{"id":"69fd2f29-b966-489e-820a-d0023862485c","arxiv_id":"2412.11695","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A CNN-transformer hybrid with frequency-based masked autoencoding and source-frequency resampling improves low-data bio-signal transfer learning on several benchmark tasks.","lead":"The paper proposes CiTrus, a convolution-transformer hybrid for transfer learning on low-data bio-signal datasets, with masked autoencoding, a frequency-based pretraining task, and a resampling-based fine-tuning method. It reports large accuracy gains on several downstream bio-signal tasks, though the gains are mixed across datasets and the model comparison protocol is not fully consistent.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CiTrus(fp), the headline frequency-pretrained model, achieves SOTA in the 10-fold CV (EMG ACC 97.9) but collapses to near chance on the original TFC split (EMG ACC 54.3, FD-B ACC 30.8); this unexplained internal inconsistency, not merely protocol asymmetry, is the load-bearing concern.","rationale":"The reader's weakest assumption focuses on per-model protocol choices (2s vs 30s pre-training) in the appendices. That concern is real but applies mainly to Table 5, where baselines were given their preferred 30s setting while CiTrus(fp) used 2s; Table 1 itself appears to use a consistent 2s protocol. The more serious issue is that CiTrus(fp), the model highlighted in the abstract and Table 13 as best on average, collapses from 97.9 ACC in Table 1 to 54.3 ACC in Table 5 on EMG, and from 79.8 to 30.8 on FD-B, while other models remain roughly consistent or improve. This is an internal inconsistency that cannot be explained by the protocol deviation, because CiTrus(fp) uses its preferred 2s setting in both tables. It suggests either a concrete implementation error in one of the two evaluations or that the headline result is not reproducible under the original TFC splits. Since the central claim of significant improvement rests on CiTrus(fp)'s performance, this discrepancy must be resolved before the paper can be assessed. The proposed test (re-running Table 5 with the Table 1 code path) would settle whether the cause is a bug or non-reproducibility. Until then, the manuscript is not verifiable, hence the move from CONDITIONAL to UNVERDICTED.","tokens_in":38021,"tokens_out":13064,"duration_ms":107745,"concrete_test":"Independently reproduce Table 5's CiTrus(fp) rows on EMG and FD-B using the Table 1 fine-tuning code path (2s pre-training, 200-timestep interpolation, linear head, same hyperparameters/seeds). If ACC returns to ~97.9 / ~79.8, the Table 5 drop is an evaluation-protocol bug and the SOTA claim holds under the 10-fold CV protocol; if ACC remains ~54 / ~31, the headline result does not reproduce on the original splits and the central claim must be restricted or withdrawn.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"Table 1 (10-fold CV, low-data regimes) reports CiTrus(fp) as best on EMG@20% (ACC 97.92, ROC 99.02, PRC 98.76) and FD-B@0.5% (ACC 79.78). Table 5, the original TFC splits evaluated with the same stated pre-training and fine-tuning choices (Appendix B: CiTrus(fp) uses 2s pre-training and 200-timestep interpolation, its preferred setting per Appendix G), reports CiTrus(fp) at ACC 54.27 / ROC 76.11 / PRC 72.11 on EMG and ACC 30.84 / ROC 53.11 / PRC 36.88 on FD-B, far below all baselines and near chance for 3-class tasks. Other models are roughly consistent or improve in Table 5 (e.g., CiTrus(p) EMG 95.12 vs 90.45 in Table 1; Ci(s) FD-B 82.9 vs 74.26). The collapse is not explained by the 2s-vs-30s protocol deviation, since CiTrus(fp) is already using its preferred 2s setting in both tables; nor by data regime, since Table 5 uses the full original training split. This suggests a bug (e.g., incorrect input length, wrong classification head) or that the Table 1 result is not reproducible under the standard TFC evaluation protocol. Because the abstract and Table 13 use CiTrus(fp)'s average performance to support the frequency-pre-training and significant-margin claims, this internal inconsistency is load-bearing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CiTrus, a convolution-transformer hybrid for transfer learning from SleepEDF EEG to low-data bio-signal classification tasks. It introduces masked autoencoding, a frequency-based spectrogram prediction variant, multimodal EEG/EOG pre-training, and a frequency-matching resampling fine-tuning strategy. Experiments compare against PatchTST, bioFAME, NLPatchTST, SimMTM, and a CNN-only variant across EMG, ECG, PPG, HAR, FD-B, Epilepsy, Gesture, and SleepEDF, under multiple data regimes and with both 10-fold cross-validation and the original TFC splits. The paper claims a significant performance margin for CiTrus, especially for the frequency-pretrained variant, and advocates 10-fold evaluation for transfer-learning benchmarks.","tokens_in":38391,"tokens_out":5224,"duration_ms":50092,"significance":"If the results hold, the paper makes useful contributions: a new hybrid architecture, two pre-training objectives, a resampling-based fine-tuning method that improves on temporal interpolation, and a more thorough evaluation protocol with cross-validation and variance reporting. The statistical reporting (standard deviations, Wilcoxon tests, fold-vs-seed variance analysis) is stronger than is typical for this literature and deserves credit. However, the headline claims currently rest on unresolved internal inconsistencies and protocol asymmetries that must be fixed before the contribution can be accepted.","major_comments":[{"comment":"Table 1 reports CiTrus(fp) as the best model on EMG@20% (ACC 97.92, ROC 99.02, PRC 98.76) and FD-B@0.5% (ACC 79.78), while Table 5 reports CiTrus(fp) at ACC 54.27 / ROC 76.11 / PRC 72.11 on EMG and ACC 30.84 / ROC 53.11 / PRC 36.88 on FD-B under the original TFC split. Both evaluations use the same stated setting for CiTrus(fp) in Appendix B: 2s pre-training windows and 200-timestep interpolation, which is the setting that Appendix G identifies as preferred for this model. The two tables therefore contradict each other for the model that underpins the abstract claim and the Table 13 averages, and the manuscript offers no explanation. The authors must either reconcile these results or replace the affected claims with results from a reproducible protocol; a near-chance result on the full TFC split for a model that is near-perfect under 10-fold CV suggests an implementation issue that cannot remain unaddressed.","section":"Section 4, Table 1 vs Table 5, Appendix B"},{"comment":"The headline comparison is not run under a single transfer protocol. For EMG and FD-B, Appendix B states that most models use 30s SleepEDF pre-training with resampling to 3000 timesteps, while SimMTM and CiTrus(fp) use 2s pre-training because of incompatibility or better performance. Appendix G then shows large protocol-dependent differences, e.g., moving from 2s to 30s pre-training improves PatchTST(s) on EMG@20% by +85.7% while changing CiTrus(fp) by -7.2% to -9.8%. If these per-model protocol choices were made after inspecting test performance, the claimed significant margin over baselines is not a controlled comparison. Please report all models under the same pre-training/fine-tuning protocol per dataset, or justify protocol choices using only validation-based model selection.","section":"Appendices B and G"},{"comment":"SleepEDF is used both as the pre-training corpus and as a downstream evaluation target, but the manuscript does not explicitly state whether the SleepEDF test set or the subjects in it were excluded from pre-training. If the test set or its subjects were included in the pre-training data, the SleepEDF downstream results and the dataset-averaged results in Table 13 are affected by information leakage. The authors should state the exact subject/window split used for pre-training versus downstream evaluation, and, if necessary, re-run the SleepEDF experiments with a properly held-out pre-training split.","section":"Section 4, Appendix A, SleepEDF evaluation"}],"minor_comments":[{"comment":"The second contribution bullet contains a typo: 'propsed' should be 'proposed'.","section":"Section 1, contributions"},{"comment":"The sentence 'during we evaluate each model across a range of data percentages' is missing a word; it should read 'during fine-tuning, we evaluate' or similar.","section":"Section 4, Experimental settings"},{"comment":"The phrase 'middle data regie' should be 'middle data regime'.","section":"Section 5, Discussion"},{"comment":"Tables 15-17 report Wilcoxon p-values only for accuracy, while the text refers to 'average performance improvement' across three metrics; please clarify whether the tests were run on each metric separately or only on accuracy.","section":"Appendix F"},{"comment":"The figure is dense and the subfigure labels are small; increasing font size and resolution would make the architecture and pre-training pipelines substantially easier to follow.","section":"Figure 1"},{"comment":"No code or data link is provided. For an empirical benchmark paper of this type, releasing the code is important for reproducibility and would also help resolve the Table 1/Table 5 discrepancy discussed above.","section":"General"},{"comment":"The caption should state explicitly that the EMG/FD-B protocol in this table deviates from Table 1 for some models (as described in Appendix B), so readers do not interpret the table as a like-for-like comparison.","section":"Table 5 caption"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistency between Table 1 and Table 5 for CiTrus(fp) is severe enough that I would want it resolved before publication, and the per-model protocol choices in Appendices B and G raise a fairness question about the central comparison. The paper has useful ideas and unusually careful variance reporting, so a major revision with a single controlled protocol and a resolution of the CiTrus(fp) discrepancy seems achievable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Xiang — the thing to know about this paper is that its headline model, CiTrus(fp), is both the best-looking model in the paper and its biggest red flag. In the 10-fold CV (Table 1) it hits 97.9 accuracy on EMG@20% and 79.8 on FD-B@0.5%. On the original TFC splits (Table 5), with what appears to be the same pre-training and fine-tuning choices, it collapses to 54.3 and 30.8 — near chance for those 3-class tasks. The paper never acknowledges this. That is not a minor quibble; it bears directly on the abstract's claim that frequency-based pre-training performs best on average.\n\nWhat the paper does well: the evaluation framework is genuinely more careful than most in this literature. Ten-fold CV, three data regimes, standard deviations across seeds and folds, Wilcoxon tests — that is real work and worth emulating. The resampling-adaptive fine-tuning idea (match downstream sampling frequency to pre-training) is simple and, in the reported tables, helps most models substantially. The hybrid CNN encoder plus PatchTST transformer is a modest combination of known pieces, but the ablation of the CNN encoder against an MLP encoder (NLPatchTST) is a fair way to isolate its effect.\n\nThe soft spots, in order of severity. First, the CiTrus(fp) collapse. If it is a bug, the paper needs to say so; if not, the Table 1 result is not reproducible under the standard protocol. Either way the claim 'significantly outperforms previous models' is not supportable as written. Second, Appendices B and G document per-model choices of pre-training window length (2s vs 30s), made after seeing performance. That is selection, and it makes the headline comparison not a controlled experiment. Third, no code is provided, which matters more than usual because the discrepancy is the kind of thing code would resolve quickly.\n\nThis is a serious empirical study with a load-bearing internal inconsistency. I would send it to peer review — the framework and the fine-tuning trick are worth refereeing — but I would require the authors to explain or fix the Table 5 numbers and release code before acceptance. For my own work, I would not cite the frequency-pre-training claim until that is settled.","headline":"Solid evaluation framework and a useful fine-tuning trick, but the headline frequency-pretrained model has an unexplained near-chance collapse on the original TFC splits.","tokens_in":38933,"tokens_out":4458,"would_cite":false,"duration_ms":38069,"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":"A CNN-transformer hybrid pretrained on EEG sets new low-data bio-signal benchmarks.","keywords":["bio-signal transfer learning","masked autoencoding","convolution-transformer hybrid","frequency-based pretraining","low-data classification","channel independence","resampling-adaptive fine-tuning","self-supervised learning"],"falsifier":"Evaluate CiTrus(fp) on the EMG and FD-B benchmark splits using the same 30-second pretraining windows and the same resampling rule as bioFAME, with no per-model exceptions; if its accuracy falls to the level of the baselines (e.g., below ~90% on EMG@20%), the claimed significant margin is an artifact of protocol selection.","tokens_in":37804,"feed_emoji":"🧠","tokens_out":5503,"duration_ms":46804,"temperature":0.7,"pith_summary":"This paper argues that a convolution-transformer hybrid, called CiTrus, can substantially improve transfer learning for low-data bio-signal classification when pretrained on a large EEG dataset with a masked-autoencoding task. The authors claim that the convolution-only part of the model already reaches state-of-the-art performance on several downstream tasks, and that adding the transformer often helps further. They also introduce a frequency-based pretraining task that predicts z-scored spectrograms of masked patches, and a fine-tuning strategy that resamples downstream data to the pretraining sampling frequency instead of interpolating it. If the claims hold, models pre-trained on EEG could reduce the amount of labeled EMG, ECG, PPG, and other bio-signal data needed for wearable health applications.","feed_headline":"Pretrained on EEG, a hybrid model wins low-data bio-signal tasks","feed_subtitle":"Frequency-aware pretraining plus sampling-rate alignment cuts labels needed for EMG, ECG, and PPG classification.","key_machinery":"The load-bearing object is CiTrus, a hybrid of a 3-layer residual convolutional encoder and a PatchTST transformer. The CNN produces patch embeddings that are masked and fed to the transformer; channel independence is achieved by concatenating channels along the batch dimension, which lets the same pretrained weights transfer to datasets with different channel counts. Two self-supervised pretraining heads are used: one reconstructs the raw time series of masked patches, and the other reconstructs z-scored mel-spectrograms, forcing the model to predict frequency content that is not available from neighboring patches. For transfer, the fine-tuning data is resampled to the pretraining sampling frequency; if it becomes longer than the pretraining window, an overlapping sliding window is applied and embeddings are averaged across windows. This alignment is what makes the pretrained frequency filters and temporal relationships reusable.","core_discovery":"The central claim is that CiTrus outperforms previous bio-signal transfer-learning models (TFC, bioFAME, SimMTM, PatchTST) by a significant margin on several low-data benchmarks. The best configuration, CiTrus with frequency-based pretraining (fp), achieves the highest average performance in the lowest and highest data regimes, including 97.92% accuracy on EMG@20%, 82.44% on ECG@0.5%, and 79.78% on FD-B@0.5%. The convolution-only encoder (Ci) trained from scratch already outperforms all baselines on several datasets, indicating that convolutional inductive bias and parameter efficiency carry much of the advantage. The paper also claims that resampling downstream data to match the pretraining frequency, with overlapping sliding windows when needed, improves fine-tuning for essentially all models, by up to about 60% on the PPG dataset, because it keeps the learned frequency representations aligned with the pretraining data.","pith_inferences":["The channel-batch concatenation trick could generalize to any multivariate time-series problem, making pretrained sensor models portable across devices with different channel counts.","The success of frequency pretraining suggests a broader principle for signals where discriminative information lives in the spectrum: self-supervised targets should be spectral, not raw, which could carry over to audio and vibration monitoring.","Because the convolution-only encoder sometimes matches the hybrid, a testable extension is to ablate the transformer depth or replace it with a lighter mixer in extreme low-data regimes to cut parameters further.","The reported margins may depend on per-model protocol choices (2s vs 30s pretraining windows, resampling rules); a controlled re-run with a single fixed protocol per dataset would give a cleaner ranking of models."],"forward_implications":["Low-data bio-signal classifiers for EMG, ECG, PPG, and fault detection can be built by pretraining once on EEG, then fine-tuning with a fraction of the usual labeled data.","Predicting the spectrogram of masked patches is a better pretraining task than raw-signal reconstruction for the lowest and highest data regimes, so future self-supervised designs should consider frequency-domain targets.","Resampling downstream data to the pretraining sampling rate, rather than interpolating to the pretraining length, is a nearly universal improvement and should become the default transfer protocol.","Convolution-based encoders outperform transformer-only encoders in low-data bio-signal settings, suggesting that weight sharing and parameter efficiency matter more than attention capacity when labels are scarce.","Pretraining helps transformer-heavy models the most, especially in low-data regimes, while fully convolutional models trained from scratch already work well on some datasets (EMG, FD-B)."],"supporting_citations":[{"why":"Defines the TFC contrastive pretraining method, the original test splits for EMG, FD-B, Epilepsy, and Gesture, and the temporal-interpolation fine-tuning approach that the paper compares against.","marker":"(Zhang et al. 2022)"},{"why":"Introduces bioFAME, the frequency-aware masked autoencoder baseline, the 30s SleepEDF pretraining setup, and the use of PatchTST as a transformer backbone that the paper adapts.","marker":"(Liu et al. 2023)"},{"why":"Introduces SimMTM, the manifold-based masked time-series pretraining baseline, and its official implementation used for comparison.","marker":"(Dong et al. 2024)"},{"why":"Proposes PatchTST, the channel-independent transformer that forms the transformer part of CiTrus and the PatchTST baseline.","marker":"(Nie et al. 2022)"},{"why":"Provides the masked autoencoding framework that the paper's pretraining heads are built on.","marker":"(He et al. 2022)"}],"fun_headline_variants":["CiTrus: Frequency-aware pretraining boosts low-data bio-signal tasks","Hybrid model with frequency masking wins low-data bio-signal benchmarks","CiTrus squeezes more from low-data bio-signals via frequency pretraining","Convolution-only CiTrus beats baselines on low-data bio-signal tasks","Sampling-rate alignment multiplies gains in bio-signal transfer learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline comparisons assume every model is evaluated under the same transfer protocol, but the paper allows per-model choices of pretraining window length and resampling in the appendix, so the claimed significant margin could reflect those choices rather than the architecture alone.","fun_headline_variants_meta":{"raw":{"variants":["CiTrus: Frequency-aware pretraining boosts low-data bio-signal tasks","Hybrid model with frequency masking wins low-data bio-signal benchmarks","CiTrus squeezes more from low-data bio-signals via frequency pretraining","Convolution-only CiTrus beats baselines on low-data bio-signal tasks","Sampling-rate alignment multiplies gains in bio-signal transfer learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1505,"prompt_tokens":992,"completion_tokens":513,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":414}},"tokens_in":608,"tokens_out":513,"duration_ms":4691,"temperature":1.0,"reasoning_tokens":414,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:40:18.407217+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate CiTrus(fp) on the EMG and FD-B benchmark splits using the same 30-second pretraining windows and the same resampling rule as bioFAME, with no per-model exceptions; if its accuracy falls to the level of the baselines (e.g., below ~90% on EMG@20%), the claimed significant margin is an artifact of protocol selection.","supporting_citations":[],"review_version":1}