{"id":"5ab49918-92d3-450c-a860-ebe740bc85de","arxiv_id":"2412.08896","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A conv-attention encoder-decoder with hand-crafted long-view morphology features reports state-of-the-art EEG and MEG spike classification, including a 13.58-point balanced-accuracy gain on a clinical MEG set.","lead":"This paper introduces a deep learning system that flags epileptic spikes in brain recordings by combining short spike-shape information with about six seconds of surrounding brain activity. It reports top results on a public EEG benchmark and on a hospital MEG dataset, suggesting that longer context helps automated spike review.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim hinges on unverified file-level long-view context: neither TUEV split disjointness nor non-overlap of the 100-wave normalization windows is specified, so temporal leakage could inflate the reported gains.","rationale":"The reader's weakest-assumption analysis correctly identifies the long-view feature pipeline as the load-bearing point: the headline SOTA numbers and the plug-and-play improvements would be inflated if any training slice's 100-wave normalization context included test time points. My reading of §3.2.1 and §3.1.1 confirms the relevant details are unspecified: the dataset description mentions 1-second segments and a predefined split, but not whether those segments are extracted from shared continuous files or whether the normalization windows are partition-safe. The MEG datasets are stated to be separate training and test collections, which reduces the risk of train-test leakage there, but the TUEV case remains open because the official TUEV corpus is organized by recordings and the paper does not state the split granularity. This is a correctable condition rather than a demonstrated flaw, so conditional acceptance remains appropriate: code and data release, plus a precise statement of the split and normalization windows, would settle it. I do not see a stronger objection than this one. Other weaknesses, such as the metric-dependent MEG comparison (LV-CadeNet's WF1 is below Labram's in Table 4) and the modest overclaim that all evaluated models improve by 'more than 8%' in balanced accuracy (EEGPT improves by 7.70 in Table 5), are real but secondary; they would not by themselves overturn the central claim if the leakage issue is resolved. The paper's strengths include repeated runs, validation-based early stopping, and systematic ablations, all of which are appropriate, but none of these protections rule out the temporal-context leakage described here.","tokens_in":16283,"tokens_out":8012,"duration_ms":92635,"concrete_test":"Obtain the official TUEV file structure and its predefined split. For every training slice, reconstruct the 100-wave normalization window and check whether any window overlaps a test slice, a test annotation time point, or even a test file; similarly check whether train and test share any recording file. For the Sanbo MEG data, verify that no 6-second context window of a Sanbo-Clinic test slice overlaps any Sanbo-CMR training time point. If all splits are file-disjoint and no context window crosses the partition boundary, the leakage hypothesis is refuted. If any overlap exists, rerun LV-CadeNet and the Table 5-6 baselines using a strict file-disjoint split and excluding overlapping context windows, and compare the resulting BAcc values (e.g., 67.32 vs 66.16 on TUEV; 86.32 vs 72.74 on MEG) with the reported numbers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central result depends on the long-view feature construction in §3.2.1. The method computes file-level features F, z-scores each complete wave against the distribution of the surrounding 100 complete waves (about 6 seconds), and then extracts each input slice X_i centered on annotation point i from the file-level matrix. For a valid evaluation, two conditions must hold: (a) each TUEV file must actually contain the surrounding 100 waves, so the data cannot be stored as isolated 1-second clips; and (b) the 100-wave normalization context for every training slice must not contain time points that belong to the test partition. The paper states only that the predefined TUEV training-test split was used (§3.1.1) and never specifies whether that split is file/session-disjoint or event-level within shared recordings. It also does not state whether the 100 surrounding waves exclude the wave being normalized, nor whether normalization windows for train and test slices can overlap. If TUEV's split assigns different events from the same continuous file to train and test, then training features for a slice are computed using statistics that include test waveforms, directly contaminating the comparison. The MEG evaluation uses distinct train/test collections, so train-test overlap is less likely there, but the same ambiguity applies within test files: a test slice is normalized using neighboring test time points, which is transductive context rather than label leakage, but it still should be declared. Because the plug-and-play gains in Tables 5-6 are the main evidence for the long-view contribution, the absence of a stated file/time-disjoint guarantee is the most load-bearing unverified premise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LV-CadeNet for EEG/MEG spike classification and detection. The method augments raw signal segments with six hand-crafted \"long-view\" morphological features, computed at file level and normalized against the surrounding 100 complete waves, and feeds the concatenated input into a new convolution-attention encoder-decoder named CadeNet. On the TUEV EEG dataset, LV-CadeNet-large reports BAcc 67.32, CKap 68.23, and WF1 83.61, slightly exceeding Labram-large, and on a Sanbo MEG test set it reports BAcc 86.32, 13.58 points above the best baseline. Ablation experiments show that adding the long-view features improves SpikeNet, Labram, EEGPT, and SimBA by more than 8 points in balanced accuracy, and architecture ablations attribute gains to convolutional temporal extraction, progressive spatiotemporal abstraction, and the decoder.","tokens_in":16570,"tokens_out":5721,"duration_ms":63230,"significance":"If the central claims hold, the paper makes a useful contribution: it demonstrates a plausible way to inject long-range context into spike analysis without naively extending raw input windows, and the architecture ablations are informative. The strengths include 30 repeated runs with reported standard deviations, evaluation on a public benchmark (TUEV), cross-architecture plug-and-play experiments, and qualitative case inspection. However, the main empirical claim currently rests on a file-level context procedure whose data availability and leakage safety are not specified, and one of the cross-model MEG comparisons is internally inconsistent with the text. The result is defensible but needs clarification and possibly re-analysis before the state-of-the-art conclusion is fully supported.","major_comments":[{"comment":"The load-bearing premise of the long-view features is that every training slice has access to the surrounding 100 complete waves (about 6 seconds) and that the normalization statistics contain no test time points. The paper states that TUEV comprises \"1-second bipolar channel segments\" and that the official training-test split was used, but it does not state whether these segments are embedded in continuous recordings, whether the official split is file/patient-disjoint or event-level within shared recordings, whether the 100-wave context excludes the wave being normalized, or whether the file-wise z-score of Section 3.1.2 is computed before splitting. These details are decisive because an event-level split within continuous files would let training features use test-set statistics and could explain the large LV gains in Tables 5 and 6. Please clarify these points, and if the current implementation permits overlap, rerun the experiments with strictly disjoint context windows.","section":"Section 3.2.1 and Section 3.1.1"},{"comment":"The text says that LV-CadeNet demonstrates \"consistent performance superiority over CadeNet\" for MEG spike detection, but Table 4 shows that LV-CadeNet is worse than CadeNet on Cohen's Kappa (46.91 vs 51.05), weighted F1 (95.96 vs 96.83), and accuracy. Only balanced accuracy improves. This contradiction weakens the claim that the long-view module is uniformly beneficial. Please report paired significance tests for the differences or explicitly restrict the claim to balanced accuracy.","section":"Table 4 and Section 4.2.1"},{"comment":"The state-of-the-art claim over Labram-large rests on differences of 1.16 points in BAcc, 0.78 points in CKap, and 0.32 points in WF1, all of which are within the reported standard deviations (for example, CKap 68.23±3.41 vs 67.45±1.95). Because 30 repeated runs are available, the paper should report paired tests or confidence intervals for these headline comparisons before asserting that LV-CadeNet-large outperforms Labram-large.","section":"Table 3"}],"minor_comments":[{"comment":"The label smoothing term uses epsilon/2, but the EEG task has N=6 classes and epsilon=0.1; standard label smoothing distributes epsilon/N over the incorrect classes. Please correct the formula or justify the binary-style form.","section":"Equation (17)"},{"comment":"Both tables contain a \"Div\" column with no entries; either fill it with the reported improvement values or remove the column.","section":"Tables 5 and 6"},{"comment":"There is a typo: \"TEUV\" should be \"TUEV\".","section":"Section 4.2.2"},{"comment":"The claim that extending the raw input window degrades performance is described only qualitatively. Please report the numeric results underlying Figure 4 so readers can compare the extended-window baseline with the long-view feature approach.","section":"Section 4.2.1 and Figure 4"},{"comment":"The assignment of 0.5/-0.5 to non-extremum points in the topological feature is described verbally; a precise rule (for example, based on proximity to the nearest maximum or minimum) would improve reproducibility.","section":"Section 3.2.1"}],"recommendation":"major_revision","confidential_remarks":"The skeptical concern about temporal leakage is real and must be resolved, but I do not see evidence of intentional leakage or circularity in the equations. The missing specifications about TUEV's storage format and split disjointness, plus the MEG inconsistency in Table 4, are fixable in revision. If the authors confirm strictly disjoint context and rerun where needed, the paper's central comparative claims could stand; if they cannot confirm disjointness, the reported LV gains would need substantial re-interpretation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the long-view feature construction is a genuinely interesting idea, and the empirical package is more thorough than most papers in this area. But the paper's central SOTA claim rests on an implementation detail—how the 100-wave normalization context relates to the train/test partition—that is left vague in §3.2.1. If that detail is not resolved, the reported gains could be inflated by temporal leakage. This is fixable, but it needs to be fixed before the numbers can be trusted.\n\nWhat's new and what's good. The six-feature long-view representation (topology, amplitude, mean amplitude, slope, half-slope, sharpness) with context normalization is, as far as I know, a new way to give spike classifiers broader context without feeding them raw longer windows. The paper shows that this feature set, as a plug-and-play addition, improves SpikeNet, Labram, EEGPT, and SimBA by large margins on TUEV and on their MEG data. The ablations are reasonably systematic: repeated runs with standard deviations, comparisons against raw extended windows, and a decomposition of the architecture into its components. The MEG result against Labram (86.32 vs 72.74 BAcc) is striking. The architecture itself—conv-attention blocks with pixel-shuffle downsampling and a single-stage decoder—is not radically new, but the combination is sensible, and the small parameter count relative to Labram-large is a plus.\n\nSoft spots. The biggest one is the one I opened with. TUEV is described in the paper as '1-second bipolar channel segments.' The long-view construction requires roughly 6 seconds of surrounding complete waves per channel. The paper never explains where that context comes from—whether they used the original continuous recordings, and, if so, whether the predefined train/test split is file/session-disjoint. The z-score normalization in §3.2.1 uses the 'distribution of values from the surrounding 100 complete waves,' and the paper doesn't state whether the wave being normalized is excluded, nor whether train and test slices can share normalization statistics. If the split assigns events from the same recording to both train and test, then the training features are computed with statistics that include test waveforms, and the plug-and-play gains in Tables 5–6 lose their force. For the MEG data the train/test collections are distinct, so the risk is lower there, but within test files the normalization is still transductive and should be declared. Second, no code or data are released, so none of this can be checked from the paper alone. That's not fatal, but it makes the missing specification harder to excuse. Third, minor: the text overclaims in places ('new paradigm,' 'fundamental breakthroughs'), and there are a few typos (e.g., 'propse').\n\nBottom line: If the train/test disjointness is confirmed, this is a solid contribution to automated spike analysis and worth publishing. If it isn't, the central comparison collapses. That makes the paper a good candidate for peer review—with specific instructions to the referees to verify the feature pipeline and ask for code/data. I'd rather see this in the review process than desk-rejected.","headline":"The long-view feature idea is genuinely interesting and the experiments are thorough, but the SOTA claim hinges on an underspecified train/test separation in the feature-normalization step that needs to be resolved before the numbers can be trusted.","tokens_in":17144,"tokens_out":4699,"would_cite":false,"duration_ms":48485,"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 convolutional-attention network with six long-view waveform features claims new best results on EEG and MEG spike analysis.","keywords":["EEG spike classification","MEG spike detection","interictal epileptiform discharges","long-view context features","convolution-attention encoder-decoder","public EEG spike corpus","balanced accuracy"],"falsifier":"Re-run the EEG and MEG comparisons with the 100-wave context window constrained never to include the labelled segment's own wave or any test-set time point; if the >8-point balanced-accuracy gains persist, the context-learning claim survives, and if they disappear, the reported gains came from leakage.","tokens_in":16107,"feed_emoji":"⚡","tokens_out":6755,"duration_ms":66265,"temperature":0.7,"pith_summary":"This paper argues that automated EEG/MEG spike analysis fails not because the models are too weak but because they look only at short windows around a candidate spike. LV-CadeNet adds a long-view channel: six morphological features computed for every complete wave in the signal and normalized against the surrounding roughly six seconds of context, concatenated with the raw segment. On the largest public EEG spike corpus, the full model claims the best balanced accuracy, Cohen's kappa, and weighted F1 among six state-of-the-art baselines; on a clinical MEG dataset, it reports balanced accuracy more than 13 points above the strongest baseline. The paper further claims the long-view features are a plug-and-play module that raises balanced accuracy by more than 8 points across four unrelated architectures.","feed_headline":"Long-view context lifts EEG/MEG spike detection by 8 points","feed_subtitle":"Adding ~6 seconds of normalized waveform context to raw EEG/MEG segments beats the strongest baselines.","key_machinery":"The load-bearing object is the seven-channel input matrix $X = [S, F]$, the raw signal segment $S$ concatenated with six long-view feature maps. Each feature map is built from complete waves, where a complete wave is a local maximum between two local minima; the features encode wave topology, amplitude, mean amplitude, left and right slopes, half-width slopes, and sharpness, and each wave's values are z-scored against the surrounding 100 waves, about six seconds of context. CadeNet's encoder stacks convolution-attention blocks in which temporal 1D convolutions and multi-head spatial attention operate on the same representation, with pixel-shuffle downsampling that moves spatial and temporal information into the feature dimension, followed by a decoder that distills the encoder features. The machinery's job is to let the network compare a candidate spike against its long-range background without diluting the input with raw noise.","core_discovery":"The central claim is that spike morphology is only meaningful in context: a waveform that looks like a spike in isolation is frequently a non-epileptic transient, and clinicians judge it against a longer stretch of surrounding activity. The paper operationalizes that context as a fixed set of six numeric features per complete wave — topology, amplitude, mean amplitude, slope, half-slope, and sharpness — z-scored over the neighboring 100 waves, and feeds them alongside the raw signal to a convolution-attention encoder-decoder. In the paper's experiments, the combined system reaches 67.32% balanced accuracy, 68.23% Cohen's kappa, and 83.61% weighted F1 on the EEG corpus, and 86.32% balanced accuracy on the clinical MEG test set, compared with 72.74% for the best MEG baseline. The same long-view features added to four unrelated architectures improve their balanced accuracy by more than 8 points, which the paper reads as evidence that the missing ingredient is context, not architecture.","pith_inferences":["Editorial inference: the decisive test for clinical adoption is a strict leave-one-subject-out or recording-level split; the reported gains could shrink if the 100-wave context window crosses file boundaries between training and test sets.","Editorial inference: because the features are defined per complete wave, the same construction could be applied to other transient biomarkers such as high-frequency oscillations without changing the network, only the wave detector.","Editorial inference: the uniform >8-point improvement across four different architectures is large enough that data leakage through the normalization window is the main alternative explanation to rule out before accepting the context-learning story."],"forward_implications":["If the long-view features are as general as claimed, existing EEG/MEG spike detectors can be upgraded by concatenating six normalized feature maps to their input rather than redesigning their architecture.","Naive extension of raw input windows hurts all tested models, so contextual information must be supplied as normalized features rather than simply longer clips.","The large balanced-accuracy gain on the imbalanced clinical MEG set suggests context features matter most when the class distribution reflects real clinical practice.","Ablation results attribute distinct gains to convolutional temporal extraction, progressively increasing spatial/temporal abstraction, and the decoder module, implying each architectural choice matters independently.","The 13.58-point MEG balanced-accuracy improvement over the strongest baseline, if it holds, would make automated spike detection substantially more usable in clinical workflows."],"supporting_citations":[{"why":"Supplies the public EEG spike corpus used for the main multiclass classification experiments.","marker":"[26]"},{"why":"Provides SpikeNet, a primary EEG baseline and a target architecture for the plug-and-play long-view features.","marker":"[5]"},{"why":"Provides Labram, the strongest EEG/MEG baseline that the method must beat and one of the models tested with and without long-view features.","marker":"[16]"},{"why":"Provides EEGPT, a foundation-model baseline whose balanced accuracy is evaluated with and without the long-view module.","marker":"[17]"},{"why":"Provides SimBA, a Mamba-based baseline used in both EEG and MEG comparisons with and without long-view features.","marker":"[38]"},{"why":"Provides EMS-Net, a prior MEG spike detector used as a MEG baseline and in the MEG plug-and-play experiments.","marker":"[9]"},{"why":"Defines the wave properties (amplitude, slope, half-width slope, sharpness) on which the six long-view features are based.","marker":"[19]"},{"why":"Supplies the extremum-detection and complete-wave segmentation method used to identify the waves that feed the feature construction.","marker":"[28]"}],"fun_headline_variants":["Long-view context boosts EEG/MEG spike detection by 8+ points","Context matters: 8-point spike detection gain from long-view features","LV-CadeNet: Long-view context lifts MEG spike detection by 13.6%","Beyond the spike: long neural context sharpens EEG/MEG diagnosis","Long-view features add 8 points to EEG/MEG spike detection accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each segment's roughly six seconds of surrounding context is cleanly available from the same recording without containing the segment's own label or any test-set time points, a guarantee the paper does not spell out.","fun_headline_variants_meta":{"raw":{"variants":["Long-view context boosts EEG/MEG spike detection by 8+ points","Context matters: 8-point spike detection gain from long-view features","LV-CadeNet: Long-view context lifts MEG spike detection by 13.6%","Beyond the spike: long neural context sharpens EEG/MEG diagnosis","Long-view features add 8 points to EEG/MEG spike detection accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1484,"prompt_tokens":1073,"completion_tokens":411,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":689,"completion_tokens_details":{"reasoning_tokens":312}},"tokens_in":689,"tokens_out":411,"duration_ms":3964,"temperature":1.0,"reasoning_tokens":312,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:27:30.327139+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the EEG and MEG comparisons with the 100-wave context window constrained never to include the labelled segment's own wave or any test-set time point; if the >8-point balanced-accuracy gains persist, the context-learning claim survives, and if they disappear, the reported gains came from leakage.","supporting_citations":[{"cited_title":"The temple university hospital eeg data corpus,","cited_arxiv_id":null,"evidence_quote":"Supplies the public EEG spike corpus used for the main multiclass classification experiments."},{"cited_title":"Development of expert-level automated detection of epileptiform discharges during Electroencephalogram interpretation,","cited_arxiv_id":null,"evidence_quote":"Provides SpikeNet, a primary EEG baseline and a target architecture for the plug-and-play long-view features."},{"cited_title":"Large brain model for learning generic representations with tremendous EEG data in BCI,","cited_arxiv_id":null,"evidence_quote":"Provides Labram, the strongest EEG/MEG baseline that the method must beat and one of the models tested with and without long-view features."},{"cited_title":"EEGPT: Pretrained transformer for universal and reliable representation of eeg signals,","cited_arxiv_id":null,"evidence_quote":"Provides EEGPT, a foundation-model baseline whose balanced accuracy is evaluated with and without the long-view module."},{"cited_title":"EMS-Net: A deep learning method for autodetecting epileptic Magnetoencephalography spikes,","cited_arxiv_id":null,"evidence_quote":"Provides EMS-Net, a prior MEG spike detector used as a MEG baseline and in the MEG plug-and-play experiments."},{"cited_title":"Toward a definition of MEG spike: Parametric description of spikes recorded simultaneously by MEG and depth electrodes,","cited_arxiv_id":null,"evidence_quote":"Defines the wave properties (amplitude, slope, half-width slope, sharpness) on which the six long-view features are based."},{"cited_title":"EMHapp: a pipeline for the automatic detection, localization and visualization of epileptic magnetoencephalographic high-frequency oscillations,","cited_arxiv_id":null,"evidence_quote":"Supplies the extremum-detection and complete-wave segmentation method used to identify the waves that feed the feature construction."}],"review_version":1}