{"id":"bbb524ff-edd1-4350-a3e2-7651d50a1393","arxiv_id":"2607.07091","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":5,"one_line_summary":"A temporal-aware cross-attention framework enables structural MRI to provide complementary diagnostic value for Alzheimer's disease without degrading stronger longitudinal clinical signals.","lead":"This paper proposes a method to combine brain MRI scans with cognitive test scores over time for Alzheimer's diagnosis, using a constrained attention mechanism to prevent noisy MRI data from degrading clinical signals. It matters because it demonstrates a practical approach to integrating weak and strong medical data modalities without performance loss.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"MRI 'complementary information' claim is undermined by the gap between the clinical-only ablation and the XGB baseline using the same clinical features.","rationale":"The reader correctly identifies label leakage as a concern, and the paper itself acknowledges this in §IV-F. However, the label leakage issue is somewhat generic to ADNI studies and is already hedged by the authors. A more load-bearing and specific concern is the architectural confound: the paper attributes the performance gap between full AT-Attn and clinical-only AT-Attn to MRI complementarity, but this comparison is within a single architecture whose clinical-only performance (0.6935) is notably weaker than XGB's clinical-only performance (0.7187) on the same features. This means the 'complementary MRI information' may be compensating for AT-Attn's suboptimal clinical feature processing rather than adding genuinely new signal. The paper's claim of 'remaining competitive with strong tabular baselines' is supported, but the stronger interpretive claim—that MRI provides 'clinically relevant complementary information'—is not cleanly demonstrated because the appropriate control (a strong clinical model + MRI) is missing. The verdict remains CONDITIONAL: the architectural contribution (temporal-aware asymmetric cross-attention) is reasonably novel and ablated, but the central interpretive claim about MRI complementarity requires the proposed check to be convincing. If the check shows XGB+MRI-features ≈ XGB-only, the paper's framing should be adjusted from 'MRI provides complementary information' to 'AT-Attn's fusion strategy recovers performance comparable to strong tabular baselines.'","tokens_in":14819,"tokens_out":3268,"duration_ms":123296,"concrete_test":"Extract the pooled MRI features (h_img-direct and h_fused from §III-E) from the trained AT-Attn model and append them as additional features to the XGB tabular baseline. If XGB+MRI-features does not meaningfully outperform XGB-only (currently 0.7187 Acc), then MRI is not providing complementary information beyond what a strong clinical model already captures, and the AT-Attn ablation gap reflects architectural differences rather than genuine MRI complementarity. Conversely, if XGB+MRI-features improves by a margin comparable to the AT-Attn ablation gap (~2.6 points), the complementary-information claim is strengthened.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that MRI provides 'complementary information' beyond clinical features. The key evidence is the ablation in Table III: full AT-Attn (0.7192 Acc) vs. clinical-only matched AT-Attn (0.6935 Acc), a 2.6-point gap attributed to MRI. However, Table II shows that XGB—using the same cognitive-scale trajectories, timestamps, and static features but NO MRI—achieves 0.7187 Acc. This means: (1) AT-Attn with MRI (0.7192) ≈ XGB without MRI (0.7187), and (2) AT-Attn without MRI (0.6935) is substantially worse than XGB without MRI (0.7187). The apparent 'complementary information' from MRI may simply compensate for AT-Attn's weaker clinical-feature processing relative to XGB, rather than reflecting genuine biomarker complementarity. If AT-Attn's clinical branch were as effective as XGB at extracting diagnostic signal from the longitudinal cognitive trajectories, MRI might add nothing. The paper does not address this confound directly in §IV-B or §IV-F. The label leakage concern raised by the reader is valid but acknowledged and generic to ADNI studies; this architectural confound is more specific to the paper's central interpretive claim.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The paper proposes AT-Attn, a temporal-aware multimodal framework for longitudinal Alzheimer's disease (AD) diagnosis. The architecture combines (1) a Change-and-Time encoder that captures inter-visit feature differences and absolute temporal position, (2) time-biased asymmetric cross-attention in which MRI features query cognitive-scale features under an additive exponential temporal penalty, and (3) gated fusion with shortcut paths that preserve direct unimodal evidence. The framework is evaluated on an MRI-retained ADNI cohort of 1,520 patients with patient-level five-fold cross-validation, against unimodal baselines, naive fusion baselines, and strong tabular baselines (XGB, LGBM, etc.). The main asymmetric AT-Attn model achieves accuracy 0.719, macro F1 0.721, ROC-AUC 0.873, and PR-AUC 0.783, outperforming naive fusion baselines and remaining competitive with tree-based tabular methods. Systematic ablations (Table III), subgroup analyses (Fig. 3), attention pattern analyses (Fig. 4), and representative corrected cases (Table IV) are provided.","tokens_in":15073,"tokens_out":1441,"duration_ms":177836,"significance":"The problem of stable integration of a weaker modality (structural MRI) with a stronger clinical signal under irregular longitudinal sampling is well-motivated and practically relevant. The architectural design is principled: the asymmetric cross-attention direction (MRI queries cognitive scales, not vice versa) is a concrete and falsifiable design choice motivated by modality strength imbalance, and the ablation confirming that removing Change-and-Time encoding causes the strongest component-level degradation (Table III) supports the temporal modeling contribution. The subgroup gain analysis (Fig. 3) showing context-dependent MRI contribution—particularly larger gains when baseline diagnosis is DEM—is a clinically plausible and informative finding. The paper is transparent about the label-leakage limitation inherent to ADNI studies (§IV-F).","major_comments":[{"comment":"§IV-B, Table II vs. Table III: The central interpretive claim that MRI provides 'complementary information' beyond clinical features is confounded by the gap between the clinical-only ablation and the XGB baseline. Full AT-Attn with MRI achieves 0.7192 Acc (Table III), which is approximately equal to XGB without MRI (0.7187 Acc, Table II). Meanwhile, AT-Attn without MRI (clinical-only matched, 0.6935 Acc, Table III) is substantially worse than XGB without MRI (0.7187). This pattern is consistent with MRI compensating for AT-Attn's weaker clinical-feature processing relative to XGB, rather than reflecting genuine biomarker complementarity. The paper does not address this confound directly. A controlled comparison—e.g., adding MRI to the XGB clinical feature set, or demonstrating that AT-Attn's clinical branch matches XGB before MRI is added—would substantially strengthen the 'complementar","section":null}],"minor_comments":[{"comment":"§III-D, Eq. (8): The text states that '-exp(Δt/τ_h) explicitly down-weights temporally distant keys, whereas exp(-Δt/τ_h) mainly gives nearby visits a small positive bonus.' The two expressions are mathematically equivalent up to sign convention in the exponent; the distinction as stated is unclear. Consider rephrasing to clarify that the adopted form is a negative penalty added to pre-softmax logits.","section":null},{"comment":"§IV-A: The cohort funnel (Fig. 2A) reports 4,813 visits with matched MRI and 1,753 without, but Table I reports 5,176 preprocessed MRI scans. The difference (363 scans) is not explained. Clarify whether these scans were excluded during visit-grid matching and why.","section":null},{"comment":"Table II: The 'MCI Sens@90Spe / Achieved Spe' column header is ambiguous. Consider splitting into two columns or clarifying in the caption that the format is 'sensitivity / achieved specificity at a threshold targeting 90% specificity.'","section":null},{"comment":"§IV-D, Fig. 3A: The subgroup sizes (n=598, 682, 240) sum to 1,520, but Fig. 3D uses different subgroup sizes (n=592, 570, 358) that also sum to 1,520. The discrepancy in CN/MCI/DEM counts between panels A and D is unexplained.","section":null},{"comment":"§III-C, Eq. (2): The normalization constant 12.0 for Δt_norm is introduced without justification. A brief note that this corresponds to a 12-month reference interval would aid interpretation.","section":null},{"comment":"§IV-F: The limitation that 'ADNI clinical diagnoses are strongly tied to cognitive and functional assessments' is acknowledged, but the paper could strengthen its framing by noting this is a known issue across ADNI-based ML studies and citing prior work that has attempted to mitigate it (e.g., by excluding diagnosis-defining scales).","section":null},{"comment":"Fig. 1: The diagram is informative but dense. The 'Learnable Time Bias' and 'Time-Biased Cross-Attention' boxes appear disconnected from the main data flow arrows. Consider adding arrows to clarify the flow from the time bias module to the cross-attention computation.","section":null},{"comment":"§IV-E, Fig. 4A: The y-axis label 'Mean Attention Mass' could be clarified as 'mean attention weight' to distinguish from the pre-softmax attention logits.","section":null},{"comment":"References: Several MICCAI 2025 citations (e.g., [29], [30], [31]) appear with volume numbers but no DOI. For a journal submission, complete bibliographic information should be provided.","section":null}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the XGB vs. clinical-only AT-Attn gap is the most substantive issue. It is not that the paper is wrong—MRI may well provide complementary information—but the current experimental design cannot distinguish between 'MRI adds genuine biomarker signal' and 'MRI compensates for AT-Attn's suboptimal clinical-feature processing.' This is a fixable problem: the authors could run XGB+MRI features (e.g., pooled MRI embeddings) or demonstrate parity between AT-Attn's clinical branch and XGB before MRI addition. The label-leakage concern is real but generic to ADNI studies and already acknowledged; it should not be the primary basis for rejection. The paper is a solid contribution if this confound is addressed."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The paper proposes AT-Attn, a temporal-aware multimodal framework for longitudinal AD diagnosis. The core architectural idea is sound: use asymmetric cross-attention where MRI features query cognitive-scale features (not the reverse), add a learnable temporal bias to the attention logits, and gate the fusion so raw MRI evidence isn't overwritten. The Change-and-Time encoding avoids explicit temporal differencing, which is a practical choice for irregular intervals. The ablations are thorough, and the subgroup analysis showing MRI gains are context-dependent (larger for DEM baseline, longer follow-up) is a nice touch. The paper is honest about its limitations, which I appreciate. The label leakage concern (cognitive scales used as both features and label criteria) is real but generic to ADNI studies and acknowledged in §IV-F. The more specific issue is this: the paper claims MRI provides 'complementary information' beyond clinical features, supported by the ablation gap between full AT-Attn (0.7192 Acc) and clinical-only AT-Attn (0.6935 Acc). But XGB, using the same clinical features and no MRI, gets 0.7187. So AT-Attn with MRI barely matches XGB without MRI, and AT-Attn without MRI is substantially worse. This means the apparent MRI 'complementarity' may just be compensating for AT-Attn's weaker clinical-feature processing relative to a gradient-boosted tree. The paper doesn't address this confound. If the clinical branch were as strong as XGB, MRI might add nothing. That said, the architectural contribution itself is legitimate. The asymmetric design, temporal bias, and gated fusion are a reasonable combination for the stated problem of stable MRI integration. The components aren't novel individually, but their specific combination for this use case is. The paper would benefit from either (a) a clinical-only AT-Attn variant that matches XGB, then showing MRI still helps, or (b) a more careful framing that doesn't overclaim MRI's independent contribution. This is a solid methods paper with an interpretive overreach. It deserves a serious referee who can push on the confound and the framing.","headline":"Solid architecture for constrained multimodal fusion, but the 'MRI adds complementary information' claim is confounded by a gap between AT-Attn's clinical branch and XGB.","tokens_in":15556,"tokens_out":532,"would_cite":false,"duration_ms":71020,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"One-way cross-attention lets noisy brain scans help—not hurt—AD diagnosis","keywords":["Alzheimer's disease diagnosis","multimodal fusion","cross-attention","longitudinal modeling","structural MRI","cognitive assessment","irregular time series","asymmetric attention"],"falsifier":"If a variant that excludes the cognitive scales most directly used in ADNI diagnostic criteria (e.g., ADAS11, ADAS13) from the input features shows no MRI contribution gain over clinical-only baselines, then the MRI value demonstrated here depends on label leakage rather than independent structural information.","tokens_in":14935,"feed_emoji":"","tokens_out":1695,"duration_ms":119349,"temperature":0.7,"pith_summary":"The paper argues that in longitudinal Alzheimer's disease diagnosis, the problem is not whether to combine brain MRI with cognitive test scores, but how. MRI is high-dimensional, noisy, and often missing at some visits; cognitive scales are more directly diagnostic. When you fuse them naively—by concatenation or symmetric cross-attention—the weaker MRI signal can distort the stronger clinical signal and degrade performance. The authors propose a constrained fusion architecture called AT-Attn built around three ideas: a Change-and-Time encoder that captures how both MRI features and cognitive scores change between visits without computing unstable temporal derivatives; an asymmetric cross-attention layer where MRI features query cognitive-scale features but not the reverse, so clinical context refines MRI representations without letting noisy MRI feedback contaminate the clinical branch; and a learnable time bias that down-weights temporally distant visits in the attention computation. A gated fusion step then blends the cross-attention-enhanced MRI with the original MRI features, and shortcut paths preserve direct unimodal evidence for the final classifier. On an ADNI cohort of 1,520 patients with patient-level five-fold cross-validation, this architecture outperforms unimodal MRI, unimodal cognitive scales, and all naive fusion baselines (late fusion, concatenation, modality dropout, missing-aware gated fusion) across accuracy, macro F1, ROC-AUC, and PR-AUC. It remains competitive with strong tabular baselines like XGBoost and LightGBM. Subgroup analysis shows MRI contribution is context-dependent: gains are largest when baseline impairment is already substantial, under moderate MRI missingness, and with longer follow-up—consistent with the clinical understanding that structural MRI becomes more informative as neurodegeneration progresses.","feed_headline":"One-way cross-attention lets noisy brain scans help—not hurt—AD diagnosis","feed_subtitle":"Letting MRI query cognitive scores but not vice versa preserves clinical signal while extracting complementary imaging value from irregular,","key_machinery":"AT-Attn combines: (1) a Change-and-Time encoder computing inter-visit feature differences with normalized elapsed time, absolute temporal position, and instantaneous state, using separate parameters per modality; (2) time-biased asymmetric cross-attention where MRI features serve as queries and cognitive-scale features as keys/values, with an additive penalty of the form −exp(|Δt|/τ_h) per attention head on pre-softmax logits, where τ_h is learnable and positive; (3) gated fusion blending raw and cross-attended MRI features via a per-timestep sigmoid gate; (4) shortcut paths exposing direct unimodal pooled features alongside fused features to the classifier; (5) a 3D ResNet-18 MRI backbone (","core_discovery":"The central claim is that structural MRI provides complementary diagnostic value in longitudinal AD staging only when integrated through a constrained, temporally-aware fusion mechanism—not when added naively. The load-bearing mechanism is the asymmetric direction of cross-attention: MRI queries cognitive scales, not the reverse, which prevents the weaker modality from corrupting the stronger one while still allowing cross-modal enrichment. Combined with temporal change encoding and a learnable exponential time penalty on attention logits, this architecture extracts clinically relevant MRI signal that naive fusion cannot.","pith_inferences":["The asymmetric principle may extend to a general design rule: in any multimodal system with a dominant and a complementary modality, cross-attention should flow from dominant to complementary, never bidirectionally, unless both modalities are comparably reliable. This is not tested in the paper but is a natural extrapolation from the ablation showing symmetric cross-attention performs worse on acc","The context-dependent MRI contribution pattern (largest gains in DEM subgroup) raises the possibility that the architecture is effectively learning to weight MRI more when cognitive scores are already abnormal—which is what clinical intuition would prescribe. A system that explicitly conditions fusion strength on estimated impairment severity might achieve the same effect more transparently.","If the label-leakage concern the authors acknowledge is significant—ADNI diagnoses are partly determined by the same cognitive scales used as inputs—then the true independent biomarker value of MRI may be higher than reported, because the model's ceiling is artificially capped by labels that partially encode the input features.","The time-biased attention pattern (nearest visit accounts for only ~25% of attention peaks, distant visits ~50%) suggests the model discovers that diagnostically relevant MRI information is distributed across the trajectory rather than concentrated at the most recent scan, which has implications for how clinical visits should be scheduled and prioritized."],"forward_implications":["If the asymmetric fusion principle generalizes, any multimodal clinical setting with unequal modality strength—where one signal is noisier or less directly diagnostic—could benefit from one-way cross-attention rather than symmetric fusion, including PET-MRI, EEG-clinical, or pathology-radiology combinations.","The learnable temporal bias on attention logits offers a general mechanism for irregularly-sampled multimodal data, potentially applicable beyond AD to any longitudinal disease staging where visit spacing varies across patients.","The finding that MRI contribution is context-dependent (larger in already-impaired patients, under moderate missingness, with longer follow-up) suggests that multimodal AI systems should report subgroup-level value rather than aggregate metrics alone, as a modality that helps on average may hurt specific patient subgroups.","The shortcut-path design—preserving direct unimodal evidence alongside fused features—implies that fusion architectures should never fully replace a modality's raw representation, only augment it, when that modality carries independent diagnostic signal."],"fun_headline_variants":["Asymmetric cross-attention stops noisy MRI from corrupting AD diagnosis","MRI queries cognitive scores but not reverse, improving longitudinal AD staging","Constrained fusion lets noisy MRI add diagnostic value in Alzheimer's staging","Time-aware gated fusion extracts AD signal from MRI that naive fusion loses","One-way attention prevents weak MRI modality from degrading AD clinical scores"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that ADNI clinical diagnoses used as training labels are not largely a deterministic function of the cognitive-scale features used as inputs. The authors acknowledge this risk themselves: if diagnoses are substantially defined by the same cognitive tests fed to the model, then the model may be fitting label leakage rather than learning independent biomarker value from MRI.","fun_headline_variants_meta":{"raw":{"variants":["Asymmetric cross-attention stops noisy MRI from corrupting AD diagnosis","MRI queries cognitive scores but not reverse, improving longitudinal AD staging","Constrained fusion lets noisy MRI add diagnostic value in Alzheimer's staging","Time-aware gated fusion extracts AD signal from MRI that naive fusion loses","One-way attention prevents weak MRI modality from degrading AD clinical scores"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":645,"prompt_tokens":555,"completion_tokens":90,"prompt_tokens_details":null},"tokens_in":555,"tokens_out":90,"duration_ms":57480,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T20:11:08.259228+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If a variant that excludes the cognitive scales most directly used in ADNI diagnostic criteria (e.g., ADAS11, ADAS13) from the input features shows no MRI contribution gain over clinical-only baselines, then the MRI value demonstrated here depends on label leakage rather than independent structural information.","supporting_citations":[],"review_version":1}