{"id":"c3f8a788-65e0-4d8f-b6db-1ed4dc20883b","arxiv_id":"2501.16409","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A spatio-temporal transformer with contrastive learning reaches 89.1% accuracy in classifying mild cognitive impairment from dynamic functional connectivity on ADNI data.","lead":"This paper trains a deep-learning model to classify mild cognitive impairment from brain scan time series and reports 89.1% accuracy on 345 ADNI subjects. The model combines a transformer that reads brain-region and time-window patterns with a contrastive learning loss that pairs similar and different patients.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Superiority claim rests on unmatched cross-study comparisons in Table 1; no baseline is run on the same ADNI data, so the reported 89.1% does not establish superiority.","rationale":"The reader's weakest assumption is exactly where the paper's advertised conclusion breaks: Table 1 is presented as evidence of superiority but is not a controlled comparison. Sample size differences alone (29–95 cases vs 120) change task difficulty; EMCI vs MCI changes label definition; atlas differences change feature geometry. Without running comparators on the identical data, the headline accuracy cannot be attributed to the method. I checked other candidate concerns: the missing positional encoding in the transformer description (Sec. 2.1.1) is a real reproducibility gap, and the contrastive pairs use diagnosis labels despite claiming reduced label dependence; however, neither is as decisive as the unmatched comparison, because the central claim is empirical superiority. The subject-level split is a sound internal practice, and the ablation gives matched internal evidence, which is good. The proposed test—re-running the strongest baseline on the same data with significance testing—would settle the concern. Since the paper can still satisfy that test, the conditional verdict remains appropriate.","tokens_in":5980,"tokens_out":6193,"duration_ms":64507,"concrete_test":"Ask the authors to run at least the strongest Table 1 competitor (e.g., Li et al. 2020 dynamic FC) on the exact same 345-subject ADNI scans, Destrieux atlas, window length 70 / step 2, and subject-level 5-fold CV, and to report mean ± std and 95% CI over folds with a paired McNemar or DeLong test against the proposed model. If the competitor is not significantly worse on ACC or AUC, the 'superiority' claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of the abstract—'demonstrate the superiority' on ADNI—depends entirely on Table 1. Each comparator row is copied from a separate publication with a different sample size (29–95 MCI/NC vs 120/225 here), a different diagnostic subgroup (EMCI vs MCI), a different atlas (AAL vs Destrieux), and different preprocessing/window choices. None of these baselines is re-run on the same 345-subject/570-scan sample under the same subject-level 5-fold protocol; no confidence intervals or significance tests are reported. The 89.1% ACC is therefore a cross-study comparison with uncontrolled confounds, and it cannot by itself demonstrate superiority over prior methods. The internal ablation (Table 2) is matched and shows that contrastive and spatio-temporal components help on this dataset, but that only establishes internal consistency, not superiority over prior work. If a strong baseline—e.g., a standard static-FC classifier or the best prior method—were run on the same data and matched the 89.1% figure, the paper's main conclusion would not hold.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a spatio-temporal transformer with a contrastive learning loss for classifying mild cognitive impairment (MCI) versus normal control (NC) from resting-state fMRI. Dynamic functional connectivity (dFC) networks are constructed with a sliding-window strategy over the Destrieux atlas time series, and a temporal block and a spatial block, each containing a transformer layer and a convolutional layer, are used to extract fused representations. A contrastive loss that pulls together dFC features from subjects with the same diagnosis and pushes apart features from different diagnoses is combined with a cross-entropy classification loss. The authors report subject-level 5-fold cross-validation on 345 ADNI subjects (570 scans) and compare with ten literature results in Table 1, claiming superiority with accuracy 89.1%, sensitivity 91.4%, specificity 87.2%, AUC 89.3%, and F1 90.3% for the full model with contrastive learning. Ablation experiments in Table 2 compare static FC, spatial-only, temporal-only, and the full model.","tokens_in":6183,"tokens_out":3292,"duration_ms":31848,"significance":"If the reported superiority held under a rigorous comparison protocol, this would be a practically useful contribution to early AD screening, and the general design of jointly modeling temporal and spatial dFC information in a transformer is interesting. The paper has clear strengths: the subject-level 5-fold cross-validation prevents scan-level data leakage; the ablation study is a reasonable internal check; and the contrastive-learning mechanism is concretely described with a joint loss. However, the central empirical claim of superiority over prior work is not yet supported by the evidence presented, because the comparisons in Table 1 are cross-study and uncontrolled. The contribution is therefore promising but requires substantial additional validation before the abstract-level claim is justified.","major_comments":[{"comment":"The central claim that the method \"demonstrate[s] the superiority\" on ADNI is not supported by the evidence. Each comparator row in Table 1 comes from a different publication with a different sample size (29 to 95 MCI/NC subjects versus 120 MCI/225 NC here), a different diagnostic subgroup (EMCI versus MCI), a different atlas (AAL versus Destrieux), and different preprocessing and window choices. None of these baselines is re-run on the same 345-subject/570-scan sample under the same subject-level 5-fold protocol, and no confidence intervals, standard deviations, or significance tests are reported. The 89.1% accuracy therefore reflects a cross-study comparison with uncontrolled confounds and cannot by itself establish superiority. The authors should run at least one or two strong baselines (e.g., a static-FC classifier, a standard dFC classifier, or the best-performing prior method) on the exact same data and protocol, and report matched results with statistical testing.","section":"Table 1 and Section 3.2"},{"comment":"The claim that contrastive learning \"reduc[es] dependency on labeled data\" is misleading in its current form. Positive and negative dFC pairs are constructed explicitly from diagnosis labels (\"same diagnosis status\" versus \"different diagnoses\"), so the contrastive loss uses the same labels as the classification loss, just in a pairwise form. The paper does not demonstrate any ability to exploit unlabeled data. This claim should either be qualified to say that the contrastive loss provides additional supervisory signal from pairwise label structure, or the method should be extended and evaluated with truly unlabeled samples.","section":"Abstract and Section 2.1.2"},{"comment":"The ablation study does not cleanly isolate the contribution of contrastive learning. Table 2 compares os-FC, S-only, T-only, and \"Ours\" (89.1%), but the \"Ours\" row includes the contrastive loss, so the improvement over S-only/T-only could be driven by contrastive learning rather than by the temporal-spatial architecture. Meanwhile, Section 3.2 attributes a 9.7% accuracy improvement to contrastive learning by comparing \"Ours\" without contrastive (82.4%, Table 1) with \"Ours (cl)\" (89.1%), but that no-contrastive full model is absent from Table 2. The ablation should include the full temporal-spatial model without contrastive learning, and the word \"significantly\" should be backed by a significance test rather than a single point comparison.","section":"Table 2 and Section 3.2"},{"comment":"The method is not sufficiently specified for reproduction. The paper gives window length L=70 and step size S=2, but does not report the number of windows T, the length of the ROI time series after preprocessing, the number of transformer layers, the number of attention heads, the embedding dimensions, the feed-forward hidden size, the pooling strategy before classification, or the specific values of the learnable hyperparameters alpha and beta in Eq. (8). These details are necessary for any independent verification of the reported 89.1% accuracy and for assessing whether the method is genuinely novel relative to existing dFC transformers.","section":"Section 2.1 and Section 3.1"}],"minor_comments":[{"comment":"The title contains a typo: \"CLASSIFFICATION\" should be \"CLASSIFICATION,\" and the abstract uses \"BLOD\" where \"BOLD\" is intended.","section":"Title and Abstract"},{"comment":"The subsection numbering is inconsistent: Sections 2.1.1 and 2.1.2 appear as subsubsections of Section 2.2 (Model), but they are numbered as if they belong to Section 2.1. They should be renumbered 2.2.1 and 2.2.2.","section":"Section 2"},{"comment":"The notation \"1[NO#]\" in Eq. (6) is unclear; the indicator function should be written explicitly, for example as 1[k ≠ i] with a definition of the pair index k, so that the contrastive loss is formally readable.","section":"Equation (6)"},{"comment":"Reference [12] (Lin et al., Frontiers in Neuroscience, 2022) is listed twice, and reference [11] is missing from the reference list; the citation numbering should be corrected throughout.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be written in a compact conference style. If the intended venue is a full journal, the authors should substantially expand the methods and evaluation. The main concern is that the headline claim of superiority is not supported by the uncontrolled comparison in Table 1; however, this is fixable through matched experiments and statistical reporting, so the paper is not beyond salvage. I would also note that the incomplete mention of a repository in the review materials was not verifiable from the manuscript itself."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this paper has a well-designed internal experiment and a clean ablation, but its central claim—'superiority' over prior MCI classification methods—is not supported by the evidence as presented. Table 1 compares the proposed model against numbers copied from other papers with different sample sizes, different atlases (AAL vs Destrieux), different diagnostic subgroups (EMCI vs MCI), and different preprocessing. No baseline is run on the same 345-subject ADNI sample under the same protocol. The 89.1% accuracy is therefore a single number in a mixed table, not a demonstrated improvement.\n\nWhat is actually new: the specific combination of a temporal transformer block, a spatial transformer block, and a supervised contrastive loss applied to dynamic functional connectivity for MCI classification is not in the cited literature. Each piece is familiar—BolT and others have done transformers for fMRI—but the assembly is a reasonable contribution. The subject-level 5-fold cross-validation is the right protocol, and there is no apparent leakage. The ablation (Table 2) is matched and shows that both the temporal and spatial blocks add value over static FC and over either block alone. That is a genuinely useful internal consistency check. The contrastive loss produces a 9.7 point accuracy improvement over the no-contrastive variant, which is a large effect and worth taking seriously.\n\nSoft spots, in order of severity. First, the unmatched comparison. This is the load-bearing issue. The authors should run at least one standard baseline (say, a linear SVM or a simple static-FC deep net) on the same data and report error bars across folds. Second, the claim that contrastive learning 'reduces dependency on labeled data' is misleading: the positive and negative pairs are built from the same diagnosis labels used for classification, so it is supervised contrastive learning, not a reduction in label reliance. Third, implementation details are thin: no code or data release, no mention of positional encodings in the transformer despite the paper's emphasis on sequential information, and a few typographical problems (BLOD for BOLD, duplicated reference [12], and odd section numbering). These are fixable.\n\nWho is this for? Researchers working on fMRI-based MCI or AD screening, especially those combining dFC with deep learning. The paper deserves a serious referee, because the internal ablation is informative and the method is plausible, but it should not be accepted in its current form. The comparison needs to be made matched and statistically grounded.\n\nRecommendation: send to review with a request for matched baselines, significance testing, and code/data release, or at least a detailed implementation appendix.","headline":"A sound internal story—subject-level CV and clean ablations—but the central 'superiority' claim rests on cross-study comparisons that don't match on data, atlas, or sample size.","tokens_in":6742,"tokens_out":1860,"would_cite":false,"duration_ms":16678,"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 spatio-temporal transformer that jointly encodes the timing and spatial layout of dynamic brain connectivity, boosted by contrastive learning, reaches 89.1% accuracy in distinguishing mild cognitive impairment from normal controls on…","keywords":["mild cognitive impairment","dynamic functional connectivity","resting-state fMRI","spatio-temporal transformer","contrastive learning","Alzheimer's disease","brain state classification","ADNI"],"falsifier":"Re-run one of the Table 1 baselines, for example the deep spatial-temporal fusion method [24], on the exact same 345 subjects, same Destrieux atlas, same sliding-window parameters ($L=70$, $S=2$), and same subject-level 5-fold split. If that baseline's accuracy is statistically indistinguishable from or above the reported 89.1%, the claim that the proposed architecture is superior would be falsified; if it remains clearly lower, the claim would gain direct support.","tokens_in":5755,"feed_emoji":"🧠","tokens_out":15483,"duration_ms":120652,"temperature":0.7,"pith_summary":"This paper sets out to show that mild cognitive impairment (MCI), the prodromal stage of Alzheimer's disease, can be distinguished from healthy aging by modeling how brain connectivity changes over time, rather than by a single static connectivity map. The authors build a transformer-based network that reads the sequence of sliding-window functional connectivity matrices twice, once as a temporal trajectory and once as a spatial layout of regions, then fuses the two views with global attention. A contrastive learning step, in which scans from the same diagnosis are pulled together and scans from different diagnoses are pushed apart, is added to reduce dependence on labeled data. On 345 subjects and 570 scans from ADNI, the full model achieves 89.1% accuracy, 91.4% sensitivity, 87.2% specificity, and 89.3% AUC. The paper claims these results show the framework outperforms earlier static and dynamic connectivity classifiers and has clinical potential for early AD identification.","feed_headline":"Fusing time and space in brain scans hits 89.1% accuracy for MCI","feed_subtitle":"A transformer reading timing and spatial layout of brain connections outperforms earlier methods on 345 ADNI subjects.","key_machinery":"The load-bearing object is the sequence of dynamic functional connectivity matrices $M = \\{M_1,\\dots,M_T\\}$ built by sliding a window of length $L=70$ with step $S=2$ over each subject's ROI-averaged time series, with each $M_t$ holding Pearson correlations among 148 Destrieux regions. The machinery is a spatio-temporal transformer block applied twice: a temporal block reads each matrix transposed, so rows are time points, while a spatial block reads rows as regions. Each block projects its input to query, key, and value matrices, computes multi-head self-attention with $\\mathrm{softmax}(QK^\\top/\\sqrt{d})V$, applies a feed-forward network with residual connections and layer normalization, then a convolutional layer. The two views are concatenated and fused by global attention before a single-layer perceptron classifies the subject. The contrastive branch defines positive pairs as dFC networks from subjects with the same diagnosis and negative pairs as those from different diagnoses, optimizing an InfoNCE-style loss with temperature $\\tau$; the total loss is $\\alpha\\mathcal{L}_{\\mathrm{cont}} + \\beta\\mathcal{L}_{\\mathrm{CE}}$ with learnable weights.","core_discovery":"The central claim is that the temporal sequence of connectivity matrices contains information that static functional connectivity discards, and that a transformer can jointly capture both the temporal and the spatial structure of that sequence. The architecture processes each windowed connectivity matrix in two orientations: a temporal block sees the time course of each region, and a spatial block sees the region-to-region coupling within each window. Both blocks combine a multi-head self-attention transformer layer with a convolutional layer, and their outputs are concatenated and passed through global attention to form a fused whole-brain representation. Training adds a contrastive loss, defined by cosine similarity between dFC representations, to the cross-entropy classification loss. The paper reports that this full model reaches 89.1% accuracy, and that the contrastive component improves accuracy by 9.7 percentage points over the same architecture trained without it.","pith_inferences":["The paper leaves untested whether the same architecture transfers to the EMCI subgroup or to independent cohorts, since the Table 1 comparisons mix EMCI and MCI samples across different studies.","A natural next step, not reported here, is to pretrain the temporal and spatial transformers on unlabeled rs-fMRI scans with the contrastive objective and then fine-tune on labeled MCI data; the reported contrastive gain suggests a fully self-supervised stage could push accuracy further.","Because window length and step size were fixed at $L=70$, $S=2$, a reader interested in the mechanism would want a sweep of these parameters: if performance is highly sensitive to them, part of the reported superiority may reflect the window configuration rather than the architecture.","A matched head-to-head re-implementation of the Table 1 baselines on the same 345-subject, Destrieux-parcellated, subject-level 5-fold split would be the direct test of the superiority claim; the paper does not provide that experiment."],"forward_implications":["If the central claim holds, MCI screening can be performed directly from the sequence of resting-state connectivity, giving 89.1% accuracy and 91.4% sensitivity on the ADNI cohort and catching most MCI cases at the prodromal stage.","The contrastive branch, which pairs scans by diagnosis and pulls same-diagnosis representations together, is reported to add 9.7 percentage points of accuracy over the same architecture without it, implying that diagnostic pairing provides a training signal beyond the label itself.","The ablation study implies the temporal and spatial branches are complementary: spatial-only reaches 76.6% accuracy, temporal-only 70.7%, and the fused model 89.1%, so the joint encoding carries the performance.","At 87.2% specificity, roughly one in eight healthy controls would be flagged for further testing, a reasonable screening operating point if the numbers transfer to independent cohorts."],"supporting_citations":[{"why":"ADNI dataset; source of all 345 subjects and 570 rs-fMRI scans used for training and evaluation.","marker":"[16]"},{"why":"Sets the FSL FEAT preprocessing pipeline (smoothing, slice-timing correction, filtering, drift removal) applied before dFC construction.","marker":"[17]"},{"why":"Companion preprocessing reference the paper follows for resting-state fMRI preparation.","marker":"[18]"},{"why":"Destrieux atlas parcellation that defines the 148 ROIs used to compute Pearson-correlation connectivity matrices.","marker":"[19]"},{"why":"Supplies the empirically chosen sliding-window length $L=70$ and step $S=2$ used to build the dFC network sequence.","marker":"[20]"},{"why":"Source of the contrastive learning formulation whose InfoNCE-style loss is added to the cross-entropy objective.","marker":"[15]"},{"why":"Table 1 baseline (static FC, AAL atlas) that the paper claims to outperform with the reported 89.1% accuracy.","marker":"[21]"},{"why":"Table 1 baseline using dynamic FC on the AAL atlas, closest dynamic-connectivity comparison on a different parcellation.","marker":"[22]"},{"why":"Deep spatial-temporal feature fusion baseline, the most direct predecessor the proposed architecture extends and compares against.","marker":"[24]"},{"why":"Recent static-FC baseline with the highest reported AUC (87.7%) among the comparisons, the number the full model must surpass.","marker":"[25]"}],"fun_headline_variants":["Transformer decodes shifting brain networks to flag MCI","Dynamic brain connectivity + transformer hits 89.1% for MCI","Spatio-temporal transformer spots MCI from brain scan sequences","Contrastive learning sharpens MCI detection from fMRI","Time-aware brain connectome analysis reaches 89.1% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that the proposed method is superior to previous work assumes that the previously published results in Table 1 are directly comparable to the new model's numbers, even though those studies used different numbers of subjects, different MCI subtypes, a different brain atlas, and different preprocessing, and no matched experiment is reported.","fun_headline_variants_meta":{"raw":{"variants":["Transformer decodes shifting brain networks to flag MCI","Dynamic brain connectivity + transformer hits 89.1% for MCI","Spatio-temporal transformer spots MCI from brain scan sequences","Contrastive learning sharpens MCI detection from fMRI","Time-aware brain connectome analysis reaches 89.1% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000223,"raw_usage":{"total_tokens":1457,"prompt_tokens":943,"completion_tokens":514,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":428}},"tokens_in":559,"tokens_out":514,"duration_ms":4885,"temperature":1.0,"reasoning_tokens":428,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:34:28.054515+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run one of the Table 1 baselines, for example the deep spatial-temporal fusion method [24], on the exact same 345 subjects, same Destrieux atlas, same sliding-window parameters ($L=70$, $S=2$), and same subject-level 5-fold split. If that baseline's accuracy is statistically indistinguishable from or above the reported 89.1%, the claim that the proposed architecture is superior would be falsified; if it remains clearly lower, the claim would gain direct support.","supporting_citations":[],"review_version":1}