{"id":"9b96fa09-143a-4967-94c5-f0d43eab3814","arxiv_id":"2506.05428","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A diffusion model with LLM-guided scoring synthesizes future sMRI features from baseline scans and reports 5-12% better early MCI conversion accuracy than baselines.","lead":"MCI-Diff generates synthetic future brain MRI features from a single baseline scan using diffusion models and an LLM scorer, then uses them to predict whether mild cognitive impairment will convert to Alzheimer's. It reports large accuracy gains on ADNI and AIBL, but the evaluation lacks code, error bars, and an independent check of the LLM's clinical plausibility.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The LLM 'clinical plausibility' criterion is undefined and unvalidated against real follow-up scans, so the central claim that MCI-Diff generates clinically plausible future sMRI representations is not yet supported.","rationale":"Table 2 shows a 5-6% absolute gain over the best baseline, while the w/o LLM-Guidance ablation loses 8 points on ADNI, larger than any baseline gap. This makes the LLM selector the single highest-leverage component in the reported result. The selection rule is the only mechanism that connects 'clinically plausible' to the generated representations, yet it is specified only as 'compared to expected values' with no definition, no FreeSurfer prediction accuracy, and no validation against observed follow-up data. The reader's conditional verdict already flags this as the weakest assumption; I agree. Other concerns (no code/data, no error bars, test-set hyperparameter selection, and the abstract's 12% claim versus Table 2's 5-6% gains) are real but secondary: they affect confidence in the numbers, whereas the undefined plausibility criterion affects whether the central claim means what it says. The proposed test, comparing LLM-selected candidates with actual follow-up features and replacing the LLM with a mean-distance ranker, would settle whether the selection is steering toward true future representations or merely toward a stereotyped training distribution. If the concern lands, the paper should be revised to claim only that LLM guidance improves classification via regularization, not that it generates clinically plausible future sMRI representations.","tokens_in":957,"tokens_out":913,"duration_ms":79812,"concrete_test":"Take a held-out set of patients with observed follow-up sMRI features. For each patient, generate N candidate features at each future time point with Algorithm 2, score them with the fine-tuned LLM, and compute each candidate's distance to the patient's actual follow-up feature/FreeSurfer measurements. Compare the LLM-selected candidate's distance with the median candidate distance and with the baseline feature's distance. If the selected candidate is not closer to the true future than the median candidate or the baseline, the 'linguistic compass' is not steering toward the actual future. As a second check, replace the LLM scorer with a trivial rule that ranks candidates by closeness of LLM-predicted FreeSurfer values to the training-set mean progression; if ADNI/AIBL ACC stays near 0.95/0.94, the clinical-language component is not the source of the gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing condition for the central claim is that the LLM guidance loop (Sec. 3.2, Algorithm 2) actually selects clinically plausible future sMRI representations. The paper never defines 'expected values' in the selection rule 'most clinically plausible measurements (compared to expected values),' and it reports no accuracy for the LLM's FreeSurfer prediction task. The LLM is instruction-tuned on paired tokenized HFCN features and FreeSurfer measurements from the same training scans; if its scoring reduces to ranking candidates by proximity to the training-set-conditional mean trajectory, then the selected candidate is the most stereotyped, not the one closest to the individual's true 6-36 month anatomy. Since the w/o LLM-Guidance ablation drops ADNI ACC from 0.950 to 0.870, most of the reported gain is carried by this unvalidated selector. The ablation shows the component changes the result, but not that it provides clinical plausibility; a shrinkage regularizer or a distance-to-mean criterion could produce the same accuracy gain. Without comparing selected candidates with actual follow-up scans, the central claim that generated features are clinically plausible future representations is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"MCI-Diff is a diffusion-based framework that generates latent-space representations of future sMRI scans (6-36 months) from a single baseline scan, with the goal of early prediction of MCI-to-AD conversion. The method combines two ideas: (i) a multi-task sequence reconstruction training scheme in which a shared denoising network learns interpolation and extrapolation tasks under a progressive difficulty schedule to handle irregular follow-up sampling, and (ii) an LLM-driven 'clinical plausibility sampling' loop in which candidate future feature vectors are quantized, tokenized, and scored by a fine-tuned language model that predicts FreeSurfer-type structural measurements, with the most plausible candidate selected at each autoregressive step. A downstream classifier is then applied to baseline plus generated features to predict pMCI versus sMCI. Evaluations on ADNI (trained on ADNI-1, tested on ADNI-2) and AIBL report ACC of 0.950 and 0.936, respectively, and the abstract claims a 5-12% improvement over state-of-the-art baselines.","tokens_in":11602,"tokens_out":13945,"duration_ms":136918,"significance":"The paper addresses a genuine and clearly framed trade-off between prediction immediacy and accuracy, and the central idea of using a language model as a scorer of clinical plausibility for generated imaging features is novel and potentially influential if validated. The multi-task interpolation/extrapolation schedule is a sensible response to irregular longitudinal sampling, and the authors deserve credit for evaluating on two independent cohorts and for reporting component ablations. It should also be acknowledged that the accuracy evaluation uses external conversion labels, so the main result is not definitionally circular; the concerns below are evidential rather than definitional. However, significance is conditional on three unmet supports: statistical grounding of the headline numbers (no error bars or significance tests, hyperparameters selected on the test set), validation of the LLM plausibility scorer against real follow-up anatomy, and a complete specification of the prediction pipeline. Until these are supplied, the central claims of clinical plausibility and of a 5-12% improvement are not yet established.","major_comments":[{"comment":"The selection rule in §3.2.2 ('the candidate with the most clinically plausible measurements (compared to expected values)') is not operationalized anywhere in the paper. Algorithm 2 calls PlausibilityScore but never specifies how the score is computed, and the 'expected values' of structural biomarkers are never defined. The manuscript also reports no accuracy for the LLM's FreeSurfer-prediction task, so there is no evidence that the LLM's ratings are informative; the only support for the central 'clinically plausible' claim is the qualitative case study in §4.5, which provides no quantitative metric. This matters because the w/o LLM-Guidance ablation (§4.3, Table 3) drops ADNI ACC from 0.950 to 0.870, so most of the reported gain is carried by this unvalidated selector. Please (a) define how the expected values are obtained, (b) report the LLM's prediction quality on a held-out set, and (c) validate the selected candidates against actual follow-up data, for example by computing distances between generated and real follow-up features or between LLM-predicted and real follow-up FreeSurfer measurements.","section":"§3.2.2 / Algorithm 2"},{"comment":"The hyperparameters T=40, Dmax=4, and N=20 are chosen from Fig. 4, which plots accuracy on the ADNI test set (ADNI-2, under the protocol described in §4.1); no validation split is mentioned. The reported test ACC of 0.950 is therefore an in-sample optimized number. In addition, Tables 2 and 3 report single numbers with no error bars, no number of random seeds, and no significance tests, and the AIBL cohort contains only 17 pMCI subjects, so the AIBL sensitivity of 0.882 rests on very few positive cases. Please report means and standard deviations over multiple runs, apply significance testing, and either tune hyperparameters on a validation split or explicitly state that the reported test-set figures are optimistic bounds from test-set selection.","section":"§4.4 / Tables 2-3"},{"comment":"The improvement statistics are inconsistent across the abstract, §4.2, and Table 2. The 'Improvement (%)' row of Table 2 mixes reference baselines: most entries are percentage-point gains over the best baseline (e.g., ADNI ACC 0.950 vs HFCN+ 0.899 = +5.1), but the AIBL AUC entry (+9.6) is computed against VGG-TS (0.914 vs 0.818), while the best baseline AUC (HFCN+, 0.853) would give only +6.1 percentage points. The abstract and conclusion claim a 5-12% improvement in 'early conversion accuracy', but the ACC columns support at most about a 7% relative improvement (AIBL: 0.936 vs 0.873); the 'up to 12%' figure comes from the AUC column and from §4.2's relative-improvement wording ('5.8-11.7%'), which uses a different definition than the table. Please adopt one metric convention and one reference baseline throughout and adjust the abstract and conclusion accordingly.","section":"Table 2 / §4.2 / Abstract"},{"comment":"The prediction model behind Eq. (3) is never specified. The paper does not describe the classifier used to compute P r(Y | φ(X_0), {Ẑ}): its architecture, how it is trained, whether it uses real, generated, or combined features, or how it handles the two datasets. Furthermore, §4.1 states that 'specific implementation details such as batch sizes, diffusion steps, and hardware configurations will be provided in the Appendix', but the manuscript contains no appendix, and no code is released; details such as the base LLM, the quantization and tokenization scheme, and the adapter are absent. The central numerical results are therefore not reproducible from the manuscript as it stands.","section":"§2.2 / §4.1"}],"minor_comments":[{"comment":"The heading 'Abalation Study' should read 'Ablation Study'.","section":"§4.3"},{"comment":"Equation (6) writes the loss with ε_θ(x_t, T_i, t) and omits the condition C, whereas Eqs. (5) and (7) include C; the notation should be made consistent.","section":"§3.1.1"},{"comment":"In Eq. (2), the noise variables z_0 and z_τ are introduced without explanation and are indexed differently from the diffusion step t used in Eqs. (4)-(6).","section":"§2.2"},{"comment":"References [26] and [27] are the same paper (Thompson et al., NeuroImage, 2005); one should be removed.","section":"References"},{"comment":"The citation '[29, 1]' for the ADNI dataset is incorrect: [29] is the Transformer paper and [1] is an MCI transfer-learning paper; neither is the ADNI data reference.","section":"§1"},{"comment":"Table 2 lists VAE as a generative baseline, but the baselines paragraph in §4.1 does not mention VAE.","section":"§4.1 / Table 2"},{"comment":"The horizontal axis label in Fig. 4(c) contains a typo ('Candiddate Set Size').","section":"Fig. 4"},{"comment":"The symbol T denotes both the set of follow-up time points in §2.1 and the number of diffusion steps in §4.4; please disambiguate.","section":"§2.1 / §4.4"},{"comment":"Because Fig. 5 is the only direct visual evidence for alignment with the real data distribution, please add a quantitative distance or similarity metric to support the qualitative claim.","section":"§4.5"}],"recommendation":"major_revision","confidential_remarks":"The most original component, the LLM guidance loop, is also the least validated, and the headline numbers appear to be tuned on the test set. For a revise-and-resubmit, I would require validation-split hyperparameter selection, error bars and significance tests, a direct check of selected candidates against actual follow-up FreeSurfer/sMRI data, and a full description of the classifier and the missing appendix details. I would also ask the authors to reconcile the improvement statistics and to verify the 'first to leverage diffusion models' claim and the ADNI citation. The paper has a likeable and timely idea, but the empirical case is not yet at the standard required for the claims made."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Briefly: the paper is worth a reader's time but not yet evidence for its central claim. The actual new idea — a multi-task diffusion model that learns interpolation and extrapolation over latent sMRI features, paired with an LLM scoring step to pick among candidate future representations — is a reasonable and novel combination. The ablations show each component contributes on ADNI and AIBL, and the goal of generating future representations from a single baseline to get longitudinal-level accuracy is a sensible touchstone. This is a method-level contribution, not a paradigm shift.\n\nThe soft spots are real. There is no code, no data, no error bars, no significance tests. Hyperparameters T, d, N are selected from the ADNI test set (Fig. 4), so the headline numbers are partly the result of tuning on the evaluation set. The abstract says 5–12% improvement, but Table 2 shows at most ~9.6% improvement (AIBL AUC); the 12% figure appears nowhere in the results. The bigger gap is the 'clinical plausibility' claim. The LLM scorer is fine-tuned on the same training data, and the selection criterion 'compared to expected values' is never defined. There is no comparison of generated features to actual follow-up scans, so the paper has not shown that the selected candidates are clinically plausible — only that they help a classifier. That is a useful property, but it is not the claimed property.\n\nNone of this is fatal to the method's potential. The idea deserves a serious referee, and the paper should be sent out. But a responsible referee should ask for code or a very detailed appendix, statistics (CIs, bootstrap, repeated runs), and a direct evaluation of generated features against real follow-up data for at least a subset of participants. The authors should also reconcile the abstract's 12% claim with the reported table.","headline":"A novel diffusion-plus-LLM pipeline for MCI conversion from a single baseline scan, but the clinical-plausibility claim is unvalidated and the reported gains lack statistics and code.","tokens_in":12109,"tokens_out":2406,"would_cite":false,"duration_ms":28570,"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 diffusion model that generates future brain representations from a single baseline scan predicts early MCI conversion with superior accuracy.","keywords":["mild cognitive impairment","MCI conversion prediction","diffusion model","sMRI","longitudinal imaging","LLM-guided sampling","clinical plausibility","Alzheimer's disease"],"falsifier":"On the held-out test set, compute the correlation between the LLM's predicted FreeSurfer volumes and the true FreeSurfer volumes from the follow-up scans; if the correlation is near zero, the linguistic compass is not measuring anything clinically real, and a control condition that randomly picks among the N candidates should recover the reported accuracy if the LLM is not doing the work.","tokens_in":11158,"feed_emoji":"🧠","tokens_out":3415,"duration_ms":41322,"temperature":0.7,"pith_summary":"This paper tries to break the long-standing trade-off in mild cognitive impairment (MCI) prognosis between making fast predictions from a single baseline brain scan and making accurate predictions by waiting for longitudinal follow-up scans. The proposed framework, MCI-Diff, synthesizes plausible future sMRI feature representations directly from the baseline scan, then feeds them to a classifier, so that a clinician can get longitudinal-grade prediction at the moment of the first scan. On ADNI and AIBL data, the authors report accuracy of 0.950 and 0.936 respectively, beating eight baselines by 5 to 12 percent. The core claim is that generation of future representations, steered by a language-model-based clinical plausibility scorer, can replace actually waiting for future scans.","feed_headline":"Future brain states from one scan predict MCI conversion at 95%","feed_subtitle":"A diffusion model with an LLM 'linguistic compass' beats methods that must wait months for follow-up scans.","key_machinery":"The central object is MCI-Diff, a conditional denoising diffusion probabilistic model whose denoiser takes masked sMRI feature representations, positional embeddings, and mask embeddings as conditions, and predicts the feature at a target timepoint. Its two load-bearing mechanisms are multi-task sequence reconstruction training (interpolation and extrapolation, progressively augmented with self-imputed data) and LLM-driven clinical plausibility sampling: generated feature candidates are quantized and tokenized, the fine-tuned LLM predicts FreeSurfer volumetric and thickness measurements from them, and the candidate whose measurements are most plausible relative to expected MCI progression is selected for the next autoregressive step. This selection loop is what the paper calls the linguistic compass, and it is the component that prevents error compounding in sequential generation.","core_discovery":"The central claim is that MCI-Diff, by autoregressively generating clinically plausible future sMRI feature representations from a single baseline scan, achieves both high prediction immediacy and accuracy that rivals or exceeds longitudinal models. The method trains a shared denoising diffusion network on both interpolation and extrapolation tasks to handle irregular follow-up sampling, and then uses a fine-tuned LLM to score candidate generated features by their predicted FreeSurfer structural measurements against expected MCI progression. The best-scoring candidate is selected at each autoregressive step, acting as a \"linguistic compass\" that steers generation toward realistic disease trajectories. On the ADNI and AIBL cohorts the full model reaches ACC of 0.950 and 0.936, outperforming cross-sectional, longitudinal, and generative baselines, with the ablation showing that removing LLM guidance drops ADNI accuracy from 0.950 to 0.870.","pith_inferences":["If the LLM's plausibility ratings are the true driver of the gain, a much smaller regression model trained directly on the same FreeSurfer targets might match the performance at a fraction of the compute; the paper does not test this.","The framework should transfer naturally to other progressive neurological conditions with serial imaging, provided a structural-biomarker prior similar to FreeSurfer exists for that disease.","A direct clinical extension the paper does not simulate is triaging MCI patients at baseline for trial enrichment, where the 95% accuracy would need to be weighed against the cost of false positives and negatives in real decision-making.","A testable stress test is to swap the fixed feature extractor (HFCN) for a different pretrained encoder and check whether the reported margin over baselines persists, since the quality of the feature space may be partly responsible for the gains."],"forward_implications":["Early MCI conversion can be predicted from a single baseline sMRI with accuracy on par with or better than models that require 6-36 months of follow-up scans.","The 5-12% improvement over eight baselines on ADNI and AIBL indicates the approach generalizes across cohorts, scanners, and imaging protocols.","Removing the LLM guidance drops ADNI accuracy from 0.950 to 0.870, so the language-model plausibility scorer is a major contributor, not a minor add-on.","Multi-task interpolation and extrapolation training makes the model robust to missing follow-up visits, which are common in real longitudinal studies.","Generating future representations from baseline data is a viable alternative to collecting additional scans for prognosis, shortening the time to actionable prediction."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model backbone on which MCI-Diff's forward and reverse processes are built.","marker":"[11]"},{"why":"Supplies the hierarchical fully convolutional network used as the pretrained feature extractor and as a strong cross-sectional baseline.","marker":"[16]"},{"why":"Provides the ADNI-1 train / ADNI-2 test protocol followed in the experiments and a dual-attention cross-sectional baseline.","marker":"[31]"},{"why":"RNN-based longitudinal analysis baseline that MCI-Diff is compared against.","marker":"[4]"},{"why":"VGG-TSwinFormer longitudinal model used as a state-of-the-art longitudinal baseline.","marker":"[12]"},{"why":"Temp-GAN generative baseline for temporal MCI conversion prediction that MCI-Diff is compared against.","marker":"[30]"}],"fun_headline_variants":["One baseline MRI, diffusion + LLM compass predict MCI at 95%","LLM-steered diffuser forecasts brain changes to spot early MCI","MCI-Diff: single scan, future brain states, 95% conversion accuracy","One MRI to predict MCI conversion: diffusion with LLM compass","LLM-guided diffusion predicts future brain states from a single MRI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gain from LLM guidance rests on the fine-tuned language model's plausibility ratings actually tracking clinical realism; if those ratings are noisy or biased, the accuracy advantage would shrink or vanish.","fun_headline_variants_meta":{"raw":{"variants":["One baseline MRI, diffusion + LLM compass predict MCI at 95%","LLM-steered diffuser forecasts brain changes to spot early MCI","MCI-Diff: single scan, future brain states, 95% conversion accuracy","One MRI to predict MCI conversion: diffusion with LLM compass","LLM-guided diffusion predicts future brain states from a single MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001408,"raw_usage":{"total_tokens":5674,"prompt_tokens":912,"completion_tokens":4762,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":4664}},"tokens_in":528,"tokens_out":4762,"duration_ms":35268,"temperature":1.0,"reasoning_tokens":4664,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:35:39.834858+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the held-out test set, compute the correlation between the LLM's predicted FreeSurfer volumes and the true FreeSurfer volumes from the follow-up scans; if the correlation is near zero, the linguistic compass is not measuring anything clinically real, and a control condition that randomly picks among the N candidates should recover the reported accuracy if the LLM is not doing the work.","supporting_citations":[{"cited_title":"Hierarchical fully convolutional network for joint atrophy localization and alzheimer’s disease diagnosis using structural mri","cited_arxiv_id":null,"evidence_quote":"Supplies the hierarchical fully convolutional network used as the pretrained feature extractor and as a strong cross-sectional baseline."},{"cited_title":"Dual attention multi-instance deep learning for alzheimer’s disease diagnosis with structural mri","cited_arxiv_id":null,"evidence_quote":"Provides the ADNI-1 train / ADNI-2 test protocol followed in the experiments and a dual-attention cross-sectional baseline."},{"cited_title":"Rnn-based longitudinal analysis for diagnosis of alzheimer’s disease","cited_arxiv_id":null,"evidence_quote":"RNN-based longitudinal analysis baseline that MCI-Diff is compared against."},{"cited_title":"Vgg-tswinformer: Transformer-based deep learning model for early alzheimer’s disease prediction","cited_arxiv_id":null,"evidence_quote":"VGG-TSwinFormer longitudinal model used as a state-of-the-art longitudinal baseline."},{"cited_title":"Temporal correlation struc- ture learning for mci conversion prediction","cited_arxiv_id":null,"evidence_quote":"Temp-GAN generative baseline for temporal MCI conversion prediction that MCI-Diff is compared against."}],"review_version":1}