. Repeated TUEP and TUSZ rows represent distinct montage configurations.

EEG Foundation Models

Electroencephalography (EEG) foundation models learn reusable representations of brain signals for transfer across people, tasks, and recording setups.

See all 100 papers

EEG-PRIME: Prototype-Aligned Representation Learning with Multi-Level Conditioning for EEG Decoding
Aug 13, 2026Shuailei Zhang, Muyun Jiang, Wei Zhang +5Cross-Subject EEG Decoding Zero-Shot Learning
College of Computing and Data Science, Nanyang Technological University, Singapore · Southeast University, China · Centre for AI in Medicine, Nanyang Technological University, Singapore
DARE-EEG: A Foundation Model for Mining Dual-Aligned Representation of EEG
May 18, 2026Yang Shao, Peiliang Gong, Qun Dai +1Masked Autoencoders Self-Supervised Pre-Training
College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics Nanjing, Jiangsu, China
ProtoEEG: Structured Prototype Alignment for EEG Visual Decoding
Oct 20, 2026Kieren Yu, Chang Huang, Ziyang Liu +1EEG-Based Visual Decoding Cross-Modal Alignment

Aug 13, 2026J/K move · Enter open · S save

EEG-PRIME: Prototype-Aligned Representation Learning with Multi-Level Conditioning for EEG Decoding

Shuailei Zhang, Muyun Jiang, Wei Zhang, Jinbo Chen+4

College of Computing and Data Science, Nanyang Technological University, Singapore · Southeast University, China · Centre for AI in Medicine, Nanyang Technological University, Singapore

Cross-Subject EEG Decoding Zero-Shot Learning EEG Decoding Prototype Learning Cross-Session EEG Decoding EEG Foundation Models

Open paper AskPDF arXiv

Electroencephalography (EEG) decoding models often generalize poorly across datasets and subjects due to domain shifts in acquisition protocols and individual neurophysiology. We propose EEG-PRIME, a two-stage EEG foundation model for cross-dataset multi-task decoding. EEG-PRIME combines masked pretraining with prototype-aligned instruction tuning to enable instruction-aware and subject-invariant decoding across diverse BCI paradigms. During pretraining, an EEG encoder learns transferable representations through masked reconstruction with frequency-cutoff spectral augmentation. During instruction tuning, EEG-PRIME incorporates task-semantic, dataset-specific, and subject-invariant conditioning. The resulting conditioning signal modulates the Q-Former through Layer-wise Query Modulation, while frozen text embeddings of class labels serve as prototypes for cosine-similarity-based prediction across heterogeneous label spaces. Experiments on sixteen datasets covering motor imagery, emotion recognition, ADHD detection, covert speech, and mental workload show consistent improvements over state-of-the-art baselines and prior EEG foundation models under cross-subject settings. On two additional held-out datasets, EEG-PRIME achieves balanced accuracy comparable to within-session calibration models without target-domain optimization, calibration, or linear probing, demonstrating promising zero-shot transfer capability.

Abstract

EEG Foundation Models

Explore similar work

EEG-PRIME: Prototype-Aligned Representation Learning with Multi-Level Conditioning for EEG Decoding

DARE-EEG: A Foundation Model for Mining Dual-Aligned Representation of EEG

ProtoEEG: Structured Prototype Alignment for EEG Visual Decoding