{"id":"14b3eff5-8ce6-4d18-89ac-bd1dd8701ffe","arxiv_id":"2506.10933","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new cross-subject SSVEP transfer method, iTRCA, and its subject-selection variant SS-iTRCA, improve recognition accuracy over TRCA and several transfer baselines, particularly with limited target calibration data.","lead":"This paper proposes a transfer learning framework, iTRCA, that combines source-subject and target-subject features to recognize SSVEP brain signals with less calibration data, plus a variant SS-iTRCA that selects similar source subjects using task-related components. Tests on three EEG datasets show accuracy gains over standard TRCA and some prior transfer methods, especially when the target subject has few training blocks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SS-iTRCA gain may be inflated by test-set-tuned selection hyperparameters: gamma and clb are fixed at 0.5 and 0.9 without a nested validation loop, and Fig. 8a shows the Benchmark peak is at clb=0.8.","rationale":"The reader's verdict is already CONDITIONAL, and the reader's rationale explicitly flags the threshold being tuned on the evaluation dataset. I agree with that condition, and my concrete check is designed to settle whether the SS-iTRCA enhancement survives a properly nested selection of gamma and clb. The main iTRCA-versus-TRCA result is supported by repeated-measures ANOVAs across three datasets and does not depend on clb, so no rejection is warranted. However, the 'SS-iTRCA further enhances' claim and the comparison with TransRCA, DGTF, and CSSFT rely on hyperparameters fixed without a validation split and on a single dataset, respectively. A nested tuning check would determine whether the reported selection gain is real or an artifact of test-set selection. Since the reader already conditioned on these points, the verdict remains unchanged.","tokens_in":15180,"tokens_out":4651,"duration_ms":51931,"concrete_test":"Recompute the Benchmark and BETA evaluations with nested leave-one-subject-out tuning: for each held-out target subject, select gamma in {0.3, 0.5, 0.7} and clb in {0.6, 0.7, 0.8, 0.9} using LOBO accuracy on the remaining subjects only, then evaluate on the held-out subject. Compare mean accuracy and the Table II deltas at Ntb=2 and Ntb=3. If the SS-iTRCA advantage over iTRCA is no longer statistically significant or the deltas shrink materially, the fixed-threshold results overstate the selection benefit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline comparisons fix the subject-selection hyperparameters at gamma=0.5 and clb=0.9 (Section III.A) without a nested validation procedure. Section III.C evaluates clb on the same Benchmark evaluation protocol and reports that SS-iTRCA accuracy peaks at clb=0.8 and declines at 0.9 (Fig. 8a), so the chosen value appears to be a post hoc selection on the evaluation set. Because all SS-iTRCA results in Figs. 4-6 and Table II use this fixed value, the reported SS-iTRCA-versus-iTRCA improvements (e.g., Table II deltas of 11.32, 9.08, and 21.19 percentage points at Ntb=2) are not unbiased estimates of the selection benefit. This is reinforced by the self-collected dataset, where SS-iTRCA does not significantly beat iTRCA. The parameter-free iTRCA-versus-TRCA claim is less affected, since it does not depend on clb; the concern targets the 'SS-iTRCA further enhances' part of the conclusion and the comparison with existing transfer methods, which is reported only on Benchmark.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes iTRCA, an SSVEP transfer-learning method that combines a subject-general feature—obtained by CCA between the target subject's template and a stack of source subjects' task-related components—with a target-specific TRCA feature in a filter-bank correlation score. A subject-selection extension, SS-iTRCA, filters source subjects by the normalized correlation between target and source TRCs. The method is evaluated with leave-one-subject-out and leave-one-block-out cross-validation on the Benchmark, BETA, and a self-collected dataset, with comparisons to TRCA and to TransRCA, DGTF, and CSSFT. The central empirical claim is that iTRCA and SS-iTRCA outperform TRCA, with the largest gains when the target has few calibration blocks, e.g., Table II deltas of 11.32, 9.08, and 21.19 percentage points at Ntb=2.","tokens_in":15477,"tokens_out":8588,"duration_ms":93152,"significance":"If the results are unbiased, the contribution is practically relevant: it offers a simple, training-efficient way to reduce SSVEP calibration data, and the similarity-based subject selection is far cheaper than accuracy-based methods such as CSSFT. The TRCA and CCA machinery is standard, and the nested LOSO/LOBO protocol is correctly applied so that test trials do not enter training. The main weaknesses are the selection of the c_lb threshold on the Benchmark evaluation set, the Benchmark-only state-of-the-art comparison, and the failure of SS-iTRCA to improve over iTRCA on the self-collected dataset; these weaken the strength of the conclusion as currently worded. No code is provided, so reproducibility is limited.","major_comments":[{"comment":"The subject-selection hyperparameters γ=0.5 and c_lb=0.9 are fixed for all reported evaluations without a nested validation loop, and the only sensitivity analysis (Fig. 8a, Benchmark) shows SS-iTRCA accuracy peaking at c_lb=0.8 and declining at 0.9. Because all SS-iTRCA results in Figs. 4–6 and Table II use c_lb=0.9, the headline SS-iTRCA-vs-iTRCA gains, including Table II deltas of 11.32, 9.08, and 21.19 percentage points at Ntb=2, are not unbiased estimates of the benefit of subject selection. Please either introduce an inner validation loop that selects γ and c_lb from training data only, or report the main results across a range of c_lb values on all three datasets and demonstrate that the chosen value is not evaluation-set-dependent.","section":"Section III.A, III.C, Eq. (19)"},{"comment":"The comparison with existing transfer learning methods TransRCA, DGTF, and CSSFT appears only on the Benchmark dataset (Fig. 9). The conclusion's statement that SS-iTRCA 'shows superior results compared to the existing transfer learning methods' is therefore broader than the evidence supports. Please either run the comparison on BETA and the self-collected dataset, or qualify the conclusion to Benchmark. In addition, the statistical test in Fig. 9a is described only as a paired t-test between iTRCA and TransRCA; please specify all pairwise comparisons and any multiple-comparison correction.","section":"Section III.D and V"},{"comment":"The conclusion states that SS-iTRCA 'further enhances the performance of iTRCA', but the self-collected dataset consistently shows no statistically significant difference between SS-iTRCA and iTRCA (Figs. 4c, 5c, 6c), a point acknowledged in Section IV.C. Since this negative result is part of the paper's own evidence, the conclusion should be restricted to the public datasets or supported by an explicit cross-dataset analysis; as written, it overstates the benefit of the selection module.","section":"Section IV.C and V"},{"comment":"The trigger rule in Eq. (16) is not justified: when all c_n_i ≤ γ, all source subjects are used even though this is precisely the situation where the source pool is least similar to the target; the stated preference for more data may conflict with the goal of avoiding negative transfer. The value γ=0.5 is fixed without a sensitivity analysis, and Fig. 8c shows that for some target subjects the entire source pool is used. Please provide an ablation of the trigger (e.g., always select, never trigger, varied γ) or a justification based on data, since this rule materially changes which source subjects enter the transfer.","section":"Section II.D, Eq. (16)"}],"minor_comments":[{"comment":"The text says the gap among SS-iTRCA, iTRCA, and TRCA is organized in Table II, but the table lists only TRCA and SS-iTRCA; an iTRCA column would make the table consistent with the text.","section":"Table II"},{"comment":"Using |c_n_i| to define similarity groups source subjects with large-magnitude negative correlations together with positively correlated subjects; the authors should justify why negative TRC correlation should count as similarity.","section":"Eq. (17)"},{"comment":"The number of sub-bands is fixed to N_m=3 without sensitivity analysis; since filter-bank settings affect SSVEP feature quality, a brief analysis or citation justifying this choice is needed.","section":"Section II.C, Eq. (13)"},{"comment":"The asterisks for significance are not tied to specific pairs or to a multiple-comparison procedure; please clarify which comparisons are tested and how the repeated-measures ANOVAs were followed up.","section":"Fig. 4"},{"comment":"Eq. (22) and its surrounding text should state explicitly that the selection time T = d + 0.5 s (gaze shifting plus data length) rather than only 'including' these components.","section":"Eq. (22)"},{"comment":"There is no data or code availability statement; releasing the self-collected dataset and MATLAB code would substantially aid reproducibility.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript header contains a DOI and appears to be a published JBHI article. If this submission is the same as the published version, the editor should confirm the relationship and whether the hyperparameter-selection concern was addressed in the published version. The c_lb issue is fixable but requires re-running experiments, which supports a major-revision rather than an accept recommendation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this SSVEP transfer learning paper. The headline: iTRCA is a credible, modest improvement over TRCA that holds across three datasets, and the subject-selection extension (SS-iTRCA) is a neat idea with a real methodological weakness: the similarity threshold looks tuned on the evaluation set.\n\nWhat's actually new: representing each source subject as a TRC instance and then learning a CCA projection between the concatenated source TRCs and the target template, with per-source weights, is a sensible instance-based transfer scheme. The TRC-based similarity metric for selecting source subjects is also new and avoids the expensive per-source recognition pipeline used in accuracy-based selection. The paper is careful with LOSO/LOBO cross-validation, and the accuracy/ITR/feature-distribution comparisons on Benchmark, BETA, and their own dataset are thorough. The computational-cost comparison showing SS-iTRCA cheaper than CSSFT is useful.\n\nThe soft spots are real but don't sink the paper. First, c_lb (the selection threshold) is set to 0.9 throughout, but Fig. 8a shows that on Benchmark accuracy peaks at c_lb=0.8 and declines at 0.9. That means the reported SS-iTRCA gain over iTRCA is not a clean estimate — the threshold appears to have been picked on the same data used for the headline numbers. A nested validation or a sensitivity analysis across datasets would fix this. Second, the comparison with TransRCA/DGTF/CSSFT is only on Benchmark, so we don't know how those methods do on BETA or the self-collected set. Third, on the self-collected set SS-iTRCA does not significantly beat iTRCA, which the authors acknowledge; it's consistent with small-sample behavior but weakens the \"SS-iTRCA further enhances\" claim as a general statement.\n\nI don't think these flaws break the central iTRCA-versus-TRCA claim, because that comparison is parameter-free and consistent. The SS-iTRCA-specific gains are probably real on public datasets, but the magnitude may be inflated by the threshold choice. The math is standard TRCA/CCA, correctly applied.\n\nThis deserves a serious referee. If I were handling it, I'd want the c_lb issue addressed before publication. I'd bring it to reading group as an example of a well-executed but slightly over-claimed transfer-learning paper.","headline":"Solid SSVEP transfer-learning paper: iTRCA is a credible improvement over TRCA, but the SS-iTRCA gains rest on a selection threshold tuned on the evaluation set.","tokens_in":15994,"tokens_out":2333,"would_cite":true,"duration_ms":24987,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A transfer-learning framework that treats each source subject's task-related component as a weighted instance, and filters out dissimilar sources before transfer, is reported to beat the standard TRCA method on three SSVEP datasets and to…","keywords":["SSVEP","brain-computer interface","transfer learning","task-related component analysis","canonical correlation analysis","negative transfer","subject selection","cross-subject recognition"],"falsifier":"Replicate the threshold sweep on a held-out target population: for each target, compute TRC similarities to all source subjects and evaluate SS-iTRCA accuracy for every $c_{lb}$ between 0 and 1. The claim predicts a peak at an intermediate threshold (the paper reports one near 0.8 on Benchmark); if the peak is absent, or if accuracy is highest when no source subject is excluded, or if the excluded low-similarity subjects would have improved accuracy, then TRC similarity is not a reliable guide to which source subjects transfer well.","tokens_in":15008,"feed_emoji":"🧠","tokens_out":12919,"duration_ms":122128,"temperature":0.7,"pith_summary":"This paper tries to establish that a new SSVEP brain-computer interface user can be calibrated with far less data if the decoder borrows other subjects' brain responses not as equal contributors but as weighted instances, each summarized by a task-related component (TRC). The proposed iTRCA framework projects source TRCs and the target subject's own template into a common latent space by canonical correlation analysis, forming a subject-general feature, while a parallel TRCA branch preserves the target's individual response; a second stage, SS-iTRCA, selects only source subjects whose TRCs are similar to the target's, which is meant to prevent negative transfer and to avoid the cost of accuracy-based selection. On the Benchmark, BETA, and self-collected datasets, the paper reports that both frameworks consistently beat TRCA in recognition accuracy and feature separation, and that SS-iTRCA also beats the existing transfer methods TransRCA, DGTF, and CSSFT. The largest reported gains come when the target subject has only two calibration blocks, up to 21.19 percentage points over TRCA.","feed_headline":"Similarity-aware subject selection lifts SSVEP-BCI accuracy","feed_subtitle":"Weights other subjects' brain responses and selects similar ones to cut calibration time without hurting accuracy.","key_machinery":"The load-bearing object is the task-related component (TRC), a one-channel summary of a subject's SSVEP response made by applying a TRCA spatial filter to that subject's averaged template. The argument rides on two operations built from it. First, CCA between the concatenated source TRCs and the target's own template produces a common latent space, so each source subject is assigned a real-valued weight (the corresponding entry of $\\hat{w}_i^{GS}$) instead of the equal weight implicit in grand averaging. Second, the same TRC is reused as a similarity metric: the normalized correlation $\\tilde{c}_i^n = |c_i^n|/\\max_n |c_i^n|$ decides which sources enter the transfer. A trigger parameter $\\gamma$ governs whether selection happens at all, and the threshold $c_{lb}$ makes the framework a continuous bridge: at $c_{lb}=0$ every source is used (iTRCA), and at $c_{lb}=1$ no source is used, so the model collapses to the target-only TRCA baseline.","core_discovery":"The central claim is that cross-subject SSVEP recognition improves when each source subject's response is represented by a single task-related component and those components are treated as weighted instances rather than as equally contributing subjects. iTRCA concatenates source TRCs into a source template $Y_i^S$, runs CCA against the target subject's individual template $\\bar{X}_i^{Ta}$ to get a source weight vector $\\hat{w}_i^{GS}$ and a target spatial filter $\\hat{w}_i^{GT}$, and forms a subject-general feature $\\rho_{1,i}$ by correlating the spatially filtered test trial with the weighted source template; a subject-specific TRCA feature $\\rho_{2,i}$ preserves the target's own response. The features are fused as $\\rho_i = \\mathrm{sign}(\\rho_{1,i})\\rho_{1,i}^2 + \\mathrm{sign}(\\rho_{2,i})\\rho_{2,i}^2$ and combined across filter-bank sub-bands. SS-iTRCA pre-selects source subjects by the normalized TRC similarity $\\tilde{c}_i^n$, dropping sources that look unlike the target. The paper reports that this consistently outperforms TRCA on Benchmark, BETA, and a self-collected dataset, and also outperforms TransRCA, DGTF, and CSSFT, with the largest gains when the target has only two calibration blocks.","pith_inferences":["Editorial inference: Because the paper's threshold sweep on Benchmark shows accuracy rising as $c_{lb}$ increases to about 0.8 and then declining, the selection rule introduces a new hyperparameter; a practical deployment would need a principled way to set it per target subject, or part of the calibration saving is eaten by threshold tuning.","Editorial inference: The near-zero benefit of selection on the 11-subject self-collected dataset suggests the similarity filter matters mainly when the source pool is large and heterogeneous; a testable prediction is that the gain grows with source-pool size and shrinks or reverses when the pool is small.","Editorial inference: Because the method's selection metric is computed before transfer, the same TRC-similarity score could be tested as a generic source-quality index for other SSVEP transfer pipelines, not only for iTRCA's feature construction.","Editorial inference: The self-collected dataset uses different stimulus frequencies and only 12 targets; if the shared latent space is encoding-specific, the transfer benefit may change with stimulus design, which could be checked by running the same framework on a code-modulated or high-frequency SSVEP dataset."],"forward_implications":["With only two target calibration blocks, SS-iTRCA raises recognition accuracy over TRCA by 11.32 percentage points on Benchmark, 9.08 on BETA, and 21.19 on the self-collected dataset, so a new user can be calibrated with far fewer trials.","Because the subject-general and subject-specific features are fused in a single correlation score, the method fits into the same filter-bank processing pipeline as TRCA, so the practical change is a preprocessing and scoring step rather than a new classifier.","The TRC-based similarity score reuses computations already performed for iTRCA, which is why selecting subjects adds little cost; the paper reports substantially lower training time than accuracy-based CSSFT with comparable inference time to TransRCA.","Feature distributions from the transfer framework are more compact and better separated than TRCA's on all three datasets, with higher CHI values, which is the internal reason given for the accuracy gains.","The framework degenerates gracefully: with no similar source subjects it becomes plain TRCA, and with all source subjects passing the threshold it becomes iTRCA, so it never forces transfer when the data argue against it."],"supporting_citations":[{"why":"Supplies task-related component analysis, the spatial-filter method that generates the source instances and the target-specific feature.","marker":"[19]"},{"why":"Supplies canonical correlation analysis, used to build the common latent space between source TRCs and the target template.","marker":"[37]"},{"why":"The cross-domain correlation maximization baseline that iTRCA extends and is compared against.","marker":"[23]"},{"why":"The domain-generalization transfer baseline used in the comparative evaluation.","marker":"[21]"},{"why":"The accuracy-based subject-selection baseline; SS-iTRCA's lower computation cost is claimed against it.","marker":"[25]"},{"why":"Provides the Benchmark dataset, one of the three evaluation sets.","marker":"[31]"},{"why":"Provides the BETA dataset, the second public evaluation set.","marker":"[32]"},{"why":"Introduces instance weighting as the conceptual basis for treating each source subject's TRC as a weighted instance.","marker":"[34]"},{"why":"Provides the filter-bank sub-band decomposition and combination rule used in the final recognition score.","marker":"[39]"}],"fun_headline_variants":["Instance-weighted transfer sharpens cross-subject SSVEP accuracy","Similarity-based subject selection cuts calibration in SSVEP BCIs","iTRCA: instance-based transfer cuts SSVEP calibration data","Less calibration data, same accuracy: similarity-aware SSVEP transfer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single task-related component per stimulus adequately summarizes a subject's SSVEP response and that the correlation between that component and the target's own template is a reliable guide to which source subjects will help, because both the transfer and the selection rule are built on that correlation.","fun_headline_variants_meta":{"raw":{"variants":["Instance-weighted transfer sharpens cross-subject SSVEP accuracy","Similarity-based subject selection cuts calibration in SSVEP BCIs","iTRCA: instance-based transfer cuts SSVEP calibration data","Less calibration data, same accuracy: similarity-aware SSVEP transfer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000849,"raw_usage":{"total_tokens":3747,"prompt_tokens":1053,"completion_tokens":2694,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":2620}},"tokens_in":669,"tokens_out":2694,"duration_ms":19622,"temperature":1.0,"reasoning_tokens":2620,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:13:57.389031+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replicate the threshold sweep on a held-out target population: for each target, compute TRC similarities to all source subjects and evaluate SS-iTRCA accuracy for every $c_{lb}$ between 0 and 1. The claim predicts a peak at an intermediate threshold (the paper reports one near 0.8 on Benchmark); if the peak is absent, or if accuracy is highest when no source subject is excluded, or if the excluded low-similarity subjects would have improved accuracy, then TRC similarity is not a reliable guide to which source subjects transfer well.","supporting_citations":[{"cited_title":"Canonical Correlation Analysis,","cited_arxiv_id":null,"evidence_quote":"Supplies canonical correlation analysis, used to build the common latent space between source TRCs and the target template."},{"cited_title":"Cross-subject transfer method based on domain generaliza- tion for facilitating calibration of SSVEP-based BCIs,","cited_arxiv_id":null,"evidence_quote":"The domain-generalization transfer baseline used in the comparative evaluation."},{"cited_title":"Cross-subject spatial filter transfer method for SSVEP-EEG feature recognition,","cited_arxiv_id":null,"evidence_quote":"The accuracy-based subject-selection baseline; SS-iTRCA's lower computation cost is claimed against it."},{"cited_title":"A benchmark dataset for SSVEP-based brain–computer interfaces,","cited_arxiv_id":null,"evidence_quote":"Provides the Benchmark dataset, one of the three evaluation sets."},{"cited_title":"BETA: A large benchmark database toward SSVEP-BCI application,","cited_arxiv_id":null,"evidence_quote":"Provides the BETA dataset, the second public evaluation set."},{"cited_title":"Instance weighting methods,","cited_arxiv_id":null,"evidence_quote":"Introduces instance weighting as the conceptual basis for treating each source subject's TRC as a weighted instance."},{"cited_title":"Filter bank canonical correlation analysis for implementing a high-speed SSVEP-based brain– computer interface,","cited_arxiv_id":null,"evidence_quote":"Provides the filter-bank sub-band decomposition and combination rule used in the final recognition score."}],"review_version":1}