{"id":"e43d766f-e3a0-49ef-85a0-7fa0f2b13aad","arxiv_id":"2506.17068","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"EpiNT, a transformer pretrained on over 2,700 hours of EEG and iEEG from 1,199 patients with masked autoencoding and a frequency-domain quantizer, matches or beats other pretrained models on six epilepsy classification benchmarks.","lead":"A new pretrained transformer, EpiNT, learns shared representations from scalp and intracranial EEG by combining masked autoencoding with a random frequency-domain quantizer. It is tested against other pretrained models on six epilepsy classification tasks and often, but not always, achieves the best F1 score.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Modality-specific preprocessing (256 Hz/12 s vs 1024 Hz/3 s, §4.8) is a perfect source-modality confound, and no downstream task crosses modalities, so the unified-harmonization claim is untested.","rationale":"The reader's weakest assumption and my stress-test converge on the same point: modality-specific preprocessing is an untested confound. My analysis adds that the absence of any cross-modal downstream task makes the problem more direct, because the six reported tasks can all be solved with within-modality features. This is the most load-bearing issue because the central contribution is not merely a strong per-modality pretraining method but 'cross-modal harmonization' and a 'unified' model, as stated in the title, abstract, and introduction. If the model simply encodes the sampling-rate/epoch-length signature, the empirical results are compatible with a much weaker claim. The concern is concrete and testable, and the authors themselves flag the preprocessing as an implicit modality annotation in the Discussion. I do not recommend rejection because the linear-probing results are consistently positive, the comparison to other pretrained models is informative, and the confound can be controlled with additional experiments. Requiring the modality-probe control and, if it succeeds, a common-preprocessing or cross-modal transfer check is the appropriate condition for acceptance.","tokens_in":19257,"tokens_out":12018,"duration_ms":135977,"concrete_test":"Freeze the pretrained EpiNT encoder; extract [cls] embeddings for a balanced sample of scalp and iEEG pretraining epochs; train a logistic-regression (or linear SVM) probe to predict which preprocessing pipeline (256 Hz/12 s vs 1024 Hz/3 s) generated each input, using patient-disjoint 5-fold cross-validation. Report AUROC against chance. If AUROC is near 0.5, the confound is not encoded and the harmonization claim survives; if AUROC is high (e.g., >0.9), the representation contains the modality cue, and the authors must add a control (e.g., resampling both modalities to a common rate or adding a same-modality cross-pipeline transfer task) before the unified-analysis claim is credible. This probe requires only the promised released weights and code.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing premise of the title and abstract is that EpiNT learns a shared, modality-invariant neurophysiological representation. The preprocessing fixed in Section 4.8 makes scalp EEG and iEEG perfectly separable by construction: scalp inputs are always 256 Hz, 12 s epochs; iEEG inputs are always 1024 Hz, 3 s epochs, both truncated to 3072 points. Every pretraining and fine-tuning example therefore carries an exact one-dimensional marker of its source modality, so the model is never forced to form a common cross-modal representation. The six downstream tasks in Table 2 are each single-modality (three scalp, three iEEG), and Figure 3's modality-transfer experiment only varies pretraining data composition; it does not test whether a representation trained on one modality transfers to a task on the other. The Discussion explicitly concedes the implicit modality annotation but offers no control. Unless the model is shown not to exploit this cue, the improved downstream F1 scores do not establish harmonized cross-modal representations; they may reflect high-capacity within-modality models plus a cheap modality tag.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes EpiNT, a channel-independent Transformer pre-trained with masked autoencoding and vector quantization. The pre-training uses a 'frequency domain mapping quantizer' that assigns pseudo-labels by cosine similarity to fixed random codebooks in a Fourier-transformed space. EpiNT is pre-trained on 2,741.1 hours of scalp EEG and iEEG from 1,199 patients and then fine-tuned on six epilepsy-related binary classification datasets spanning scalp and intracranial signals. The authors report that EpiNT outperforms randomly initialized models and several pre-trained baselines, and they present ablations on pre-training data composition, codebook size, number of quantizers, mask ratio, and architectural choices, along with interpretability and stability analyses.","tokens_in":19469,"tokens_out":7622,"duration_ms":69265,"significance":"The potential significance is high: a single pre-trained backbone for both scalp and intracranial neurophysiology could benefit epilepsy research, where labeled data are scarce. The paper has notable strengths: a large curated multi-modal pre-training corpus, plans to release code and weights, extensive ablation and comparison experiments, and quantitative stability analysis via CKA. However, the central cross-modal harmonization claim is not yet supported by the experimental design, because the downstream evaluation never crosses modalities and the preprocessing fixes a deterministic modality marker. If the authors add a control for this confound and a genuine cross-modal transfer experiment, the contribution could be valuable to the field.","major_comments":[{"comment":"Section 4.8 and Discussion: The modality-specific preprocessing (scalp EEG downsampled to 256 Hz and cut into 12-s epochs; iEEG resampled to 1024 Hz and cut into 3-s epochs, both truncated to 3072 points) gives every input a deterministic source-modality tag. Because all six downstream datasets are single-modality, the improved F1 scores in Table 3 do not demonstrate that EpiNT learns a shared, modality-invariant neurophysiological representation; the model could simply use the sampling-rate/epoch-length cue as a cheap feature. The Discussion explicitly concedes this 'implicit annotation' but offers no experiment to rule it out. To support the title's and abstract's harmonization claim, the authors should add a cross-modal transfer test (e.g., pre-train on one modality and evaluate on the other modality's tasks) and/or a control with a common preprocessing pipeline for both modalities. This is load-bearing: without it, the results support only within-modality pre-training benefits.","section":"Section 4.8 and Discussion"},{"comment":"Abstract and Section 2.5, Table 3: The abstract claims EpiNT 'outperformed both randomly initialized models and other pre-trained methods on six downstream classification tasks,' but under full-parameter fine-tuning EpiNT achieves the highest F1 on only four of the six datasets (TASMC-UCLA, CHB-MIT, FNUSA, MAYO), not on CUK-IMHANS or HUH. Section 2.5 itself states 'four out of six evaluation tasks.' The abstract should be corrected or qualified. Furthermore, several top scores are within one standard deviation of the runner-up (e.g., linear-probing MAYO 0.930±0.027 vs VQ-MTM 0.929±0.018), so the use of the term 'outperformed' should be accompanied by explicit significance tests for every comparison claimed.","section":"Abstract and Section 2.5, Table 3"},{"comment":"Section 2.2 and Figure 3: The experiment described as 'Impact of Pre-training Data Modality' does not actually test cross-modal transfer. It pre-trains on different data compositions (scalp only, iEEG only, mixed) and then evaluates on within-modality downstream tasks; the resulting performance differences do not tell us whether a representation learned on one modality transfers to a task on the other. The text's conclusion that 'combining both scalp and iEEG data during pre-training is crucial for optimal model performance and effective transfer of learned features across different types of brain activity recordings' is not supported by this experiment. The authors need a direct cross-modal evaluation or should rephrase the claim.","section":"Section 2.2 and Figure 3"}],"minor_comments":[{"comment":"The abstract and Section 2.1 disagree on the number of downstream tasks: the abstract says 'six downstream classification tasks,' while Section 2.1 says 'five clinically relevant binary classification tasks across six datasets.' Please reconcile.","section":"Abstract and Section 2.1"},{"comment":"Section 4.3: The default mask ratio m is not specified for the final model; Section 4.9 reports only a range (0.3–0.7). Please provide the exact hyperparameters used for the main results (mask ratio, codebook size, number of quantizers, model dimension, number of layers, etc.) in the main text or a table.","section":"Section 4.3"},{"comment":"Table 3 has formatting issues: numerical entries are not aligned with their column headers, and the statistical significance asterisks are not accompanied by a description of the test procedure or multiple-comparison correction.","section":"Table 3"},{"comment":"The Discussion states that 'models trained from scratch outperformed pre-trained models on the CHB-MIT dataset,' which appears to contradict Section 2.5 and Table 3, where EpiNT (pre-trained) has the highest F1 on CHB-MIT in all three fine-tuning regimes. Please clarify which comparison is intended.","section":"Discussion"},{"comment":"There are numerous typographical and formatting errors, including 'A PREPRINT - SEPTEMBER13, 2025' (missing space), 'outpus' (§4.6), 'taks' (Figure 1 caption), 'neuronphysio- logical' (§4.11), and 'acrossing' (§4.2).","section":"Throughout"},{"comment":"Equation (16) uses inconsistent notation: the normalized codebook vector is called \\hat{c}_k in Eq. (15) but \\tilde{c}_k in the surrounding text.","section":"Section 4.7"},{"comment":"The paper claims the frequency domain mapping quantizer 'captures crucial frequency features,' but the random vector h_proj is frozen and the codebook is random, so the mechanism is closer to a random projection in the frequency domain. The claim should be tempered or supported by further analysis.","section":"Section 4.7"}],"recommendation":"major_revision","confidential_remarks":"The authors are honest about the implicit modality annotation in the Discussion, which is a positive sign, but the mismatch between the abstract's strong claim and the actual evidence is a serious issue. If they can add a cross-modal transfer experiment or a preprocessing control, the paper could be publishable. There is also a framing issue: the paper might be better positioned as a pre-training method for single-modality epilepsy tasks rather than as a 'harmonization' method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful pretraining paper with one genuine new piece, and a central claim that outruns the evidence. The genuine new piece is Eq. (14), a frequency-domain mapping quantizer that replaces random matrix projection with a random circular convolution plus cosine-similarity codebook matching. It's simple, and the ablation (Model 1 vs Model 7) shows it beats time-domain random projection on their downstream tasks. The evaluation is serious by the standards of the field: six epilepsy datasets, cross-subject splits, baselines retrained on the same pretraining corpus, ablations on codebook size, mask ratio, and pretraining volume, plus promised code.\n\nThe soft spots are real, though not fatal. The abstract says EpiNT outperformed other methods on six downstream tasks, but under full fine-tuning it is top in four of six; the wins on HUH and CUK-IMHANS are not there. Several of the reported advantages are within one standard deviation of VQ-MTM or MAE-EEG, so the 'consistently outperformed' language is doing more work than the table supports. The bigger issue is the title and framing: 'harmonization' is not actually demonstrated. All scalp data are resampled to 256 Hz and cut into 12 s epochs, all iEEG to 1024 Hz and 3 s epochs. That is a perfect one-dimensional marker of source modality, and no downstream task requires the model to process both modalities at once. The authors candidly admit in the Discussion that they 'implicitly differentiated between signal sources through modality-specific data pre-processing strategies,' but they do not test whether the gains survive removing that cue. The Figure 3 modality-transfer result is suggestive, but it is still consistent with the model using a cheap modality tag plus within-modality feature learning. So the cross-modal harmonization claim is untested, not false.\n\nFor a reader building an EEG/iEEG pretraining pipeline, this paper is worth reading for the quantizer idea and the careful ablation methodology. It deserves peer review, but a referee should ask for a control that removes the preprocessing confound (for example, matching sampling rates and epoch lengths across modalities, or a modality-swap experiment) and a tightened abstract.\n\nMy recommendation: send to review, and expect revision.","headline":"A solid engineering contribution that overstates its cross-modal harmonization: the frequency-domain quantizer is worth a look, but the unified-modality claim is untested because preprocessing hands the model a perfect source tag.","tokens_in":20027,"tokens_out":3746,"would_cite":true,"duration_ms":35064,"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 single Transformer, pre-trained on scalp and intracranial EEG with a frequency-domain vector quantizer and masked autoencoding, reported the best F1 among compared models across six epilepsy classification tasks.","keywords":["EEG pre-training","intracranial EEG","vector quantization","masked autoencoder","frequency-domain mapping","epilepsy classification","Transformer foundation model","cross-modal transfer"],"falsifier":"Re-run pre-training with matched preprocessing across modalities — for instance, scalp EEG also resampled to 1024 Hz and cut into 3-second epochs — and fine-tune on the same six tasks; if the gap over single-modality pre-training shrinks or disappears, the reported cross-modal advantage is carried by the modality-specific preprocessing cue rather than by the frequency-domain quantizer.","tokens_in":19028,"feed_emoji":"🧠","tokens_out":10701,"duration_ms":101542,"temperature":0.7,"pith_summary":"The paper aims to establish that one Transformer, EpiNT, can serve as a shared backbone for both scalp EEG and intracranial EEG (SEEG and ECoG) across multiple epilepsy-related classification tasks. Its recipe is to treat each recording channel independently, patch it, mask most patches, and pre-train the encoder to predict discrete labels assigned by a frozen, randomly initialized quantizer that maps patches in the frequency domain. Pre-trained on 2,741.1 hours of clinical recordings from 1,199 patients, EpiNT reported the highest F1 among compared models under linear probing on all six downstream tasks and on four of six under full fine-tuning. The reason to care is practical: a single pre-trained model that transfers across montage, amplitude, and frequency differences could reduce the need for task- and patient-specific training in epilepsy diagnostics.","feed_headline":"Pretrained transformer beats rivals on six epilepsy tasks","feed_subtitle":"A frequency-domain codebook plus masked autoencoding lets one backbone transfer across scalp EEG, SEEG, and ECoG.","key_machinery":"The load-bearing mechanism is the frequency-domain mapping quantizer: a frozen, randomly initialized set of codebooks whose entries are assigned to patches by circular convolution with a random vector, computed as the inverse DFT of the product of the patch spectrum and the random vector's spectrum, followed by cosine-similarity matching and normalization. Because circular convolution in time is Hadamard multiplication in frequency, this quantizer scores each patch by its spectral content without learned frequency filters. Multiple quantizers (Q) each supply a pseudo-label, and the pre-training loss is the average cross-entropy of predicting those labels for masked patches; the decoder is only a single linear layer. The paper shows this mechanism's necessity by ablation: quantization beats direct L2 reconstruction, and frequency-domain mapping beats time-domain random projection.","core_discovery":"EpiNT's central claim is that predicting which entries of a fixed frequency-domain codebook a masked signal patch belongs to is a better pre-training objective for epilepsy neurophysiology than reconstructing the raw waveform, and that doing this on single channels makes the model robust to montage variability. The pre-training corpus mixes scalp EEG (2,548.3 hours) and iEEG (165.5 hours) from 1,199 patients; after fine-tuning, the model achieved the best F1 among the compared methods on tasks that include SOZ localization (TASMC-UCLA, up to 0.925), seizure prediction (CHB-MIT, 0.913), pathology classification (FNUSA 0.886, MAYO 0.965), IED detection, and neonatal seizure detection. Under linear probing, where only a classification head is trained on top of the frozen backbone, EpiNT posted the best F1 on all six tasks; under full fine-tuning it was best on four of six. The paper interprets the consistent linear-probing results as evidence that the frozen backbone, not the fine-tuning head, carries transferable representations.","pith_inferences":["If the modality-specific preprocessing (256 Hz and 12-second epochs for scalp EEG, 1024 Hz and 3-second epochs for implanted EEG) is the real carrier of cross-modal transfer, then training on a version of the corpus with matched sampling rates and epoch lengths should shrink the reported advantage over single-modality pre-training.","Because the quantizer is a fixed random feature map, the codebook need not be learned; one could pre-specify codebooks concentrated on clinically relevant bands such as delta, theta, ripple, and fast ripple, and test whether band-targeted quantization sharpens downstream performance.","Channel-independent modeling discards spatial relations between electrodes; a natural extension is to add a small channel-attention head at fine-tuning time and ask whether SOZ localization, which depends on electrode location, improves.","The non-monotonic scaling observed on TASMC-UCLA and MAYO suggests data composition matters more than raw hours; a controlled study that varies the scalp-to-iEEG ratio at fixed total hours would clarify when mixing modalities helps."],"forward_implications":["A frozen EpiNT backbone read out by a linear probe delivers the best F1 of all compared models on all six downstream tasks, so the pre-training itself, not the tuning procedure, is doing the transfer work.","Pre-training on a mixture of scalp and intracranial data beats either single-modality corpus, especially for iEEG tasks, so cross-modal exposure contributes beyond total data volume.","Quantizing in the frequency domain outperforms both direct raw-signal reconstruction and time-domain vector quantization, pointing to spectral structure as the invariant the model should preserve.","High masking ratios destroy clinically relevant sparse events in epilepsy signals; raising the mask ratio drops FNUSA F1 from 0.805 to 0.631, the opposite of what image MAE would predict.","Multiple small codebooks outperform a single large codebook, suggesting the quantizer's capacity should be spread across complementary random frequency projections."],"supporting_citations":[{"why":"Supplies the reconstruction-based MAE-EEG baseline that EpiNT's vector-quantized objective is designed to beat.","marker":"[16]"},{"why":"Applies vector quantization to tokenize EEG into discrete sequences, motivating the discrete-representation pre-training paradigm.","marker":"[18]"},{"why":"Provides VQ-MTM, the random-projection vector quantization pre-training method that EpiNT's frequency-domain quantizer directly extends and competes against.","marker":"[19]"},{"why":"Presents Brant, a spectral-encoding intracranial EEG foundation model that motivates frequency-aware design and serves as a comparison baseline.","marker":"[21]"},{"why":"Supplies the masked autoencoder paradigm from vision whose masking-ratio behavior EpiNT finds inverted for sparse neurophysiological events.","marker":"[29]"},{"why":"Provides the TUEP corpus, one of the largest scalp EEG pre-training sources in the 2,741.1-hour corpus.","marker":"[38]"},{"why":"Provides the TUSZ seizure corpus, the dominant source of scalp EEG pre-training hours.","marker":"[39]"},{"why":"Provides the TASMC-UCLA annotated iEEG sleep dataset used for SOZ localization evaluation.","marker":"[42]"},{"why":"Provides the CHB-MIT scalp EEG dataset used for seizure prediction evaluation.","marker":"[43]"},{"why":"Provides the multicenter iEEG dataset used for FNUSA and MAYO pathology classification evaluation.","marker":"[46]"}],"fun_headline_variants":["Frequency-domain codebook pretraining unifies EEG and iEEG","Pretrained transformer unifies scalp and intracranial EEG","One transformer handles both scalp and intracranial EEG tasks","Codebook pretraining beats rivals on all epilepsy tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the two preprocessing pipelines — scalp EEG downsampled to 256 Hz with 12-second epochs and implanted EEG resampled to 1024 Hz with 3-second epochs — are a neutral harmonization step, not a signature the model can detect and exploit to tell the two recording types apart instead of learning shared neurophysiology.","fun_headline_variants_meta":{"raw":{"variants":["Frequency-domain codebook pretraining unifies EEG and iEEG","Pretrained transformer unifies scalp and intracranial EEG","One transformer handles both scalp and intracranial EEG tasks","Codebook pretraining beats rivals on all epilepsy tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001043,"raw_usage":{"total_tokens":4387,"prompt_tokens":950,"completion_tokens":3437,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":3373}},"tokens_in":566,"tokens_out":3437,"duration_ms":25643,"temperature":1.0,"reasoning_tokens":3373,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:13:20.471515+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run pre-training with matched preprocessing across modalities — for instance, scalp EEG also resampled to 1024 Hz and cut into 3-second epochs — and fine-tune on the same six tasks; if the gap over single-modality pre-training shrinks or disappears, the reported cross-modal advantage is carried by the modality-specific preprocessing cue rather than by the frequency-domain quantizer.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the reconstruction-based MAE-EEG baseline that EpiNT's vector-quantized objective is designed to beat."},{"cited_title":"Vector quantization pretraining for EEG time series with random projection and phase alignment","cited_arxiv_id":null,"evidence_quote":"Provides VQ-MTM, the random-projection vector quantization pre-training method that EpiNT's frequency-domain quantizer directly extends and competes against."},{"cited_title":"Brant: Foundation model for intracranial neural signal.Advances in Neural Information Processing Systems, 36, 2024","cited_arxiv_id":null,"evidence_quote":"Presents Brant, a spectral-encoding intracranial EEG foundation model that motivates frequency-aware design and serves as a comparison baseline."},{"cited_title":"Big data resources for EEGs: Enabling deep learning research","cited_arxiv_id":null,"evidence_quote":"Provides the TUEP corpus, one of the largest scalp EEG pre-training sources in the 2,741.1-hour corpus."},{"cited_title":"The temple university hospital seizure detection corpus.Frontiers in Neuroinformatics, 12:83, 2018","cited_arxiv_id":null,"evidence_quote":"Provides the TUSZ seizure corpus, the dominant source of scalp EEG pre-training hours."},{"cited_title":"Annotated interictal discharges in intracranial EEG sleep data and related machine learning detection scheme.Scientific Data, 11(1):1354, 2024","cited_arxiv_id":null,"evidence_quote":"Provides the TASMC-UCLA annotated iEEG sleep dataset used for SOZ localization evaluation."},{"cited_title":"Multicenter intracranial EEG dataset for classification of graphoelements and artifactual signals.Scientific Data, 7(1):179, 2020","cited_arxiv_id":null,"evidence_quote":"Provides the multicenter iEEG dataset used for FNUSA and MAYO pathology classification evaluation."}],"review_version":2}