{"id":"d80fb810-dc3c-441c-b916-7ea5aa0ee7e8","arxiv_id":"2509.02746","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Mamba-based EEG model with self-supervised pretraining reaches AUROC 0.72 for seizure detection on the TUSZ corpus.","lead":"This paper trains a Mamba-based neural network on EEG recordings to detect seizures, reporting AUROC 0.72 on a held-out test set after self-supervised reconstruction pretraining. It is an early step toward EEG foundation models, but the result is modest, lacks baselines and error bars, and the 'foundation model' label overstates the scope.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pretraining gain is confounded by unequal training budgets; without matching total optimization steps, the 0.64→0.72 AUROC improvement cannot be attributed to self-supervised pretraining.","rationale":"The paper's central empirical claim is that self-supervised reconstruction pretraining improves seizure detection. The comparison in §3.2 is confounded by unequal training budgets: the pretrained model receives additional optimization steps during pretraining, so any improvement could stem from longer training rather than the learned representations. This is the single most load-bearing concern because it directly undermines the 'importance of pretraining' claim, which is a central contribution of the paper. Other limitations (no baselines, no error bars, overbroad 'foundation model' label) are significant but less fundamental. If the test proposed above resolves the confound, the paper would still need error bars and baselines, but its main claim would be on firmer ground. The reader's CONDITIONAL verdict already appropriately hedges on this concern, so no change in verdict is needed.","tokens_in":4327,"tokens_out":7188,"duration_ms":74413,"concrete_test":"Train a from-scratch model on the detection task for a total number of gradient steps equal to the sum of the pretraining and fine-tuning steps used for the pretrained model. If this budget-matched from-scratch model achieves AUROC within 0.02 of 0.72, the pretraining-specific benefit is not supported; if it remains near 0.64, the pretraining claim is provisionally validated. Report the mean and standard deviation over at least 3 seeds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 reports AUROC 0.72 with pretraining and 0.64 without, and the Discussion states this 'boosts the downstream AUROC significantly, pointing to the importance of this self-supervised pretraining approach.' However, the paper provides no information about the number of epochs, optimization steps, learning rate schedules, or total compute budget for either condition. If the from-scratch model was trained for fewer iterations than the pretrained model's pretraining plus fine-tuning, the apparent gain could be due solely to additional training budget, not to the reconstruction task's learned representations. This is the key load-bearing claim because the paper's narrative and abstract emphasize pretraining as the source of the model's effectiveness. Without controlling for this confound, the 0.08 AUROC gap is uninterpretable. Additionally, no error bars or multiple seeds are reported, so we cannot tell if the difference is statistically meaningful.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 'Mentality,' a Mamba-based encoder-decoder for EEG, trained in two stages: self-supervised reconstruction on the Temple University Hospital EEG Seizure Corpus (TUSZ) v2.0.1, followed by supervised seizure detection. The authors report a held-out AUROC of 0.72 with pretraining and 0.64 when training from scratch, and claim that self-supervised pretraining significantly improves downstream detection. They also describe a spectral loss that reduces reconstruction MSE, and provide qualitative interpretability analyses via channel saliency and first-layer weights.","tokens_in":4593,"tokens_out":1964,"duration_ms":24461,"significance":"If the claims hold, the paper would be a useful positive data point for Mamba-style state-space models applied to EEG, and for self-supervised pretraining on physiological signals. The use of a large public corpus with patient-disjoint train/test splits is a genuine strength, and the held-out AUROC is a plausible quantitative signal. However, the central comparison between pretrained and from-scratch models is currently under-controlled, and the absence of uncertainty estimates, strong baselines, and full experimental details weakens the contribution. The work is preliminary but the core idea is worth pursuing; the manuscript needs additional experiments and reporting before the claims can be accepted.","major_comments":[{"comment":"The central claim that self-supervised pretraining improves AUROC from 0.64 to 0.72 is confounded by training budget. The paper does not report the number of epochs, optimization steps, learning rate schedules, or total compute for either the pretrained-plus-fine-tuned model or the from-scratch model. If the from-scratch model was trained for fewer iterations, the gain could be due to additional training budget rather than the representations learned by the reconstruction task. Please report matched training conditions, including total steps and wall-clock time, or run an ablation where the from-scratch model receives the same total number of optimization steps as pretraining plus fine-tuning.","section":"Section 3.2 and Discussion"},{"comment":"Only a single AUROC value is reported for each condition, with no confidence intervals, no standard deviations across seeds, and no repeated runs. Because the difference is only 0.08 AUROC, it is not possible to assess whether this is statistically meaningful. Please provide at least 3–5 independent runs with reported mean and standard deviation, and ideally a bootstrap confidence interval on the held-out test set.","section":"Section 3.2"},{"comment":"The spectral loss is described only in words ('computes the loss in the Fourier domain') and no equation or weighting hyperparameter is provided. Since the paper explicitly credits the spectral loss for a four-fold MSE reduction, this term is load-bearing for the pretraining result. Please define the loss precisely, including the relative weight of MSE and spectral components, and report the hyperparameter value used.","section":"Section 2.2"},{"comment":"The architecture description omits several details needed to reproduce the model: number of Mamba blocks, hidden dimensions at each downsampling level, downsampling factors, kernel sizes for the double convolution blocks, and the size of the final hidden representation. For a paper proposing a new architecture, these choices are not presentation trivia; they affect the validity of the comparison and the usefulness of the results. Please include a table of architectural hyperparameters and training details.","section":"Section 2.2 and Table-free results"}],"minor_comments":[{"comment":"The data section states the test subset contains '43 patients, with 469 seizure events' but does not report the number of seizure versus non-seizure windows in the train or test sets. Reporting the window-level class balance is important for interpreting a single AUROC value, especially if the test set is highly imbalanced.","section":"Section 2.1"},{"comment":"The reconstruction MSE values (0.0063 and 0.025) are reported without units or normalization. Since raw EEG amplitude scales with preprocessing, it would be clearer to also report a normalized metric such as relative MSE or signal-to-noise ratio.","section":"Section 3.1"},{"comment":"The interpretability analysis is qualitative. Figure 3 shows a saliency map for one seizure sample; it would strengthen the paper to quantify saliency consistency across samples or to validate the highlighted channels against clinical annotations.","section":"Section 3.3"},{"comment":"The paper is formatted as a conference submission and contains no code or data availability statement. Given the emphasis on reproducibility in deep learning research, please add a statement about releasing code and model checkpoints.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper's central finding is plausible but not yet convincing because the pretraining comparison lacks matched training budgets and uncertainty quantification. The missing spectral loss equation and architectural details also impede reproducibility. These are fixable within the scope of the paper, so I recommend major revision rather than rejection. I would also flag that the 'foundation model' framing is stronger than what the experiments support, since only a single downstream task is evaluated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The punchline: this is a modest, incremental paper. The architecture is a Mamba-U-Net hybrid with an EEGNet-style first layer, and the pretraining task is reconstruction with MSE plus spectral loss. The paper does not deliver a foundation model, and the 0.72 AUROC is not compared to any strong baseline. But the spectral loss result is real and simple: removing it increases reconstruction MSE fourfold. That is the most valuable piece of the paper.\n\nWhat is new: applying Mamba blocks to EEG in an encoder-decoder with U-Net-style down/upsampling is a legitimate extension of known work (SaShiMi, EEGNet, U-Net), but it is not a new technique. The paper does not cite prior Mamba-for-EEG work, so novelty relative to its own reference list is thin.\n\nThe soft spots. Section 3.2 reports AUROC 0.72 pretrained versus 0.64 from scratch, and the Discussion attributes the gap to pretraining. But no training budgets, epochs, or schedules are given for either condition. The from-scratch model could simply be undertrained. That is a load-bearing confound because the abstract and Discussion lean on pretraining. No error bars or multiple seeds either, so the 0.72 versus 0.64 gap might not be stable. The comparison is only against the same architecture without pretraining, not against EEGNet or other standard detectors, so the clinical or practical significance is unclear.\n\nThe Discussion is mostly future work, and the \"foundation model\" framing is not supported by one task on one corpus. Also, the paper header says \"Published as a conference paper at ICLR 2024\" while the arXiv version is dated September 2025. That may be a template artifact, but it should be fixed or explained.\n\nBottom line: this is a reasonable workshop-level result that needs stronger evaluation before it can support the claims made. The central question—whether Mamba pretraining helps EEG—is worth testing properly, and the spectral loss finding is useful. A serious referee should push for matched training budgets, multiple seeds, baselines, and a narrower title. It deserves referee time, but as is it needs substantial revision.","headline":"A modest Mamba-on-EEG paper with one useful observation (spectral loss helps reconstruction) and a central pretraining claim that is confounded by unstated training budgets.","tokens_in":5031,"tokens_out":2095,"would_cite":false,"duration_ms":23694,"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":"A Mamba-based EEG model reaches 0.72 AUROC on seizure detection after self-supervised pretraining.","keywords":["EEG","Mamba","selective state space models","seizure detection","self-supervised pretraining","foundation models","TUSZ","spectral loss"],"falsifier":"Train the same model from scratch with the same number of epochs, same optimizer and learning-rate schedule, and same compute budget as the pretrained-then-fine-tuned model; if the from-scratch AUROC rises to about 0.72, the reported pretraining gain is explained by training conditions rather than the reconstruction objective.","tokens_in":4252,"feed_emoji":"🧠","tokens_out":5660,"duration_ms":60559,"temperature":0.7,"pith_summary":"The paper tries to establish that a Mamba-based selective state-space model can act as a foundation model for EEG: pretrain an encoder-decoder to reconstruct raw seizure and non-seizure windows, then fine-tune the latent representation for seizure detection. On a held-out set of 43 patients from the TUSZ corpus, the pretrained model reaches a detection AUROC of 0.72, compared with 0.64 when trained from scratch. The paper also shows that adding a spectral (Fourier-domain) loss to plain mean-squared error improves reconstruction, lowering MSE from 0.025 to 0.0063. A reader should care because EEG is noisy, high-dimensional, and labor-intensive to read, and a transferable pretrained model could support diverse neurological diagnoses and recording setups.","feed_headline":"Mamba EEG model hits 0.72 AUROC for seizure detection","feed_subtitle":"Self-supervised pretraining on raw EEG lifts detection AUROC from 0.64 to 0.72 on held-out patients.","key_machinery":"The load-bearing object is the stacked Mamba block—a selective state-space layer with layer normalization and a residual connection—embedded in a SaShiMi/U-Net-style encoder-decoder. Mamba is a sequence model that selectively decides what to retain at each time step, allowing it to handle long-range temporal dependencies in EEG at linear cost. The block does two jobs: during pretraining it must reconstruct the full signal through symmetric downsampling and upsampling, and after fine-tuning its latent representation feeds the seizure classifier. The initial CNN and channel-mixing layer translate raw channels into frequency-based features, and the spectral loss keeps the reconstruction faithfu","core_discovery":"The central claim is that self-supervised reconstruction pretraining is what makes a Mamba EEG model useful for seizure detection. The authors build a U-Net-like encoder-decoder whose core layers are Mamba blocks, preceded by a per-channel convolutional layer that learns frequency filters up to 50 Hz and a channel-mixing layer. After pretraining on 10-second TUSZ windows with an MSE-plus-spectral loss, they attach two linear layers to the downsampled latent and train a seizure classifier, reaching AUROC 0.72 on a held-out test set of 43 patients; the same architecture trained from scratch reaches only 0.64. Channel-wise saliency on a seizure sample highlights channels T4 and P4, which the au","pith_inferences":["A matched-budget comparison—same number of epochs, same optimizer schedule, and same compute for from-scratch and pretrained runs—would settle whether pretraining itself, rather than longer or different training, causes the 0.08 AUROC gain; the paper does not provide that control.","Mamba's linear-time inference should make continuous hour-scale EEG processing feasible, but the paper only tests 10-second windows.","Explicitly using the learned state-space dynamics, rather than only the final hidden latent, could improve both performance and interpretability; the paper itself notes those dynamics are not currently used.","The proposed channel-masked pretraining could be tested directly by random channel dropout during training and measuring AUROC on partial-channel inputs, mimicking wearable EEG setups."],"forward_implications":["If the 0.72 AUROC holds, self-supervised reconstruction on unlabeled EEG is enough to pretrain useful representations for seizure detection.","The fourfold MSE change implies that frequency-domain supervision should be retained in follow-up work; without it reconstruction degrades substantially.","Because downstream classification needs only two linear layers on the latent, the same pretrained backbone could be fine-tuned cheaply for other EEG tasks.","The channel-mixing architecture suggests possible adaptation to different montages, though the paper only tests the standard 19-channel setup.","Expanded to larger corpora spanning other conditions, the model could become a general EEG foundation model rather than a seizure detector."],"supporting_citations":[{"why":"Supplies the Mamba selective state-space blocks that the model stacks to model temporal dynamics.","marker":"Gu & Dao, 2023"},{"why":"Provides the TUSZ seizure corpus used for pretraining and evaluation, including patient splits and annotations.","marker":"Shah et al., 2018"},{"why":"Inspires the initial per-channel CNN and channel-weight interpretability approach for EEG.","marker":"Lawhern et al., 2018"},{"why":"Supplies the SaShiMi-style U-Net architecture with state-space blocks that the encoder-decoder follows.","marker":"Goel et al., 2022"},{"why":"Provides the U-Net downsampling/upsampling and residual concatenation structure used throughout the architecture.","marker":"Ronneberger et al., 2015"}],"fun_headline_variants":["Self-supervised Mamba EEG boosts seizure AUROC to 0.72","Mamba EEG pretraining lifts seizure detection AUROC by 0.08","Mamba EEG: Self-supervised pretraining raises seizure AUROC to 0.72","EEG foundation model: Mamba hits 0.72 AUROC after self-supervision"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The paper credits self-supervised pretraining with improving the model's seizure-detection score from 0.64 to 0.72, but it does not report matching training budgets, number of passes over the data, or optimizer schedules for the two runs, so the improvement may come from confounded training conditions rather than the pretraining task itself.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised Mamba EEG boosts seizure AUROC to 0.72","Mamba EEG pretraining lifts seizure detection AUROC by 0.08","Mamba EEG: Self-supervised pretraining raises seizure AUROC to 0.72","EEG foundation model: Mamba hits 0.72 AUROC after self-supervision"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000857,"raw_usage":{"total_tokens":3534,"prompt_tokens":697,"completion_tokens":2837,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":441,"completion_tokens_details":{"reasoning_tokens":2747}},"tokens_in":441,"tokens_out":2837,"duration_ms":19569,"temperature":1.0,"reasoning_tokens":2747,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:25:51.768979+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same model from scratch with the same number of epochs, same optimizer and learning-rate schedule, and same compute budget as the pretrained-then-fine-tuned model; if the from-scratch AUROC rises to about 0.72, the reported pretraining gain is explained by training conditions rather than the reconstruction objective.","supporting_citations":[{"cited_title":"Eegnet: a compact convolutional neural network for eeg-based brain--computer interfaces","cited_arxiv_id":null,"evidence_quote":"Inspires the initial per-channel CNN and channel-weight interpretability approach for EEG."},{"cited_title":"It's raw! audio generation with state-space models, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the SaShiMi-style U-Net architecture with state-space blocks that the encoder-decoder follows."}],"review_version":1}