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

arXiv:2507.21037 · cs.LG · Submitted 2025-07-28 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Brain foundation model-guided source-selective domain adaptation for cross-subject EEG decoding".

Jane: The gist 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…

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, wrapping up this paper "Brain foundation model-guided source-selective domain adaptation for cross-subject EEG decoding," the authors Jinzhou Wu et al. are proposing a way to handle the mess of inter-subject variability in motor imagery decoding >

Jane: They’re using a brain foundation model to first find the best source subjects, and then they perform relevance-weighted distribution alignment over those specific sources >

Lu: The implication here is that by selectively leveraging relevant source subjects and combining both feature-level and decision-level alignment, you can get better cross-subject EEG decoding performance than many existing baselines >

Meng: It moves the focus from trying to use every single source subject to intelligently picking the best ones, which makes the whole process faster for real deployment >

Lalam: They showed that this approach significantly outperforms most baselines with a p value less than zero point zero five on both public MI-EEG benchmarks >

Jane: What it means in simpler terms is that we can use pre-trained brain knowledge to filter out the subjects who are going to cause problems, which leads to faster model tuning and more practical deployment for BCI applications >

Conclusion: Tom: So we're wrapping up this talk on "Brain foundation model-guided source-selective domain adaptation for cross-subject EEG decoding" and what it actually means for us in AI deployment >

Jane: It boils down to using a brain foundation model to first find the best source subjects, and then aligning those sources carefully with our target data without having any target labels >

Lu: The core idea is that inter-subject variability, which is usually a huge problem, can be managed if you use pre-trained brain knowledge as a starting point for selecting compatible sources >

Meng: But I gotta ask, how much extra work does this source selection add to the training cost? Because if it adds too much overhead at scale, it's just theoretical stuff on paper >

Lalam: The complexity analysis shows that without source selection, the cost scales with N squared multiplied by B squared times d plus K, but with selection, it drops to something closer to (N/S) squared for that target >

Tom: That reduction in scaling is interesting. It means we can actually train these models on bigger pools of subjects without the training time exploding exponentially >

Jane: And the results on those two public MI-EEG benchmarks show performance gains, like eighty-six point sixteen percent and seventy-eight point forty-one percent, which beats some representative cross-subject methods >

Lu: The key is that they use a Cauchy–Schwarz divergence to measure compatibility both at the feature level and the decision level, which addresses different kinds of noise in the data >

Meng: So if I'm shipping this into a real BCI device tomorrow, does it actually hold up when the recording setup is completely different from what was trained on? >

Jane: That’s a limitation they point out. The source selection depends on the geometry of that pretrained brain model, and that might not capture all kinds of recording mismatches >

Tom: True. It works really well under controlled conditions with healthy participants and binary tasks, but we need to see it tested on more messy, real-world data sets >

Lu: The implication is that this BFM guided approach gives us a way to use massive amounts of pre-existing brain knowledge to prune the search space before we start the heavy distribution alignment work >

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

State Key Laboratory of Mechanical Transmission for Advanced Equipment, Chongqing University · Department of Computer Science, Vrije Universiteit Amsterdam Department of Computer Science at Vrije Universiteit Amsterdam Machine Learning Group at UiT–The Arctic University of Norway College of Computing and Data Science at Nanyang Technological University

cs.LG

Submitted: 2025-07-28

Updated: 2026-10-07

Comments: 22 pages, 11 figures

Code: https://github.com/wjz114/BFM-MSDA

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 75/100

The gist: The gist Cross-subject motor-imagery electroencephalography (MI-EEG) decoding remains challenging because substantial inter-subject variability can cause both negative transfer from poorly matched

Key concepts

Brain Foundation Model (BFM): A large, pre-trained model like LaBraM trained on vast amounts of heterogeneous EEG data. It learns rich, subject-level representations of brain activity. This learned representation acts as a prior to estimate how similar different subjects are, guiding the selection process for compatible source subjects.
Source–Target Compatibility (CS Divergence): A mathematical measure, specifically Cauchy–Schwarz divergence, used to quantify the distance between the feature distributions of source and target EEG data. This calculation helps determine which source subjects are most relevant and transferable to the target domain.
Relevance-Weighted Dual Alignment Strategy: A two-stage alignment process applied to selected sources. Feature-level alignment adjusts raw signal representations, while decision-level alignment aligns the final class predictions. This dual approach ensures that both the underlying features and the resulting decisions are well-matched between source and target data.

Terminology

Adaptation — 3× in this paper · explained in 1 episode
Alignment — 2× in this paper
Brain Foundation Model — 2× in this paper
Domain Adaptation — 2× in this paper · explained in 5 episodes
Foundation Model — 2× in this paper · explained in 3 episodes
Benchmark — 1× in this paper · explained in 2 episodes
Calibration — 1× in this paper · explained in 7 episodes

Summary

The gist 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.

Problem Definition

The study investigates subject-independent MI decoding in an unsupervised domain adaptation setting, where the goal is to retrieve a compact set of source subjects most compatible with the target and adapt them without using target labels. The objective is to learn a model M = g ◦ f, where the feature extractor f: R C×T → R d maps raw MI-EEG signals to a latent representation with dimension d, and the classifier g: R d → K outputs predicted class for MI tasks, such that M generalizes effectively to Dt in an unsupervised DA setting.

Brain Foundation Model Guided Source Retrieval

The framework first utilizes representations learned by a pretrained brain foundation model (BFM) to estimate source–target compatibility and retrieve target-relevant sources. Specifically, the method uses LaBraM as a unified foundation model pretrained on roughly 2,500 hours of heterogeneous EEG data spanning 20 datasets to extract subject-level EEG representations. This process involves segmenting EEG recordings into fixed-length temporal patches and aggregating trial-level embeddings into subject-level representations for subsequent similarity estimation. Source selection is guided by computing the Cauchy–Schwarz (CS) divergence between source and target feature distributions, ds,t = DCS (p(hs)∥p(ht)).

Relevance-Weighted Dual Alignment Strategy

Subsequently, a relevance-weighted dual alignment strategy is applied to the selected sources and the unlabeled target domain. This strategy involves two levels of alignment:

Feature-level alignment (FLA): This uses Cauchy–Schwarz (CS) divergence to reduce discrepancies in marginal feature distributions, formulated as LSTFLA = X N s=1 ωs DCS(ps(z)∥pt(z)) + 2 N(N − 1) X N s=1 X N s′>s DCS(ps'(z)∥pt(z)).
Decision-level alignment (DLA): This uses conditional Cauchy–Schwarz (CCS) divergence to align class-dependent decision distributions, formulated as LDLA = X N s=1 ωs DCS(ps(yz)∥pt(yz)) + 2 N(N − 1) X N s=1 X N s′>s DCS(ps'(yz)∥pt(yz)).

Source relevance is incorporated into both alignment terms so that more transferable source subjects contribute more strongly to adaptation.

Experimental Validation and Findings

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. The key contributions include a BFM-guided source retrieval strategy that uses pretrained brain representations as a transferability prior, reducing negative transfer and computational burden. Furthermore, the controlled analysis with a large source pool demonstrates that target-aware source retrieval improves scalability while mitigating negative transfer. The results show that BFM distance reliably identifies source subjects with higher transferability after controlling for target difficulty and that the combination of FLA and DLA yields cumulative gains over individual terms, indicating complementary effects.

Complexity and Scalability

The per-epoch cost of BFM-MSDA is dominated by pairwise CS and CCS divergences, leading to a complexity of O N 2B2 (d + K) + NBd. Without source selection, the cost scales as O S 2B2 (d + K), which is reduced by the selection rule to approximately (N/S) 2 for that target. This scaling is consistent with empirical timing showing that training time per epoch rises super-linearly as more distant sources are admitted. The framework exhibits stable and consistent performance across a wide range of hyperparameters α and β once the alignment losses are activated.

Limitations and Future Directions

The present findings should be interpreted within the scope of a controlled cross-subject evaluation involving healthy participants and binary left/right MI tasks. A limitation is that source selection depends on the geometry of the pretrained BFM embedding space, which may not capture only MI-discriminative structure. Future work should address evaluating robustness to mismatch across heterogeneous recording setups, extending the framework to multi-class MI paradigms and clinical populations, and investigating continual source retrieval strategies that accommodate repository growth over time.

The proposed BFM-MSDA framework first retrieves targetrelevant sources in a pretrained neural representation space and then performs relevance-weighted distribution alignment over the selected sources. The experimental results on two public MI-EEG benchmarks show that the proposed framework improves cross-subject decoding over representative baselines. The large-source-pool experiments further demonstrate that poorly matched sources can increase computational cost and degrade transfer. This study presented a BFM-guided source-selective MSDA framework for cross-subject EEG decoding. The proposed BFMMSDA significantly outperforms most baselines (p < 0.05) in average performance on both datasets, demonstrating superior generalization across subjects. The results highlight the advantage of selectively leveraging relevant source subjects and conducting both feature-level and decision-level alignment to mitigate inter-subject variability. By effectively filtering out irrelevant subjects, the proposed framework enables faster model tuning and more practical deployment in real-world BCI applications. This study presented a BFM-guided source-selective MSDA framework for cross-subject EEG decoding. The proposed BFMMSDA significantly outperforms most baselines (p < 0.05) in average performance on both datasets, demonstrating superior generalization across subjects. The results highlight the advantage of selectively leveraging relevant source subjects and conducting both feature-level and decision-level alignment to mitigate inter-subject variability. By effectively filtering out irrelevant subjects, the proposed framework enables faster model tuning and more practical deployment in real-world BCI applications. This study presented a BFM-guided source-selective MSDA framework for cross-subject EEG decoding. The proposed BFMMSDA significantly outperforms most baselines (p < 0.05) in average performance on both datasets, demonstrating superior generalization across subjects. The results highlight the advantage of selectively leveraging relevant source subjects and conducting both feature-level and decision-level alignment to mitigate inter-subject variability. By effectively filtering out irrelevant subjects, the proposed framework enables faster model tuning and more practical deployment in real-world BCI applications. This study presented a BFM-guided source-selective MSDA framework for cross-subject EEG decoding. The proposed BFMMSDA significantly outperforms most baselines (p < 0.05) in average performance on both datasets, demonstrating superior generalization across subjects. The results highlight the advantage of selectively leveraging relevant source subjects and conducting both feature-level and decision-level alignment to mitigate inter-subject variability. By effectively filtering out irrelevant subjects, the proposed framework enables faster model tuning and more practical deployment in real-world BCI applications. This study presented a BFM-guided source-selective MSDA framework for cross-subject EEG decoding. The proposed BFMMSDA significantly outperforms most baselines (p < 0.05) in average performance on both datasets, demonstrating superior generalization across subjects. The results highlight the advantage of selectively leveraging relevant source subjects and conducting both feature-level and decision-level alignment to mitigate inter-subject variability. By effectively filtering out irrelevant subjects, the proposed framework enables faster model tuning and more practical deployment in real-world BCI applications. This study presented a BFM-guided source-selective MSDA framework for cross-subject EEG decoding. The proposed BFMMSDA significantly outperforms most baselines (p < 0.05) in average performance on both datasets, demonstrating superior generalization across subjects. The results highlight the advantage of selectively leveraging relevant source subjects and conducting both feature-level and decision-level alignment to mitigate inter-subject variability. By effectively filtering out irrelevant subjects, the proposed framework enables faster model tuning and more practical deployment in real-world BCI applications. This study presented a BFM-guided source-selective MSDA framework for cross-subject EEG decoding. The proposed BFMMSDA significantly outperforms most baselines (p < 0.05) in average performance on both datasets, demonstrating superior generalization across subjects. The results highlight the advantage of selectively leveraging relevant source subjects and conducting both feature-level and decision-level alignment to mitigate inter-subject variability. By effectively filtering out irrelevant subjects, the proposed framework enables faster model tuning and more practical deployment in real-world BCI applications. This study presented a BFM-guided source-selective MSDA framework for cross-subject EEG decoding. The proposed BFMMSDA significantly outperforms most baselines (p < 0.05) in average performance on both datasets, demonstrating superior generalization across subjects.

Improvements for AI systems

Improved source selection via BFM distance: The system can retrieve source subjects most relevant to the current target by using representations learned by a pretrained brain foundation model to estimate source–target compatibility, which is supported by Spearman’s rank correlation results showing that closer BFM embeddings reliably identify source subjects with higher transferability.
Joint feature and decision-level alignment: The framework can jointly align feature distributions and decision distributions across selected sources and the target by employing Cauchy–Schwarz (CS) divergence for feature-level alignment and conditional Cauchy–Schwarz (CCS) divergence further aligns class-dependent decision distributions.
Relevance-weighted adaptation: The system can mitigate negative transfer by applying a relevance-weighted dual alignment strategy, where source relevance is incorporated into both alignment terms so that more transferable source subjects contribute more strongly to adaptation.