HyBRA-EEG: Hyperbolic Global-Local Brain-Region Aggregation for Self-Supervised EEG Representation Learning

Xin Liu; Feng Lu; Wei Li

cs.LGOct 12, 2026

HyBRA-EEG: Hyperbolic Global-Local Brain-Region Aggregation for Self-Supervised EEG Representation Learning

Authors: Xin Liu, Feng Lu, Wei Li

Abstract

Electroencephalography (EEG) representation learning is challenged by noisy signals, inter-subject variability, and heterogeneous electrode montages in different datasets and recording settings. Recent self-supervised EEG pretraining methods improve transferability by converting heterogeneous recordings into unified token or sequence representations, but channel-aware designs may still encode montage- or dataset-specific regularities. We propose HyBRA, a self-supervised framework based on hyperbolic global-local brain-region aggregation. HyBRA maps heterogeneous channel observations into a fixed regional token space by combining electrode coordinates, region priors, and feature-dependent hyperbolic distances. A channel-evidence injection branch preserves fine-grained discriminative cues during aggregation. The resulting tokens are processed by a Region-Time Transformer and pretrained with masked reconstruction of latent features and region-level signals. HyBRA outperforms the compared methods in average balanced accuracy on six benchmark datasets. Under a controlled six-fold evaluation on OpenBMI-MI, HyBRA matches CSBrain in balanced accuracy, improves AUROC by 1.84 percentage points and AUPRC by 5.23 percentage points, and uses 96.8% fewer parameters. Matched comparisons on TUAB and TUEV show higher balanced accuracy for hyperbolic matching than for Euclidean distance or cosine similarity.

Figures & tables

Figure 1: Conceptual comparison between a schematic uniform regional aggregation baseline and HyBRA. The baseline assigns equal importance to channels within each region and may attenuate informative activity. HyBRA combines feature-dependent hyperbolic channel-to-region matching with a direct channel-evidence pathway to form fused region tokens. Waveforms are schematic.

Figure 2: Overview of HyBRA-EEG. Coordinate-aware channel patches enter parallel prior-constrained hyperbolic matching and channel-evidence injection paths. Their sum forms fixed region tokens, which are modeled by the Region-Time Transformer. Self-supervised pretraining predicts masked latent features and reconstructs region-level signals.

Phase	Dataset	Type	# Ch.	# Samp.	Dur.
Pre-train	BETA (2s)	SSVEP	60	2,400	2s
	BETA (3s)	SSVEP	60	8,800	3s
	Tsinghua SSVEP	SSVEP	60	8,400	5s
	Cho2017	MI	64	10,520	3s
	PhysioNet-MI	MI	64	9,270	4s
	FACED	EMO	30	33,880	3s

Table 1: Datasets used for pretraining and downstream fine-tuning. All signals are resampled to 200 Hz.

Method	TUAB			TUEV
Method	Balanced Acc.	AUPRC	AUROC	Balanced Acc.	Cohen’s Kappa	Weighted-F1
EEGNet	0.7642 $\pm$ 0.0036	0.8299 $\pm$ 0.0043	0.8412 $\pm$ 0.0031	0.3876 $\pm$ 0.0143	0.3577 $\pm$ 0.0155	0.6539 $\pm$ 0.0120
EEGConformer	0.7758 $\pm$ 0.0049	0.8427 $\pm$ 0.0054	0.8445 $\pm$ 0.0038	0.4074 $\pm$ 0.0164	0.3967 $\pm$ 0.0195	0.6983 $\pm$ 0.0152
SPaRCNet	0.7896 $\pm$ 0.0018	0.8414 $\pm$ 0.0018	0.8676 $\pm$ 0.0012	0.4161 $\pm$ 0.0262	0.4233 $\pm$ 0.0181	0.7024 $\pm$ 0.0104
ContraWR	0.7746 $\pm$ 0.0041	0.8421 $\pm$ 0.0104	0.8456 $\pm$ 0.0074	0.4384 $\pm$ 0.0349	0.3912 $\pm$ 0.0237	0.6893 $\pm$ 0.0136
CNN-Transformer	0.7777 $\pm$ 0.0022	0.8433 $\pm$ 0.0039	0.8461 $\pm$ 0.0013	0.4087 $\pm$ 0.0161	0.3815 $\pm$ 0.0134	0.6854 $\pm$ 0.0293

Table 2: Downstream results on six EEG datasets, reported as mean

\pm

standard deviation. For TUAB and TUEV, task-specific baselines ( Lawhern et al. 2018 ; Song et al. 2023 ; Jing et al. 2023 ; Yang et al. 2023b ; Peh et al. 2022 ; Li et al. 2022 ; Song et al. 2021 ) and EEG foundation models form separate comparison groups; for the remaining datasets, all EEG foundation models are compared together. Bold and underlined values denote the best and second-best results within each group.

	OpenBMI-MI		MentalArithmetic
Method	BAcc	Params	BAcc	Params
HyBRA	81.17 $\pm$ 4.36	1.45 (1.0 $\times$ )	76.63 $\pm$ 5.01	1.45 (1.0 $\times$ )
CSBrain	81.04 $\pm$ 3.51	44.86 (30.9 $\times$ )	65.38 $\pm$ 9.64	28.06 (19.3 $\times$ )
LaBraM	70.13 $\pm$ 2.84	5.82 (4.0 $\times$ )	69.64 $\pm$ 5.08	5.82 (4.0 $\times$ )
BIOT	69.08 $\pm$ 3.18	3.19 (2.2 $\times$ )	64.16 $\pm$ 7.03	3.19 (2.2 $\times$ )
EEGPT	68.10 $\pm$ 3.54	25.32 (17.4 $\times$ )	55.12 $\pm$ 3.68	25.32 (17.4 $\times$ )

Table 3: BAcc (%) and parameter counts on OpenBMI-MI and MentalArithmetic. BAcc is reported as mean

\pm

standard deviation over six subject-group folds. Parameters are in millions, with ratios relative to HyBRA in parentheses. Bold values indicate the highest BAcc or smallest model.

		Number of brain regions
Dataset	Metric	5	10	11
TUAB	BAcc (%)	81.82 $\pm$ 0.57	81.17 $\pm$ 0.74	82.24 $\pm$ 0.21
TUAB	Kappa (%)	63.59 $\pm$ 1.17	62.34 $\pm$ 1.51	64.38 $\pm$ 0.45
OpenBMI-MI	BAcc (%)	72.25 $\pm$ 4.82	80.90 $\pm$ 4.32	81.17 $\pm$ 4.36
OpenBMI-MI	Kappa (%)	44.51 $\pm$ 9.64	61.81 $\pm$ 8.64	62.33 $\pm$ 8.71

Table 4: Effect of brain-region granularity on TUAB and OpenBMI-MI. The 5-, 10-, and 11-region partitions use coarse lobar groups, bilateral lobar groups, and bilateral groups with a midline token, respectively. Results are mean

\pm

standard deviation over five random seeds for TUAB and six subject-group folds for OpenBMI-MI. Bold values indicate the best result.

Module	Setup	TUAB	TUEV
Coordinate-Aware Patch Encoding	w/o coordinate encoding	0.8057 $\pm$ 0.0051	0.6555 $\pm$ 0.0340
Hyperbolic Region Mixer	Euclidean distance-based matching	0.8134 $\pm$ 0.0078	0.6509 $\pm$ 0.0214
	Cosine similarity-based matching	0.8164 $\pm$ 0.0035	0.5989 $\pm$ 0.0206
	w/o brain-region tokenization	0.7980 $\pm$ 0.0073	0.6166 $\pm$ 0.0175
	w/o channel evidence injection	0.8123 $\pm$ 0.0049	0.6468 $\pm$ 0.0428
Region-Time Transformer	region-only attention	0.8137 $\pm$ 0.0028	0.6814 $\pm$ 0.0089

Table 5: Ablation BAcc on TUAB and TUEV, reported as mean

\pm

standard deviation over five seeds. The Euclidean and cosine variants replace only the matching metric. The variant without brain-region tokenization applies the same axial Transformer and SSL objectives directly to channel-time tokens, while the variant without SSL pretraining trains HyBRA from random initialization. Bold values indicate the best result.

Figure 3: Region attribution and token perturbation on OpenBMI-MI. (a) Subject-averaged normalized absolute Integrated Gradients (IG) attribution for each motor imagery class. (b) Subject-level BAcc after replacing the most-attributed, randomly selected, or least-attributed region tokens in each trial with the corresponding fold-specific training means. Shaded bands indicate bootstrap 95% confidence intervals across 54 subjects.

Figure 4: t-SNE visualization of downstream representations on BCIC-IV 2a. Each point denotes one EEG trial, and colors indicate motor-imagery classes. Each panel reports the Silhouette coefficient and Davies–Bouldin index computed on the displayed two-dimensional embedding.

Electroencephalography

Electroencephalography (EEG) studies brain-generated electrical activity to understand cognition and support diagnosis, monitoring, and brain-computer interfaces.

See all 183 papers

MSBraM: A Multi-scale Self-supervised Brain Foundation Model for Hierarchical EEG Dynamics Learning
Jul 23, 2026Tao Zhou, Jing Han, Lingyu Shu +1Electroencephalography Self-Supervised Pre-Training
Hunan University Changsha, China
RECTOR: Masked Region-Channel-Temporal Modeling for Affective and Cognitive Representation Learning
Jun 13, 2026Jinhan Liu, Mahsa ShoaranRepresentation Learning Self-Supervised Time Series Representation Learning
Institutes of Electrical and Micro Engineering and Neuro-X, EPFL, Geneva, Switzerland.
EEGDM: Learning EEG Representation with Latent Diffusion Model
Aug 28, 2025Shaocong Wang, Tong Liu, Yihan Li +6Electroencephalography EEG Signal Generation
Tsinghua University · Tianjin University

Jul 23, 2026J/K move · Enter open · S save

MSBraM: A Multi-scale Self-supervised Brain Foundation Model for Hierarchical EEG Dynamics Learning

Tao Zhou, Jing Han, Lingyu Shu, Zixing Zhang

Hunan University Changsha, China

Electroencephalography Self-Supervised Pre-Training Hierarchical Representation Learning Multiscale Time-Series Modeling EEG Foundation Models

Open paper AskPDF arXiv

Self-supervised foundation models have recently shown strong potential for electroencephalogram (EEG)-based analysis. However, existing approaches struggle to capture the inherently multi-scale temporal structure of EEG signals, where local neural patterns and long-range dependencies jointly encode task-relevant information. This limitation hampers cross-scale representation learning and generalization across diverse downstream tasks. To address this challenge, we propose MSBraM, a Multi-Scale self-supervised Brain foundation Model designed to learn hierarchical EEG representations. MSBraM follows a two-stage pretraining framework. First, a multi-scale neural tokenizer discretizes raw EEG signals into semantic codes at different temporal resolutions via vector-quantized reconstruction. Second, the model is pretrained to predict masked codes using a curriculum multi-scale masking strategy, progressively integrating fine-grained local patterns with global temporal context. We pretrain MSBraM on over 2,400 hours of EEG data and evaluate it across 10 downstream tasks on 12 public datasets. Extensive experiments show that MSBraM achieves superior performance on other state-of-the-art pretrained models, demonstrating strong generalization and transferability. These results indicate that explicitly modeling multi-scale temporal dynamics is critical for effective EEG foundation models.

Abstract

Electroencephalography

Explore similar work

MSBraM: A Multi-scale Self-supervised Brain Foundation Model for Hierarchical EEG Dynamics Learning

RECTOR: Masked Region-Channel-Temporal Modeling for Affective and Cognitive Representation Learning

EEGDM: Learning EEG Representation with Latent Diffusion Model