Brain foundation model-guided source-selective domain adaptation for cross-subject EEG decoding

Jinzhou Wu; Baoping Tang; Qikang Li; Yi Wang; Cheng Li; Shujian Yu

cs.LGJul 28, 2025

Brain foundation model-guided source-selective domain adaptation for cross-subject EEG decoding

Authors: Jinzhou Wu, Baoping Tang, Qikang Li, Yi Wang, Cheng Li, Shujian Yu

Organizations: State Key Laboratory of Mechanical Transmission, College of Mechanical and Vehicle Engineering, Chongqing University, Chongqing 400044, China · Department of Computer Science, Vrije Universiteit Amsterdam, 1081 HV Amsterdam, The Netherlands · Machine Learning Group, UiT–The Arctic University of Norway, 9019 Tromsø, Norway

Abstract

Cross-subject motor-imagery electroencephalography (MI-EEG) decoding remains challenging because substantial inter-subject variability can cause both negative transfer from poorly matched source subjects and persistent distribution discrepancies between source and target domains. Existing multi-source domain adaptation methods often incorporate all available source domains or estimate source relevance using signal-level or task-specific representations, while distribution alignment is frequently performed only at the feature level. These limitations may introduce irrelevant source knowledge and fail to preserve class-discriminative structures across subjects. In this study, we propose a brain foundation model-guided multi-source domain adaptation framework (BFM-MSDA) for cross-subject MI-EEG decoding. The method first utilizes representations learned by a pretrained brain foundation model to estimate source--target compatibility and retrieve target-relevant sources. Subsequently, a relevance-weighted dual alignment strategy is applied to the selected sources and the unlabeled target domain. Specifically, Cauchy--Schwarz (CS) divergence is used to reduce discrepancies in marginal feature distributions, while conditional Cauchy--Schwarz (CCS) divergence further aligns class-dependent decision distributions. Source relevance is incorporated into both alignment terms so that more transferable source subjects contribute more strongly to adaptation. Experiments on two public MI-EEG benchmarks achieve average accuracies of 86.16% and 78.41%, outperforming representative cross-subject decoding and domain adaptation methods. Additional experiments with a large source pool further demonstrate that target-aware source retrieval improves scalability while mitigating negative transfer. These results highlight the importance of informed source selection under cross-subject distribution shift.

Figures & tables

Figure 1: Overview of the proposed BFM-MSDA framework for subject-independent MI decoding: (a) source selection with a pretrained BFM, where trial-level embeddings $\mathbf{h}_{1},\ldots,\mathbf{h}_{S}$ and $\mathbf{h}_{T}$ yield pairwise CS distances $d_{S_{i},T}$ used to retain target-relevant sources; (b) MSDA with feature-level alignment (FLA) via weighted CS on $p(z)$ and decision-level alignment (DLA) via conditional CS on $p(y|z)$ over selected source–target and source–source pairs; (c) testing, where the trained feature extractor and classifier predict MI labels.

Methods	Subject									Avg. acc.	Kappa
Methods	1	2	3	4	5	6	7	8	9	Avg. acc.	Kappa
EEGNet	72.92	64.58	79.17	78.47	86.11	74.31	83.33	79.86	82.64	77.93 ∗∗ $\pm$ 6.52	0.5586
ShallowConvNet	75.73	74.89	78.47	76.39	81.25	85.43	87.50	83.33	76.39	79.93 ∗∗ $\pm$ 4.62	0.5986
Conformer	81.59	68.53	88.07	82.04	74.68	78.54	86.97	82.59	83.74	80.75 ∗∗ $\pm$ 6.10	0.6150
JAN	79.69	76.32	84.25	75.69	86.81	84.72	89.19	90.28	81.94	83.21 ∗ $\pm$ 5.25	0.6642
DJP-MMD	79.78	72.22	89.58	81.94	85.42	83.33	88.89	89.58	86.28	84.11 ∗ $\pm$ 5.64	0.6823

Table 1: Cross-subject accuracies (mean

\pm

std. in %) and kappa on Dataset I. Best performances are highlighted in bold. ∗ represents

p<0.05

and ∗∗ represents

p<0.01

Methods	Subject										Avg. acc.	Kappa
Methods	1	2	3	4	5	6	7	8	9	10	Avg. acc.	Kappa
EEGNet	66.57	60.96	82.70	70.14	73.89	59.05	55.68	63.56	88.29	65.06	68.59 ∗∗ $\pm$ 10.41	0.3718
ShallowConvNet	73.92	67.63	75.05	71.55	75.23	63.65	70.56	67.70	78.02	66.37	70.97 ∗∗ $\pm$ 4.60	0.4194
Conformer	77.62	71.01	78.80	74.12	78.99	66.83	74.09	71.09	81.92	69.69	74.42 ∗∗ $\pm$ 4.82	0.4883
JAN	78.27	71.38	80.72	73.75	79.46	76.84	73.05	70.30	87.35	70.28	76.14 ∗∗ $\pm$ 5.47	0.5228
DJP-MMD	76.31	74.43	78.26	71.85	73.62	77.94	70.22	71.43	82.45	71.52	74.80 ∗∗ $\pm$ 3.89	0.4961

Table 2: Cross-subject accuracies (mean

\pm

std. in %) and kappa on Dataset II. Best performances are highlighted in bold. ∗ represents

p<0.05

and ∗∗ represents

p<0.01

Group	Subject										Avg. acc.
Group	1	2	3	4	5	6	7	8	9	10	Avg. acc.
Random Group 1	66.42	63.57	80.65	66.50	79.51	67.83	59.52	66.36	89.52	56.23	69.61 ∗∗ $\pm$ 10.37
Random Group 2	73.85	74.31	79.27	73.03	75.65	70.95	71.18	73.54	85.42	74.85	75.21 ∗ $\pm$ 4.29
Random Group 3	72.38	70.15	71.33	72.27	72.92	73.65	78.63	71.81	88.71	74.75	74.66 ∗ $\pm$ 5.45
Random Group 4	71.38	72.81	79.52	65.15	78.81	71.57	69.82	70.55	89.17	65.67	73.44 ∗∗ $\pm$ 7.24
Random Group 5	72.94	67.35	73.17	67.24	75.64	68.32	70.72	74.36	88.46	66.75	72.50 ∗∗ $\pm$ 6.47

Table 3: Cross-subject performances of different source groups from the GigaDB dataset. Best performances are highlighted in bold. ∗ represents

p<0.05

and ∗∗ represents

p<0.01

Figure 2: Effect of selection threshold on performance and training time.

Dataset	Mean Spearman $\bar{\rho}$	$p$ -value	Targets
Dataset I	$-0.304$	$0.016$	9
Dataset II	$-0.265$	$0.018$	10

Table 4: Mean within-target Spearman correlation

\bar{\rho}

between BFM distance and single-source transferability.

\bar{\rho}

is the average of the per-target rank correlations

\rho_{t}

. The

p

-value is from a blocked permutation test of

\bar{\rho}

Panel of Figure 3 — Figure 3: BFM distance against single-source transfer accuracy after centring accuracy within each target. Each point represents an ordered source–target pair, and the solid line is a least-squares fit. Larger BFM distance is associated with lower transferability on both (a) Dataset I and (b) Dataset II.

Panel of Figure 4 — Figure 4: Cross-subject classification accuracies of ablation variants. (a) Dataset I. (b) Dataset II.

Dataset	Baseline	CS/CCS	KDE-KL
Dataset I	77.93 $\pm$ 6.52	86.16 $\pm$ 5.37	79.34 $\pm$ 8.46
Dataset II	68.59 $\pm$ 10.41	78.41 $\pm$ 6.57	71.37 $\pm$ 9.24

Table 5: Cross-subject performance comparison between CS and KDE-based KL divergences.

Panel of Figure 5 — Figure 5: Heat maps and MDS visualizations of pairwise distances for Datasets I and II. (a) Heat map of inter-subject distances for Dataset I. (b) MDS visualization of inter-subject distances for Dataset I. (c) Heat map of inter-subject distances for Dataset II. (d) MDS visualization of inter-subject distances for Dataset II.

Figure 6: t-SNE visualization of target domain features extracted by the feature extractor for subject 1 from Dataset I (top row) and Dataset II (bottom row). (a) and (d) baseline EEGNet without DA, (b) and (e) MSDA without source selection, and (c) and (f) full BFM-MSDA framework.

Panel of Figure 7 — Figure 7: EEG topography for subject 1 from Dataset I with different MI tasks. (a) Raw EEG signals averaged across all trials. (b) Grad-CAM of extracted features after BFM-MSDA.

Panel of Figure 8 — Figure 8: Parameter sensitivity analysis. (a) Influence of balancing factors for Dataset I. (b) Influence of balancing factors for Dataset II. (c) Influence of transition point $\tau_{0}$ for Dataset I. (d) Influence of transition point $\tau_{0}$ for Dataset II.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Analysis	$\bar{\rho}$	$p$	Blocks
I	Within source	$-0.283$	$0.026$	9
I	Pooled (72 pairs)	$-0.350$	$0.002$	–
II	Within source	$-0.132$	$0.250$	10
II	Pooled (90 pairs)	$-0.182$	$0.088$	–

Appendix

Table C.1: Supplementary rank correlations between BFM distance and single-source transferability under alternative pair groupings. Negative values indicate that nearer sources transfer better.

Figure D.1: Target accuracy against the number of selected sources on Dataset I, without domain adaptation. BFM-guided selection is above random selection at every

k

below the full pool and peaks at

k=6

. The dotted line marks the best accuracy random selection attains at any

k

. Error bars are the standard error across targets.

Figure D.2: Target accuracy against the number of selected sources on Dataset II, without domain adaptation. Accuracy falls

4.3

points from its maximum at

k=6

k=9

. Error bars are the standard error across targets.

Target	LaBraM-based selection		CBraMod-based selection		Overlap
	Selected sources	Acc. (%)	Selected sources	Acc. (%)	(%)
S1	S3, S4, S7	83.19	S3, S4, S8	84.27	50.0
S2	S1, S4, S6, S8	76.53	S1, S3, S5, S8	74.32	33.3
S3	S2, S6, S8	91.74	S4, S6, S8	89.08	50.0
S4	S1, S5, S8	83.26	S1, S3, S5, S6	82.15	40.0
S5	S3, S4, S7	89.37	S4, S7, S9	86.94	50.0

Appendix

Table F.1: Adaptive soft-gating source sets and downstream BFM-MSDA accuracies on Dataset I when source–target distances come from LaBraM or from CBraMod. Overlap is the Jaccard index. LaBraM accuracies match Table 1 .

Figure F.1: Inter-subject CS distances on Dataset I. (a) CBraMod embeddings. (b) LaBraM embeddings.

Explore similar work

Apr 21, 2026eess.SP

Foundation Model Guided Dual-Branch Co-Adaptation for Source-Free EEG Decoding

Source-free domain adaptation (SFDA) provides a practical solution to cross-subject EEG decoding by adapting source-pretrained models to unlabeled target domains without accessing source data. However, existing SFDA methods rely solely on the limited internal knowledge of source-pretrained models, leading to inferior cross-domain generalization and unreliable pseudo-labels. Although EEG Foundation Models (FMs) pretrained on large-scale data exhibit strong generalizability, their potential in SFDA remains largely unexplored. To this end, we propose FUSED, a Foundation-guided Source-free EEG Decoding framework that integrates a large-scale FM with a compact Specialist Model (SM) via dual-branch co-adaptation. Specifically, we introduce a Co-adaptation mechanism equipping both branches with linear and prototype views, enabling cross-branch pseudo-label generation. Additionally, we design a Consensus Filtering Mechanism that exploits the FM's inherent stability to identify high-quality samples, along with a Two-Stage Pseudo-Label Refinement scheme to suppress error accumulation through cross-branch arbitration. Finally, we calibrate the FM's decision boundaries via mutual information maximization with the SM, followed by knowledge distillation from FM to SM, forming a principled calibrate-then-distill pipeline. To our knowledge, FUSED is the first work to leverage EEG FMs within the SFDA framework for cross-subject EEG decoding. Extensive experiments across three EEG paradigms, including motor imagery, emotion recognition, and SSVEP, demonstrate consistent state-of-the-art performance, validating the effectiveness of foundation-guided synergy for robust and privacy-preserving EEG decoding.

Peiliang Gong, Han Zhang, Zhen Jiang +5

College of Computing and Data Science, Nanyang Technological University, 50 Nanyang Avenue, 639798, Singapore · College of Artifical Intelligence and Automation, Hohai University, Changzhou 213022, China · College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China +2

Cross-Subject EEG Decoding Source-Free Domain Adaptation EEG Decoding EEG Foundation Models

Ask PDF

Sep 27, 2026cs.HC

EEG-Fusion: Failure-Informed Source-Free Expert Routing for Robust Motor Imagery EEG Decoding

Subject-independent motor-imagery (MI) EEG decoding can exhibit subject-level failures even when average performance appears acceptable: under subject shift, a decoder can become an overconfident near-one-class predictor. This is especially problematic in source-free deployment, where target-user labels are unavailable during adaptation and expert selection. We present \textit{EEG-Fusion}, a failure-informed decision-level fusion framework that treats source-free MI decoding as label-free reliability estimation over heterogeneous experts. EEG-Fusion applies subject-wise Euclidean alignment and normalization-only test-time adaptation, then routes each target subject to a neural, covariance-based, or physiological-feature expert using a reliability gate trained on source-held-out folds to predict expert performance and collapse risk from label-free stream diagnostics. The gate uses confidence, entropy, prediction diversity, expert agreement, and predicted class balance; collapse is measured as the maximum predicted class fraction. In 9-fold leave-one-subject-out (LOSO) evaluation with three seeds, relative to a no-alignment raw EEGNet source-free anchor, EEG-Fusion improves subject macro-F1 from 0.417 to 0.529 on BCI IV-2a local protocol, from 0.314 to 0.482 on BNCI2014-001, and from 0.607 to 0.708 on BNCI2014-004; corresponding collapse-index reductions are 0.199, 0.227, and 0.169. In a 9-subject Cho2017 external subset, EEG-Fusion improves macro-F1 from 0.516 to 0.630. These results suggest that label-free reliability estimation can reduce subject-level failure modes in source-free MI-EEG deployment.

Abdul Basit, Saim Rehman, Muhammad Shafique

eBRAIN Lab, Division of Engineering New York University (NYU) Abu Dhabi, Abu Dhabi, UAE

Source-Free Domain Adaptation MI Classification EEG Decoding Mixture-of-Experts Inference

Ask PDF

Jul 3, 2026cs.LG

Stacked LoRA for Subject-Adaptive EEG Foundation Models in Motor Imagery Decoding

Electroencephalography (EEG) decoding for brain-computer interfaces (BCIs) faces a major challenge: substantial inter-subject variability limits effective cross-subject generalization. Consequently, practical systems still rely largely on subject-specific models trained from scratch and requiring individual recalibration. EEG foundation models have recently emerged as a promising alternative; however, even large pretrained models cannot simply be used as fixed feature extractors and still require additional adaptation before they can be reliably applied to downstream tasks. In this work, we address this challenge through targeted adaptation strategies. Building on recent EEG foundation models such as REVE, LaBraM, and LUNA, we examine the impact of different low-rank adaptation strategies on motor imagery classification. We propose a framework that structurally decouples subject-invariant knowledge from subject-specific neural signatures: the low-rank update at each adapted layer is split into a Global adapter, trained jointly across all subjects, and Subject-Specific adapters, each absorbing individual variability. To assess the contribution of each path, we compare three adaptation strategies: (i) subject-specific LoRA (ii) global LoRA and (iii) stacked LoRA, combining both Global and Subject Specific adapters. Experiments on BCI Competition IV-2a, PhysioNet Motor Imagery, and the clinical Zuo2025 benchmark show that Stacked LoRA effectively mitigates inter-subject variability, achieving the best accuracy in the large majority of backbone and dataset combinations. Our analysis further reveals that the optimal balance between the global and subject-specific paths depends on the target population: a shared adapter is sufficient for large, diverse cohorts, whereas subject-specific adaptation is decisive under the high inter-session variability of clinical recordings.

Aymen Sarhane, Fouad Lbakali, Mouad Souissi +2

IMT Atlantique, Lab-STICC, Brest, France

Cross-Subject EEG Decoding Brain-Computer Interfaces Neural Decoding EEG Decoding Low-Rank Adaptation EEG Foundation Models

Ask PDF