{"id":"8298b87a-12d4-4edc-8a19-a3fc007e2bc0","arxiv_id":"1908.08973","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"On 2056 hours of intracranial EEG, variance and lag-1 autocorrelation showed no reliable pre-seizure increase across 105 seizures in 28 patients.","lead":"This study used long-term brain recordings from 28 people with epilepsy to test whether a generic warning signal, called critical slowing down, appears before seizures. The authors report no consistent increase in variance or lag-1 autocorrelation before 105 seizures.","discovery_kind":"replication","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Control baseline is contaminated by other pre-seizure intervals, so the ROC test may miss a real CSD signal.","rationale":"I read the paper as a negative result under a specific operationalization: distributional separation of two CSD indicators in prespecified pre-seizure windows versus all other data. The claim would be secure if the control distribution were a true baseline. The weakest point is that it is not: other seizures' pre-seizure windows remain in the control. This is exactly the reader's weakest_assumption. It is not a fatal flaw—it can be resolved by excluding those windows. The surrogate test mitigates temporal randomness but not this systematic contamination. Circadian modulation is also present, but because the authors report seizure onsets roughly uniform over 24h and the ROC uses full-day baselines, the baseline-contamination issue is more directly load-bearing. Therefore I agree with the reader's conditional verdict; no change is needed, but the concern should be addressed before the conclusion is treated as definitive.","tokens_in":8841,"tokens_out":5242,"duration_ms":59246,"concrete_test":"For each subject and each site, recompute AROC with the control distribution restricted to windows at least Tpre=4h away from every seizure onset (in addition to the existing 60min post-seizure exclusion), and rerun the 19-surrogate significance procedure. If the mean AROC and the number of significant sites remain at the reported levels, the contamination concern is empirically closed; if a sizable set of sites becomes significant with positive AROC, the paper's 'no evidence' conclusion would need to be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on comparing indicator values in a Tpre=4h pre-seizure window with 'the remaining data'. Section II states that only 'the 60 min interval after the onset of a seizure' is excluded from the baseline. In subjects with multiple seizures (105 total), the baseline therefore includes the 4h windows preceding every other seizure. If CSD actually raised variance or lag-1 autocorrelation before seizures, those elevations would appear in both the signal distribution and the control distribution, pulling AROC toward zero and reducing the number of sites passing the surrogate threshold. The seizure-time surrogates are generated from the same contaminated data, so they do not provide a clean null for this particular bias. The negative result is thus conditional on a control assumption that is not stated or tested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper tests the hypothesis that critical slowing down (CSD), measured as increases in variance and lag-1 autocorrelation, precedes human epileptic seizures. The authors analyze 2056 h of multichannel intracranial EEG from 28 subjects with 105 seizures. For each recording site, they compute sliding-window estimates of variance and lag-1 autocorrelation, compare the distributions of these indicators in pre-seizure windows (T_pre = 4, 2, 1, 0.5 h) with the remaining data using the area under the ROC curve, and assess significance with a seizure-time surrogate test. They report that only about one-seventh of brain sites pass the surrogate test, that the majority of significant sites show a 'critical speeding up' rather than slowing down, and that there is no clear spatial clustering in the seizure onset zone. They conclude that there is no evidence for CSD prior to these seizures and that bifurcation-induced tipping may be too simplistic for the epileptic brain.","tokens_in":9013,"tokens_out":10708,"duration_ms":96291,"significance":"The manuscript addresses an important and controversial question with an unusually large clinical dataset (28 subjects, 105 seizures, 2056 h of iEEG). The use of seizure-time surrogates is a methodological strength, and a robust null result would be a valuable contribution to both the epilepsy and critical-transition literatures. However, the negative conclusion is conditional on several control assumptions that are not tested or not clearly described: the ROC baseline includes pre-seizure windows of other seizures, the strong 24 h modulation is identified but not removed, and the reported 'critical speeding up' result appears inconsistent with the one-sided surrogate test defined in the Methods. These issues require reanalysis before the central claim is fully supported.","major_comments":[{"comment":"The control distribution in the ROC analysis is contaminated by other pre-seizure periods. The manuscript states that only the 60 min interval after seizure onset is excluded, so for the 105 seizures, the 'remaining data' include the T_pre windows preceding every other seizure. If a genuine pre-seizure increase in ρ or σ² exists, those elevated values appear in both the signal distribution and the control distribution, compressing AROC toward zero and reducing the number of sites that pass the surrogate threshold. The seizure-time surrogate test does not repair this bias because it uses the same contaminated baseline. Please re-run the analysis after excluding all pre-seizure windows from the control distribution, or using a time-of-day/state-matched control, and report whether the distribution of AROC and the number of significant sites change.","section":"Section II, ROC analysis"},{"comment":"There is an internal inconsistency in the definition of the surrogate test. Section II defines a significant indication of CSD as 'AROC > 0 for original seizure times and AROC exceeds the maximum one obtained with 19 the seizure time surrogates', which is a one-sided test for positive AROC. Section III, however, reports for the sites that passed the surrogate test that 'in the majority of cases both ρ and σ2 rather point to a \"critical speeding up\"', which requires negative AROC values in that set. Please clarify whether the surrogate test was actually two-sided; if so, specify the two thresholds and report the counts of significant positive and negative AROC values separately. If not, the 'critical speeding up' statement is unsupported by the test as defined.","section":"Section III, Fig. 3 vs Section II"},{"comment":"The 24 h rhythm is identified as a dominant contributor to the variability of both indicators, yet it is not removed or stratified in the ROC analysis. The power spectral densities in Fig. 2 show strong contributions at about 24 h for both ρ and σ², and the Conclusion names the sleep-wake cycle as a potential confounding variable. Because this slow modulation inflates the variance of both the pre-seizure and baseline distributions, it can reduce the sensitivity of the ROC comparison even if seizures are uniformly distributed over the day. Please repeat the analysis after removing the 24 h component (e.g., by subtracting a subject-specific diurnal average computed from interictal data or by stratifying the comparison by time of day) and report the resulting AROC distributions and surrogate-test outcomes.","section":"Section III, Fig. 2 and Conclusion"},{"comment":"The preprocessing that omits the upper 0.5% of all σ² values is applied to the very indicator under study. If a genuine pre-seizure elevation of variance falls in the upper tail, this step removes the signal of interest. The paper provides no robustness analysis for this threshold, and it does not report whether the excluded windows are enriched in the pre-seizure periods. Please either justify the threshold with a quantitative check or repeat the analysis for several thresholds (e.g., no exclusion, 0.1%, 1%) and show that the negative result is unchanged.","section":"Section II, variance preprocessing"}],"minor_comments":[{"comment":"There are small typographical errors; for example, 'There is a ongoing debate' in the abstract, 'An alternatively, data-driven approach' in the Introduction, and 'The absolute value of AROC is than confined' in Section II.","section":"Throughout"},{"comment":"No data or code availability statement is provided; for a quantitative null result of this kind, a statement about data availability and analysis code would improve reproducibility.","section":"Data and code availability"},{"comment":"The surrogate test uses only 19 surrogates, yielding a minimum p-value of 0.05; the manuscript does not discuss how the granularity of this permutation test affects the interpretation of 'one-seventh of brain sites' passing.","section":"Section II, surrogate test"},{"comment":"The statement that seizures occurred equally distributed over the 24 h cycle is based on all 105 seizures pooled across subjects; a per-subject circadian distribution would be more informative given the subject-level analysis elsewhere in the paper.","section":"Section III, seizure timing"},{"comment":"The phrase 'similar to what has been described before [24,39]' would benefit from a brief description of how the previously reported fraction compares quantitatively with the 1-2% found here.","section":"Section III, prior results"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is from a group with a long track record on this data, and the self-citations are appropriate for data provenance. The main issue is not novelty or circularity but the robustness of the negative conclusion. I would encourage the editor to require the reanalysis described in the major comments before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the Wilkat, Rings, Lehnertz paper on critical slowing down before seizures. My take: the central negative claim is broadly credible, but the abstract states it a bit more cleanly than the analysis warrants.\n\nWhat is actually new: not the core idea. Mormann 2005 and Milanowski and Suffczynski 2016 already questioned CSD before seizures. The contribution is scale and statistical discipline: 28 subjects, roughly 1650 brain sites, 105 seizures, 2056 hours of iEEG, evaluated with seizure-time surrogates and ROC-based sensitivity and specificity rather than anecdotal pre-seizure trajectories. The methods are clearly described, and the authors run sensible robustness checks: several pre-seizure window lengths, local derivatives of the indicators, and spatial localization with hypergeometric tests to correct for electrode oversampling. They also identify the strong 24-hour modulation of both indicators and discuss it honestly rather than hiding it.\n\nSoft spots: the stress-test concern about the ROC baseline is real. The control distribution is 'the remaining data' with only the 60 minutes after each seizure excluded. The four-hour pre-seizure windows of other seizures therefore stay in the control for any subject with multiple seizures. If a genuine pre-seizure signal existed, those elevated values would appear in both distributions and pull the AROC toward zero. The seizure-time surrogates inherit the same contaminated baseline, so they do not fully repair the bias. I do not think this sinks the negative result; the contaminated fraction of control data is modest, and the main cost is statistical power rather than validity. But it does make the conclusion conditional: no evidence under a somewhat diluted comparison.\n\nThe second issue is the circadian rhythm. The authors show large 24-hour contributions to indicator variability and note that seizures are somewhat more common in the morning, yet the ROC comparison never controls for time of day. If pre-seizure windows cluster at a particular circadian phase, they are being compared against a control set drawn from all phases, which can mask or mimic a pre-seizure effect. The authors flag sleep-wake as a confound but do not quantify its impact.\n\nMinor: no code or data are provided, which limits independent checking, though the methods section is reimplementation-friendly.\n\nOverall this is an honest, well-executed negative study, consistent with earlier doubts, and the conclusion is defensible under the authors' operationalization. The seizure-prediction field benefits from a systematic null like this. It deserves a serious referee; the right referee will ask for a clean control baseline and circadian-phase-matched comparisons before treating the null as settled. I would send it to review.","headline":"A systematic, careful negative study: CSD indicators show no reliable increase before 105 seizures, though the contaminated ROC control baseline and unremoved circadian rhythm make the null less clean than the abstract implies.","tokens_in":9484,"tokens_out":5542,"would_cite":true,"duration_ms":51334,"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":"Using intracranial EEG from 28 people and 105 seizures, this paper finds no evidence that variance or lag-1 autocorrelation—the standard critical-slowing-down indicators—increase before seizure onset.","keywords":["critical slowing down","early warning signals","epileptic seizures","intracranial EEG","lag-1 autocorrelation","variance","seizure-time surrogates","ROC analysis"],"falsifier":"Look for whether the conclusion changes when the control set is cleaned: rerun the ROC analysis excluding every 4-h window immediately before any seizure from the \"remaining data,\" rather than only the 60 min after each seizure. If variance or lag-1 autocorrelation then shows $A_{\\mathrm{ROC}}>0$ beyond the seizure-time surrogate threshold at substantially more sites, the paper's negative result would be overturned by exactly the kind of pre-seizure signal it claims to rule out.","tokens_in":8650,"feed_emoji":"🧠","tokens_out":9008,"duration_ms":80885,"temperature":0.7,"pith_summary":"The authors set out to test whether variance and lag-1 autocorrelation—the two standard early-warning indicators for critical slowing down, the slower recovery and larger fluctuations expected near a tipping point—increase before human epileptic seizures. They analyzed 2056 hours of intracranial EEG from 28 people with epilepsy, covering 105 seizures, and compared pre-seizure windows with all remaining data using ROC analysis and seizure-time surrogates. They report no consistent, statistically significant increase in either indicator before seizures; at the minority of sites that passed the surrogate test, the indicators more often decreased than increased. If this negative result is correct, it undercuts a widely cited class of seizure-prediction claims based on generic critical-slowing-down signals and redirects attention to network-level and multi-day processes.","feed_headline":"No evidence of critical slowing down before 105 seizures","feed_subtitle":"Two standard early-warning indicators do not rise before seizures; the tipping-point story lacks support.","key_machinery":"The argument rests on a surrogate-based ROC evaluation. For every brain site, the two indicators are computed in sliding 20.48-s windows; the area under the receiver-operating-characteristic curve measures how well indicator values in assumed pre-seizure periods (4, 2, 1, or 0.5 h) separate from the remaining data, while 19 seizure-time surrogates per subject—random permutations of the intervals between seizures—define the null distribution for significance. This machinery is what lets the authors distinguish a real pre-seizure change from chance fluctuations and from the strong 24-h circadian modulation they observe in both indicators.","core_discovery":"The central discovery is a null result stated on the paper's own terms: across 1647 intracranial brain sites in 28 subjects, neither variance nor lag-1 autocorrelation of the EEG shows a robust, statistically significant increase in the 4 h before seizure onset relative to the rest of the recording. Roughly one-seventh of sites passed a seizure-time surrogate test, but most of those significant sites showed decreases (\"critical speeding up\"), and the few sites with CSD-like increases could be traced to spatial oversampling in two to three subjects. The authors conclude that these data do not support the bifurcation-induced tipping picture of seizure generation and that univariate linear CSD indicators are not reliable generic early warning signals for human seizures.","pith_inferences":["A direct check the authors do not perform: repeating the analysis with all pre-seizure windows excluded from the control baseline could reveal whether the null result is partly an artifact of the contaminated reference set.","The strong circadian modulation they report suggests a testable alternative route to seizure prediction: use multiday rhythms of interictal activity or network measures as covariates, since those are the processes their spectral analysis shows dominate the variability of the CSD indicators.","If the result generalizes, the human epileptic brain becomes a counterexample in the wider early-warning-signal debate, showing that a system can undergo a major transition to an extreme event without leaving the statistical fingerprints of bifurcation-induced critical slowing down."],"forward_implications":["If the result holds, seizure-warning devices that rely solely on rising variance or lag-1 autocorrelation lack empirical support in this dataset.","The absence of spatial specificity to the seizure-onset zone argues against a local bifurcation driving the transition into a seizure and supports network-level accounts of seizure emergence.","Daily modulation of the indicators means that any future early-warning test must control for time-of-day and sleep-wake state, or risk false positives and missed detections.","The same surrogate-ROC framework can be applied to other candidate seizure-predictor indicators, not just the two CSD indicators tested here.","The negative result narrows but does not eliminate the possibility of seizure prediction; it redirects attention to noise-induced or rate-dependent tipping, for which generic early warnings are not expected."],"supporting_citations":[{"why":"Supplies the seizure-time surrogate null-hypothesis test used to judge significance.","marker":"[49]"},{"why":"Earlier study reporting seizures start without common signatures of critical transition; this paper extends that finding.","marker":"[24]"},{"why":"Established the insufficiency of univariate indicators for seizure prediction and gives the baseline this study must be compared against.","marker":"[39]"},{"why":"Provides the ROC-analysis methodology for quantifying sensitivity and specificity of the indicators.","marker":"[47]"},{"why":"Fluctuation-dissipation theorem linking variance and recovery time, the theoretical basis for CSD indicators.","marker":"[13]"},{"why":"Canonical formulation of variance and lag-1 autocorrelation as generic early-warning signals, the claim being tested.","marker":"[8]"},{"why":"Evidence for multi-day rhythms modulating seizure risk, used to support the confounding-variable interpretation.","marker":"[55]"},{"why":"Current review framing seizure prediction as a network problem, cited in the conclusion that network approaches are more promising.","marker":"[41]"}],"fun_headline_variants":["105 seizures show no critical slowing down","No early-warning signs in 105 seizures","Critical slowing down absent before seizures","Seizure warning indicators fail a large test","Tipping point warning not seen before seizures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The control baseline still contains the 4-h pre-seizure windows of other seizures—only the 60 min after each seizure was removed—so a true pre-seizure signal would contaminate both distributions and bias the ROC area toward zero, making the null result easier to obtain.","fun_headline_variants_meta":{"raw":{"variants":["105 seizures show no critical slowing down","No early-warning signs in 105 seizures","Critical slowing down absent before seizures","Seizure warning indicators fail a large test","Tipping point warning not seen before seizures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001282,"raw_usage":{"total_tokens":5173,"prompt_tokens":815,"completion_tokens":4358,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":431,"completion_tokens_details":{"reasoning_tokens":4293}},"tokens_in":431,"tokens_out":4358,"duration_ms":30710,"temperature":1.0,"reasoning_tokens":4293,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:24:30.174696+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Look for whether the conclusion changes when the control set is cleaned: rerun the ROC analysis excluding every 4-h window immediately before any seizure from the \"remaining data,\" rather than only the 60 min after each seizure. If variance or lag-1 autocorrelation then shows $A_{\\mathrm{ROC}}>0$ beyond the seizure-time surrogate threshold at substantially more sites, the paper's negative result would be overturned by exactly the kind of pre-seizure signal it claims to rule out.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the seizure-time surrogate null-hypothesis test used to judge significance."},{"cited_title":"Milanowski \\ and\\ author P","cited_arxiv_id":null,"evidence_quote":"Earlier study reporting seizures start without common signatures of critical transition; this paper extends that finding."},{"cited_title":"Mormann , author T","cited_arxiv_id":null,"evidence_quote":"Established the insufficiency of univariate indicators for seizure prediction and gives the baseline this study must be compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the ROC-analysis methodology for quantifying sensitivity and specificity of the indicators."},{"cited_title":"Kubo ,\\ title title The fluctuation-dissipation theorem , \\ @noop journal journal Rep","cited_arxiv_id":null,"evidence_quote":"Fluctuation-dissipation theorem linking variance and recovery time, the theoretical basis for CSD indicators."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Evidence for multi-day rhythms modulating seizure risk, used to support the confounding-variable interpretation."},{"cited_title":"Kuhlmann , author K","cited_arxiv_id":null,"evidence_quote":"Current review framing seizure prediction as a network problem, cited in the conclusion that network approaches are more promising."}],"review_version":1}