{"id":"e2c20b62-9f51-4ef5-9eab-653a408c5c82","arxiv_id":"2508.08216","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"ITSA aligns individual EEG tangent spaces and fuses regularized spatial patterns with Riemannian geometry to improve cross-subject BCI generalization.","lead":"The authors introduce ITSA, a preprocessing method that aligns EEG data across people by recentering, distributing, and rotating each subject's features, then fuses two feature types for classification. They report improved cross-subject brain-computer interface performance, with parallel feature fusion working best.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The transfer claim rests on the supervised rotational alignment not using the held-out subject's labels; the abstract does not disclose this, so LOSO improvements may be leakage-driven.","rationale":"The reader's weakest-assumption analysis correctly identifies the ambiguity in the abstract regarding whether rotational alignment uses target labels or target data. This is indeed the most load-bearing issue: the paper's headline is a cross-subject transfer claim under leave-one-subject-out validation, and any use of the held-out subject's labels in the alignment would invalidate that claim. It is not a matter of novelty or architectural detail; it is a matter of experimental validity. We also considered whether the hybrid RCSP-Riemannian fusion is sufficiently novel or whether the 'parallel fusion' result is robust across electrode configurations, but these are secondary if the pre-alignment leaks target information. The proposed concrete test would settle the concern by inspecting the algorithm or running a simple ablation with source-only rotations. Since the full text is not available to us, we cannot perform this check now. Therefore, the current UNVERDICTED status remains appropriate; no verdict change is justified until the method details are examined.","tokens_in":681,"tokens_out":2930,"duration_ms":38266,"concrete_test":"In the paper's method section, trace the data flow of the supervised rotational alignment step (likely in Algorithm 1 or Section 3). Determine whether the held-out subject's labels or test-set covariance matrices are used to compute the rotation R_i. If they are, rerun the LOSO experiment with R_i fixed using only source-subject information (e.g., R_i = identity after recentering, or R_i as the mean of source-subject rotations). If mean classification accuracy changes by more than a few percentage points, the cross-subject improvement is attributable to target leakage rather than to ITSA. Additionally, check whether the distribution-matching step uses held-out-subject statistics; if so, repeat with source-only statistics and compare.","verdict_should_be":"UNCHANGED","load_bearing_attack":"From the abstract alone, the central claim is that ITSA, combined with RCSP-Riemannian fusion, significantly improves leave-one-subject-out cross-subject performance. For that claim to hold, the subject-specific recentering, distribution matching, and especially the 'supervised rotational alignment' must be fit using only source-subject data. If the held-out subject's labels (or test-set covariance statistics) enter the rotation fit, then the LOSO protocol is violated and the reported generalization is an artifact of target leakage. The abstract does not state which data and labels are used for the supervised rotation, nor whether the distribution-matching step uses target-subject statistics. A second, related risk is that aggressive distribution matching (e.g., aligning each subject's covariance mean to a common template) can collapse class-discriminative directions, improving overall domain similarity while reducing class separability. The paper provides no equations or algorithm trace in the abstract to distinguish these possibilities, so correctness remains genuinely unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Individual Tangent Space Alignment (ITSA), a pre-alignment strategy for cross-subject EEG transfer learning that combines subject-specific recentering, distribution matching, and supervised rotational alignment. This is coupled with a hybrid architecture fusing Regularised Common Spatial Patterns (RCSP) with Riemannian geometry in parallel and sequential configurations. The central claim is that, under leave-one-subject-out cross-validation, ITSA yields significant performance improvements across subjects and conditions, with the parallel fusion configuration outperforming the sequential one. The abstract provides no quantitative results, algorithmic details, or statistical evidence. As the full text was not available for review, the assessment is based on the abstract alone.","tokens_in":914,"tokens_out":2148,"duration_ms":26642,"significance":"Cross-subject and cross-montage EEG transfer learning is an important problem for practical BCIs, as it promises to reduce calibration time and improve generalisation across individuals. If ITSA's claimed improvements are real and the method is correctly evaluated, the work would be a useful contribution to the Riemannian-geometry transfer literature. However, the abstract's lack of quantitative results and unspecified supervision source makes the significance conditional and unverifiable at this stage. The paper also commits to releasing code, which would be a welcome reproducibility asset.","major_comments":[{"comment":"The term 'supervised rotational alignment' is load-bearing for the LOSO transfer claim, but the abstract does not state which data and which labels are used in this alignment step. If the rotation is fit using the held-out subject's labels (or target test statistics), the claimed cross-subject generalisation would be an artifact of information leakage and the LOSO protocol would be violated. Please specify precisely: (i) what data (source subjects only?) are used to estimate the rotation, (ii) what the supervision signal is, and (iii) whether any target-subject statistics (covariances, means, labels) are touched before evaluation.","section":"Abstract, 'supervised rotational alignment'"},{"comment":"The central claim is supported by no numbers, effect sizes, confidence intervals, or statistical tests in the abstract. 'Significant' is asserted but not demonstrated. To make the claim assessable, the paper must report concrete comparisons against standard baselines (e.g., no alignment, common spatial patterns, Riemannian alignment) for the LOSO protocol, including at least classification accuracy or AUC with variance across subjects. This is not a stylistic issue: without quantitative evidence, the central claim cannot be evaluated.","section":"Abstract, 'significant performance improvements'"},{"comment":"The distribution-matching step, if not designed carefully, could collapse class-discriminative information while improving domain similarity. The abstract does not explain what distribution is matched (marginal covariance? class-conditional? which moments?) or how the matching interacts with class separability. Since the method claims to improve 'class separability' while 'maintaining geometric structure', a formal or empirical argument is needed that recentering and distribution matching do not wash out discriminative directions. Please provide equations or algorithm traces for these steps.","section":"Abstract, 'distribution matching'"}],"minor_comments":[{"comment":"The comparison between parallel and sequential fusion is stated without any quantitative support. If the full paper includes such comparisons, the abstract should at least cite a table or a representative accuracy figure to make the claim credible.","section":"Abstract, 'parallel fusion approach shows the greatest enhancement'"},{"comment":"The terms 'data conditions' and 'electrode configurations' are vague. Please define what is varied (e.g., number of channels, montage type, motor imagery vs. SSVEP, artefact level) and what 'robust' means operationally (e.g., small performance drop across configurations).","section":"Abstract, 'varying data conditions and electrode configurations'"},{"comment":"This is a strength. However, the abstract should also mention the exact evaluation metrics and the number of subjects/datasets to allow the reader to gauge statistical power.","section":"Abstract, 'code will be made publicly available'"}],"recommendation":"major_revision","confidential_remarks":"The full text was not available to the referee; only the abstract and reader's take were provided. The abstract is too thin to verify any of the main claims, and the 'supervised rotational alignment' creates a genuine leakage risk. I recommend the editor require the authors to clarify the alignment protocol and provide quantitative LOSO results before further consideration. If the full paper already contains these details, the revision should simply point to them explicitly in the abstract."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: based solely on the abstract, this is a reasonable, incremental EEG transfer method, and the key claim about leave-one-subject-out generalization stands or falls on a detail that is not disclosed—whether the 'supervised rotational alignment' uses the held-out subject's labels. That's not necessarily a flaw, but it's the first thing a referee must check.\n\nWhat's actually new: the specific combination of subject-specific recentering, distribution matching, and supervised rotation as a pre-alignment step, fused with RCSP and Riemannian geometry. The components are all known, and the abstract openly says it's a hybrid. The claimed performance gains are plausible, and the paper does the right thing by committing to LOSO and promising code.\n\nWhat I can't evaluate: there's no empirical detail in the abstract, no equations, no effect sizes, no ablation. So I can't say whether the method works. The abstract also doesn't say what the rotation is fit to. The stress-test worry is legitimate: if target-subject labels or test-set covariance statistics enter the rotation, the LOSO claim is leakage, not transfer. A second, subtler risk is that aggressive distribution matching can collapse class-discriminative directions, which would explain 'improvements' in domain similarity without real classification gains.\n\nThat said, none of these are fatal by themselves. The paper appears to be a coherent engineering contribution in an active subfield. The fact that only an abstract is available to me (the full text is not in this submission) makes this a review of the abstract, not the paper. If the full text exists and the method details are there, this deserves a serious referee. If the abstract is all there is, it's not enough.\n\nWho this is for: people working on cross-subject BCI calibration and Riemannian EEG methods. They would want to see whether the alignment step is label-free.\n\nRecommendation: accept for peer review, but with the explicit instruction to require a precise specification of the training/alignment protocol—specifically which data and labels are used at each step—and an ablation with and without the supervised rotation. If the authors can show the rotation is fit only on source subjects, the contribution is real though modest.","headline":"Plausible EEG transfer method; the whole cross-subject claim depends on an undisclosed detail about what the supervised rotation fits, so it needs a careful referee.","tokens_in":1321,"tokens_out":2569,"would_cite":false,"duration_ms":29754,"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 proposes Individual Tangent Space Alignment (ITSA), a pre-alignment step that recentres, matches, and rotates EEG covariance matrices, and shows it improves cross-subject classification in a hybrid RCSP-Riemannian architecture.","keywords":["EEG","brain-computer interface","transfer learning","Riemannian geometry","tangent space alignment","common spatial patterns","cross-subject generalization","motor rehabilitation"],"falsifier":"Remove all target-subject labels from the supervised rotational-alignment stage and re-run leave-one-subject-out cross-validation. If the reported accuracy gain disappears or drops sharply, the method is not label-free cross-subject transfer.","tokens_in":645,"feed_emoji":"🧠","tokens_out":6647,"duration_ms":72931,"temperature":0.7,"pith_summary":"The paper tries to establish that cross-subject EEG variability can be overcome by aligning each subject's covariance geometry before classification. It introduces ITSA, which recentres, matches distributions, and applies supervised rotation to covariance matrices, then feeds the aligned features into a hybrid architecture combining Regularised Common Spatial Patterns with Riemannian geometry. Under leave-one-subject-out cross-validation, the method reports significant improvements across subjects and conditions, with the parallel fusion variant performing best. The motivation is practical: if true, BCIs calibrated on one group could generalise to new users with little or no recalibration.","feed_headline":"Aligning EEG geometries lets models transfer between subjects","feed_subtitle":"New BCI users could skip long calibration sessions if this transfer method holds.","key_machinery":"ITSA: a subject-specific pre-alignment applied to EEG covariance matrices before feature extraction—recentring each subject's SPD matrices, matching their distributions, and applying a supervised rotation to bring class-discriminative directions into a common frame. The hybrid architecture then combines RCSP-derived spatial filters with Riemannian geometry features in parallel and sequential pathways, so the classifier sees both filter-bank projections and manifold-geodesic distances while preserving the SPD structure.","core_discovery":"The central claim is that EEG covariance matrices, viewed as points on a Riemannian manifold of symmetric positive-definite (SPD) matrices, can be transferred across subjects by a three-stage pre-alignment: subject-specific recentering, distribution matching, and supervised rotational alignment. After this ITSA step, a hybrid classifier that fuses RCSP spatial-filter features with Riemannian tangent-space features achieves higher leave-one-subject-out classification accuracy than either branch alone. The paper reports that the parallel fusion of these two feature streams outperforms a sequential configuration, and that performance holds across varying data conditions and electrode configurat","pith_inferences":["The 'supervised' in supervised rotational alignment is the pivotal ambiguity: if it consumes target-subject labels, then the leave-one-subject-out experiment is not a true label-free transfer test. A label-free version of the rotation would be the decisive extension.","The same recentre-match-rotate rubric could transfer to any modality whose features live in a Riemannian manifold of covariance matrices—such as EMG or local field potentials—potentially broadening the method's reach beyond EEG.","The success of the parallel fusion over sequential suggests that early, irreversible fusion of spatial-filter and tangent-space features destroys useful information; an ablation study that varies fusion depth would make this explicit.","A natural next experiment is online or closed-loop evaluation: the paper's static leave-one-subject-out metrics would need to hold under non-stationarity and real-time constraint for the rehabilitation application to be realised."],"forward_implications":["New users could receive a working BCI from an existing model without a long calibration session, since the pre-alignment absorbs subject-specific offsets.","The method's reported robustness to electrode configurations means recordings from different montages could be pooled or used interchangeably.","Parallel fusion of RCSP and Riemannian features suggests that spatial-filter and manifold representations carry complementary class information; keeping both helps more than chaining them.","If the approach transfers across subjects and conditions in motor rehabilitation settings, it could enable personalised music-based interventions that adapt in real time to a patient's neural state."],"supporting_citations":[],"fun_headline_variants":["EEG manifold alignment transfers models across subjects without recalibration","Fusing RCSP and Riemannian features improves cross-subject EEG decoding","Parallel fusion of spatial and tangent-space features aids EEG transfer","Skip recalibration: ITSA aligns EEG geometries across new users","EEG covariance alignment boosts cross-subject BCI with no per-user training"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The entire transfer gain rests on the premise that after recentering, distribution matching, and supervised rotation, different subjects' covariance structures share a common Riemannian geometry, and that the supervised alignment does not rely on labels from the target subject.","fun_headline_variants_meta":{"raw":{"variants":["EEG manifold alignment transfers models across subjects without recalibration","Fusing RCSP and Riemannian features improves cross-subject EEG decoding","Parallel fusion of spatial and tangent-space features aids EEG transfer","Skip recalibration: ITSA aligns EEG geometries across new users","EEG covariance alignment boosts cross-subject BCI with no per-user training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000336,"raw_usage":{"total_tokens":1682,"prompt_tokens":710,"completion_tokens":972,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":883}},"tokens_in":454,"tokens_out":972,"duration_ms":9914,"temperature":1.0,"reasoning_tokens":883,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:33:21.358343+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Remove all target-subject labels from the supervised rotational-alignment stage and re-run leave-one-subject-out cross-validation. If the reported accuracy gain disappears or drops sharply, the method is not label-free cross-subject transfer.","supporting_citations":[],"review_version":1}