{"id":"4a4eafbf-7f92-4e8b-b83b-9cd65da4cff2","arxiv_id":"2502.09203","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A tutorial and placement study confirming that Euclidean alignment works best after temporal filtering and before spatial filtering in EEG transfer learning pipelines, based mainly on one motor imagery dataset.","lead":"This paper reviews a widely used preprocessing step called Euclidean alignment, which makes EEG signals from different people look more alike before a decoder is trained. It also reports new experiments suggesting where in the processing pipeline this step should be placed for motor imagery brain-computer interfaces.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Placement rule rests on a single motor-imagery dataset with five subjects and no significance testing; the 13-paradigm table supports EA's effectiveness, not the placement recommendation.","rationale":"The reader's weakest assumption—that the placement rule is extrapolated from a single MI benchmark and a heterogeneous literature table—is the most load-bearing concern for the paper's central claim. I considered other potential issues: the EA algorithm assumes an invertible mean covariance matrix (not discussed for N < c), and the deep learning experiment conflates explicit band-pass filtering with EA placement. However, these are secondary to the placement recommendation, which is the paper's main prescriptive contribution. The math in Section II-A is correct, and the experiment is well-designed for demonstrating that, on this dataset, EA between temporal and spatial filtering helps. But the paper generalizes this to a rule without cross-dataset validation or statistical tests. The 13-paradigm table does not address placement; it only shows that some form of EA usually helps, which is a different claim. The absence of a limitations section and the strong wording in the abstract and conclusions elevate the concern. The reader's CONDITIONAL verdict is appropriate: the paper is a useful tutorial and the recommendation is plausible, but independent multi-paradigm validation or a more cautious framing is needed. My stress-test does not change that verdict, so I recommend UNCHANGED.","tokens_in":24213,"tokens_out":7006,"duration_ms":66659,"concrete_test":"Re-run the Section II-B six-pipeline comparison on at least one additional motor-imagery dataset (e.g., BCI Competition IV Dataset 2a, using pairwise or four-class tasks) and on one non-MI paradigm (e.g., P300 ERP from BCI Competition III Dataset II) with the same preprocessing choices, reporting per-subject means and standard deviations and a paired significance test (e.g., Wilcoxon signed-rank) between TF-EA-RCSP-wAR and TF-RCSP-EA-wAR. If the recommended placement does not consistently outperform the after-spatial placement in these datasets, the placement rule in Figure 2 should be restricted to the original dataset or revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central prescriptive claim—that EA should be placed between temporal filtering and spatial filtering in the traditional pipeline (Figure 2) and between temporal filtering and the deep learning module (Figure 5)—is supported only by experiments in Sections II-B and II-C on BCI Competition IV Dataset 1, using a subset of five subjects with matched left/right hand classes. No significance tests, confidence intervals, or standard deviations are reported; conclusions rest on averages over 30 repeats (traditional) and 3 repeats (deep learning). The Table III survey of 13 paradigms contains w/ vs w/o EA comparisons from heterogeneous studies with different datasets, classifiers, and preprocessing, but it does not test placement. Therefore the placement rule is not established outside this single dataset, and even within motor imagery it needs replication. The provided rationale (temporal filtering improves covariance quality; spatial filtering benefits from aligned inputs) is plausible but post hoc. If the optimal placement differs for other MI datasets (e.g., four-class MI, different montages) or for ERP/SSVEP pipelines, the rule as stated would mislead users. The paper itself acknowledges the narrow experiment by saying 'we confirmed through experiments' on this dataset, yet the abstract and conclusions present the placement as part of EA's 'correct usage' without this caveat.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper revisits Euclidean alignment (EA) for transfer learning in EEG-based BCIs. It restates the EA algorithm, proves in Eq. (3) that EA whitens the mean covariance matrix of each domain, and presents new experiments on BCI Competition IV Dataset 1 (five-subject, two-class motor imagery) to argue that EA should be placed between temporal filtering and spatial filtering in traditional pipelines and after temporal filtering before deep learning modules. The paper also surveys 13 BCI paradigms in which EA has been reported to improve accuracy, compares EA with related alignment methods such as Riemannian alignment and label alignment, and discusses extensions including adversarial robustness and privacy preservation. The central prescriptive claim is that the recommended placement constitutes the 'correct usage' of EA.","tokens_in":24403,"tokens_out":4413,"duration_ms":43877,"significance":"If the placement rule holds, it offers a simple and actionable guideline for BCI practitioners and may explain inconsistent results in prior EA applications. The paper's literature compilation in Table III and the method comparison in Table V are useful resources, and the derivation of the core EA property in Eq. (3) is clear and correct. The paper also explicitly credits related work and quotes independent evaluations, which strengthens its tutorial value. However, the experimental support for the placement rule is narrow and lacks statistical treatment; at this stage the contribution is best characterized as a tutorial with a preliminary recommendation rather than a fully validated design rule.","major_comments":[{"comment":"The central placement recommendation is supported only by averages over 30 random repeats on one two-class motor-imagery dataset with five subjects, with no error bars, no per-subject variance, and no significance tests. The text uses 'always' (Sec. II-B, item 1) and 'almost always' (item 2) without statistical support. Please add per-subject results with confidence intervals or paired significance tests (e.g., Wilcoxon signed-rank across the five subjects), or substantially soften the claims to 'in our experiments on this dataset.'","section":"Sec. II-B, Fig. 4"},{"comment":"The deep-learning comparison reports only three repeats per subject and no variance. The EA-EEGNet average (48.00%) is actually lower than the EEGNet-only average (49.33%), so the conclusion that 'TF-EA-EEGNet performed much better, suggesting the necessity to apply temporal filtering explicitly before EA' requires a significance check; with five subjects these differences may be within noise.","section":"Sec. II-C, Table II"},{"comment":"Table III documents that EA improves classification accuracy within each individual study, but those studies use different datasets, classifiers, and preprocessing pipelines, and they do not compare alternative EA placements. The abstract and conclusions claim that 'numerous experiments from 13 different BCI paradigms demonstrated its effectiveness and efficiency,' but those are literature results, not new experiments, and they do not validate the placement rule. Please separate the effectiveness claim from the placement claim and avoid implying that Table III supports the placement recommendation.","section":"Sec. II-E, Table III"},{"comment":"The stated reasons for the recommended placement—that temporal filtering improves covariance-matrix quality and that EA before RCSP helps spatial filters generalize across subjects—are plausible but post hoc and are not tested by any ablation. A direct test (e.g., comparing covariance-matrix quality or RCSP filter consistency with and without EA at each position) would strengthen the causal claim; otherwise, the rationale should be explicitly labeled as a hypothesis.","section":"Sec. II-B, rationale paragraphs"}],"minor_comments":[{"comment":"There are typos: 'br ain-computer' in the abstract and 'Euclidian' in the Figure 3(a) caption; please correct them.","section":"Abstract and captions"},{"comment":"Reference [107] spells 'General Data Protection Regulation' as 'General Data Protection Regularization', and reference [118] has 'Jouranl' for 'Journal'; please fix these.","section":"References"},{"comment":"The claim that Figure 6 shows a 'clear diagonal pattern' is based on visual inspection; consider quantifying diagonal dominance with a numerical metric.","section":"Sec. II-D"},{"comment":"The champion and runner-up BEETL approaches are both cited as [72], which is the competition summary paper; please cite the actual team papers or clarify the attribution.","section":"Sec. II-E"},{"comment":"The phrase 'semi positive-definite' should be 'positive semidefinite'.","section":"Sec. II-A"},{"comment":"The 'Online or Offline' row is difficult to read because the repeated 'Both' entries are not separated from the following column; consider reformatting the table for clarity.","section":"Table V"},{"comment":"The statement 'EA is essential in adversarial training' is too strong; the cited ABA T algorithm uses EA, but 'essential' is not established by the experiments presented.","section":"Sec. IV-D"}],"recommendation":"major_revision","confidential_remarks":"The paper is heavily self-referential: a large fraction of the cited applications and extensions are from the author's own group, including [13], [14], [15], [103], [104], and [105]. This is not by itself a flaw, but the BEETL citation error (both champion and runner-up attributed to [72]) suggests the reference list needs careful checking. The main question for the editor is whether a prescriptive 'correct usage' claim can be supported by a single five-subject dataset with no significance testing; even if the tutorial aspects are valuable, the placement rule should be presented with appropriate confidence bounds or replicated on additional datasets before it is stated as a general guideline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nDongrui Wu's revisit of Euclidean alignment is a tutorial with a placement experiment bolted on. The math is correct, and the placement advice—EA after temporal filtering and before spatial filtering—is sensible and matches how the pipeline works. The paper does a good job organizing the 13-paradigm evidence that EA generally helps, and the sections on label alignment and future directions are useful.\n\nWhat's actually new is thin. The placement rule was already proposed in the author's own tutorial [14]; this paper adds one confirmatory experiment on a single motor-imagery dataset (BCI Competition IV Dataset 1, five subjects) with no significance tests or error bars. The text says EA 'always improved' performance, but that's averages over 30 repeats; the reader never sees the spread. The stress-test note is right: Table III shows EA effectiveness across paradigms, but it does not compare placements. So the prescriptive claim that this is the 'correct usage' rests on one benchmark. The rationale (temporal filtering cleans the covariances, spatial filtering benefits from aligned inputs) is plausible but post hoc.\n\nThe self-citation load is heavy, but that's expected when the method's inventor reviews their own method. It only becomes a problem if those citations prop up the placement rule. They do, mildly: the abstract and conclusions present the placement without the one-dataset caveat, even though Section II-B is more careful.\n\nWho is this for? BCI practitioners who want a concise guide to EA, label alignment, and related methods, and researchers looking for a map of where EA has been applied. They'll get a well-organized tutorial, but they should not treat the placement rule as established beyond motor imagery on one dataset.\n\nRecommendation: send it to peer review. The topic is important, the tutorial is genuinely useful, and the placement claim, though under-supported, deserves referee scrutiny. I would ask the author for significance testing, per-subject error bars, and either more datasets or a softened rule.","headline":"Useful tutorial from EA's inventor, but the central placement rule rests on a single five-subject dataset with no significance tests; treat it as guidance, not established fact.","tokens_in":24949,"tokens_out":3119,"would_cite":false,"duration_ms":27315,"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":"Euclidean alignment should be inserted between temporal and spatial filtering in EEG transfer learning pipelines.","keywords":["EEG","brain-computer interface","transfer learning","Euclidean alignment","label alignment","motor imagery","data alignment","preprocessing"],"falsifier":"A direct counter-test would be to run the same three configurations (no EA, EA between temporal and spatial filtering, EA after spatial filtering) on a different motor imagery dataset such as BCI Competition IV Dataset 2a with its nine subjects and four classes; if placing EA after spatial filtering matches or beats the recommended middle placement on average across subjects, the rule would be refuted. A second check would compare deep pipelines (TF-EA-EEGNet vs. EA-EEGNet with band-pass filtering inside the network) on an ERP or SSVEP dataset to see whether the placement advantage also holds beyond motor imagery.","tokens_in":23958,"feed_emoji":"🧠","tokens_out":7160,"duration_ms":59279,"temperature":0.7,"pith_summary":"Euclidean alignment (EA) whitens each subject's EEG trials so that every domain's mean covariance matrix becomes the identity, making data from different subjects more similar for transfer learning. This paper argues that to get the full benefit, EA must be inserted at a specific point in the decoding pipeline: between temporal filtering and spatial filtering for classical pipelines, and immediately after temporal filtering for deep learning models. The paper confirms this placement on a five-subject motor imagery benchmark, where the recommended pipeline reaches 81.6% plug-and-play accuracy with zero labeled target data, and it surveys results from 13 BCI paradigms that each show accuracy gains from EA. If the placement rule holds, BCI researchers can adopt a standard, free preprocessing step that reduces calibration effort and improves cross-subject decoding.","feed_headline":"EEG alignment wins when placed mid-pipeline, not after","feed_subtitle":"Cross-subject motor imagery accuracy jumps from chance to 66% with zero calibration in deep models.","key_machinery":"The central object is Euclidean alignment: for a domain with $N$ trials $X_n \\in \\mathbb{R}^{c\\times t}$, compute the mean covariance matrix $\\bar{R} = \\frac{1}{N}\\sum_{n=1}^{N} X_n X_n^\\top$ and transform each trial by $\\tilde{X}_n = \\bar{R}^{-1/2}X_n$, so the domain's mean covariance becomes the identity matrix $I$. The alignment is unsupervised, uses two closed-form formulas, and works in the Euclidean space, so any subsequent Euclidean-space classifier can be applied. The paper's placement analysis uses this whitening identity as the mechanism that reduces inter-subject distribution shift, and the experiments test where in the pipeline the transformation matrix should be computed: after temporal filtering (to suppress outliers) and before spatial filtering (so the filter is trained on aligned data).","core_discovery":"On the paper's own terms, the central discovery is that Euclidean alignment should be placed between the temporal filtering and spatial filtering blocks in the traditional BCI transfer learning pipeline, and between the temporal filtering block and the deep learning module in deep pipelines. The paper's new experiments on BCI Competition IV Dataset 1 show that EA placed there almost always outperforms both no-EA and EA placed after spatial filtering, because temporal filtering improves the quality of the covariance matrices that EA whitens, and aligning before spatial filtering lets filters like RCSP be learned on distributionally similar data. The paper also compiles results from 13 BCI paradigms to argue that EA is a broadly applicable, efficient, unsupervised preprocessing step, and recommends it as a standard component for cross-subject models.","pith_inferences":["The 13-paradigm evidence in Table III pools results from different datasets, classifiers, and papers; a controlled multi-dataset benchmark under identical preprocessing would make the placement rule directly comparable across paradigms.","The placement rationale implies that any preprocessing that improves the covariance estimate (for example, artifact removal or channel selection) should be ordered before EA, while any step that assumes aligned data should come after it.","If the reported convergence speedup from aligned training holds for larger models, EA could reduce the compute and data needed to pretrain large EEG foundation models, which the paper identifies as an open direction."],"forward_implications":["Motor imagery BCIs can operate with no per-subject calibration: the best traditional pipeline reaches 81.6% accuracy with zero labeled target trials, and accuracy rises only slowly as labeled target data are added.","For deep decoding, explicit temporal filtering before EA is necessary: EEGNet-only and EA-without-filtering stay near chance, while TF-EA-EEGNet averages 66.3% accuracy across subjects with no calibration.","The pipeline TF-CAR-EA-RCSP-wAR becomes a strong default baseline for cross-subject motor imagery classification, outperforming the same pipeline with no EA or with EA after spatial filtering."],"supporting_citations":[{"why":"Introduces the EA algorithm and its two closed-form formulas; the paper's recommended placement builds directly on this method.","marker":"[13]"},{"why":"Proposed the placement of EA between temporal and spatial filtering in the traditional TL pipeline; this paper's experiments test and confirm that recommendation.","marker":"[14]"},{"why":"Riemannian alignment, the inspiration for EA, whose multi-paradigm gains provide the baseline comparison in Table I.","marker":"[16]"},{"why":"Introduces label alignment, the supervised extension of EA; the paper contrasts EA vs LA when label spaces match or differ.","marker":"[15]"},{"why":"Systematic evaluation of EA with deep learning showing improved accuracy and faster convergence; supports the paper's recommendation that EA be standard for cross-subject deep models.","marker":"[30]"},{"why":"EEGNet is the deep model used in the paper's deep TL experiments, so the placement comparison is measured on this architecture.","marker":"[40]"},{"why":"Regularized common spatial pattern is the spatial filter in the traditional pipelines whose quality determines the placement's benefit.","marker":"[38]"},{"why":"Weighted adaptation regularization is the classifier in the traditional pipeline evaluations, making the accuracy comparisons concrete.","marker":"[39]"}],"fun_headline_variants":["EEG alignment works best mid-pipeline, not after","Place Euclidean alignment between filtering stages","Mid-pipeline EA beats post-hoc for EEG transfer","Align EEG after temporal, before spatial filtering","Mid-pipeline EEG alignment yields zero-calibration BCI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that one motor imagery dataset (BCI Competition IV Dataset 1, five subjects with matched left/right classes) plus a heterogeneous table of literature results across 13 paradigms is enough to establish both the placement rule and its broad generality.","fun_headline_variants_meta":{"raw":{"variants":["EEG alignment works best mid-pipeline, not after","Place Euclidean alignment between filtering stages","Mid-pipeline EA beats post-hoc for EEG transfer","Align EEG after temporal, before spatial filtering","Mid-pipeline EEG alignment yields zero-calibration BCI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000379,"raw_usage":{"total_tokens":1973,"prompt_tokens":863,"completion_tokens":1110,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":1038}},"tokens_in":479,"tokens_out":1110,"duration_ms":11181,"temperature":1.0,"reasoning_tokens":1038,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T22:17:51.618453+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct counter-test would be to run the same three configurations (no EA, EA between temporal and spatial filtering, EA after spatial filtering) on a different motor imagery dataset such as BCI Competition IV Dataset 2a with its nine subjects and four classes; if placing EA after spatial filtering matches or beats the recommended middle placement on average across subjects, the rule would be refuted. A second check would compare deep pipelines (TF-EA-EEGNet vs. EA-EEGNet with band-pass filtering inside the network) on an ERP or SSVEP dataset to see whether the placement advantage also holds beyond motor imagery.","supporting_citations":[{"cited_title":"EEGNet: a compact convolutional neural net work for EEG-based brain-computer interfaces,","cited_arxiv_id":null,"evidence_quote":"EEGNet is the deep model used in the paper's deep TL experiments, so the placement comparison is measured on this architecture."}],"review_version":1}