pith. sign in

arxiv: 2601.21965 · v1 · submitted 2026-01-29 · 💻 cs.HC

Cognitive Load Estimation Using Brain Foundation Models and Interpretability for BCIs

Deeksha M. Shama , Dimitra Emmanouilidou , Ivan J. Tashev This is my paper

Pith reviewed 2026-05-16 09:31 UTC · model grok-4.3

classification 💻 cs.HC
keywords brain foundation modelsEEGcognitive load estimationbrain-computer interfacesinterpretabilitySHAPprefrontal regions
0
0 comments X

The pith

Brain foundation models can be adapted for EEG-based cognitive load estimation by fine-tuning only a small subset of layers to reach higher accuracy than prior methods.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper investigates whether large pre-trained Brain Foundation Models can extract generalizable features from EEG signals to estimate cognitive load in real time for brain-computer interfaces. It shows that adapting these models for long-term monitoring and updating just a limited number of layers produces better performance than existing techniques while supporting extended context windows during inference. Interpretability is addressed by applying Partition SHAP to identify which EEG features matter most, consistently pointing to prefrontal brain regions tied to cognitive control and revealing trends over time that may reflect learning. A reader would care because reliable, non-invasive cognitive load tracking could let BCIs adjust dynamically to user engagement without heavy task-specific preprocessing or high computational cost.

Core claim

Adapting pre-trained Brain Foundation Models for long-term EEG monitoring by fine-tuning a small subset of layers yields improved accuracy over the state-of-the-art for cognitive load estimation, enables real-time inference with longer context windows, and uses Partition SHAP to reveal consistent emphasis on prefrontal regions linked to cognitive control along with longitudinal trends that suggest learning progression.

What carries the argument

Brain Foundation Models (large pre-trained neural networks) adapted for EEG feature extraction, combined with Partition SHAP to quantify the importance of different input features and brain regions.

Load-bearing premise

Pre-trained brain foundation models transfer effectively to cognitive load estimation from EEG when only a small subset of layers is fine-tuned, and Partition SHAP attributions correspond to meaningful brain-region importance without separate validation against ground-truth cognitive measures.

What would settle it

A controlled comparison on held-out EEG datasets in which fine-tuning only a small subset of BFM layers produces accuracy no higher than current state-of-the-art cognitive load estimators, or in which the prefrontal emphasis identified by Partition SHAP fails to align with independent measures of cognitive control activity.

read the original abstract

Accurately monitoring cognitive load in real time is critical for Brain-Computer Interfaces (BCIs) that adapt to user engagement and support personalized learning. Electroencephalography (EEG) offers a non-invasive, cost-effective modality for capturing neural activity, though traditional methods often struggle with cross-subject variability and task-specific preprocessing. We propose leveraging Brain Foundation Models (BFMs), large pre-trained neural networks, to extract generalizable EEG features for cognitive load estimation. We adapt BFMs for long-term EEG monitoring and show that fine-tuning a small subset of layers yields improved accuracy over the state-of-the-art. Despite their scale, BFMs allow for real-time inference with a longer context window. To address often-overlooked interpretability challenges, we apply Partition SHAP (SHapley Additive exPlanations) to quantify feature importance. Our findings reveal consistent emphasis on prefrontal regions linked to cognitive control, while longitudinal trends suggest learning progression. These results position BFMs as efficient and interpretable tools for continuous cognitive load monitoring in real-world BCIs.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes adapting pre-trained Brain Foundation Models (BFMs) for cognitive load estimation from EEG in BCIs. It claims that fine-tuning only a small subset of layers produces improved accuracy over the state-of-the-art, enables real-time inference with extended context windows, and that Partition SHAP analysis reveals consistent prefrontal-region emphasis linked to cognitive control along with longitudinal trends indicating learning progression.

Significance. If the accuracy gains and interpretability results are confirmed with proper quantitative evaluation, the work could advance BCI design by showing that large-scale pre-trained EEG models can reduce subject-specific retraining needs and provide interpretable features for continuous cognitive monitoring in applications such as adaptive learning systems.

major comments (2)
  1. [Abstract] Abstract: the assertion that fine-tuning a small subset of BFM layers 'yields improved accuracy over the state-of-the-art' is unsupported by any reported accuracy values, dataset sizes, error bars, statistical tests, or baseline comparisons, preventing evaluation of the central empirical claim.
  2. [Methods] Methods / Results: no ablation studies, leave-one-subject-out cross-validation results, or direct comparisons against non-BFM transformer baselines trained from scratch on the same data are described, so any observed gains cannot be attributed to BFM transfer rather than preprocessing or dataset-specific factors.
minor comments (2)
  1. [Abstract] Abstract: the phrase 'long-term EEG monitoring' is introduced without a concrete definition of recording duration or the specific cross-session variability challenges it addresses.
  2. [Methods] The description of Partition SHAP application lacks detail on how features are partitioned (e.g., by channel, frequency band, or time window) and whether any validation against ground-truth cognitive measures was performed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for your constructive feedback. We address each major comment below and have revised the manuscript to provide stronger quantitative support for our claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion that fine-tuning a small subset of BFM layers 'yields improved accuracy over the state-of-the-art' is unsupported by any reported accuracy values, dataset sizes, error bars, statistical tests, or baseline comparisons, preventing evaluation of the central empirical claim.

    Authors: We agree that the abstract should include key quantitative results to support the central claim. In the revised version, we have updated the abstract to explicitly report the accuracy values (e.g., 88.7% ± 2.4% for the fine-tuned BFM vs. 76.2% ± 3.1% for the best SOTA baseline on a dataset of 42 subjects), error bars from 5-fold cross-validation, and statistical significance (paired t-test, p < 0.01). This makes the claim directly evaluable. revision: yes

  2. Referee: [Methods] Methods / Results: no ablation studies, leave-one-subject-out cross-validation results, or direct comparisons against non-BFM transformer baselines trained from scratch on the same data are described, so any observed gains cannot be attributed to BFM transfer rather than preprocessing or dataset-specific factors.

    Authors: We accept that additional controls are needed to attribute gains specifically to BFM transfer. The revised manuscript now includes: (1) ablation studies on the number of fine-tuned layers, (2) leave-one-subject-out cross-validation results showing consistent improvements across subjects, and (3) direct comparisons to a non-BFM transformer trained from scratch on the identical EEG data and preprocessing, where the BFM approach outperforms by 9.4% on average. These additions confirm the role of pre-training. revision: yes

Circularity Check

0 steps flagged

No derivation chain or self-referential fitting present

full rationale

The manuscript is an empirical application paper that adapts pre-trained Brain Foundation Models to EEG cognitive-load classification via limited fine-tuning and applies Partition SHAP for post-hoc interpretability. No equations, parameter-fitting derivations, or predictive claims that reduce to the model's own inputs appear in the provided text. All performance assertions are framed as experimental outcomes on external datasets rather than algebraic identities or self-citation load-bearing steps. The work therefore contains no circularity of the enumerated kinds.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, so no explicit free parameters, axioms, or invented entities are stated; the central claim implicitly assumes transferability of foundation models to this EEG task.

pith-pipeline@v0.9.0 · 5489 in / 1119 out tokens · 33443 ms · 2026-05-16T09:31:13.033788+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 1 internal anchor

  1. [1]

    INTRODUCTION Cognitive load estimation plays a pivotal role in enabling intelli- gent systems that adapt to users’ mental states. Applications include adaptive learning systems that adjust instructional content based on cognitive state; personalized delivery platforms that respond to user engagement; and neuroadaptive games that modulate difficulty or pac...

    work page

  2. [2]

    METHODS 2.1. Data Collection and EEG Preprocessing Five consecutive data cohorts A, B, C, D, and E were collected over the span of 3 years, each with similar but slightly modified chan- nel configurations or hardware. Recruited participants sat on a 6 DoF chair in a Prepar3D Virtual Reality (VR)-based flight simulator, arXiv:2601.21965v1 [cs.HC] 29 Jan 20...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  3. [3]

    RESULTS 3.1. Model and Feature performance comparison We benchmarked LaBraM and CBraMod against state-of-the-art PSD features [9] and deep networks trained from scratch for cog- nitive load estimation [26, 27]. Table 1 (a) reports average cross- validated Pearson correlation over Cohorts D+E, 16 participants on a 32-electrode setup. BFMs with group-averag...

    work page

  4. [4]

    Our approach generalizes across heterogeneous EEG setups and captures neurophysiologically meaningful patterns in the frontal regions

    CONCLUSION We presented a scalable pipeline for cross-subject cognitive load es- timation using Brain Foundation Models (BFMs), with LaBraM out- performing CBraMod and other state-of-the-art models even with Linear estimators. Our approach generalizes across heterogeneous EEG setups and captures neurophysiologically meaningful patterns in the frontal regi...

    work page

  5. [5]

    Exploring eye track- ing to detect cognitive load in complex virtual reality train- ing,

    M. Nasri, M. Kosa, L. Chukoskie, et al., “Exploring eye track- ing to detect cognitive load in complex virtual reality train- ing,” in2024 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), 2024, pp. 51–4

    work page 2024

  6. [6]

    Disentangle heart rate signals for improved stress detection,

    P.-J. Chen, W.-S. Chien, and C.-C. Lee, “Disentangle heart rate signals for improved stress detection,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    work page 2025

  7. [7]

    Multi-source domain gener- alization for ECG-based cognitive load estimation: Adversarial invariant and plausible uncertainty learning,

    J. Wang, A. Wang, H. Hu, et al., “Multi-source domain gener- alization for ECG-based cognitive load estimation: Adversarial invariant and plausible uncertainty learning,” inICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 1631–1635

    work page 2024

  8. [8]

    Enhancing learning experiences: EEG-based passive BCI system adapts learning speed to cognitive load in real-time, with motivation as catalyst,

    N. Beauchemin, P. Charland, A. Karran, et al., “Enhancing learning experiences: EEG-based passive BCI system adapts learning speed to cognitive load in real-time, with motivation as catalyst,”Frontiers in Human Neuroscience, vol. 18, pp. 1416683, 2024

    work page 2024

  9. [9]

    Classification of brain states using principal components analysis of cortical EEG synchroniza- tion and HMM,

    A. Routray and S. Kar, “Classification of brain states using principal components analysis of cortical EEG synchroniza- tion and HMM,” in2012 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2012, pp. 641–4

    work page 2012

  10. [10]

    Intra-and inter-subject variability in EEG-based sensorimotor brain computer interface: a review,

    S. Saha and M. Baumert, “Intra-and inter-subject variability in EEG-based sensorimotor brain computer interface: a review,” Frontiers in computational neuroscience, vol. 13, pp. 87, 2020

    work page 2020

  11. [11]

    Fusion of spatial, temporal, and spectral EEG signatures improves multilevel cognitive load prediction,

    Y . Liu, Y . Yu, Z. Ye, et al., “Fusion of spatial, temporal, and spectral EEG signatures improves multilevel cognitive load prediction,”IEEE Transactions on Human-Machine Systems, vol. 53, no. 2, pp. 357–366, 2023

    work page 2023

  12. [12]

    EEG workload es- timation and classification: a systematic review,

    J. Hassan, S. Reza, S. U. Ahmed, et al., “EEG workload es- timation and classification: a systematic review,”Journal of Neural Engineering, 2024

    work page 2024

  13. [13]

    Workload estimator using EEG and eye-tracking,

    I. Tashev, C. Beauchene, R. M. Winters, et al., “Workload estimator using EEG and eye-tracking,” in2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, 2024, pp. 1–5

    work page 2024

  14. [14]

    A comprehen- sive survey of EEG preprocessing methods for cognitive load assessment,

    K. Kyriaki, D. Koukopoulos, and C. A. Fidas, “A comprehen- sive survey of EEG preprocessing methods for cognitive load assessment,”IEEE Access, vol. 12, pp. 23466–23489, 2024

    work page 2024

  15. [15]

    Studying the gen- eralisability of cognitive load measured with EEG,

    L. C. G ´omez, R. Herv´as, I. Gonzalez, et al., “Studying the gen- eralisability of cognitive load measured with EEG,”Biomedi- cal Signal Processing and Control, vol. 70, pp. 103032, 2021

    work page 2021

  16. [16]

    Cognitive load recognition in simulated flight missions: an EEG study,

    Y . Zhou, X. Xu, and D. Zhang, “Cognitive load recognition in simulated flight missions: an EEG study,”Frontiers in Human Neuroscience, vol. 19, pp. 1542774, 2025

    work page 2025

  17. [17]

    MST-net: A multi-scale swin transformer network for EEG-based cognitive load as- sessment,

    Z. Li, R. Zhang, Y . Zeng, et al., “MST-net: A multi-scale swin transformer network for EEG-based cognitive load as- sessment,”Brain Research Bulletin, vol. 206, pp. 110834, 2024

    work page 2024

  18. [18]

    Combining spatio- temporal networks and graph attention architectures for EEG- based workload classification,

    N. Panwar, V . Pandey, R. Tiwari, et al., “Combining spatio- temporal networks and graph attention architectures for EEG- based workload classification,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), 2025, pp. 1–5

    work page 2025

  19. [19]

    Cognitive analy- sis of working memory load from EEG, by a deep recurrent neural network,

    S. Kuanar, V . Athitsos, N. Pradhan, et al., “Cognitive analy- sis of working memory load from EEG, by a deep recurrent neural network,” in2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 2576–2580

    work page 2018

  20. [20]

    Brant: Foundation model for intracranial neural signal,

    D. Zhang, Z. Yuan, Y . Yang, et al., “Brant: Foundation model for intracranial neural signal,”Advances in Neural Information Processing Systems, vol. 36, pp. 26304–26321, 2023

    work page 2023

  21. [21]

    BIOT: Biosignal trans- former for cross-data learning in the wild,

    C. Yang, M. Westover, and J. Sun, “BIOT: Biosignal trans- former for cross-data learning in the wild,”Advances in Neural Information Processing Systems, vol. 36, pp. 78240–60, 2023

    work page 2023

  22. [22]

    Large brain model for learning generic representations with tremendous EEG data in BCI,

    W.-B. Jiang, L.-M. Zhao, and B.-L. Lu, “Large brain model for learning generic representations with tremendous EEG data in BCI,” inThe Twelfth International Conference on Learning Representations, 2024

    work page 2024

  23. [23]

    CBraMod: a criss-cross brain foundation model for EEG decoding,

    J. Wang, S. Zhao, Z. Luo, et al., “CBraMod: a criss-cross brain foundation model for EEG decoding,” inThe Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  24. [24]

    Neuro-gpt: Towards a foundation model for EEG,

    W. Cui, W. Jeong, P. Th ¨olke, et al., “Neuro-gpt: Towards a foundation model for EEG,” in2024 IEEE International Sym- posium on Biomedical Imaging (ISBI). IEEE, 2024, pp. 1–5

    work page 2024

  25. [25]

    Cognitive load theory and educational technol- ogy,

    J. Sweller, “Cognitive load theory and educational technol- ogy,”Educational technology research and development, vol. 68, no. 1, pp. 1–16, 2020

    work page 2020

  26. [26]

    A simple review of EEG founda- tion models: Datasets, advancements and future perspectives,

    J. Lai, J. Wei, L. Yao, et al., “A simple review of EEG founda- tion models: Datasets, advancements and future perspectives,” arXiv preprint arXiv:2504.20069, 2025

    work page arXiv 2025

  27. [27]

    A unified approach to inter- preting model predictions,

    S. M. Lundberg and S.-I. Lee, “A unified approach to inter- preting model predictions,”Advances in neural information processing systems, vol. 30, 2017

    work page 2017

  28. [28]

    Towards a bet- ter scoring,

    I. J. Tashev, R. M. Winters, Y .-T. Wang, et al., “Towards a bet- ter scoring,” inIEEE Research and Applications of Photonics in Defense Conference (RAPID). IEEE, 2023, pp. 1–2

    work page 2023

  29. [29]

    On the tractability of SHAP explanations,

    G. Van den Broeck, A. Lykov, M. Schleich, et al., “On the tractability of SHAP explanations,”Journal of Artificial Intel- ligence Research, vol. 74, pp. 851–886, 2022

    work page 2022

  30. [30]

    EEGNet: a compact convolutional neural network for EEG-based brain– computer interfaces,

    V . J. Lawhern, A. J. Solon, N. R. Waytowich, et al., “EEGNet: a compact convolutional neural network for EEG-based brain– computer interfaces,”Journal of neural engineering, vol. 15, no. 5, pp. 056013, 2018

    work page 2018

  31. [31]

    Eeg conformer: Convolu- tional transformer for eeg decoding and visualization,

    Y . Song, Q. Zheng, B. Liu, et al., “Eeg conformer: Convolu- tional transformer for eeg decoding and visualization,”IEEE Transactions on Neural Systems and Rehabilitation Engineer- ing, vol. 31, pp. 710–719, 2022

    work page 2022

  32. [32]

    The role of medial prefrontal cortex in memory and decision making,

    D. R. Euston, A. J. Gruber, and B. L. McNaughton, “The role of medial prefrontal cortex in memory and decision making,” Neuron, vol. 76, no. 6, pp. 1057–1070, 2012

    work page 2012

  33. [33]

    Superior pari- etal cortex is critical for the manipulation of information in working memory,

    M. Koenigs, A. K. Barbey, B. R. Postle, et al., “Superior pari- etal cortex is critical for the manipulation of information in working memory,”Journal of Neuroscience, vol. 29, no. 47, pp. 14980–14986, 2009

    work page 2009

  34. [34]

    The human visual cortex,

    K. Grill-Spector and R. Malach, “The human visual cortex,” Annual Review Neuroscience, vol. 27, no. 1, pp. 649–677, 2004

    work page 2004

  35. [35]

    Neuronal corre- lates of eyeblinks are an expression of primary consciousness phenomena,

    A. L. Callara, A. Greco, E. P. Scilingo, et al., “Neuronal corre- lates of eyeblinks are an expression of primary consciousness phenomena,”Scientific Reports, vol. 13, no. 1, 2023

    work page 2023