{"id":"7e2a3811-52d7-420b-9ab9-965b8a8efc48","arxiv_id":"2412.15224","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A multi-branch Transformer with mutual distillation between raw EEG and wavelet bands achieves the best reported accuracy and balanced accuracy on CHSZ and TUSZ seizure subtype classification.","lead":"This paper proposes a Transformer model that mixes raw EEG signals with their frequency-band wavelets during training and uses mutual distillation between them. The authors report improved seizure subtype classification on two public EEG datasets, aiming to help when labeled data are scarce.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation does not establish the headline superiority claim: key CHSZ/TUSZ margins are within one standard deviation and no patient-level significance test is reported.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the reported performance gaps may not represent true superiority because of patient-level variance and the absence of significance testing. The central claim of the paper is empirical, so the validity of the cross-patient evaluation is the most load-bearing condition. The manuscript reports only means and standard deviations, and for several key comparisons the margins are within one standard deviation. The small numbers of patients per fold, especially in the rare TNSZ and TCSZ classes, make balanced accuracy very sensitive to which patients land in which fold. A patient-level permutation test is the natural check: it directly tests whether the observed metric difference could arise under patient shuffling. If the test fails to reject, the claimed 'outperformed' conclusion is not supported by the current evidence. The verdict should remain CONDITIONAL because the method may still be effective, but the reported experiments are insufficient to confirm the central claim.","tokens_in":15653,"tokens_out":9220,"duration_ms":91948,"concrete_test":"Using the released GitHub code, run the exact three-fold cross-patient splits and compute MBMD BCA minus WaveletTransformer (4-layer) BCA on CHSZ, then perform a paired patient-level permutation test: shuffle patient-to-fold assignments 10,000 times, recompute the same metric difference under the null, and report the two-sided p-value. If p > 0.05, the CHSZ superiority claim is not statistically established. Also verify in the splitter that no two windows from the same patient appear in different folds; any violation makes the comparison invalid regardless of the p-value.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MBMD Transformer outperforms traditional and deep baselines. The evidence is a set of means and standard deviations in Tables II–V, but the paper does not state whether the ten repeats are independent cross-validation splits or training seeds on fixed splits, and it performs no significance test. Under three-fold cross-patient validation with CHSZ=27 and TUSZ=68 patients, the effective comparison unit is the patient/fold, not the 4-second window or the training run. Several decisive margins are smaller than the reported variability: on CHSZ, MBMD BCA is 0.684 ± 0.066 versus WaveletTransformer (4-layer) at 0.670 ± 0.055; on TUSZ, MBMD ACC 0.746 ± 0.024 is tied with CE-stSENet 0.745 ± 0.070; and in Table III MBMD only ties BAKE on TUSZ ACC (0.746 vs 0.747). Because the minority classes (TNSZ=15, TCSZ=16 windows in CHSZ) make per-fold balanced accuracy highly volatile, a favorable fold assignment can produce the reported gap by chance. If the cross-validation splitter is window-level rather than patient-level, overlapping 50%-overlap windows leak between train and test, further invalidating the comparison. The paper claims 'cross-patient' splits, but no patient-ID bookkeeping is shown, so the correctness of the split is unverified from the manuscript alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Multi-Branch Mutual-Distillation (MBMD) Transformer for cross-subject EEG-based seizure subtype classification. The architecture replaces even-numbered encoder blocks of a Vision Transformer with multi-branch encoder blocks that process wavelet-decomposed frequency bands through separate expert FFNs, and a mutual-distillation loss transfers soft-label knowledge between the raw EEG branch and the wavelet branches. Experiments on the CHSZ and TUSZ datasets compare the method against traditional machine learning baselines, several deep learning models, five self-distillation approaches, and ablations, reporting that MBMD Transformer achieves the best or near-best accuracy, balanced accuracy, and weighted F1 on both datasets. The authors claim this is the first work on knowledge distillation for EEG-based seizure subtype classification and make their code publicly available.","tokens_in":15964,"tokens_out":5230,"duration_ms":49742,"significance":"If the empirical superiority claim is statistically reliable, the MBMD Transformer is a useful contribution to small-data EEG classification and represents a plausible first application of self-distillation to seizure subtype classification. The paper has notable strengths: it provides open-source code, compares against multiple baselines and self-distillation methods, includes ablation studies for the distillation loss and attention mechanism, and examines parameter sensitivity. The proposed mechanism—using deterministic wavelet decomposition to create auxiliary branches that mutually distill with the raw signal—is conceptually clear and avoids the circularity that would arise if the targets were fitted to the labels. However, the central claim rests on performance gaps that are often smaller than the reported variability, and the evaluation protocol is not specified tightly enough to establish statistical significance. As presented, the contribution is promising but not yet fully validated.","major_comments":[{"comment":"The central claim that MBMD Transformer \"outperformed several traditional machine learning and state-of-the-art deep learning approaches\" is not supported by any significance test, and several decisive comparisons show overlapping standard deviations: for example, on CHSZ the ACC is 0.650±0.071 for MBMD versus 0.632±0.073 for WaveletTransformer, and on TUSZ the ACC is 0.746±0.024 for MBMD versus 0.745±0.070 for CE-stSENet. Because the ultimate comparison unit is the patient or the fold, not the 4-second window or the training run, the paper should report per-fold results and perform paired tests (e.g., permutation or Wilcoxon signed-rank) at the patient/fold level. Without such tests, the reported margins are within the noise and do not establish superiority.","section":"IV-A, Tables II–III"},{"comment":"The sentence \"All reported results were the average of ten repeats\" is ambiguous: it is not stated whether each repeat is an independent three-fold cross-patient split with different patient assignments, or repeated training on the same fixed splits with different weight initializations. These two protocols yield very different interpretations of the reported standard deviations. The authors must specify the repeat protocol and, if the repeats are not independent patient splits, re-analyze the data with repeated patient-level splits to obtain valid estimates of patient-level variability.","section":"IV-A (experimental settings)"},{"comment":"Because the EEG is segmented into 4-second windows with 50% overlap, a genuine cross-patient evaluation requires that all windows belonging to the same patient be kept in the same fold. The manuscript states that three-fold cross-patient validation was used, but it does not provide patient-ID bookkeeping, per-fold patient counts, or a description of the splitting code. If overlapping windows from the same recording are split between training and test, the reported results would be inflated by patient leakage. Please clarify the exact splitting procedure and report per-fold patient and window counts for both datasets.","section":"IV-A (preprocessing and cross-validation)"},{"comment":"Equation (2) defines Lkl(pt||ps) with a leading minus sign, which makes it the negative of the Kullback-Leibler divergence. Minimizing this quantity in the overall loss of Eq. (7) would encourage divergence rather than agreement between the teacher and student distributions. This sign error mis-specifies the proposed training objective. The correct definition should be Lkl(pt||ps) = (1/K) Σ pt log(pt/ps), without the minus sign. The authors should correct Eq. (2) and verify that the reported experiments use the intended positive KL divergence.","section":"III-C, Eq. (2)"},{"comment":"The CHSZ dataset contains only 15 TNSZ and 16 TCSZ samples, and Table I appears to report the number of windows rather than patients. Under three-fold cross-validation, each fold will contain very few examples of these minority classes, making per-fold balanced accuracy highly volatile. The reported BCA means can therefore be dominated by a single favorable fold assignment. The paper should report per-fold balanced accuracy and class-wise accuracies, or provide patient-level bootstrap confidence intervals, so that the reader can judge whether the reported differences are stable.","section":"Table I and IV-B"}],"minor_comments":[{"comment":"The text uses \"FNN layers\" when referring to the feedforward network in the multi-branch encoder block; the rest of the paper uses \"FFN\". Please standardize the terminology.","section":"III-B"},{"comment":"Figures 7, 8, and 9 contain garbled labels and legends (e.g., Unicode placeholder sequences), making the plots unreadable. Please regenerate these figures with proper font embedding so that the axis labels and legend entries are legible.","section":"Figures 7–9"},{"comment":"The comparison with WaveletTransformer (4-layer) mentions that MBMD uses four encoder blocks while the baseline uses 24 encoders, but no parameter counts or computational costs are given. A parameter count and runtime comparison would strengthen the claim that MBMD \"uses the encoders more effectively.\"","section":"IV-B"},{"comment":"The claim that this is \"the first work on knowledge distillation for EEG-based seizure subtype classification\" is difficult to verify and is not supported by a systematic literature search. Consider softening this claim or providing a more comprehensive comparison with prior distillation-based EEG methods.","section":"Abstract and Contributions"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a plausible architecture and a thorough experimental apparatus, but the headline claim of superiority is not yet statistically substantiated. The repeat protocol is ambiguous, the cross-patient split bookkeeping is missing, and Eq. (2) has a mathematical sign error that affects the formal definition of the training loss. All of these issues are fixable within the scope of a revision. I recommend major revision and would be willing to review a revised version that adds significance tests and clarifies the evaluation protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The architecture is a sensible and clearly-described mashup of known components: multi-branch FFNs (MoE-style) for six wavelet bands, shared attention with the raw EEG, and a bidirectional mutual-distillation loss between raw and wavelet predictions. The wavelets are deterministic transforms, so the distillation target is not circular; it is a self-distillation training strategy, and the authors frame it honestly. The writing is clear, and the experiments are more thorough than most: comparisons with four traditional and five deep baselines, five self-distillation methods, and ablations on the loss, attention mechanism, block placement, temperature, and number of branches. Credit where due.\n\nThe soft spot is the one the stress test flags, and it lands. The headline claim is that MBMD outperforms SOTA, but the decisive margins are mostly inside one standard deviation. On CHSZ, MBMD BCA is 0.684±0.066 vs WaveletTransformer's 0.670±0.055; on TUSZ, MBMD ACC 0.746±0.024 is essentially tied with CE-stSENet 0.745±0.070 and BAKE 0.747±0.028. There is no significance test anywhere, and the paper does not say whether the ten repeats are independent cross-validation splits or different training seeds on fixed splits. With 27 and 68 patients, the effective sample size is the patient or the fold, not the window or the run. The minority classes in CHSZ are 15 and 16 windows, so per-fold balanced accuracy is extremely volatile. The paper claims patient-disjoint splits but shows no patient-ID bookkeeping; if the 50%-overlap windows are not kept together, leakage alone could explain the gap.\n\nThat said, the internal ablations are encouraging. Adding mutual distillation consistently helps over single-direction distillation, and the multi-branch block helps over plain ViT. So the training strategy is likely useful even if the SOTA claim is not proven. One small reproducibility wrinkle: the code link points to 'EBE-Transformer' rather than MBMD.\n\nWho is this for? Someone working on low-data EEG classification will find the distillation setup worth trying. It deserves a serious referee, but the authors need to add patient-level significance tests, clarify the repeated runs, and verify the split integrity. I would send it out, with the expectation of major revision.","headline":"The architecture is sensible and the experiments are thorough, but the reported statistics do not support the headline superiority claim.","tokens_in":16478,"tokens_out":2561,"would_cite":false,"duration_ms":24240,"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":"This paper claims that its Multi-Branch Mutual-Distillation Transformer classifies EEG seizure subtypes across patients better than prior machine-learning and deep-learning approaches on two public datasets.","keywords":["EEG","seizure subtype classification","Transformer","knowledge distillation","self-distillation","wavelet packet decomposition","cross-subject","mutual distillation"],"falsifier":"Re-running the reported three-fold cross-patient protocol while auditing that every window from a given patient and recording is confined to a single fold, then measuring the balanced accuracy gap between MBMD Transformer and WaveletTransformer on CHSZ and TUSZ; if the gap disappears under leakage-free splits or the variance across the three folds is comparable to the gap itself, the claimed superiority is not established.","tokens_in":15474,"feed_emoji":"🧠","tokens_out":7331,"duration_ms":58104,"temperature":0.7,"pith_summary":"This paper claims that a modified Vision Transformer, the Multi-Branch Mutual-Distillation (MBMD) Transformer, classifies EEG seizure subtypes across different patients better than existing traditional machine-learning and deep-learning baselines, while needing only small labeled datasets. The method decomposes each raw EEG trial into six wavelet frequency bands, feeds the raw signal and all six wavelets through shared attention layers with separate expert feedforward branches, and trains them with a bidirectional knowledge-distillation loss so the raw-data path and the wavelet branches teach each other. On the CHSZ and TUSZ public datasets, the authors report the highest accuracy, balanced accuracy, and weighted F1 scores among the compared methods, and ablation studies indicate that both the mutual-distillation loss and the learned wavelet-attention weights contribute to the gains. If correct, the work offers a way to train clinically useful seizure subtype classifiers from the limited labeled data available in practice.","feed_headline":"Mutual-distillation transformer beats prior EEG seizure subtype models","feed_subtitle":"Pairs raw EEG with six wavelet frequency bands and distills knowledge both ways to classify seizures across patients.","key_machinery":"The load-bearing mechanism is the multi-branch encoder block paired with wavelet packet decomposition. Each raw EEG trial is decomposed into six frequency bands — delta (0–4 Hz), theta (4–8 Hz), alpha (8–16 Hz), beta (16–32 Hz), gamma (32–64 Hz), and a residual band — and all six wavelets plus the raw signal share the same linear projection and multi-head attention layers, while each wavelet gets its own expert feedforward network inside the even-numbered blocks. A branch-wise wavelet attention mechanism learns an adaptive weight vector (softmax-normalized, with L1 regularization) that is used to ensemble the expert outputs for the raw-data path. Training couples the raw-data classifier and the six wavelet classifiers through a mutual-distillation loss: the sum of KL divergences in both directions, at temperature T, so each branch acts as both student and teacher. At test time only the raw EEG is fed in, with the ensemble of expert feedforwards used for classification.","core_discovery":"On its own terms, the paper establishes that a transformer architecture can be made to learn cross-subject seizure subtype classification from small labeled EEG data by forcing collaboration between the raw signal and its wavelet decompositions. Specifically, the authors replace every even-numbered encoder block of a vanilla Vision Transformer with a multi-branch encoder block in which six wavelet-specific expert feedforward networks process the raw EEG simultaneously, and a wavelet attention mechanism learns per-branch weights that are made sparse by L1 regularization. A mutual-distillation loss, defined as the average of the KL divergences in both directions between the raw-data prediction and each branch prediction at temperature T, transfers knowledge back and forth. In experiments on CHSZ (27 patients) and TUSZ (68 patients), the MBMD Transformer achieved the best ACC, BCA, and weighted F1 among nine baseline classifiers (SVM, RC, LR, GBDT, EEGNet, TIE-EEGNet, CE-stSENet, ViT, and WaveletTransformer) and among five prior self-distillation methods adapted to a ViT backbone. Ablations show that the bidirectional distillation term outperforms the one-way version and that the learned attention beats simple averaging and a gate network, and the authors state this is the first application of knowledge distillation to EEG-based seizure subtype classification.","pith_inferences":["The same mutual-distillation scheme could transfer to other frequency-sensitive BCI tasks such as sleep staging or motor imagery, where sub-band decompositions also carry discriminative information.","The reported insensitivity to the number of branches suggests the gain may come more from the distillation objective than from the specific band definitions; a test with randomized band splits on a fixed six-branch model would separate the two.","The chosen hyperparameters (T=6, lambda=0.01) are dataset-specific; a validation-set sweep would likely be needed when applying the model to new montages or recording hardware."],"forward_implications":["The method could be adopted for clinical EEG archives where labeled seizures are scarce, because it does not require a large pretrained teacher.","The bidirectional distillation loss is a drop-in training objective that could improve other transformer-based EEG classifiers without changing their inference path.","Reducing the branch count from six to two or three retains most of the benefit, so the model can be made cheaper for deployment.","Being the first distillation-based approach for seizure subtype classification, it opens that task to the wider toolbox of distillation methods."],"supporting_citations":[{"why":"Provides the CHSZ dataset and the three-fold cross-patient evaluation protocol used in all experiments.","marker":"[24]"},{"why":"Provides the TUSZ (V1.5.2) dataset used as the second evaluation benchmark.","marker":"[37]"},{"why":"Supplies the wavelet packet decomposition scheme that generates the six frequency-band auxiliary branches.","marker":"[26]"},{"why":"Supplies the Vision Transformer backbone that the multi-branch encoder blocks modify.","marker":"[27]"},{"why":"Supplies the mutual-learning idea of peer models teaching each other, which the bidirectional distillation loss implements.","marker":"[17]"},{"why":"Supplies the KL-divergence-with-temperature formulation used in the distillation loss.","marker":"[16]"},{"why":"Supplies the 'multi-view' theory that motivates why ensemble and distillation should improve test accuracy.","marker":"[36]"},{"why":"One of the self-distillation baselines (BYOT), adapted to a ViT backbone for comparison in Table III.","marker":"[31]"},{"why":"Another self-distillation baseline (ONE), adapted to a ViT backbone for comparison in Table III.","marker":"[32]"},{"why":"Supplies the 41 features used for the traditional machine-learning baselines in the comparisons.","marker":"[21]"}],"fun_headline_variants":["First mutual-distillation transformer for EEG seizure subtyping","Small-data transformer distills EEG seizure subtypes","Raw plus wavelet branches distill knowledge for seizure subtype labeling","Bidirectional distillation improves cross-subject EEG seizure classification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that the three-fold cross-patient splits built from overlapping 4-second windows keep each patient's data entirely within one fold, and that 27 and 68 patients are enough for the reported means and standard deviations to be stable; if windows leak across folds or the patient counts make the variance estimates unreliable, the reported performance gaps may not hold.","fun_headline_variants_meta":{"raw":{"variants":["First mutual-distillation transformer for EEG seizure subtyping","Small-data transformer distills EEG seizure subtypes","Raw plus wavelet branches distill knowledge for seizure subtype labeling","Bidirectional distillation improves cross-subject EEG seizure classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000969,"raw_usage":{"total_tokens":4137,"prompt_tokens":973,"completion_tokens":3164,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":3102}},"tokens_in":589,"tokens_out":3164,"duration_ms":20993,"temperature":1.0,"reasoning_tokens":3102,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:37:34.059290+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-running the reported three-fold cross-patient protocol while auditing that every window from a given patient and recording is confined to a single fold, then measuring the balanced accuracy gap between MBMD Transformer and WaveletTransformer on CHSZ and TUSZ; if the gap disappears under leakage-free splits or the variance across the three folds is comparable to the gap itself, the claimed superiority is not established.","supporting_citations":[{"cited_title":"TIE-EEGNet: Temporal information enhanced EEGN et for seizure subtype classification,","cited_arxiv_id":null,"evidence_quote":"Provides the CHSZ dataset and the three-fold cross-patient evaluation protocol used in all experiments."},{"cited_title":"The Temple university h ospital seizure detection corpus,","cited_arxiv_id":null,"evidence_quote":"Provides the TUSZ (V1.5.2) dataset used as the second evaluation benchmark."},{"cited_title":"W A VELET2VEC: A filter bank masked autoencoder for EEG-base d seizure subtype classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the wavelet packet decomposition scheme that generates the six frequency-band auxiliary branches."},{"cited_title":"An image is worth 16x16 words: Transformers for image recognition at scale,","cited_arxiv_id":null,"evidence_quote":"Supplies the Vision Transformer backbone that the multi-branch encoder blocks modify."},{"cited_title":"Deep mut ual learn- ing,","cited_arxiv_id":null,"evidence_quote":"Supplies the mutual-learning idea of peer models teaching each other, which the bidirectional distillation loss implements."},{"cited_title":"Be yo ur own teacher: Improve the performance of convolutional neural n etworks via self distillation,","cited_arxiv_id":null,"evidence_quote":"One of the self-distillation baselines (BYOT), adapted to a ViT backbone for comparison in Table III."},{"cited_title":"Knowledge distillation by o n-the-fly native ensemble,","cited_arxiv_id":null,"evidence_quote":"Another self-distillation baseline (ONE), adapted to a ViT backbone for comparison in Table III."},{"cited_title":"Source-free domain adaptat ion (SFDA) for privacy-preserving seizure subtype classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the 41 features used for the traditional machine-learning baselines in the comparisons."}],"review_version":1}