REVIEW 4 major objections 5 minor 24 references
Graph Adapter of EEG Foundation Models for Parameter Efficient Fine Tuning
T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read By freezing the BENDR temporal backbone and fine-tuning only a graph neural network adapter, EEG-GraphAdapter raises F1 by up to 16.1% on two clinical EEG tasks.
desk verdict Plausible PEFT idea for EEG with a frozen backbone, but the experiments are undermined by segment-level cross-validation and likely TUAB/TUEG overlap. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the EEG-GraphAdapter (EGA), a two-layer graph neural network inserted before the frozen BENDR encoder. It treats each of the 19 EEG electrodes as a node in a fully connected graph whose edge weights are the geodesic distances between sensor positions, so each node's embedding is a message-passing aggregation of signals from all other sensors. This is what carries the spatial information that BENDR's temporal convolutions are claimed to lack. The GNN variants (GCN averaging, GraphSAGE sampling, GAT attention-weighted) define three concrete instances of the same mechanism.
What would settle it
Check the TUAB evaluation subjects against the 69,652 TUEG pre-training samples: if any TUAB patient appears in pretraining, exclude them, retrain the frozen encoder, and re-run EGA-GraphSAGE; if the 16.1% F1 gain disappears, the central claim fails. A cleaner test is to evaluate EGA on an EEG abnormality set recorded at a different institution.
Extended reading notes
Core claim
The central claim is that a lightweight graph adapter can inject sensor-level spatial structure into a frozen temporal EEG encoder and improve downstream classification. EGA builds a fully connected weighted graph over the 19 channels, with edge weights given by geodesic sensor distances, and applies a two-layer GNN (GCN, GraphSAGE, or GAT) to the input before BENDR. Only the adapter is trained. The paper reports that EGA-GAT raises F1 from 0.7581 to 0.8548 and AUROC from 0.9407 to 0.9661 on MDD, while EGA-GraphSAGE raises F1 from 0.4328 to 0.5027 and AUROC from 0.7377 to 0.7459 on TUAB. It also notes that the best GNN architecture differs by task, so the adapter's value is real but the right message-passing scheme is task-specific.
Load-bearing premise
The load-bearing premise is that the TUAB evaluation patients were never included in the TUEG data on which BENDR was pre-trained, since both come from the same hospital EEG collection; the paper does not state this exclusion, and if it fails, the TUAB improvements may be memorization rather than adaptation.
Editorial extensions
If this is right
- Freezing the backbone prevents catastrophic forgetting of pretrained EEG representations while the adapter learns task-specific spatial structure, so the combined model keeps the benefits of large-scale pretraining on small labeled sets.
- Trainable parameters drop from 6,459,257 for BENDR to roughly 1M–2M for the EGA versions, which should lower per-task storage and serving cost even if wall-clock speedup is currently limited to about 17% by data-loading overhead.
- Because the adapter is task-agnostic in placement, the same EGA design could be re-tested with other temporal encoders without redesigning the graph module.
- The inconsistent best GNN across tasks implies that adapter architectures need to be selected per downstream task rather than universal: GAT suited MDD and GraphSAGE suited TUAB in these experiments.
Reading between the lines
- Editorial extension: the same adapter could be tested with data-driven functional-connectivity weights instead of fixed geodesic distances; if connectivity-based adjacency outperforms geodesic weights, the spatial prior itself is the active ingredient.
- Editorial extension: the paper's speedup finding (far fewer parameters but only 17% faster) suggests the bottleneck is input preprocessing; parallel data loading would be the next experiment needed to realize the efficiency claim in wall-clock terms.
- Editorial extension: an immediate validation would be re-running the TUAB comparison with the pretraining corpus explicitly disjoint from evaluation patients; this is the natural next check of the 16.1% gain.
- Editorial extension: for small clinical datasets where labels are scarce, EGA's freezing of the backbone suggests it could combine with few-shot or transfer methods beyond simple fine-tuning, though the paper does not test this.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes EEG-GraphAdapter (EGA), a parameter-efficient fine-tuning method that inserts a GNN-based adapter before a frozen BENDR backbone, so that only the adapter is trained on downstream EEG classification tasks. The method is evaluated on two binary classification tasks, MDD and TUAB, comparing EGA with GCN, GraphSAGE, and GAT variants against a fully fine-tuned BENDR baseline. The central claim is that EGA improves F1-score by up to 16.1% while updating far fewer parameters than the baseline, and the paper also reports runtime and parameter-count analyses.
Significance. If the results hold, EGA is a useful contribution: it combines a pre-trained temporal EEG encoder with a lightweight spatial graph module, a natural and inexpensive design for the small labeled datasets typical in clinical EEG. The authors are transparent about the modest observed speedup (Section 3.3), which is a welcome departure from overselling parameter-count reductions. However, the empirical validation is not yet convincing. The TUAB result is threatened by likely overlap between the TUEG pre-training corpus and the TUAB evaluation set, and both evaluations use segment-level, not subject-level, cross-validation. The reported gains also come from post hoc selection of the best GNN per task, with no variance estimates. The MDD experiment is independent of pre-training but still suffers from the leakage risk in its split scheme. The central claim therefore needs stronger experimental evidence before the paper can be accepted.
major comments (4)
- [Section 3.1.1 and Section 3.1.2, Table 3] The manuscript pre-trains BENDR on TUEG v2.0.1 (69,652 samples) and evaluates on TUAB [16], which is derived from the same Temple University Hospital EEG collection. The text never states that TUAB subjects were excluded from the TUEG pre-training data, nor that the downstream cross-validation is subject-disjoint from pre-training. If the same patients appear in both corpora, the frozen BENDR backbone can carry memorized subject-specific patterns, while the fully fine-tuned baseline may overwrite those patterns. In that case the reported 16.1% F1 improvement of EGA-GraphSAGE over the baseline would not demonstrate the value of spatial graph adaptation. The authors must either show explicit subject-level de-duplication or re-run the TUAB experiment with a pre-training set that provably excludes the TUAB test subjects.
- [Section 3.1.2, Table 1] The cross-validation is performed on segments rather than on subjects. For MDD, 63 subjects yield 126 samples from eye-open and eye-closed 60-second segments; for TUAB, 40 subjects yield 517 segments. A k-fold split over segments will almost certainly place segments from the same subject in both the training and test folds, creating information leakage and inflating every reported metric. The paper should report subject-exclusive folds, along with fold-level means and standard deviations, for both tasks.
- [Tables 2 and 3] Each model is summarized by a single point estimate without error bars or significance tests. For example, the TUAB F1 improvement of EGA-GraphSAGE over the baseline is 0.5027 versus 0.4328; without variance estimates or paired statistical tests, this difference may be within noise. In addition, the best GNN is selected per task after observing test performance (GAT for MDD, GraphSAGE for TUAB), so the reported gains are partly the result of selection on the test set. The authors should either pre-specify the model-selection rule or use an internal validation split for choosing the GNN variant, and then report results on a held-out test set.
- [Abstract and Introduction] The paper claims that EGA reduces data requirements for downstream tasks, but no experiment varies the amount of labeled data or compares EGA with other PEFT methods such as LoRA or standard adapter layers. The only empirical support is full-data comparisons on two tasks. To substantiate the data-efficiency motivation, the authors should either add low-data experiments or temper the claim to what the experiments actually show.
minor comments (5)
- [Section 3.1.3] The sentence 'We implement EGA as the extension of BENDR, which is implemented in1 with PyTorch backend' contains a dangling reference marker 'in1' with no visible footnote; this needs to be fixed.
- [Section 2.1.3] The description of the trainable linear layer that adjusts sequence length is underspecified: it is unclear whether the linear projection is shared across EEG channels and how its output is fed into the graph convolution. Adding the exact tensor shapes for the GNN input and output would improve reproducibility.
- [Tables 2 and 3] The tables report AUROC to four decimal places and F1 to four decimal places as well, but the text often cites only two or three digits; please use consistent precision throughout.
- [Section 4] The related-work section would benefit from a direct comparison with other PEFT approaches for EEG, rather than only with general GNN and LLM-GNN adapters, so that the novelty of EGA is clearer.
- [Figure 1] The figure caption does not define all components, such as the Linear Aggregator, Classifier, and the exact placement of the EGA module; a more detailed caption or annotation would help readers follow the architecture description.
Circularity Check
No significant circularity: EGA's performance claims rest on empirical comparisons, not on a derivation that reduces to its inputs.
full rationale
The paper's central claim is that a GNN adapter (EGA) inserted before a frozen BENDR backbone improves downstream F1 on MDD/TUAB relative to a fully fine-tuned BENDR baseline. This is an experimental result; no equation in the paper defines the reported improvement in terms of the adapter's construction. The backbone (BENDR) and the Linear Aggregator are taken from external prior work [1], the geodesic edge weights adopt EEG-GCNN's external design [6], and the three GNN variants (GCN/GraphSAGE/GAT) are standard external architectures [11-13]. There is no load-bearing self-citation: the only author-affiliated reference [24] is the mdx compute platform and is not part of the PEFT argument. The TUAB/TUEG overlap raised by the reader is a potential data-independence/leakage concern about whether pre-training already saw TUAB test subjects; that would be an evaluation flaw, not circularity, because the reported EGA gain does not follow by construction from the pre-training input. Accordingly, no circular step can be quoted with a specific reduction, and the paper should not be scored for circularity.
Assumptions & free parameters
free parameters (4)
- GNN hidden layer size =
64
- learning rate =
1e-5
- fine-tuning epochs =
7
- GAT attention heads =
1
assumptions (4)
- domain assumption A fully connected graph with geodesic distance edge weights captures spatial relationships among EEG sensors.
- domain assumption The frozen BENDR backbone preserves the temporal representations needed for the tasks, so only the spatial adapter needs training.
- domain assumption Preprocessing steps (50 Hz notch, 0.1-100 Hz bandpass, resampling to 256 Hz) do not introduce systematic bias.
- domain assumption 60-second segments of the same subject are conditionally independent for k-fold evaluation.
Cite this review
Pith. "Pith review of Graph Adapter of EEG Foundation Models for Parameter Efficient Fine Tuning." pith.science (2026). https://pith.science/paper/GBOPMP3D
@misc{pith2026241116155,
author = {Pith},
title = {Pith review of: Graph Adapter of EEG Foundation Models for Parameter Efficient Fine Tuning},
year = {2026},
howpublished = {\url{https://pith.science/paper/GBOPMP3D}},
note = {Machine review of arXiv:2411.16155}
}
read the original abstract
In diagnosing neurological disorders from electroencephalography (EEG) data, foundation models such as Transformers have been employed to capture temporal dynamics. Additionally, Graph Neural Networks (GNNs) are critical for representing the spatial relationships among EEG sensors. However, fine-tuning these large-scale models for both temporal and spatial features can be prohibitively large in computational cost, especially under the limited availability of labeled EEG datasets. We propose EEG-GraphAdapter (EGA), a parameter-efficient fine-tuning (PEFT) approach designed to address these challenges. EGA is integrated into a pre-trained temporal backbone model as a GNN-based module, freezing the backbone and allowing only the adapter to be fine-tuned. This enables the effective acquisition of EEG spatial representations, significantly reducing computational overhead and data requirements. Experimental evaluations on two healthcare-related downstream tasks-Major Depressive Disorder (MDD) and Abnormality Detection (TUAB)-show that EGA improves performance by up to 16.1% in F1-score compared with the backbone BENDR model, highlighting its potential for scalable and accurate EEG-based predictions.
Reference graph
Works this paper leans on
-
[16]
Lopez, S., Suarez, G., Jungreis, D., Obeid, I., Picone, J. (2015). Automated identification of abnormal adult EEGs. In 2015 IEEE signal processing in medicine and biology symposium (SPMB) (pp. 1-5). IEEE
work page 2015
-
[1]
Kostas, D., Aroca-Ouellette, S., Rudzicz, F. (2021). BENDR: Using transformers and a con- trastive self-supervised learning task to learn from massive amounts of EEG data. Frontiers in Human Neuroscience, 15, 653659
work page 2021
-
[2]
U., Kreiman, G., Katz, B., Cases, I., Barbu, A
Wang, C., Subramaniam, V., Yaari, A. U., Kreiman, G., Katz, B., Cases, I., Barbu, A. (2023) BrainBERT: Self-supervised representation learning for intracranial recordings. In The Eleventh International Conference on Learning Representations
work page 2023
-
[3]
Pati, A., Mewada, D., Samanta, D. (2023). Meta-Learning for Subject Adaptation in Low-Data Environments for EEG-Based Motor Imagery Brain-Computer Interfaces
work page 2023
-
[4]
Wang, Y., Wang, J., Wang, W., Su, J., Hou, Z. G. (2023). Calibration-Free Transfer Learning for EEG-Based Cross-Subject Motor Imagery Classification. In 2023 IEEE 19th International Conference on Automation Science and Engineering (CASE) (pp. 1-6). IEEE
work page 2023
-
[5]
Coben, R., Mohammad-Rezazadeh, I., Cannon, R. L. (2014). Using quantitative and analytic EEG methods in the understanding of connectivity in autism spectrum disorders: a theory of mixed over-and under-connectivity. Frontiers in human neuroscience, 8, 45
work page 2014
-
[6]
Wagh, N., Varatharajah, Y. (2020). Eeg-gcnn: Augmenting electroencephalogram-based neu- rological disease diagnosis using a domain-guided graph convolutional neural network. In Machine Learning for Health (pp. 367-378). PMLR
work page 2020
-
[7]
Demir, A., Koike-Akino, T., Wang, Y., Haruna, M., Erdogmus, D. (2021). EEG-GNN: Graph neural networks for classification of electroencephalogram (EEG) signals. In 2021 43rd Annual International Conference of the IEEE Engineering in Medicine Biology Society (EMBC) (pp. 1061-1067). IEEE
work page 2021
Show all 24 references
-
[8]
Sun, X., Xu, Y., Zhao, Y., Zheng, X., Zheng, Y., Cui, L. (2023). Multi-Granularity Graph Convolution Network for Major Depressive Disorder Recognition. IEEE Transactions on Neural Systems and Rehabilitation Engineering
2023
-
[9]
Yin, Y., Zheng, X., Hu, B., Zhang, Y., Cui, X. (2021). EEG emotion recognition using fusion model of graph convolutional neural networks and LSTM. Applied Soft Computing, 100, 106954
2021
-
[10]
Wang, P., Zheng, H., Dai, S., Wang, Y., Gu, X., Wu, Y., Wang, X. (2024). A Survey of Spatio- Temporal EEG data Analysis: from Models to Applications. arXiv preprint arXiv:2410.08224
2024 arXiv
-
[11]
N., Welling, M
Kipf, T. N., Welling, M. (2022). Semi-Supervised Classification with Graph Convolutional Networks. In International Conference on Learning Representations
2022
-
[12]
Hamilton, W., Ying, Z., Leskovec, J. (2017). Inductive representation learning on large graphs. Advances in neural information processing systems, 30
2017
-
[13]
Veli ˇckovi´c, P., Cucurull, G., Casanova, A., Romero, A., Li `o, P., Bengio, Y. (2018). Graph Attention Networks. In International Conference on Learning Representations
2018
-
[14]
Baevski, A., Zhou, Y., Mohamed, A., Auli, M. (2020). wav2vec 2.0: A framework for self- supervised learning of speech representations. Advances in neural information processing systems, 33, 12449-12460
2020
-
[15]
Mumtaz, W. (2016). MDD patients and healthy controls EEG data (new). Figshare, Dataset. doi: 10.6084/m9.figshare.4244171.v2 14 Toyotaro Suzumura, Hiroki Kanezashi, and Shotaro Akahori
2016 doi
-
[17]
Yuan, Z., Zhang, D., Chen, J., Gu, G., Yang, Y. (2024). BrainWave: A Brain Signal Foundation Model for Clinical Applications. arXiv preprint arXiv:2402.10251
2024 arXiv
-
[18]
Chien, H. Y. S., Goh, H., Sandino, C. M., Cheng, J. Y. (2022) MAEEG: Masked Auto-encoder for EEG Representation Learning. In NeurIPS 2022 Workshop on Learning from Time Series for Health
2022
-
[19]
Mohammadi Foumani, N., Mackellar, G., Ghane, S., Irtza, S., Nguyen, N., Salehi, M. (2024). Eeg2rep: enhancing self-supervised EEG representation through informative masked inputs. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 5544-5555)
2024
-
[20]
J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W
Hu, E. J., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W. (2022). LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations
2022
-
[21]
Gui, A., Ye, J., Xiao, H. (2024). G-adapter: Towards structure-aware parameter-efficient trans- fer learning for graph transformer networks. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 11, pp. 12226-12234)
2024
-
[22]
He, X., Bresson, X., Laurent, T., Perold, A., LeCun, Y., Hooi, B. (2023). Harnessing Explana- tions: LLM-to-LM Interpreter for Enhanced Text-Attributed Graph Representation Learning. International Conference on Learning Representations
2023
-
[23]
Huang, X., Han, K., Yang, Y., Bao, D., Tao, Q., Chai, Z., Zhu, Q. (2024). Can GNN be Good Adapter for LLMs?. In Proceedings of the ACM on Web Conference 2024 (pp. 893-904)
2024
-
[24]
mdx: A Cloud Plat- form for Supporting Data Science and Cross-Disciplinary Research Collaborations,
Toyotaro Suzumura, Akiyoshi Sugiki, Hiroyuki Takizawa et al. , “mdx: A Cloud Plat- form for Supporting Data Science and Cross-Disciplinary Research Collaborations,” in 2022 IEEE Intl Conf on Cloud and Big Data Computing , 2022, pp. 1–7, doi: 10.1109/DASC/PiCom/CBDCom/Cy55231.2...
2022
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.