{"id":"89b12e61-83e2-4291-a8c3-03a8abfa6055","arxiv_id":"2411.13172","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A dynamic time warping method that aligns single EEG trials to the conventional average before averaging produces enhanced ERP waveforms with larger component amplitudes and lower trial-to-average differences.","lead":"The paper proposes a method to average brain-wave (EEG) responses to repeated stimuli by first stretching each response in time to match the group average, then filtering and averaging. The authors report that this reduces the blurring caused by trial-to-trial timing differences and increases the measured size of ERP peaks such as P200.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Internal-fit metrics and template bias: the claimed improvement in ERP amplitude may reflect alignment of noise and template leakage rather than recovery of the true neural response.","rationale":"The reader's weakest assumption identifies the same core issue: the evaluation rests on internal-fit metrics and amplitude increases that are anticipated consequences of warping and template alignment, without ground truth validation. My review confirms this is the load-bearing concern. The paper is otherwise methodologically reasonable: the algorithm is clearly described, uses public data, and reports cross-validated classification with train/test separation. However, the central quantitative claims (26-29% larger P200 median peak/amplitude, lower RMS/MAD) do not establish that the enhanced average is more accurate rather than merely different. A simulation with known ground truth would settle the question. The absence of comparison to prior methods (Woody, Casarotto, Kotas) is a secondary issue that would also need addressing in a revision, but the internal-fit circularity concern is the single most load-bearing objection. No ad hominem, no theatrics: the concern is about the evaluation logic, not the authors' integrity. I recommend keeping the CONDITIONAL verdict, as the paper's usefulness as a preprocessing variant is plausible, but the central claim needs stronger validation.","tokens_in":18670,"tokens_out":1519,"duration_ms":14622,"concrete_test":"Simulate a ground-truth ERP with known component amplitudes and trial-to-trial latency jitter plus additive EEG-like noise, and run both conventional and filtered DTW-based averaging on simulated trials. Compare the estimated component amplitude to the known true amplitude at varying jitter levels. If the filtered DTW method recovers the true amplitude better than conventional averaging in simulation, the concern about circular internal-fit metrics is weakened. If it only inflates amplitude without reducing error to ground truth, the central claim fails. This test is feasible because the method is fully specified in Section 3.4.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that the filtered DTW-based average reduces amplitude attenuation of ERP components and improves the averaged ERP waveform. The quantitative evidence is internal-fit: RMS/MAD distances between individual trials and the average (Eq. 4, Eq. 5) and P200 peak/amplitude (Table 1). These metrics are biased in favor of the proposed method by construction. First, the warped and filtered trial s_f^t[n] is constructed to minimize distance to the conventional average r[n] (Section 3.4), so a smaller RMS/MAD to that same average is partly a circularity consequence. However, the paper's RMS/MAD is measured against the DTW-based average r_w[n] or r_f[n], not r[n]; the concern is that any averaging of warped signals has lower variance and thus smaller trial-to-average distance when the warping absorbs trial-to-trial differences, including noise. Second, and more load-bearing, the P200 amplitude comparison is vulnerable to template leakage: the reference for alignment is the conventional average r[n], so large P200 peaks in individual trials are preferentially kept and aligned, while the averaging of the (possibly blurred) conventional average as template can inject its shape into the result. The larger P200 amplitudes are exactly what a warping-averaging scheme that selects for the template's features would produce, even if the true ERP is unchanged. The paper does not validate against ground truth or an external criterion; the classification result (Table 2) is small, on different data, and not statistically tested. This internal-fit evaluation is the load-bearing weakness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a modified dynamic time warping (DTW) procedure for ERP averaging. Each single trial is warped to the conventional trial average r[n] via a DTW path with a restricted step set; time steps that do not advance the reference index are removed so that the warped trial has the same duration as the reference, and a Kaiser-window low-pass filter is applied to suppress spurious high-frequency components introduced by the warping. The filtered warped trials are averaged to form the proposed 'filtered DTW-based average' r_f[n]. The scheme is evaluated on the Simanova EEG dataset using RMS and MAD distances between held-out trials (subset S2) and the three averages built from subset S1 (Eqs. (4)-(5)), P200 delay/peak/amplitude measured under 10-fold cross-validation (Table 1), and a four-class linear SVM classification example on the Birbaumer slow-cortical-potential dataset (Table 2). The authors report consistently smaller RMS/MAD distances for the DTW-based averages, P200 median peak and amplitude about 26% and 29% larger for the filtered DTW-based average than for the conventional average, and a small classification accuracy gain (59.52% vs. 58.30%). The central claim is that the scheme reduces the amplitude attenuation caused by latency jitter and thus improves the averaged ERP waveform.","tokens_in":18897,"tokens_out":15607,"duration_ms":141142,"significance":"If the claimed effect is genuine, the proposed scheme would be a practical, low-complexity extension of conventional ERP averaging: it uses publicly available datasets, the S1/S2 and 10-fold cross-validation designs go beyond purely in-sample reporting, the sampling-rate sensitivity check (Fig. 10) is a sensible robustness probe, and the equal-length reconstruction addresses a known obstacle to DTW-based averaging [31]. However, the validated significance is limited. The quantitative evidence is internal-fit: smaller trial-to-average RMS/MAD and larger P200 amplitude are both consequences one would expect from template-guided warping even if the true neural response were unchanged, so the headline claim needs a ground-truth simulation or an external criterion. The paper also does not benchmark against the closest prior jitter-correction methods that it cites (Woody [19], Wang [17], Casarotto [21], Kotas [22]), and the reported gains in Tables 1 and 2 are not tested for statistical significance. These gaps are fixable within the manuscript's scope.","major_comments":[{"comment":"The quantitative case for the central claim rests entirely on internal-fit metrics: RMS/MAD distances between individual trials and the averages (Eqs. (4)–(5); Figs. 8–9) and the P200 peak and amplitude of the average (Table 1). These metrics are co-defined by the alignment procedure: each trial is warped toward the conventional average r[n] (Eq. (1)), and the re-averaged signal r_f[n] (Eq. (3)) is therefore closer in shape to the trials and exhibits sharper peaks at the component latencies present in r[n]. The reported P200 median peak and amplitude gains of 26.44% and 29.30% (Section 4.1) are exactly what a template-guided warp would produce when the template contains a P200 feature, irrespective of whether any true neural signal is recovered; the same mechanism can amplify features that in r[n] originate from S1 noise. The paper's own conclusion (Section 5) acknowledges that further analysis of the 'gain in peak and amplitude values of ERP waves' is needed. To separate the claimed jitter-compensation effect from this template bias, the authors should add a ground-truth simulation (known ERP waveform with controlled latency jitter and additive noise) showing that r_f[n] reduces amplitude estimation error relative to conventional averaging, and an external-criterion check in which larger amplitude is not automatically treated as better. The S1/S2 and 10-fold designs are genuinely out-of-sample for the RMS/MAD computation, but out-of-sample distance reduction is still an internal-fit measure: a variance-reduced average is closer to arbitrary trials without necessarily being closer to the true ERP.","section":"§3.4–§4.1, Eqs. (1)–(5), Table 1"},{"comment":"The Introduction reviews the closest prior work—the Woody adaptive filter [19], the DTW-based averaging of Casarotto et al. [21], the nonlinear cycle alignment of Kotas et al. [22], and the warp-averaging method of Wang et al. [17]—but the results section contains no benchmark against any of them. Since these methods target the same problem (jitter and latency-jitter-induced attenuation in averaged ERPs) and the paper claims that the proposed scheme 'improves the averaged ERP waveforms', the lack of comparison leaves the incremental contribution unquantified and does not allow the reader to judge whether the reported RMS/MAD reductions and P200 amplitude increases are specific to the proposed filtered DTW scheme or generic to warping-based approaches. A comparison on the same folds and the same metrics (Eqs. (4)–(5) and the Table 1 protocol) would settle this.","section":"§1, §4 (comparison baseline)"},{"comment":"The headline P200 gains are reported without any significance test, and it is unclear whether the delay/peak/amplitude values in Table 1 are measured on the training-fold averages or on the left-out test folds: the text says 'one of them is left out whereas the remaining folds are employed at each evaluation step' but does not say which set is measured. The fold-to-fold standard deviations are extremely small (e.g., 0.246 µV vs. 0.235 µV for peak), yet no paired test across the 10 folds is reported. In addition, the P200 analysis covers only one subject (subject 02), one channel (channel 2), and the spoken modality, so the claim that the scheme improves ERP components in general is supported by a single case.","section":"§4.1, Table 1"},{"comment":"The classification result is a 1.22-percentage-point accuracy difference (59.52% vs. 58.30%) on a small dataset (the authors state that the least-populated class has 67 trials after artefact removal and splitting). No confidence interval, significance test, or repeated cross-validation variability is reported, and the single confusion matrix (Fig. 12) is not accompanied by the corresponding conventional-average matrix for direct comparison. As presented, Table 2 does not establish even a 'slight consistent improvement', and the wording in Section 4.2 goes beyond the evidence.","section":"§4.2, Table 2"}],"minor_comments":[{"comment":"The abstract's phrase 'latency jitter, jitter, and amplitude' is redundant (latency jitter is a form of jitter), and 'at the sight of the experiments carried out' should read 'in view of the experiments carried out'.","section":"Abstract, §2.1"},{"comment":"In Section 3.2, 'Excerpts of 1.2 ms build the trials that are analysed' should presumably be '1.2 s', given that the same section reports trial durations of 3.5 and 5.5 s.","section":"§3.2"},{"comment":"Equations (4) and (5) are typeset in a corrupted form in the manuscript (the RMS expression is incomplete), and the superscript notation linking the three averages (conventional, w, f) to the symbol * should be clarified.","section":"§3.5, Eqs. (4)–(5)"},{"comment":"The modified DTW reconstruction is described in prose only; a short pseudocode block specifying the treatment of consecutive (0,1) steps and the final sample-repetition rule would make the algorithm exactly reproducible.","section":"§3.4"},{"comment":"The statement that the RMS/MAD behaviour 'is found not only for the data shown... but for all the data analysed from [48]' is unverifiable from the results shown; aggregate statistics over all channels, subjects, modalities, and categories (or a link to the derived results) are needed to support it.","section":"§4.1"},{"comment":"The sampling-rate sensitivity analysis (Fig. 10) covers only the RMS metric; extending it to MAD and P200 amplitude would strengthen the claim that sampling rate does not alter the behaviour.","section":"§4.1, Fig. 10"},{"comment":"Minor typos: the Fig. 9 caption contains 'valúes', and reference [52] contains 'Naturea398' (should be 'Nature 398').","section":"Fig. 9 caption; Ref. [52]"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the principal risk is that the central empirical claim relies on metrics that are co-defined by the alignment procedure; the 26–30% P200 amplitude gain should not be taken at face value without a ground-truth simulation or a comparison with existing jitter-correction methods. Note also that this manuscript is the published version in Biomedical Signal Processing and Control 87 (2024) 105531; if a new submission is contemplated, prior publication should be disclosed. The most valuable revision would add (1) a simulation with a known ERP corrupted by controlled jitter and noise, (2) benchmarks against Woody [19], Wang [17], Casarotto [21], and Kotas [22], and (3) significance tests for Tables 1 and 2. The bibliographic coverage and funding disclosures are otherwise adequate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a modest but serious methods paper. The new bit is a path-restricted DTW variant—removing (0,1) steps so warped trials can't stretch beyond the reference—plus a low-pass filter before re-averaging. That combination is concrete enough to be useful, and the algorithm is described clearly enough to re-implement. The two public EEG datasets are appropriate, and the S1/S2 split for the RMS/MAD comparisons is a real attempt at generalization rather than pure in-sample fitting. Credit where due: the paper does not pretend to open a new field; it sells a preprocessing tweak.\n\nNow the soft spots. The evaluation is almost entirely internal-fit—RMS/MAD between trials and the average, and P200 peak/amplitude. The stress-test note worries about template leakage, and there is something to that: aligning to the conventional average can inflate peak height in a way that looks like recovery. But I'd stop short of calling it circular. The held-out test trials are not warped, and larger aligned peaks are exactly what any latency-correction method is supposed to produce. The real problem is that internal-fit measures are treated as validation when they should be sanity checks. There is no simulation with known jitter, no ground-truth ERP, and—more importantly—no comparison against the closest existing methods (Woody, Casarotto, Kotas, Wang). The references cover the relevant prior work; the problem is not missing citations, it's that those cited methods never appear as baselines. Without at least one of those, the central claim that this improves on conventional averaging is backed only by the method's own logic, which is not nothing but is not evidence. The classification result is a minor extra: 59.52 vs 58.30 on one dataset, no significance test. It's fine as an illustration, but it should not be used to claim BCI benefit. The writing has a few rough patches, but the math is straightforward and no internal contradiction jumps out.\n\nWho it's for: signal-processing people working on ERP averaging, especially in BCI or clinical ERP pipelines where DTW-based alignment might be a drop-in alternative. It won't change anyone's research program, but it's a clean datapoint.\n\nRecommendation: I'd accept it for peer review, but with a low bar for revision. Ask the authors to add a simulated-jitter experiment with known ground truth and to compare against Woody and at least one DTW-based prior method. If they can't or won't, the internal-fit evidence is too weak to justify publication on its own.","headline":"A workmanlike DTW-based ERP averaging variant with a real held-out split, but internal-fit metrics and missing baselines mean the central claim needs more evidence before I'd trust it.","tokens_in":19522,"tokens_out":5061,"would_cite":false,"duration_ms":53357,"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":"This paper claims that a filtered dynamic-time-warping average of EEG trials recovers ERP component amplitude lost to latency jitter, yielding larger P200 peaks and lower trial-to-average error than conventional averaging.","keywords":["event-related potentials","electroencephalography","dynamic time warping","latency jitter","ERP averaging","P200 component","brain-computer interface","signal alignment"],"falsifier":"Run the method on simulated EEG trials with a known ground-truth ERP embedded in noise with controlled latency jitter, and compare the filtered DTW-based average against the true signal: if the enhanced average does not recover the true amplitude and latency at least as accurately as conventional averaging (or if it shows a systematic bias toward the template), the central claim falls. A second, cheaper check is to apply the same scheme to pure noise trials with no ERP; if it still produces larger 'peaks' and smaller RMS distances, the metrics are not evidence of a better neural estimate.","tokens_in":18400,"feed_emoji":"🧠","tokens_out":5777,"duration_ms":49832,"temperature":0.7,"pith_summary":"Event-related potentials (ERPs) are usually extracted by averaging many EEG trials time-locked to a stimulus, but trial-to-trial differences in latency, jitter, and amplitude blur the average and shrink its peaks. This paper proposes a modified dynamic time warping (DTW) procedure that warps each trial toward the conventional average, restricts the warping path so signals are not arbitrarily lengthened, low-pass filters the warped trials, and then re-averages them. The paper reports that the resulting filtered DTW-based average reduces attenuation of ERP components, giving P200 median peak and amplitude about 26% and 29% larger than conventional averaging on the main public EEG dataset used, together with smaller RMS and maximum-absolute-difference distances between individual trials and the average. A sympathetic reader would care because cleaner averages could support more reliable ERP measurements in brain-computer interfaces and clinical evaluations, where amplitude and latency of components such as P200 or P300 are interpreted as neural markers.","feed_headline":"Warped EEG averaging lifts P200 peaks by roughly a quarter","feed_subtitle":"A filtered dynamic-time-warping average cuts jitter blur and lowers trial-to-average error in ERP analysis.","key_machinery":"The machinery is a modified dynamic time warping (DTW) scheme used to align each single trial $s_t[n]$ to the conventional average $r[n]$ taken as reference. The local cost matrix uses the absolute difference $C_t(i,j)=|r[i]-s_t[j]|$, and the optimal warping path is found under conventional monotonicity and endpoint constraints; then the path is restricted by removing the steps that do not advance the reference index, so the reconstructed signal is no longer than the reference, and any residual length difference is padded by repeating the last sample. A Kaiser-window low-pass FIR filter with cutoff at the original signal's maximum frequency is applied to each warped trial to suppress the high-frequency artefacts introduced by the warping, and the filtered warped trials are averaged: $r_f[n]=\\frac{1}{T}\\sum_{t=1}^T s_t^f[n]$. This restricted-path step is what distinguishes the scheme from ordinary DTW and what makes the averaged output defined at the same length as the reference.","core_discovery":"The central claim is that the proposed filtered DTW-based averaging scheme lessens the influence of latency jitter and jitter on averaged ERPs, reducing the attenuation of component amplitudes and yielding an averaged waveform that is closer, by RMS and MAD measures, to the individual trials. On the main EEG dataset the P200 median peak and amplitude of the filtered DTW-based average are 26.44% and 29.30% larger than the conventional average, while the coefficient of variation of peak and amplitude is smaller; similar qualitative behaviour holds for other channels, modalities, and categories analysed. The authors also show that the approach transfers to a classification setting, where signal models built from filtered DTW-based averages give slightly higher accuracy (59.52% versus 58.30%) than conventional averaging in a linear SVM task on slow cortical potentials from healthy and ALS subjects.","pith_inferences":["The paper's quantitative case rests on internal-fit metrics: if a warped average is evaluated by how close it is to the (possibly blurred) template average from which the warping was derived, part of the reported improvement is a circularity. A direct test would compare the enhanced average against a known ground-truth ERP in simulated data with controlled jitter.","Because the method uses the conventional average as the initial template, it inherits any bias of that template; iterating the warping-and-averaging procedure with the filtered DTW average as the new reference could either sharpen the estimate further or amplify artefacts.","The same alignment-plus-filter strategy could be applied to other time-locked neurophysiological signals, such as local field potentials or MEG, and to later components like P300, where latency jitter is known to be larger.","The classification gain is small and on a single dataset; a larger, multi-subject BCI study would clarify whether the sharper models translate into practically meaningful improvements in brain-computer interface performance."],"forward_implications":["On the main EEG dataset, the filtered DTW-based average yields P200 median peak and amplitude roughly 26% and 29% larger than the conventional average, indicating that the attenuation caused by jitter is partially recovered.","RMS and MAD distances between individual trials and the average are consistently smaller for the DTW-based and filtered DTW-based averages than for the conventional average, across the channels, modalities, and categories tested.","The reduced coefficient of variation of P200 peak and amplitude under the filtered DTW-based scheme suggests the enhanced average is more stable across cross-validation folds.","In a simple linear-SVM classification task on slow cortical potentials of healthy and ALS subjects, models built from filtered DTW-based averages outperform conventional-average models (59.52% vs 58.30% accuracy).","The qualitative behaviour of the scheme appears robust to sampling rate, since RMS comparisons at 500 Hz, 250 Hz, and 175 Hz show the same ordering among the three averaging methods."],"supporting_citations":[{"why":"Provides the main public EEG dataset on which the ERP averaging experiments are run.","marker":"[48]"},{"why":"Supplies the classical template-based latency compensation approach that the proposed scheme extends by allowing nonlinear warping.","marker":"[19]"},{"why":"Presents the closest prior approach, response-time corrected averaging, from which the proposed method differs by not imposing a polynomial timing model.","marker":"[20]"},{"why":"Is the prior iterative DTW averaging that the proposed scheme refines by adding a restricted path and filtering.","marker":"[21]"},{"why":"Is the prior cost-matrix modification for averaging nonlinearly aligned cycles that this scheme builds on.","marker":"[22]"},{"why":"Documents problems with DTW that make standard DTW unusable for averaging, motivating the modifications proposed here.","marker":"[31]"},{"why":"Describes across-trial latency variability in ERPs, motivating the need for jitter-resistant averaging.","marker":"[14]"},{"why":"Shows that latency jitter attenuates ERP amplitude, the phenomenon the enhanced average targets.","marker":"[18]"}],"fun_headline_variants":["DTW averaging lifts ERP peaks by 26%","Warped trials sharpen ERP waveforms and accuracy","Dynamic time warping reduces ERP jitter blur","Time-warped EEG averages boost P200 amplitude"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that smaller distances between individual trials and an average, and larger P200 peaks, actually mean the enhanced average is closer to the true neural response — otherwise the method could simply be aligning noise and inflating peaks relative to a blurred template.","fun_headline_variants_meta":{"raw":{"variants":["DTW averaging lifts ERP peaks by 26%","Warped trials sharpen ERP waveforms and accuracy","Dynamic time warping reduces ERP jitter blur","Time-warped EEG averages boost P200 amplitude"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000143,"raw_usage":{"total_tokens":1151,"prompt_tokens":903,"completion_tokens":248,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":187}},"tokens_in":519,"tokens_out":248,"duration_ms":3338,"temperature":1.0,"reasoning_tokens":187,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:44:31.926193+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on simulated EEG trials with a known ground-truth ERP embedded in noise with controlled latency jitter, and compare the filtered DTW-based average against the true signal: if the enhanced average does not recover the true amplitude and latency at least as accurately as conventional averaging (or if it shows a systematic bias toward the template), the central claim falls. A second, cheaper check is to apply the same scheme to pure noise trials with no ERP; if it still produces larger 'peaks' and smaller RMS distances, the metrics are not evidence of a better neural estimate.","supporting_citations":[{"cited_title":"Simanova, M","cited_arxiv_id":null,"evidence_quote":"Provides the main public EEG dataset on which the ERP averaging experiments are run."},{"cited_title":"Woody, Characterization of an adaptive filter for the analysis of variable latency neuroelectric signals, Med","cited_arxiv_id":null,"evidence_quote":"Supplies the classical template-based latency compensation approach that the proposed scheme extends by allowing nonlinear warping."},{"cited_title":"Gibbons, J","cited_arxiv_id":null,"evidence_quote":"Presents the closest prior approach, response-time corrected averaging, from which the proposed method differs by not imposing a polynomial timing model."},{"cited_title":"Numerical Reconstruction of 2D Magnetic Focusing Experiments","cited_arxiv_id":"2005.13841","evidence_quote":"Is the prior iterative DTW averaging that the proposed scheme refines by adding a restricted path and filtering."},{"cited_title":"Kotas, T","cited_arxiv_id":null,"evidence_quote":"Is the prior cost-matrix modification for averaging nonlinearly aligned cycles that this scheme builds on."},{"cited_title":"Mouraux, G.D","cited_arxiv_id":null,"evidence_quote":"Describes across-trial latency variability in ERPs, motivating the need for jitter-resistant averaging."},{"cited_title":"Ouyang, W","cited_arxiv_id":null,"evidence_quote":"Shows that latency jitter attenuates ERP amplitude, the phenomenon the enhanced average targets."}],"review_version":1}