{"id":"350345da-1526-4e36-b01c-3f170264d037","arxiv_id":"2506.11455","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"A 4D Swin Transformer with a convolutional decoder predicts the next 10 resting-state fMRI volumes from 32 prior volumes, reporting masked MSE of 0.035 on 20 held-out HCP subjects.","lead":"This paper describes a Swin Transformer model that predicts the next 7.2 seconds of whole-brain fMRI activity from the previous 23 seconds, voxel by voxel. The authors report low error and high similarity, but the evaluation lacks trivial baselines and uses a normalization step that may leak future information.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported masked MSE of 0.035 is not interpretable without a persistence baseline; under heavy BOLD autocorrelation, copying the last input volume may achieve comparable error, undermining the central prediction claim.","rationale":"The reader's REJECT verdict is well supported, but their stated weakest assumption (full-scan z-scoring leakage) is not the single most load-bearing issue. Data leakage through global normalization is real but likely modest because the future segment is only 10 of 1200 time points. The more serious problem is the absence of any trivial baseline. The absolute MSE number is only meaningful relative to what a simple autocorrelation-based predictor achieves. The paper's shifted-MSE control is logically insufficient: for a perfect persistence model, shifted MSE would be zero, so the observed shifted MSE being larger than the prediction MSE does not rule out persistence. Given the low temporal frequencies retained by band-pass filtering, a copy-last-frame baseline could plausibly achieve MSE at or below the reported 0.035. This is a concrete, easily testable omission that directly targets the central claim. If the baseline test fails (persistence is much worse), the paper's numbers would gain credibility; if it passes, the contribution is trivial. The reader's verdict of REJECT therefore remains unchanged, but the emphasis should shift from the z-scoring leakage to the missing baseline comparison.","tokens_in":12523,"tokens_out":6791,"duration_ms":71407,"concrete_test":"On the same held-out 20-subject test set, compute the masked MSE of (i) a persistence baseline that repeats the 32nd input brain state for all 10 future outputs, and (ii) a mean baseline that repeats the spatial mean of the 32 input frames. Report the overall and per-time-step MSE curves alongside Figure 3A. If either baseline achieves an average masked MSE less than or equal to 0.035, the model provides no measurable predictive benefit over trivial BOLD autocorrelation. If the persistence MSE is substantially higher (e.g., >0.2), the central claim survives this particular objection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the model 'successfully predicted' future brain states rests on an absolute masked MSE of 0.035, but the paper reports no trivial forecasting baseline. The shifted-MSE control in Section II-D and Section III-B does not substitute: a persistence model that outputs the previous true volume would have shifted MSE of 0, so the finding that shifted MSE (0.076) exceeds the prediction MSE (0.035) only demonstrates the predictions are not exact copies of the immediately preceding frame. It does not establish that the model beats persistence. Since the BOLD signals are band-pass filtered (0.01-0.1 Hz) and spatially smoothed, adjacent time points are highly correlated; a simple 'copy the 32nd input frame' predictor could plausibly achieve masked MSE near or below 0.035, especially for the first few predicted time points where the paper reports MSE below 0.01. Without this comparison, the reported accuracy does not demonstrate that the model learned meaningful spatiotemporal dynamics, even if the full-scan z-scoring leakage is set aside.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a voxel-level fMRI brain-state forecasting architecture consisting of a 4D Swin Transformer (SwiFT) encoder and a convolutional decoder with skip connections. Using 100 unrelated HCP subjects (80 training, 20 testing), the model takes 32 consecutive brain volumes (23.04 s) and predicts the next 10 volumes (7.2 s). Preprocessing includes smoothing, band-pass filtering, z-scoring, and downsampling to 48x48x48. The model is trained with a combined masked MSE and SSIM loss. The authors report an overall masked MSE of 0.035 and SSIM of 0.954 on the held-out test set, show that shuffling input order increases MSE dramatically, and report a shifted MSE of 0.076. Regional analysis using the Yeo17 atlas shows variation across brain regions.","tokens_in":12782,"tokens_out":3238,"duration_ms":33808,"significance":"If the quantitative claim is validated, the paper would demonstrate that a transformer-based model can learn fine-grained spatiotemporal brain dynamics at voxel resolution, with potential applications in reducing scan time and in brain-computer interfaces. The manuscript has several strengths: the architecture is described in detail, evaluation is performed on held-out subjects, multiple control analyses are attempted, and the qualitative and regional analyses are informative. However, the central quantitative claim is not yet supported because the evaluation lacks trivial forecasting baselines, the normalization procedure may leak future information, and overlapping windows are treated as independent samples. These issues need to be addressed before the prediction accuracy claim can be accepted.","major_comments":[{"comment":"The z-score normalization is applied to each voxel's full time course before sliding-window sampling, so the global mean and standard deviation used to scale the input windows are computed from data that include the 10 future time points being predicted. This makes the forecasting task easier and can artificially lower the reported masked MSE. Please re-run the evaluation with normalization parameters estimated only from training windows (or from a preceding calibration segment) and report the resulting metrics.","section":"II-A, II-C"},{"comment":"The headline claim that the model 'successfully predicted the next 10 brain states with an overall masked MSE of 0.035' is not interpretable without a trivial forecasting baseline. Because the BOLD data were band-pass filtered (0.01-0.1 Hz) and spatially smoothed (6 mm FWHM), adjacent volumes are highly correlated; a persistence predictor that outputs the last input volume could plausibly achieve masked MSE at or below 0.035, especially for the first predicted states where Fig. 3A shows MSE below 0.01. The shifted MSE (0.076) only demonstrates that the predictions are not identical to the immediately preceding true volume; it does not compare against persistence. Please include per-time-point baselines such as copy-last-frame, per-voxel mean, and linear extrapolation.","section":"III-B"},{"comment":"The sliding-window procedure generates heavily overlapping samples from the same 20 test subjects, and all samples are pooled as independent observations in the reported MSE, boxplots, and regional statistics. This pseudo-replication overstates the amount of independent evidence and narrows apparent confidence. Please aggregate metrics at the subject level (e.g., average over windows within each subject before pooling) and report the number of windows per subject, or use non-overlapping test windows.","section":"II-C, III-B"}],"minor_comments":[{"comment":"The shuffled-input MSE of 4890 is reported without explaining its scale relative to the z-scored prediction MSE; please clarify the units and why the increase is so large, as this comparison is otherwise difficult to interpret.","section":"III-B"},{"comment":"The SSIM constants c1 and c2 are described as 'the default value for SSIM calculation not scaled by the pixel intensity range'; since the data are z-scored, please state explicitly how the constants relate to the intensity range of the transformed data.","section":"II-C"},{"comment":"The text reads 'three 3D transpose convolutional layers' and should read 'three 3D transposed convolutional layers'.","section":"II-B"},{"comment":"The phrase 'novel architecture' may overstate the contribution relative to the close adaptation of the existing SwiFT encoder; consider qualifying the novelty as the application and decoder design for forecasting.","section":"I, V"},{"comment":"Training is described as running for 24 epochs without early stopping or a validation-based model selection criterion; please state whether the final model was selected by training loss convergence or by held-out validation performance.","section":"III-A"}],"recommendation":"major_revision","confidential_remarks":"The central evaluation protocol needs to be reworked before the quantitative claim can be accepted. If persistence baselines and leakage-free normalization change the conclusions substantially, the paper's main contribution would need to be reframed. The architecture and qualitative analyses are promising, but the current evidence does not yet support the headline prediction accuracy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the core architecture is a legitimate and clean extension of SwiFT: a 4D Swin encoder plus a convolutional decoder with skip connections, applied to voxel-level brain-state forecasting. That is new relative to the cited literature, and the qualitative results look genuine. Second, the central quantitative claim is not currently supported because the evaluation protocol has three fixable but serious flaws.\n\nWhat the paper does well: it includes a shuffle control showing the model relies on temporal order rather than static spatial patterns, and a shifted-MSE check indicating the predictions are not exact copies of the immediately preceding volume. The regional analysis is thorough, and the authors are transparent about preprocessing choices.\n\nThe soft spots are the load-bearing ones. The headline MSE of 0.035 has no trivial baseline. Because the BOLD data are bandpass filtered (0.01–0.1 Hz) and spatially smoothed, adjacent time points are highly autocorrelated. A persistence predictor that simply outputs the last input volume could plausibly get close to that MSE, especially for the first few predicted time points where they report errors below 0.01. The shifted-MSE comparison is not a substitute; it only rules out copy-paste of the previous true volume, not the last input volume.\n\nSecond, the z-score normalization uses the full time series, including the future segment that is being predicted. That leaks the global mean and variance of the future into the model, making the forecasting task artificially easier. The reported MSE is therefore better than what a properly leakage-free protocol would give.\n\nThird, the test statistics treat heavily overlapping sliding windows from 20 held-out subjects as independent samples. That overstates the reliability of the per-time-point boxplots and the overall MSE. Subject-level aggregation or a nested analysis would be needed.\n\nThese problems are fixable, and the paper deserves a serious referee rather than a desk rejection. The architecture is interesting and the qualitative demonstrations are reasonable, but the quantitative claim needs a persistence baseline, leakage-free normalization, subject-level statistics, and ideally comparison with a simpler model. With those revisions, the result could be credible.","headline":"A plausible Swin-Transformer forecasting architecture, but the reported accuracy is uninterpretable without a persistence baseline, and the z-score normalization leaks future statistics.","tokens_in":13297,"tokens_out":3150,"would_cite":false,"duration_ms":34587,"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 4D Swin Transformer encoder with a convolutional decoder can predict future voxel-level resting-state fMRI brain states, forecasting the next 10 volumes with a masked mean squared error of 0.035 and a structural…","keywords":["brain state prediction","resting-state fMRI","Swin Transformer","voxel-level","BOLD signal","spatiotemporal modeling","structural similarity","forecasting"],"falsifier":"Re-run the prediction with a causal normalization in which each window's voxel statistics are computed only from the 32 input time points, and evaluate on non-overlapping windows. If the masked MSE rises to the level of the shifted MSE or of a last-volume persistence baseline, then the reported 0.035 is inflated by the leakage.","tokens_in":12332,"feed_emoji":"🧠","tokens_out":4703,"duration_ms":47021,"temperature":0.7,"pith_summary":"The paper proposes that a 4D Shifted Window (Swin) Transformer encoder coupled with a convolutional decoder can predict future resting-state fMRI brain states at full voxel resolution. Using 32 consecutive brain volumes as input, the model forecasts the next 10 volumes (7.2 seconds) with a masked mean squared error of 0.035 and a structural similarity score of 0.954 on held-out subjects. If correct, this would show that transformers can learn fine-grained spatiotemporal brain dynamics rather than only regional or task-related patterns. The authors also report that shuffling the input order destroys performance, arguing the model relies on temporal dependencies, and that predictions degrade with forecast horizon. This matters because accurate voxel-level forecasting could shorten fMRI scan times and inform brain-computer interfaces.","feed_headline":"Swin transformer forecasts 10 brain states from 32 past ones","feed_subtitle":"Voxel-level resting-state fMRI prediction hits masked MSE 0.035 and SSIM 0.954 on held-out subjects.","key_machinery":"The load-bearing component is the 4D SwiFT encoder, a Swin Transformer variant that partitions each input volume-time block into 3x3x3x4 patches and applies shifted-window self-attention in four hierarchical stages, with patch merging reducing resolution as the model deepens. On top of this encoder sits a decoder built from three 3D transposed convolutional layers plus two U-Net-style skip connections, which combine temporal and channel dimensions so that the decoder's output channels are treated as predicted time points. Training minimizes masked MSE plus twice the SSIM loss, so the model is explicitly optimized to preserve both voxel intensities and local spatial structure.","core_discovery":"The central claim is that the spatiotemporal organization of the resting human brain can be learned at voxel resolution well enough to extrapolate it forward in time. On 20 held-out HCP subjects, the model takes 32 consecutive brain states (23.04 s) and predicts the next 10 states (7.2 s), achieving an overall masked MSE of 0.035 and an SSIM of 0.954 averaged over all test samples and predicted time points. The paper further reports that randomly shuffling the 32 input states raises the masked MSE to 4890, and that comparing predictions to the previous ground-truth state (shifted MSE) yields roughly twice the prediction error, which the authors take as evidence that the model generates new states rather than copying the last input volume. Regional analysis shows errors concentrate in limbic areas such as the orbitofrontal cortex, which the authors attribute to variability and signal artifacts.","pith_inferences":["The reported 0.035 masked MSE likely understates the true prediction error because each voxel's time series is z-scored using its full duration, including the future segment; a causal normalization evaluated on non-overlapping windows is needed to know how much of the accuracy is real.","If the leakage is confirmed, the qualitative similarity of predicted and ground-truth volumes may remain, but the model's advantage over a persistence baseline would shrink; the shifted-MSE control is a step in this direction but not a complete one.","The regional error pattern could be tested directly: comparing prediction error in OFC/temporal pole against a local signal-quality map (e.g., temporal SNR) would separate artifact-driven errors from genuine unpredictability.","The architecture's capacity suggests a natural next experiment: pretrain the encoder on a larger HCP cohort or on task fMRI, then fine-tune for prediction, which might extend the reliable horizon beyond 7.2 seconds."],"forward_implications":["Voxel-level forecasting of resting-state BOLD activity is achievable with a transformer, extending prior region-based predictions to full 3D volumes.","Prediction error grows with horizon, so short-horizon forecasts (first five states) are much more reliable than later ones, consistent with accumulating uncertainty.","Regions in dorsal attention and control networks predict better than limbic regions, suggesting the model exploits structured intrinsic dynamics where they exist.","The same encoder-decoder could forecast other 4D biomedical volumes, provided the evaluation avoids information leakage."],"supporting_citations":[{"why":"Supplies the SwiFT encoder architecture whose hierarchical shifted-window attention models spatiotemporal fMRI patterns.","marker":"[18]"},{"why":"Provides the Swin Transformer shifted-window mechanism that makes high-resolution 4D attention tractable.","marker":"[22]"},{"why":"Inspires the skip connections between encoder and decoder that preserve fine spatial detail for voxel-level outputs.","marker":"[26]"},{"why":"Supplies the resting-state fMRI dataset (100 unrelated HCP subjects) used for training and testing.","marker":"[24]"},{"why":"Provides the SSIM metric and loss used to evaluate and optimize structural similarity of predicted volumes.","marker":"[27]"},{"why":"Earlier region-level transformer brain-state prediction that this work extends to voxel level.","marker":"[21]"},{"why":"Defines the minimal preprocessing pipelines whose output is used as the starting point for further cleaning.","marker":"[25]"}],"fun_headline_variants":["Swin Transformer forecasts voxel-level brain states from fMRI","Predicting brain activity 7.2s ahead with Swin Transformer","32 brain states in, 10 out: Swin Transformer predicts dynamics","Swin Transformer maps past fMRI to future brain states"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that z-scoring each voxel's full time series — including the future segment to be predicted — before building the input/output windows does not leak future information into the model, and that it correctly measures forecasting ability.","fun_headline_variants_meta":{"raw":{"variants":["Swin Transformer forecasts voxel-level brain states from fMRI","Predicting brain activity 7.2s ahead with Swin Transformer","32 brain states in, 10 out: Swin Transformer predicts dynamics","Swin Transformer maps past fMRI to future brain states"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000734,"raw_usage":{"total_tokens":3284,"prompt_tokens":946,"completion_tokens":2338,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":2266}},"tokens_in":562,"tokens_out":2338,"duration_ms":17362,"temperature":1.0,"reasoning_tokens":2266,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:06:56.467658+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the prediction with a causal normalization in which each window's voxel statistics are computed only from the 32 input time points, and evaluate on non-overlapping windows. If the masked MSE rises to the level of the shifted MSE or of a last-volume persistence baseline, then the reported 0.035 is inflated by the leakage.","supporting_citations":[{"cited_title":"SwiFT: Swin 4D fMRI Transformer,","cited_arxiv_id":null,"evidence_quote":"Supplies the SwiFT encoder architecture whose hierarchical shifted-window attention models spatiotemporal fMRI patterns."},{"cited_title":"Swin transformer: Hierarchical vision transformer using shifted windows,","cited_arxiv_id":null,"evidence_quote":"Provides the Swin Transformer shifted-window mechanism that makes high-resolution 4D attention tractable."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Inspires the skip connections between encoder and decoder that preserve fine spatial detail for voxel-level outputs."},{"cited_title":"Predicting Human Brain States with Transformer,","cited_arxiv_id":null,"evidence_quote":"Earlier region-level transformer brain-state prediction that this work extends to voxel level."}],"review_version":1}