{"id":"1e98dd0a-b967-410e-9e8f-e86dcdde4477","arxiv_id":"1908.00492","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"On the CHB-MIT EEG database, variance, energy, nonlinear energy, Shannon entropy, and wavelet-derived variance, energy, kurtosis, and line length best separate seizure from normal epochs, with 4.77-13.51% relative improvement in Bayesian error over the seizure prior.","lead":"This paper reviews how computers extract features from EEG brain signals to automatically detect epileptic seizures, and adds a small experiment ranking which features matter most. It found that simple amplitude and energy measures, especially on wavelet-transformed signals, separate seizure from normal EEG better than frequency features.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'significant' feature claim rests on a post-hoc 4.5% improvement threshold with no error bars; the reported Bayes-error improvements are tiny absolute differences that could vanish under record-level resampling.","rationale":"The reader's weakest assumption identifies the arbitrary 4.5% threshold and the absence of error bars or statistical tests on the Bayes-error estimates. My independent reading of Section 5.1 and Table 7 confirms that this is the most load-bearing weakness: the absolute error reductions are about 0.001-0.002, the baseline is only 1.74%, and the seizure epochs are highly autocorrelated because of the 75% overlap between consecutive 4-second windows. The 'significant' label therefore depends on estimation noise and a post-hoc cutoff, not on a demonstrated statistical separation. A subject-level bootstrap or record-level permutation test would settle the point directly. The review portions of the paper, including the feature descriptions and the literature tables, are useful and not affected by this concern. The central experimental claim, however, should be treated as exploratory unless the uncertainty analysis is supplied. Since the reader already recommended CONDITIONAL, my assessment does not move the verdict.","tokens_in":26797,"tokens_out":4582,"duration_ms":49209,"concrete_test":"Compute a subject-level block bootstrap: resample the 24 patients (or 48 records) with replacement, recompute the KDE Bayes errors and improvement rates for all features in Table 7 and Figure 1, and construct 95% confidence intervals. Then check how many of the eight headline features have a lower confidence bound above 4.5%, and whether the CFS top-five subset (EnergyD1R, VarianceD5L, VarianceD1L, EnergyD5L, VarianceD1R) appears in, say, at least 50% of bootstrap replicates. Resampling at the patient/record level is essential because adjacent epochs overlap by 75%, so epoch-level resampling would grossly understate variance. If the confidence intervals cover zero or the 4.5% threshold, the claim that these features significantly capture seizures is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The experimental load-bearing claim is that variance, energy, nonlinear energy, Shannon entropy (raw), and variance, energy, kurtosis, line length (DWT) 'significantly capture seizures', supported by a 4.77-13.51% improvement in Bayesian error over the baseline err0 (Section 5.1). Two coupled weaknesses make this claim unsupported. First, the significance criterion is an unvalidated post-hoc threshold: Section 5.1 says 'The features with an improvement rate higher than 4.5% were considered as significant.' No null distribution is defined, and the threshold is applied after inspecting the results. Second, the improvement rates correspond to tiny absolute Bayes-error changes (e.g., Table 7a: errb = 0.0160 vs err0 = 0.0174 for variance, delta = 0.0014), computed from a KDE with bandwidth 1.06 sigma N^{-1/5} and no uncertainty estimate. With 4,677 seizure epochs that are not independent (4-second windows slide by 1 second, so adjacent epochs overlap by 75%; records were randomly selected without a reported seed), the effective sample size is far below 4,677. Without bootstrap confidence intervals or a permutation test, a delta of roughly 0.001 in error cannot be distinguished from KDE estimation noise. Therefore the feature ranking and the CFS subset in Section 5.2 are not established as statistically significant.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a review of feature extraction and performance evaluation for automated epileptic seizure detection from EEG signals, supplemented by an original evaluation experiment. The review portions provide mathematical definitions of time-, frequency-, and time-frequency-domain features, discuss their computational complexity, summarize published detection results on the CHB-MIT and Bonn databases, and review epoch-based and event-based metrics. In the experimental portion, the authors use the CHB-MIT database to compute the Bayes error rate of each feature individually, with a class-prior baseline (err0 = 0.0174), and define an improvement rate as (err0 - errb)/err0. Features with improvement rates above 4.5% are declared significant; these include variance, energy, nonlinear energy, and Shannon entropy on raw EEG, and variance, energy, kurtosis, and line length on DWT coefficients. A correlation-based feature selection (CFS) then yields an optimal subset of five features, all computable in O(N) time. The paper concludes that these features 'significantly capture the seizures' with an improvement of 4.77-13.51% in Bayesian error from the baseline.","tokens_in":27134,"tokens_out":2559,"duration_ms":27743,"significance":"If the experimental claims were properly supported, this would be a useful contribution: the feature review with explicit mathematical formulas and complexity analysis is valuable as a reference, and the idea of benchmarking individual features against the class-prior Bayes error is a more principled comparison than raw classification accuracy on imbalanced data. The literature tables organized by database are also a convenient resource. However, the central quantitative claim of 'significant' features is not yet established because the significance criterion is post hoc and the reported improvements are not accompanied by any uncertainty quantification. The strengths of the review alone justify publication if the experimental conclusion is either repaired or softened to a descriptive finding.","major_comments":[{"comment":"The statement 'The features with an improvement rate higher than 4.5% were considered as significant' is a post hoc threshold with no null distribution, no correction for multiple testing, and no justification. The word 'significantly' in the abstract and conclusions is therefore unsupported. A permutation test or a bootstrap null distribution over the class labels (or over records) is needed to determine whether the observed improvement rates exceed what would be expected by chance. This is load-bearing because the entire list of eight significant features and the subsequent redundancy analysis in Section 5.2 depend on this threshold.","section":"Section 5.1, sentence after Table 7"},{"comment":"The reported improvements are tiny absolute differences in Bayes error with no error bars or confidence intervals. For example, Table 7a lists variance with errb = 0.0160 versus err0 = 0.0174, a difference of 0.0014, and energy with a similar difference. The kernel density estimate uses a fixed bandwidth h ≈ 1.06 σ N^(-1/5) and a single random record selection. Moreover, the 4-second epochs slide by 1 second, so adjacent epochs overlap by 75%; the effective sample size is therefore far smaller than the reported 4,677 seizure epochs. Without record-level bootstrap or cluster-robust resampling, a difference of about 0.001 in error cannot be distinguished from estimation noise. The feature ranking and the CFS subset in Section 5.2 are not established as statistically significant.","section":"Section 5.1, Eq. (26) and Table 7"},{"comment":"The experiment uses a single random selection of records with no reported seed or repetition. The choice of records can strongly affect the estimated densities and Bayes errors, especially for the minority seizure class. The authors should repeat the record selection (e.g., 100 random splits) and report the distribution of improvement rates, or use a record-wise bootstrap. Without this, the results may reflect the idiosyncrasies of the chosen 48 records rather than a general property of the features.","section":"Section 5, Table 5 and text 'We randomly chose two records from each case'"}],"minor_comments":[{"comment":"The word 'exmaple' should be 'example'.","section":"Section 2.4"},{"comment":"The heading 'F requency-domain features' contains an extra space and should be 'Frequency-domain features'.","section":"Section 3.2.2 heading"},{"comment":"The y-axis label 'rate' should specify that the values are percentages, consistent with the definition of rate in Eq. (26).","section":"Figure 1"},{"comment":"The citation to the Persyst website lacks an access date and a full bibliographic entry; it is listed only as 'Accessed: 2019-4-25' in a footnote, which is fine, but the reference list does not include a formal entry.","section":"Reference [per]"},{"comment":"The definition of SVD entropy uses a matrix A without specifying how it is constructed from the epoch X; the subsequent sentence describes the delay method and channel rows, but the notation for the temporal construction (e.g., embedding dimension and lag) is not stated explicitly.","section":"Section 2.1, item 13"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the review component is solid, but the experimental section overclaims significance. The authors should either add a proper statistical evaluation (bootstrap/permutation with record-level resampling) or explicitly soften the conclusions to descriptive observations about the observed improvement rates. The post hoc 4.5% threshold and the single random record selection are fixable within the scope of a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the review portion is the real value: the authors collect the common time/frequency/time-frequency features used in EEG seizure detection, give each a clean mathematical definition, add computational complexity, and summarize prior performance tables on CHB-MIT and Bonn. That is a useful reference. Second, the headline experimental claim—that variance, energy, nonlinear energy, Shannon entropy, and a few wavelet features \"significantly\" capture seizures—is not actually supported by the analysis. The 4.5% improvement-rate threshold in Section 5.1 is post hoc and undefined; the Bayes errors differ from baseline by only 0.001 to 0.002 absolute; and there are no error bars, no seed on the random record selection, and no accounting for the 75% overlap between consecutive 4-second epochs that makes the 4,677 seizure epochs far fewer than they look. The stress-test note is right: a 0.0014 error difference cannot be distinguished from KDE estimation noise.\n\nThat said, the experiment isn't worthless. Computing per-feature Bayes error on a public dataset is a new measurement, and the CFS analysis suggesting a cheap five-feature set (mostly variance and energy from DWT) is a reasonable hypothesis to test. But the paper overstates it as \"significant\" without a null distribution, and the description in Section 5.1 acknowledges the threshold only as a cutoff. That's a load-bearing weakness for the abstract's main claim.\n\nProportionally: the review is 70% of the paper and it's solid. The feature-definition section is the kind of thing I'd point a student to. The complexity table is more careful than most such reviews. The experimental section is a preliminary ranking, not a demonstrated significance result. The authors also correctly avoid the accuracy-only trap that plagues this literature and they use a public database, which is good practice.\n\nWho should read it: anyone entering EEG seizure detection who wants a compact feature catalog with math and complexity, and anyone thinking about feature selection methodology. The experimental conclusion should be cited, if at all, as a descriptive finding pending uncertainty quantification.\n\nMy recommendation: send it to peer review, but expect major revision. The review content deserves archival publication; the experimental claims need to be either tempered to \"potentially useful\" or backed by bootstrap/permutation intervals and a pre-specified threshold. I would not recommend desk rejection. If I were a referee I'd push for that revision.","headline":"Useful review, thin experimental claim: the feature catalog and complexity analysis are worth citing, but the 'significant features' result needs uncertainty quantification and a pre-specified threshold.","tokens_in":27615,"tokens_out":4737,"would_cite":true,"duration_ms":44495,"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":"Amplitude and variance EEG features, not frequency features, separated seizures from normal signals in the authors' Bayes-error experiment on the CHB-MIT database.","keywords":["epileptic seizure detection","EEG","feature extraction","Bayes error rate","correlation-based feature selection","wavelet transform","Shannon entropy","nonlinear energy"],"falsifier":"Resample the 4,677 seizure epochs and 263,424 normal epochs to build bootstrap intervals for each feature's improvement rate; if the intervals for the eight claimed features overlap zero, or if changing the kernel bandwidth $h$ or the threshold around 4.5% changes which features qualify, the reported ranking fails. A reader could run this on the same CHB-MIT records and the reported procedure.","tokens_in":26647,"feed_emoji":"🧠","tokens_out":4766,"duration_ms":46310,"temperature":0.7,"pith_summary":"This review paper tries to establish which EEG features, taken one at a time, genuinely separate epileptic seizure epochs from normal epochs, and which of those remain informative after redundancy is removed. Using the public CHB-MIT scalp EEG database, it estimates each feature's Bayes error rate via non-parametric density estimation and compares it with the prior-only baseline. The authors report that variance, energy, nonlinear energy, and Shannon entropy on the raw signal, plus variance, energy, kurtosis, and line length on wavelet coefficients, significantly improve separation. A correlation-based feature selection then narrowed the list to five variance/energy wavelet features computable in $O(N)$ time. A sympathetic reader would take this as evidence that amplitude and variability features deserve priority in automatic seizure detection, and that reporting accuracy alone is misleading on imbalanced EEG data.","feed_headline":"Five EEG features carry seizure detection at O(N) cost","feed_subtitle":"Variance, energy, nonlinear energy, and Shannon entropy beat the prior-only baseline on scalp EEG, say the authors.","key_machinery":"The carrying mechanism is the improvement rate over a prior-probability baseline: $\\mathrm{err}_0 = P(C_2)$, and $\\mathrm{rate} = (\\mathrm{err}_0 - \\mathrm{err}_b)/\\mathrm{err}_0 \\times 100\\%$. The Bayes error $\\mathrm{err}_b = \\int \\min_i P(C_i|x) p(x) dx$ is computed with Gaussian kernel density estimates $p(x|C_i)$ using bandwidth $h \\approx 1.06 \\hat{\\sigma} N_i^{-1/5}$. Features whose improvement rate exceeds 4.5% are called significant, and then the CFS merit score $\\mathrm{Merit}_F = k \\bar{r}_{fc}/\\sqrt{k + k(k-1)\\bar{r}_{ff}}$ selects non-redundant subsets. This machinery lets the paper judge single features without training classifiers and compare them on the same footing.","core_discovery":"On the paper's own terms, the central discovery is a ranking of single-feature discriminative power for EEG seizure detection: the Bayes error rate drops by 4.77–13.51% relative to the prior-only baseline for the eight named features, while most statistical, entropy, and frequency-domain features do not move the error at all. In particular, energy computed on the D1 wavelet coefficients of the left hemisphere achieved the largest improvement (13.51%). The redundancy analysis further shows that five wavelet-domain variance/energy features—energy on D1 of the right side, variance on D5 and D1 of the left, energy on D5 of the left, and variance on D1 of the right—form an optimal subset under the CFS merit score, and summing their computations costs $O(N)$.","pith_inferences":["The 4.5% threshold is presented without a statistical justification; the ranking should be read as descriptive of this dataset until bootstrap or permutation intervals are supplied.","Because the paper attributes the weak frequency-domain results to artifacts, a cleaner recording setting or explicit artifact-removal preprocessing might change the ranking and deserves a direct test.","The five-feature subset could be tested as a fixed, low-dimensional input for lightweight classifiers on held-out patients from the same database, which the review itself does not do.","Applying the same single-feature Bayes-error protocol to other public EEG databases would show whether the amplitude/variance dominance generalizes beyond the CHB-MIT recordings."],"forward_implications":["Automatic seizure detectors can be built around variance, energy, nonlinear energy, and Shannon entropy from the raw signal, or variance, energy, kurtosis, and line length from wavelet coefficients, with a Bayes-error advantage of 4.77–13.51% over the baseline.","The five non-redundant features selected by CFS all cost $O(N)$ per epoch, so real-time screening is plausible on modest hardware.","Accuracy as a headline metric is misleading for imbalanced seizure data; epoch-based sensitivity/specificity and event-based good detection rate and false positives per hour should be reported together.","Frequency-domain features such as spectral entropy, peak frequency, intensity-weighted mean frequency, and intensity-weighted bandwidth showed no improvement over baseline on this dataset, so they should not be relied on alone for scalp EEG seizure detection."],"supporting_citations":[{"why":"Supplies the CHB-MIT Scalp EEG database used for the feature-significance experiment.","marker":"[GAG+00]"},{"why":"Provides the Bayes error rate background and the kernel density estimation approach used to compute the errors.","marker":"[Fuk90]"},{"why":"Introduces the non-parametric kernel density estimate used for the likelihood $p(x|C_i)$.","marker":"[Par62]"},{"why":"Defines the correlation-based feature selection merit score used to analyze feature redundancy.","marker":"[HS97]"},{"why":"Defines nonlinear energy, one of the significant raw-signal features in the reported results.","marker":"[Kai90]"},{"why":"Defines Shannon entropy, another significant raw-signal feature.","marker":"[Sha48]"},{"why":"Establishes that the Bayes classifier achieves the minimum possible classification error, the theoretical yardstick for the experiment.","marker":"[DGL13]"},{"why":"Introduces line length as an efficient seizure-related feature, used both as a raw-signal and a wavelet-domain feature in the review.","marker":"[EET+01]"}],"fun_headline_variants":["Energy on D1 wavelets cuts Bayes error by 13.5%","Eight EEG features capture seizures; five are optimal","Wavelet features cost O(N) for seizure detection","Most EEG features fail; eight beat baseline"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's ranking rests on treating a 4.5% improvement rate as the significance threshold and on trusting the kernel-density Bayes error estimates without error bars or statistical tests; if that threshold is arbitrary or the density estimates are unstable on the 4,677 available seizure epochs, the list of significant features is not supported.","fun_headline_variants_meta":{"raw":{"variants":["Energy on D1 wavelets cuts Bayes error by 13.5%","Eight EEG features capture seizures; five are optimal","Wavelet features cost O(N) for seizure detection","Most EEG features fail; eight beat baseline"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000341,"raw_usage":{"total_tokens":1888,"prompt_tokens":965,"completion_tokens":923,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":858}},"tokens_in":581,"tokens_out":923,"duration_ms":9413,"temperature":1.0,"reasoning_tokens":858,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:51:20.848436+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Resample the 4,677 seizure epochs and 263,424 normal epochs to build bootstrap intervals for each feature's improvement rate; if the intervals for the eight claimed features overlap zero, or if changing the kernel bandwidth $h$ or the threshold around 4.5% changes which features qualify, the reported ranking fails. A reader could run this on the same CHB-MIT records and the reported procedure.","supporting_citations":[],"review_version":1}