{"id":"ae260410-60e0-44d9-b013-ab2bb1891e94","arxiv_id":"2411.16155","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A frozen EEG temporal backbone can be adapted to spatial sensor relationships by fine-tuning a small graph adapter, improving F1 on MDD and abnormality detection.","lead":"EEG-GraphAdapter (EGA) adds a graph neural network layer in front of a frozen, pre-trained EEG model, fine-tuning only that adapter to learn spatial relationships between sensors. On two clinical EEG tasks it reports higher F1 scores than fully fine-tuning the backbone, while updating far fewer parameters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TUAB evaluation may be contaminated by pre-training on overlapping TUEG subjects; the headline 16.1% improvement is unsupported without subject-level de-duplication.","rationale":"After reading the paper, I agree with the reader that the most load-bearing assumption is the independence of the TUAB evaluation set from the TUEG pre-training corpus. The paper pre-trains BENDR on TUEG (§3.1.1) and evaluates on TUAB (§3.1.2), which is known to be a derived subset of the Temple University Hospital EEG corpus. Without explicit subject-level exclusion, the 16.1% F1 improvement on TUAB cannot be trusted because the frozen BENDR may already encode information about the test subjects. This is not an objection to the EGA idea, but to the strength of the evidence; the MDD experiment provides some independent support, though with only 126 samples and no error bars. The reader's conditional verdict is appropriate: the paper can be accepted if the authors demonstrate that the TUAB subjects were excluded from pre-training or rerun the experiment with a clean split. Other concerns, such as the absence of error bars and the post-hoc selection of the best GNN variant per task, are secondary to this data-integrity issue. If the overlap check comes back negative, the TUAB result would be credible, though the variance remains unknown. Therefore, I do not change the reader's verdict.","tokens_in":9200,"tokens_out":8182,"duration_ms":72059,"concrete_test":"Obtain the patient IDs (or session IDs) for the 40 TUAB subjects used in Section 3.1.2 and query them against the TUEG pre-training metadata for the 69,652 samples in Section 3.1.1. If any IDs overlap, re-train BENDR on TUEG with all TUAB subjects held out, rerun the Table 3 comparison, and check whether EGA-GraphSAGE still beats the baseline by about 16% F1. If the improvement vanishes, the TUAB evidence is an artifact of pre-training leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To establish that EGA improves over the baseline on TUAB by 16.1% in F1 (Table 3), the TUAB test set must be independent of the TUEG corpus used to pre-train BENDR. Section 3.1.1 states that BENDR was pre-trained from scratch on TUEG v2.0.1 (69,652 samples). Section 3.1.2 then evaluates on TUAB [16], a dataset from the same Temple University Hospital EEG collection. The paper does not state that TUAB subjects were removed from the TUEG pre-training set, nor that the downstream cross-validation is subject-wise disjoint from pre-training. Since TUAB is a subset of TUEG, overlap is likely. If the same patients are present in both, the frozen BENDR can carry memorized subject-identifying patterns that correlate with the normal/abnormal label. EGA, which keeps BENDR frozen, may preserve this signal, while the fully fine-tuned baseline may overwrite it, producing an advantage unrelated to spatial graph adaptation. This would invalidate the TUAB result as evidence for EGA. The MDD result (Table 2) remains independent, but the paper's headline 'up to 16.1%' rests on the contaminated comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":9434,"tokens_out":4093,"duration_ms":41911,"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":[{"comment":"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":"Section 3.1.1 and Section 3.1.2, Table 3"},{"comment":"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.","section":"Section 3.1.2, Table 1"},{"comment":"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.","section":"Tables 2 and 3"},{"comment":"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.","section":"Abstract and Introduction"}],"minor_comments":[{"comment":"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":"Section 3.1.3"},{"comment":"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.","section":"Section 2.1.3"},{"comment":"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":"Tables 2 and 3"},{"comment":"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.","section":"Section 4"},{"comment":"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.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The core idea is reasonable and the paper is concise, but the experimental validation needs substantial rework before publication. The TUAB overlap concern is likely fixable by re-running with subject-level disjointness, and the MDD experiment needs subject-wise folds and error bars. I would also expect a comparison with at least one other PEFT baseline to support the claims of efficiency and generality."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is simple and reasonably new: put a GNN adapter in front of a frozen BENDR encoder so you tune about a million parameters instead of 6.4 million. That combination—EEG + frozen foundation model + GNN adapter—isn't in the prior work I know, and the paper cites the nearest NLP analogues (G-Adapter, GraphAdapter, TAPE) honestly. The architecture description is clear, and the parameter counts are reported transparently. They also admit that the wall-clock speedup is only about 17% despite the big parameter reduction, which is more candid than most such papers.\n\nThe experimental support, however, is not strong enough for the headline claims. The biggest problem is that the cross-validation splits segments, not subjects. For MDD, 63 subjects yield 126 segments; for TUAB, 40 subjects yield 517 segments. In both cases, 10-fold or 5-fold on segments means the same subject's data almost certainly appears in both training and test folds. EEG models are good at memorizing subject-specific patterns, so the reported F1 gains could just reflect that. The stress-test concern about TUAB contamination is also real: the paper pre-trains BENDR on TUEG v2.0.1 and evaluates on TUAB, both from Temple University Hospital, and it never states that TUAB patients were excluded from pre-training. If they weren't, the frozen BENDR can carry subject-specific signal that the adapter preserves but full fine-tuning might overwrite, which would inflate the EGA advantage. The MDD result is independent of pre-training, but it still inherits the segment-level leakage.\n\nThere are also no error bars, and the best GNN variant is selected per task after seeing the test results (GAT for MDD, GraphSAGE for TUAB). With no variance estimates, differences of 12.8% and 16.1% could easily be noise. I'd want a subject-wise split, multiple seeds, and a proper held-out model selection before trusting either number.\n\nThe paper is a legitimate engineering contribution and the writing is clear, but as a scientific claim about EGA improving EEG classification, the current evidence is weak. It's fixable: rerun with subject-wise folds, report confidence intervals, and either exclude TUAB from pre-training or show de-duplication. If I were an editor, I'd send it to a referee because the topic is timely and the fix is well-defined, but I'd expect major revisions before it could be published.","headline":"Plausible PEFT idea for EEG with a frozen backbone, but the experiments are undermined by segment-level cross-validation and likely TUAB/TUEG overlap.","tokens_in":9934,"tokens_out":3743,"would_cite":false,"duration_ms":36921,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["parameter-efficient fine-tuning","EEG foundation models","graph neural networks","BENDR","spatial sensor graph","MDD classification","TUAB abnormality detection","GNN adapter"],"falsifier":"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.","tokens_in":9004,"feed_emoji":"🧠","tokens_out":5167,"duration_ms":43687,"temperature":0.7,"pith_summary":"The paper tries to establish that spatial information between EEG sensors can be added to a frozen pretrained temporal model cheaply and effectively. It proposes EEG-GraphAdapter (EGA), a GNN module placed before BENDR, and fine-tunes only the adapter while keeping BENDR's 6.46M parameters fixed. On two binary clinical tasks, this improves F1 by up to 12.8% on MDD and 16.1% on TUAB compared with fully fine-tuning BENDR, while reducing trainable parameters to roughly 1M–2M. If correct, EGA would make large EEG foundation models usable for small healthcare datasets by avoiding both heavy fine-tuning and overfitting.","feed_headline":"A small graph adapter boosts a frozen EEG model's F1 by 16%","feed_subtitle":"Tuning only the GNN adapter, not the 6.4M-parameter BENDR backbone, raises abnormality-detection scores on TUAB and MDD.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Defines BENDR, the frozen temporal backbone and the baseline to beat.","marker":"[1]"},{"why":"Supplies the geodesic-distance weighted graph construction used for EGA's adjacency matrix.","marker":"[6]"},{"why":"Provides the GCN message-passing variant evaluated as one EGA adapter.","marker":"[11]"},{"why":"Provides the GraphSAGE variant that achieves the best TUAB result.","marker":"[12]"},{"why":"Provides the GAT variant that achieves the best MDD result.","marker":"[13]"},{"why":"Supplies the MDD dataset used in the first downstream evaluation.","marker":"[15]"},{"why":"Supplies the TUAB abnormality-detection dataset used in the second downstream evaluation.","marker":"[16]"}],"fun_headline_variants":["Tiny graph adapter lifts frozen EEG model's F1 by 16%","Graph adapter fine-tunes EEG foundation model with minimal cost","One small GNN module boosts EEG model's F1 by 16% on TUAB","Parameter-efficient EEG: graph adapter beats full fine-tuning","EEG graph adapter adds spatial smarts to frozen temporal model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Tiny graph adapter lifts frozen EEG model's F1 by 16%","Graph adapter fine-tunes EEG foundation model with minimal cost","One small GNN module boosts EEG model's F1 by 16% on TUAB","Parameter-efficient EEG: graph adapter beats full fine-tuning","EEG graph adapter adds spatial smarts to frozen temporal model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000236,"raw_usage":{"total_tokens":1502,"prompt_tokens":941,"completion_tokens":561,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":469}},"tokens_in":557,"tokens_out":561,"duration_ms":5697,"temperature":1.0,"reasoning_tokens":469,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:28:42.600319+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines BENDR, the frozen temporal backbone and the baseline to beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the geodesic-distance weighted graph construction used for EGA's adjacency matrix."},{"cited_title":"N., Welling, M","cited_arxiv_id":null,"evidence_quote":"Provides the GCN message-passing variant evaluated as one EGA adapter."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GraphSAGE variant that achieves the best TUAB result."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GAT variant that achieves the best MDD result."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the TUAB abnormality-detection dataset used in the second downstream evaluation."}],"review_version":1}