{"id":"9a9d0202-c392-44ef-a341-c593c3829eb8","arxiv_id":"2607.22770","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A Transformer that injects site, scan-sequence, and modality-availability embeddings, regularized by a reference-model trust region, improves multi-center dementia subtype classification (mean macro AUC 85.62%).","lead":"This paper introduces COME, a deep-learning framework that encodes scanner protocol, data source, and missing-modality information as extra tokens to diagnose dementia subtypes from MRI, PET, and clinical data. The authors report gains over prior baselines across seven cohorts and show model scores track amyloid, tau, and severity measures.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Trust-region reference model is periodically copied from the meta-enhanced main model, so it cannot serve as the claimed bias-free anchor; the spurious-correlation-prevention mechanism is unsupported.","rationale":"The reader's weakest assumption identifies a real internal inconsistency. Section III-C defines the trust region as the space of representations learned exclusively from core clinical inputs, but Section III-D overwrites the reference model with the meta-enhanced main model every k epochs. After the first hard update, the reference's parameters are not those of a no-meta model; they have been optimized with the meta-knowledge-enhanced objective and hence can carry site/acquisition/missing-modality correlations. The global and local alignment losses then enforce proximity to this anchor, so the stated mechanism for preventing spurious correlations is not established. This is not a disagreement with consensus; it is a mismatch between the paper's conceptual claim and its implementation. I considered the alternative concern that cross-center evaluation does not specify how cohort embeddings are handled for unseen centers; that is a genuine reproducibility gap and should also be addressed, but the trust-region issue is more directly load-bearing because it targets the paper's own explanation of why the method is robust. A linear probe on the reference model's representations is a direct, low-cost check: if the reference encodes source information above chance after copies, the 'bias-free' anchor is contaminated. The paper's empirical results, including multi-level biomarker validation, are valuable independent support, but the claimed mechanism remains unsupported. The conditional verdict is appropriate; no change.","tokens_in":21230,"tokens_out":6480,"duration_ms":73720,"concrete_test":"During the constrained optimization phase, train a linear probe to predict the source cohort (or, for single-center evaluation, the acquisition-sequence cluster) from the frozen reference model's mean-pooled representations ẑG. If the probe's accuracy is significantly above chance after the first hard update (k=2), the reference model has absorbed site/sequence information from the meta-enhanced main model, confirming that the trust-region anchor is not 'free from heterogeneity-induced biases.' If accuracy remains at chance throughout training, the periodic copy does not measurably contaminate the anchor and the concern would be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism claim is that the trust-region reference model anchors the meta-enhanced model to a representation space 'learned exclusively from core clinical inputs... free from heterogeneity-induced biases' (Sec. III-C). The implementation, however, initializes θref ← θmain after the no-meta initialization phase and then refreshes θref ← θmain every k=2 epochs throughout the constrained phase (Sec. III-D). After the first refresh, θref's parameters are copies of the meta-enhanced main model that has been optimized with the full objective L = Lcls + λ(Lg + Ll). Even though the reference model receives no meta-knowledge tokens at inference, its weights have been shaped by the meta-enhanced training signal and can encode the very site/sequence/modality-missing correlations the trust region is supposed to exclude. Lg and Ll then actively pull the main model toward this contaminated anchor. Thus the described mechanism—preventing spurious correlations via an independent clinical-only anchor—does not follow from the procedure. The periodic copies may act as a useful temporal regularizer, and the reported empirical gains may still be real, but the paper's explanation for OOD robustness is unsupported as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes COME, a Transformer-based framework for multi-center dementia etiology diagnosis. It injects three types of meta-knowledge embeddings—acquisition-sequence, cohort-source, and modality-coverage—into the image and clinical feature streams, and regularizes training with a trust-region objective built on a reference model. The authors evaluate on seven cohorts for in-domain classification and on leave-one-center-out and cross-sequence settings for out-of-domain generalization, reporting state-of-the-art macro-averaged AUC with a 4.29-point gain over the strongest baseline. They also provide biomarker- and CDR-based validation of model predictions. The central mechanistic claim is that the trust-region anchor prevents spurious correlations by keeping the meta-enhanced model close to a representation space learned exclusively from core clinical inputs and free of heterogeneity-induced biases.","tokens_in":21633,"tokens_out":3041,"duration_ms":38254,"significance":"If the mechanism were supported, this would be a practically valuable contribution to multi-center medical imaging: it tackles a real deployment problem (site and modality heterogeneity) with an explicit modeling strategy and demonstrates results across seven public cohorts, including OOD evaluation. The paper's strengths include the breadth of the multi-center evaluation, the inclusion of cross-sequence generalization, and the external validation against amyloid, tau, FDG, and DaTscan biomarkers and CDR scores. The ablations and hyperparameter sensitivity analyses are useful. However, the load-bearing explanation for why the method works—the trust-region anchor being free of heterogeneity-induced biases—is not supported by the training procedure as written, and the leave-one-center-out protocol leaves a key inference detail unspecified.","major_comments":[{"comment":"The trust-region anchor is not independently trained. The reference model is initialized from the no-meta phase, but then θref ← θmain every k=2 epochs throughout the constrained phase. After the first refresh, θref is a lagged copy of the meta-enhanced main model optimized with L = Lcls + λ(Lg+Ll). Thus the reference model's weights have been shaped by meta-knowledge gradients, even though it receives no meta tokens at inference. The claim in Sec. III-C that the trust region is 'the space of representations learned exclusively from core clinical inputs... free from heterogeneity-induced biases' is therefore not established. The reported gains may arise from temporal averaging or other effects, but the spurious-correlation-prevention mechanism is unsupported. Please retrain an independent reference model with Lcls only (never overwritten by the main model) and compare, or revise the mech","section":"Sec. III-C and III-D, Eq. (11)"},{"comment":"The cross-center evaluation is leave-one-center-out, but cohort-aware embeddings P^s_cohort are trainable per source. For a held-out center, no such embedding was trained. The paper does not state how the model obtains a cohort embedding for the held-out center at inference—whether it is omitted, averaged over training sources, or set to a learned default. This choice changes the input distribution and directly affects the validity of the reported OOD results. Specify the exact protocol and, if an embedding is synthesized, justify it.","section":"Sec. IV-F1 and Sec. III-B1b"},{"comment":"The 'No Enhancement' ablation is described as meta information not used during inference. But the model in that ablation was trained with meta-knowledge tokens and trust-region constraints; removing the tokens only at inference changes the input distribution and may not reflect what a model trained without meta-knowledge would learn. To support the claim that meta-knowledge injection itself improves robustness, report a variant trained without meta-knowledge embeddings from the start, alongside the inference-only removal variant.","section":"Table VI and Sec. IV-G1"}],"minor_comments":[{"comment":"The caption contains garbled text: 'AℬTa u' should be 'Aβ and tau'. Please fix.","section":"Fig. 6 caption"},{"comment":"'Y a Zhang' has an extra space; should be 'Ya Zhang'.","section":"Author affiliations"},{"comment":"N is introduced as the patch token number after fusion, but the notation is overloaded: earlier N_p was used for image patches. Please clarify the dimensions of the fused sequence.","section":"Sec. III-B2, Eq. (5)"},{"comment":"The statistical test is a paired two-sided t-test over five seeds. Given only five seeds, a paired permutation test or reporting effect sizes with confidence intervals would strengthen the claim.","section":"Sec. IV-E"}],"recommendation":"major_revision","confidential_remarks":"The core issue is the unsupported trust-region anchor claim. If the authors can provide an experiment with an independently trained reference model and clarify the held-out cohort embedding protocol, the paper could become publishable. The empirical results are promising, but the current manuscript overstates the mechanism."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front: the empirical results are worth taking seriously, and the paper's mechanistic explanation for its out-of-domain gains does not hold as written. That makes it a conditional accept for me.\n\nWhat's new: COME injects acquisition semantics (standardized offline by an LLM), cohort source identifiers, and modality-coverage indicators as tokens into a Transformer, and regularizes the meta-enhanced model with a reference model. The specific combination is new in the cited literature. The authors evaluate on seven public cohorts with careful ablations, and the external biomarker and CDR validation is genuinely useful. The representation analysis showing that acquisition tokens cluster into meaningful protocol groups is a nice sanity check. The engineering and evaluation effort is solid.\n\nWhere it gets soft: the trust-region anchor. The paper defines the trust-region as representations learned exclusively from core clinical inputs, free from heterogeneity-induced biases. But the reference model is initialized from the main model after the no-meta phase and then refreshed every k=2 epochs with the meta-enhanced main model's parameters. After the first refresh, it is a lagged copy of the model optimized with the full objective. It does not receive meta tokens at inference, but its weights have been shaped by them. So it cannot serve as the independent bias-free anchor the text claims. The periodic copying might act as a useful temporal regularizer, and the reported gains may still hold, but the spurious-correlation-prevention mechanism is unsupported as described. The stress-test note is right.\n\nSecond gap: the leave-one-center-out OOD evaluation never says how cohort embeddings are obtained for a held-out center. The model has no embedding for that source. Zeroing it out, using a learned default, or something else matters for the OOD claim. That's a small fix but load-bearing.\n\nFree parameters are handled reasonably; they test λ and k and show moderate sensitivity. No code or data release, but the public cohorts and harmonization pipeline make reproducibility plausible if they ship scripts.\n\nBottom line: send it to peer review. The reviewers should press for a truly independent reference model or an honest description of the constraint as a temporal regularizer, and for a clear specification of the OOD embedding handling. If the authors resolve those, this would be a solid TMI paper.","headline":"A serious multi-center dementia diagnosis paper with strong empirical results, but the trust-region anchor is self-referential and the OOD protocol has an unspecified gap; send to review.","tokens_in":22065,"tokens_out":3180,"would_cite":true,"duration_ms":33306,"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 single diagnostic model can handle the heterogeneity of seven dementia cohorts by explicitly encoding scan protocol, cohort source, and modality availability as learned tokens, regularized by a trust-region reference model.","keywords":["dementia etiology diagnosis","multi-center learning","data heterogeneity","meta-knowledge injection","trust-region optimization","spurious correlation","multimodal fusion","Transformer"],"falsifier":"Keep the reference model frozen at the end of the initialization phase (no periodic θref ← θmain refreshes) and compare cross-center AUC to the default schedule; if performance does not drop, the periodic refresh is not what prevents spurious correlations, and if the reference and main representations become near-identical, the constraint is vacuous.","tokens_in":21150,"feed_emoji":"🧠","tokens_out":5963,"duration_ms":57458,"temperature":0.7,"pith_summary":"This paper tries to show that a single diagnostic model can handle the messy reality of multi-center dementia data—where scanners, scan protocols, population mixes, and even which modalities are available differ from site to site—if the model is told, explicitly, what kind of site and scan it is looking at. The proposed framework, COME, encodes acquisition-protocol descriptions, cohort identifiers, and modality-presence indicators as learned tokens and injects them into a Transformer, then uses a reference model trained without those tokens to keep the enhanced model from latching onto site-specific shortcuts. On seven independent cohorts the method reports a mean macro-averaged AUC of 85.62%, a 4.29-point gain over the strongest baseline, and better cross-center and cross-sequence generalization. The authors argue this makes real-world, explainable dementia diagnostics more feasible.","feed_headline":"Dementia diagnosis with site and scan tokens hits 85.6% mean AUC","feed_subtitle":"A single model told which scanner and site it sees outperforms baselines on seven cohorts and generalizes to unseen sites.","key_machinery":"Meta-knowledge enhancement: three types of heterogeneity-aware embeddings—acquisition-aware (LLM-standardized sequence descriptions encoded by a medical text model), cohort-aware (trainable per-source tokens), and modality-coverage (present/missing tokens per modality)—are injected into a shared Transformer via cross-attention to refine image features before multimodal fusion. The trust-region constraint is the counterweight: a reference model without these embeddings, refreshed every k epochs, anchors dual-level alignment—a minibatch contrastive loss on pooled patient representations (global) and a cosine-affinity MSE on patch self-similarity matrices (local)—preventing the meta-injection f","core_discovery":"COME claims that the obstacle to scale-up multi-center dementia diagnosis is not simply lack of data but unmodeled heterogeneity: acquisition protocols, cohort demographics, and modality coverage encode non-biological variation that pooled models tend to memorize. The paper's solution is to make that heterogeneity explicit. Acquisition sequence names are standardized by a large language model into technical descriptions, then encoded by a medical text encoder; cohort source and modality presence/absence become trainable tokens. These meta-knowledge embeddings are injected through cross-attention refinement layers so image features are modulated conditioned on the source and protocol. To prev","pith_inferences":["If the trust-region anchor is periodically overwritten with the enhanced model's weights, the reference may drift toward the same spurious correlations the constraint is meant to remove; the paper does not compare against a never-refreshed reference, so the mechanism's causal contribution is not fully isolated.","The method's reliance on an LLM to standardize sequence names adds an offline step whose errors (if any) would propagate; the paper's radiologist check is a one-time audit, not a scalable guarantee for new sites.","The approach suggests a general recipe for heterogeneity in medical imaging beyond dementia: any multi-center pooled training could benefit from explicit acquisition and cohort tokens plus an unenhanced anchor, provided the anchor stays clean.","A natural testable extension: vary the refresh interval k and reference initialization to see at what point the constraint becomes vacuous (reference ≈ main) and OOD performance degrades."],"forward_implications":["Pooling multi-center dementia data with explicit protocol/source/modality tokens yields better in-domain diagnosis than single-center or multi-task baselines, especially for rare etiologies and small cohorts.","The gain transfers to unseen centers and unseen MRI sequences, so models trained this way are more likely to work when deployed at a new hospital.","Constraining meta-injection with the reference model is necessary: unconstrained injection can hurt, so the regularization, not just the extra information, drives robustness.","Model predictions track established biomarkers and clinical severity (CDR, amyloid/tau), suggesting the learned representations are biologically meaningful, not just statistically tuned.","The two-phase training schedule (diagnosis-first anchor, then constrained enhancement) is important for the benefit; removing or altering it degrades performance."],"fun_headline_variants":["Site and scan tokens boost dementia diagnosis by 4.3 AUC","Explicit scanner site tokens sharpen dementia diagnosis","Meta-knowledge encoding yields robust dementia diagnosis across centers","Site-aware AI model generalizes across dementia cohorts","Multi-center dementia diagnosis hits 85.6% mean AUC with meta-tags"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The trust-region reference model is assumed to represent only core clinical signals, but because it is periodically overwritten with the main model's parameters every k epochs, after the first update it is a lagged copy of the meta-enhanced model rather than an independent meta-free anchor; if that anchor already contains site- and modality-specific biases, the constraint cannot remove them.","fun_headline_variants_meta":{"raw":{"variants":["Site and scan tokens boost dementia diagnosis by 4.3 AUC","Explicit scanner site tokens sharpen dementia diagnosis","Meta-knowledge encoding yields robust dementia diagnosis across centers","Site-aware AI model generalizes across dementia cohorts","Multi-center dementia diagnosis hits 85.6% mean AUC with meta-tags"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001264,"raw_usage":{"total_tokens":5022,"prompt_tokens":761,"completion_tokens":4261,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":4179}},"tokens_in":505,"tokens_out":4261,"duration_ms":26676,"temperature":1.0,"reasoning_tokens":4179,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T06:11:06.433838+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Keep the reference model frozen at the end of the initialization phase (no periodic θref ← θmain refreshes) and compare cross-center AUC to the default schedule; if performance does not drop, the periodic refresh is not what prevents spurious correlations, and if the reference and main representations become near-identical, the constraint is vacuous.","supporting_citations":[],"review_version":1}