{"id":"32ab777e-5dec-46b7-9001-4ac7faf42b25","arxiv_id":"2504.17888","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"In most patients, combinations of interictal iEEG band power cycles at ultradian, circadian, and multiday timescales explained most of the variance in seizure duration.","lead":"Seizure duration in people with epilepsy varies with cyclical patterns in brain signals recorded between seizures. This study finds that combining multiple cycles at different timescales can statistically explain much of that variation in individual patients.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Permutation test in S3.4 fixes the already-selected variables, so it misses the 12-model search; the reported 'above chance' p-values likely overstate the evidence for the central association.","rationale":"The reader's weakest assumption already identified this exact issue, and I agree. The permutation test in S3.4 is the only support for the statistical significance of the central claim, and conditioning on selected variables is a genuine validity problem, not a matter of consensus. A corrected full-null permutation is straightforward and would settle it. I would keep the CONDITIONAL verdict: the biological hypothesis is plausible and the methods are described carefully, but the quantitative strength of the association is not yet established without selection-adjusted inference or out-of-sample validation. The absence of code and data (Section 3) makes independent verification harder, though that alone would not change the verdict. This is a methodological concern, not a challenge to the authors' integrity.","tokens_in":28365,"tokens_out":5720,"duration_ms":60283,"concrete_test":"Rerun the full pipeline inside each permutation: for every shuffled response, repeat the S1 cycle inclusion and S3.2 12-model search (group-LASSO, 10-fold CV, LOOCV, same best-model rule) and record the best adjusted R2. Compare each subject's observed best adjusted R2 to this full-null distribution. If the proportion of subjects with p<0.05 drops from ~80% toward ~5%, the reported above-chance claim is an artifact of selection-unadjusted permutation testing; if it remains high, the association survives the corrected test.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—that combinations of band-power cycles are 'above chance-level' in explaining seizure duration—rests entirely on the permutation test in Supplementary S3.4. In each iteration the response (seizure duration) is shuffled, but the variables are fixed to the already-selected 'best' model. The earlier model-selection stage (S3.2) that chose among 12 candidate models, and the S1 prescreening of cycles, are not repeated. Because the selected variable set is a function of the original response, conditioning on it invalidates the permutation null: the observed adjusted R2 is drawn from the upper tail of the selection distribution, whereas the null distribution is generated from a fixed, response-independent variable set. This makes p-values anti-conservative and the abstract's '~80% above chance' an overstatement. The reported adjusted R2 values are also in-sample fits after group-LASSO variable selection, so they quantify in-sample association rather than predictive validity; however, the Discussion appropriately avoids claiming prospective prediction. This concern is about the statistical evidence for the association, not the biological hypothesis itself.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript investigates whether interictal iEEG band-power cycles, extracted via multivariate empirical mode decomposition (MEMD), are associated with seizure duration (used as a proxy for severity) in 20 subjects with refractory focal epilepsy. For each subject, the authors build linear-circular regression models relating log seizure duration to the phases of selected intrinsic mode functions (IMFs), with group-LASSO variable selection and a comparison of 12 candidate models. They report that 70% of subject-specific models achieve adjusted R2 values over 60%, and that ~80% of these are 'above chance' by a permutation test. They further describe which cycle periods (ultradian, circadian, multidien) contribute per subject and include a secondary analysis with a different severity marker (top-delta).","tokens_in":28567,"tokens_out":5750,"duration_ms":57100,"significance":"If the reported association is real, it would extend current understanding of how seizure severity is modulated over multiple timescales and could inform future chronotherapeutic approaches. The study uses a reasonable multi-cohort dataset, applies a decomposition method that does not assume stationarity, and includes a useful secondary severity-marker analysis. The authors also appropriately avoid claiming prospective prediction. However, the headline statistical claim depends on a permutation test that conditions on the already-selected model, and the reported adjusted R2 values are in-sample fits after model selection; both issues currently overstate the strength of the evidence for the central association.","major_comments":[{"comment":"The permutation test in S3.4 shuffles the seizure-duration response while keeping the band power cycles from the already-selected model fixed. Because the variable set was chosen by the model-selection procedure in S3.2 on the original data (12 candidate models, group-LASSO per model, choice by adjusted R2 and LOOCV RMSE), the null distribution does not reflect the full selection process. Under the null of no association between seizure duration and any band power cycle, the selected model's adjusted R2 is drawn from the upper tail of the selection distribution, whereas the permutation null generates adjusted R2 values from a fixed, response-independent variable set. The resulting p-values are therefore anti-conservative, and the abstract's claim that ~80% of the high-R2 models are 'above chance-level' is not supported by this test. Please rerun the permutation procedure with the response shuffled before model selection (repeating group-LASSO and the 12-model comparison), or provide a formal argument for why conditioning on the selected variables is valid.","section":"Supplementary S3.4; Supplementary S3.2"},{"comment":"The reported adjusted R2 values are in-sample fits computed on the same data used for group-LASSO variable selection and model selection. They are therefore optimistically biased estimates of the population association, and the claim that '70% of the models had a higher than 60% adjusted R2' describes the training data rather than a generalizable relationship. The manuscript should either provide bias-corrected estimates (e.g., via nested cross-validation or a full permutation scheme that repeats the selection) or state explicitly in the abstract and Results that the adjusted R2 values are in-sample. The term 'predicted seizure durations' in Fig. 3b and in the Supplementary (e.g., S3.3) should be replaced with 'fitted values' to avoid implying out-of-sample accuracy.","section":"Section 4.2; Fig. 4a; Supplementary S3.3"}],"minor_comments":[{"comment":"For precision, '70% of the models' should be reported as '14 of 19 subject-specific models', and 'around 80%' should be given with the exact count (e.g., 11 of 14) rather than a rounded percentage.","section":"Abstract; Fig. 4a"},{"comment":"The manuscript states that FDR correction was not applied to permutation tests; since 19 models are tested, the reported 'above chance' proportion may include false positives due to multiple comparisons. Please report how many of the permutation p-values would remain significant after a multiplicity correction.","section":"Supplementary S3.4"},{"comment":"The statement 'Code and data will be made available upon acceptance' is not reproducible; the authors should provide a link to the analysis code and processed data in a public repository at the time of submission or revision.","section":"Code and data availability"},{"comment":"Minor formatting issues include inconsistent rendering of 'adjustedR2' (should be 'adjusted R²') and inconsistent capitalization of 'P-values'; Figure 4a has some subject labels that are partially obscured and would benefit from a clearer layout.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The permutation-test issue is substantial and load-bearing for the abstract's central statistical claim. It is fixable within the manuscript's scope by re-running the null procedure to include the full model-selection pipeline, but that re-analysis may change the reported proportion of 'above chance' models. The biological hypothesis itself is not fundamentally challenged; the revision should focus on providing a valid statistical test and appropriately tempering the claims in the abstract and Results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about arXiv:2504.17888. First, it is a legitimate extension of the group's earlier work on multiscale cycles in interictal iEEG: they ask whether those cycles are associated with seizure duration, used here as a proxy for severity. That is a real empirical question, and they report subject-specific models with high in-sample adjusted R2. Second, the headline 'above chance' claim rests on a permutation test that does not account for their own model selection. That is a genuine soft spot, not a manufactured one.\n\nWhat is new: prior studies linked interictal cycles to seizure occurrence and evolution; linking them to duration/severity in a within-subject, multi-timescale framework is a reasonable step. The data come from three centers, preprocessing is detailed, and the linear-circular regression formulation is appropriate. They are also appropriately cautious in the Discussion: they call seizure duration a proxy, note the short recordings, and explicitly say they are not claiming prospective prediction. That honesty counts.\n\nThe soft spot is in the inference, not the biology. In Supplementary S3.4 they shuffle seizure duration but keep the already-selected variables fixed, then rerun only group-LASSO/regression on that fixed set. The preceding search over 12 candidate models per subject (and the S1 prescreening of cycles) is not part of the null. Because the chosen variable set is a function of the original response, the permutation null is too narrow; the observed adjusted R2 values are drawn from the upper tail of the selection distribution. So the abstract's 'around 80% above chance' is likely an overstatement. The adjusted R2 values themselves are in-sample fits after group-LASSO, so they overstate explanatory strength. The paper needs a corrected permutation procedure (refitting selection in each shuffle) or out-of-sample validation to support the current strength of the claim. The absence of code and data also prevents independent verification, though they say it will be available upon acceptance.\n\nI don't think this kills the paper. The biological hypothesis is plausible, and the descriptive finding—that subject-specific combinations of cycles correlate with duration—could survive a more honest test. But as written, the quantitative claims are stronger than the evidence. A serious referee should ask for a corrected null and ideally a small out-of-sample component before publication.\n\nThis is worth engaging with: the epilepsy monitoring community will find the question important, and the methods are mostly transparent. I would not cite it in its current form until the statistics are fixed. Bring it to reading group as a case study in selective inference.","headline":"A plausible extension of the group's cycle work to seizure duration, but the permutation test misses model selection and the 'above chance' claim is likely overstated.","tokens_in":29176,"tokens_out":3435,"would_cite":false,"duration_ms":30981,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Seizure duration is not arbitrary: it tracks patient-specific cycles in interictal brain-signal power that run from minutes to days, and combinations of those cycles explain most of its variability.","keywords":["seizure duration","seizure severity","interictal iEEG","band power cycles","empirical mode decomposition","circadian rhythm","ultradian rhythm","linear-circular regression"],"falsifier":"Re-run the permutation test by shuffling seizure durations before the model-selection step, repeating the full twelve-model search on every shuffled dataset and comparing each subject's real best adjusted $R^2$ to the distribution of shuffled best values; if the real values sit inside that distribution, the above-chance claim collapses. A second independent check is to fit each subject's model on the first half of their seizures and predict the second half; near-zero predictive $R^2$ would show the association does not generalise beyond the fitted seizure set.","tokens_in":28161,"feed_emoji":"🧠","tokens_out":8378,"duration_ms":76487,"temperature":0.7,"pith_summary":"This paper aims to establish that the length of an epileptic seizure, used here as a proxy for its severity, is partly set by slow rhythmic fluctuations in the brain's ongoing activity. Analysing days-long intracranial EEG from 20 people with epilepsy, the authors decompose median band power in five frequency bands into narrow-band cycles with periods from minutes to several days. They find that no single cycle predicts seizure duration well on its own, but a combination of cycles chosen per patient explains most of the variation: 70% of subject-specific models reached an adjusted $R^2$ above 60%, and roughly 80% of those passed a permutation test at $p<0.05$. If the association is real, interictal (between-seizure) brain rhythms could serve as a basis for forecasting seizure severity and for timing treatments to low-severity periods.","feed_headline":"Seizure length tracks brain rhythms from minutes to days","feed_subtitle":"Combinations of interictal EEG power cycles explain over 60 percent of seizure-duration variability in most patients.","key_machinery":"The load-bearing object is the band-power cycle: the paper takes 30-second median log band power from five frequency bands ($\\delta$, $\\theta$, $\\alpha$, $\\beta$, $\\gamma$), applies Multivariate Empirical Mode Decomposition (MEMD), a signal-decomposition method that jointly splits the multivariate power signals into aligned narrow-band intrinsic mode functions, and uses the Hilbert transform to assign each cycle an instantaneous phase at every seizure's onset. Those phases enter a linear-circular cosine regression, $\\log(\\text{duration}) = \\mu + \\sum_k \\beta_k \\cos(\\phi_k - \\phi_{k,0}) + \\epsilon$, which is fitted as ordinary regression on $\\cos\\phi_k$ and $\\sin\\phi_k$ pairs. Group-LASSO selects which cycles matter, leave-one-out cross-validation and adjusted $R^2$ choose among twelve candidate models per subject, and a 500-iteration permutation test decides whether the chosen model beats chance.","core_discovery":"Within an individual patient, the duration of a seizure depends in part on the phase of several concurrently running cycles in interictal iEEG band power, with each patient having its own mixture of ultradian, circadian, and multi-day cycles. The paper establishes this in two stages: a rank linear-circular correlation shows weak to moderate links between seizure duration and the phase of individual cycles, and then linear-circular regression with group-LASSO variable selection, comparing twelve candidate models per subject, shows that combinations of cycles explain most of the variability. Concretely, 70% of the selected models across 20 subjects had an adjusted $R^2$ greater than 0.60, and about 80% of those were above chance by a permutation test. The authors present the relationship as associative rather than causal: the band-power cycles are most plausibly readouts of underlying oscillatory processes that also shape seizure severity, and the detected cycle mix is subject-specific.","pith_inferences":["A stricter test the authors did not run, shuffling seizure durations before the twelve-model search rather than after model selection, would isolate how much of the reported significance owes to the screening step; I would expect the above-chance fraction to shrink, and that corrected fraction is the number to quote.","If the detected cycles are markers of metabolic or molecular oscillators, as the authors speculate, then adding non-EEG observables such as sleep stage or blood-glucose fluctuations into the same linear-circular framework is a natural extension that might lift accuracy for the patients whose models failed.","The subject-specific cycle mix suggests a clinical workflow distinct from seizure forecasting: instead of asking whether a seizure will occur, one could ask, conditional on occurrence, how severe it will be, which could change acute treatment thresholds; the paper stops at association, so a prospective predictive study is the logical next step."],"forward_implications":["Seizure severity, measured as duration, is modulated by subject-specific cycles in interictal brain activity rather than being purely stochastic.","Interictal recordings, the stretches between seizures, carry exploitable information about how long a subsequent seizure will last.","There is no universal severity cycle: the models draw on ultradian, circadian, and slower (about 2–4 day) cycles in a patient-specific mix, so any forecasting approach must be personalised.","Because duration is only a proxy for severity, the same modelling pipeline translates to other electrographic severity markers, and the paper illustrates this for one delta-band peak marker.","Timing treatments to phases of predicted low severity, chronotherapy, becomes a plausible clinical direction if the associations hold in longer ambulatory recordings."],"supporting_citations":[{"why":"Supplies the rank linear-circular correlation coefficient used to test whether seizure duration tracks the phase of individual band-power cycles.","marker":"(Mardia, 1976)"},{"why":"Introduces empirical mode decomposition, the ancestor method whose multivariate extension the paper uses to extract band-power cycles.","marker":"(Huang et al., 1998)"},{"why":"Defines multivariate EMD, the method that jointly decomposes the five frequency-band power signals into aligned narrow-band modes.","marker":"(Rehman and Mandic, 2010)"},{"why":"Provides the SWEC-centre intracranial EEG recordings and clinical seizure annotations included in the analysis.","marker":"(Burrello et al., 2019)"},{"why":"Supplies the group-LASSO procedures used to select which sine and cosine phase terms enter each subject's model.","marker":"(Breheny and Huang, 2015; Yuan and Lin, 2006)"},{"why":"Precedent for combining cycles of multiple timescales in a regression model of seizure outcomes.","marker":"(Proix et al., 2021)"},{"why":"Basis for analysing log seizure duration rather than raw duration as the response variable.","marker":"(Cook et al., 2016; Schroeder et al., 2022a)"},{"why":"Prior evidence that subject-specific band-power cycle timescales relate to how seizures evolve, motivating the same cycles as potential severity markers.","marker":"(Panagiotopoulou et al., 2022)"}],"fun_headline_variants":["Seizure duration tied to multiple brain cycle timescales","EEG power cycles explain most seizure duration variance","Interictal brain rhythms modulate seizure length","Seizure severity tracks subject-specific EEG cycles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The above-chance claim rests on a permutation test that shuffles seizure durations while keeping the already-selected band-power cycles fixed, so it does not account for the prior search over twelve candidate models per subject; if that selection step inflates the null distribution, the reported significance is overstated.","fun_headline_variants_meta":{"raw":{"variants":["Seizure duration tied to multiple brain cycle timescales","EEG power cycles explain most seizure duration variance","Interictal brain rhythms modulate seizure length","Seizure severity tracks subject-specific EEG cycles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001146,"raw_usage":{"total_tokens":4809,"prompt_tokens":1054,"completion_tokens":3755,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":670,"completion_tokens_details":{"reasoning_tokens":3694}},"tokens_in":670,"tokens_out":3755,"duration_ms":28304,"temperature":1.0,"reasoning_tokens":3694,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:31:30.142576+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the permutation test by shuffling seizure durations before the model-selection step, repeating the full twelve-model search on every shuffled dataset and comparing each subject's real best adjusted $R^2$ to the distribution of shuffled best values; if the real values sit inside that distribution, the above-chance claim collapses. A second independent check is to fit each subject's model on the first half of their seizures and predict the second half; near-zero predictive $R^2$ would show the association does not generalise beyond the fitted seizure set.","supporting_citations":[{"cited_title":"Leguia, Thomas K","cited_arxiv_id":null,"evidence_quote":"Precedent for combining cycles of multiple timescales in a regression model of seizure outcomes."}],"review_version":1}