acceptodds
Under review as a conference paper at ICLR 2027

ADWM: A Latent World Model for Modeling Alzheimer’s Disease Progression from Functional Imaging

Abstract

Understanding dynamic changes in brain function is essential for characterizing disease progression in Alzheimer’s disease (AD). As a volumetric functional imaging modality, positron emission tomography (PET) imaging provides a direct view of cerebral metabolism and can capture disease-related metabolic alterations across brain regions. To model these evolving patterns, we introduce ADWM, a latent world model that learns the dynamics of dementia progression from sparse longitudinal observations. ADWM represents the PET modality and standardized uptake value ratio (SUVR) values in an aligned latent space, organized by a weakly supervised disease progression index (DPI). A shared neural ordinary differential equation (ODE) models population-level latent evolution along the DPI, while a separate nonnegative progression rate links disease stage to chronological time under longitudinal supervision. Conditioned on the initial state, ODE prediction, patient context, and visit-level treatment information, a residual diffusion model learns the distribution of individual latent changes beyond the shared progression trajectory. By combining a population-level progression prior with individual-level stochastic refinement, ADWN enables conditional simulation of latent disease states and regional metabolic profiles from sparse observations. This framework provides a functional-imaging-based world model for characterizing both shared and individualized AD progression dynamics, offering a unified approach to simulating disease evolution and its associated metabolic changes.

Primary area:
applications to neuroscience & cognitive science
Keywords:
PET imaging, Alzheimer's Disease Progression, World Model, Latent Dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_yWYcOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper operates on frozen joint PET–SUVR embeddings, organizes a neural ODE using a clinically generated LLM-teacher disease progression index, and refines its endpoints with conditional residual flow matching. Source state, patient context, and treatment scenarios condition the regional SUVR output. The paper also claims a nonnegative progression-rate link to calendar time, although the central presented inference algorithm requires a supplied target DPI.

Critical strengths
  1. The shared-dynamics decomposition is interesting. Besides, frozen encoders and patient-level splitting help isolate transition modeling. The explicit distinction between observational treatment conditioning and causal treatment effects is suitable in this case.
  2. Internal ablations show a meaningful state-prediction signal. Showing the unfavorable cognitive trade-off makes the empirical account more informative than an unqualified claim of improvement .
Critical weaknesses

1.The prediction setup is unclear, especially how the target disease stage is obtained at test time. If it relies on future-visit information, the results do not establish forecasting from current information alone. 2.The advance over existing Neural ODE and generative disease progression models is not convincing. The ablations and baseline comparisons do not clearly establish why each component is needed, and imaging gains do not translate into consistent cognitive prediction gains.

Questions for authors
  1. How is the target disease stage determined at test time? How does the model perform with only current-visit information and a forecast horizon?
  2. What does the ODE-plus-residual design add over direct conditional generation under matched conditions?
  3. Are the main gains consistent across patients and forecast horizons?
2 comments
AuthorsOpenReview ↗

Serious Concerns Regarding Review Accuracy, Technical Understanding, and Evaluation Fairness

We respectfully express our concern that several questions raised in this review appear to be based on interpretations that do not accurately reflect the methodology, objectives, and technical contributions presented in our manuscript. The relevant technical details and explanations have already been provided in the paper, making it difficult for us to identify the specific deficiencies or unresolved concerns underlying these questions.

In particular, regarding the three questions concerning (1) target disease stage determination at test time, (2) the contribution of the ODE-plus-residual design over direct conditional generation, and (3) performance consistency across patients and forecast horizons, we find that these questions are not sufficiently connected to the specific methodological formulations and experimental settings described in our work. We have already clarified the corresponding modeling objectives, framework design, and evaluation methodology in the manuscript. We would therefore appreciate more specific references to the aspects that the reviewer considers unclear or insufficiently justified, so that we can provide meaningful technical responses rather than reiterating explanations already presented in the paper.

Furthermore, we notice an apparent inconsistency between the reviewer's second question and their own assessment in the Strengths section. The reviewer has already acknowledged the methodological advantages and contributions of our proposed framework, yet subsequently questions what the ODE-plus-residual design contributes. While we welcome further discussion of its effectiveness, we would appreciate clarification regarding which specific contributions the reviewer considers insufficiently demonstrated, particularly in light of their earlier recognition of the framework's strengths.

More importantly, we are concerned that the review may not fully reflect the actual technical content and intended contributions of our work. We respectfully ask the reviewer to carefully revisit the relevant sections and evaluate our approach based on the methodology and experimental evidence presented, rather than assumptions that may not apply to our framework. Given the importance of a rigorous and informed peer-review process, we believe that the assessment and assigned score should be grounded in a careful and accurate examination of the paper's technical content, modeling objectives, and contributions.

We sincerely hope that the reviewer will reconsider their evaluation and assigned score after reviewing the relevant methodological details and clarifications. We remain open to addressing specific technical concerns and welcome constructive discussion grounded in the actual content of our manuscript.

Reviewer_yWYcOpenReview ↗

reply

I thank the reply. However, I still remain concerns about the evidence for the method’s added value and the prediction setup so I remain the initial decision.

Reviewer_24u8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

ADWM combines frozen PET–SUVR embeddings, an LLM-derived disease-stage index, a neural ODE prior, and conditional residual flow matching to predict metabolic states at supplied target stages. ADNI experiments evaluate latent-state similarity, regional SUVR prediction, and downstream cognitive prediction.

Critical strengths
  • Internal ablations support residual refinement for metabolic-state prediction. Adding flow matching reduces SUVR MAE from 0.102 to 0.094 and SUVR-change MAE from 0.103 to 0.087 (Table 1), providing evidence that the residual model improves upon the deterministic prior.
Critical weaknesses
  • The claimed connection between disease stage and chronological time is not established. The abstract and contributions emphasize a learned nonnegative progression-speed function, but the methods and algorithms specify integration to a supplied target DPI without defining this function or its training objective. Appendix C.2 describes separate stage-axis and time-axis settings rather than their learned connection. Consequently, the presented formulation does not substantiate forecasting disease stage and metabolism at an unknown future visit from baseline information.
  • The incremental contribution is insufficiently isolated. Section 4.3 explicitly states that AlzheimerODE uses different prediction spaces, test samples, and normalization, precluding a controlled comparison. The closely related cLDM-ODE, which already combines diffusion and neural ODEs for AD forecasting, is not evaluated. Table 1 also lacks a matched flow-only model and a time-conditioned counterpart, leaving the benefit of the proposed disease-stage prior unestablished.
Questions for authors
  • How are target DPIs obtained for Tables 1–2? Are they derived from target-visit clinical measurements, including cognition or biomarkers? Please specify the progression-speed model and distinguish baseline-only forecasting results from results conditioned on observed future stage and treatment.
  • Under identical inputs, subjects, observable targets, and forecast horizons, how does ADWM compare with cLDM-ODE, AlzheimerODE, and matched flow-only and time-conditioned variants?
  • Does the proposed disease-stage coordinate outperform a simple index constructed from clinical scales? What evidence establishes its reliability beyond the LLM’s prompted scoring rules?
2 comments
AuthorsOpenReview ↗

To Reviewer 24u8

To Reviewer 24u8, we appreciate the reviewer's assessment of our work. These questions provide a valuable opportunity to clarify our methodological choices and contributions.

Response to Q1, We thank the reviewer for the opportunity to clarify how the Disease Progression Index (DPI) is used in our framework. We use LLMs, guided by constrained prompts and clinical information such as cognitive assessments and clinical reports, to construct an approximate disease progression axis for AD. Each clinical observation is assigned a corresponding DPI value.

When a target DPI is estimated from target-visit measurements, the resulting prediction is conditioned on the observed future disease stage. The coarse-grained DPI limits the amount of clinical detail passed to the model, but it still conveys information from the target visit. We will specify how the target DPIs in Tables 1–2 were obtained, including whether target-visit cognitive assessments or biomarkers were used. We will also clarify which results, if any, represent baseline-only forecasts and which are conditioned on observed future stage or treatment.

The current framework has no nonnegative progression-speed function linking DPI to chronological time. The ODE is integrated along the DPI axis, and all predicted images in our experiments are indexed by DPI. DPI represents a coarse-grained disease progression state, and its relationship with elapsed chronological time can vary across patients.

We considered chronological-time integration during development. Real-world longitudinal AD datasets typically have short follow-up periods, irregular and asynchronous visits, and limited sample sizes. In the chronological-time formulation we considered, training pairs would generally come from visits within the same patient. In our dataset, the number of such pairs is less than 1% of the number of cross-patient pairs available under the DPI-based formulation, and they typically cover only short segments of individual disease trajectories. Pairing observations across patients by estimated disease stage allows the model to learn from more pairs spanning a broader range of progression states.

We will revise the Abstract, Introduction, and Conclusion to describe the implemented DPI-based formulation and remove the descriptions of chronological-time integration and a progression-speed function from the Appendix.

Response to Q2,

We thank the reviewer for emphasizing the importance of fair and consistent baseline comparisons. The current experiments do not establish how ADWM compares with cLDM-ODE, AlzheimerODE, or matched flow-only and time-conditioned variants under identical inputs, subjects, observable targets, and forecast horizons.

The AD-related ODE method included in Table 2 was evaluated in a different space from our framework, as discussed in Section 4.2.3. We included its published metrics as an indirect reference. A controlled head-to-head comparison would require evaluation in the same 120-region SUVR space under the conditions specified by the reviewer.

Our primary experimental validation therefore uses internal ablations to examine the contributions of individual components within a consistent modeling framework. A comparison of the chronological-time and DPI-based formulations as originally considered would also change the training-pair construction strategy, the amount of available training data, and the prediction task. These differences would make it difficult to attribute a performance difference to the choice of integration coordinate alone.

Our primary contribution is to establish the feasibility of learning AD progression dynamics through a brain imaging world model. We revisited the relevant literature and evaluation strategies and have yet to identify an existing baseline that can be evaluated in our setting without substantial methodological modifications. We would welcome suggestions for a suitable baseline or feasible evaluation protocol and would make every effort to incorporate additional experiments within the rebuttal period.

Response to Q3,

DPI is intended to preserve a clinically plausible ordering of AD progression states by combining information from clinical reports and cognitive assessments with prior knowledge embedded in LLMs. It provides a coarse-grained summary of the available clinical information. The current experiments do not establish whether DPI improves prediction over a simple index constructed from clinical scales.

Table 5 shows that the estimated DPI values were broadly consistent with the distribution of the clinical observations used in the evaluation, with no rejected assignments. This provides evidence of plausibility within the reported evaluation. Establishing reliability beyond the prompted scoring rules would require validation against an independent clinical reference, and determining whether DPI improves prediction would require a matched comparison with a scale-based index.

Thank you for the candid clarifications. They confirm that baseline-only forecasting, matched baseline comparisons, and the added value of DPI remain unestablished. The proposed revisions improve the accuracy of the paper’s claims, but my main concerns remain unresolved. I therefore maintain my original rating.

Reviewer_zR6HOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ADWM to model Alzheimer’s disease progression from PET imaging and regional SUVR measurements. It combines an LLM-derived disease progression index (DPI), a neural ODE, and a residual flow-matching model. The idea is to learn shared disease progression and then account for individual differences using patient and medication information. The model improves several imaging metrics, but its benefit for cognitive prediction is less clear.

Critical strengths
  1. Distinguishing disease stage from elapsed time is a useful starting point, given that patients progress at different rates.
  2. The regional SUVR predictions make the results easier to interpret than latent-space errors alone. Patient-level splitting and the persistence baseline are also appropriate choices.
  3. The authors report results that do not favor the full model. In particular, they acknowledge that better imaging predictions do not necessarily improve MMSE prediction.
Critical weaknesses
  1. It is unclear what the model can predict without knowing the future disease stage. The formulation requires a target DPI, but the manuscript also describes forecasting at future visits. These are different tasks. The claimed progression-speed function connecting DPI to calendar time is not adequately defined or evaluated.
  2. The paper spends much more space on ablations than on showing how well ADWM performs against other methods. The main comparison should come before the component analysis. More importantly, the authors state that ADWM and AlzheimerODE use different evaluation protocols, so their numbers cannot establish which method performs better. A matched benchmark, including a time-aware Transformer and a generator without the ODE prior, is needed.
  3. The method section is lengthy relative to the experimental evidence. Several equations describe standard components, such as normalization, regression losses, and flow matching. These definitions are useful for implementation, but some could move to the appendix. The main text should explain what is new and provide stronger evidence that those choices matter.
  4. The DPI labels need more validation. Providing the teacher prompt does not establish that its scores reliably represent disease stage. The authors should identify the teacher model, report how many labels were accepted or rejected, and compare the labels with independent clinical assessment or a simpler staging rule. They should also explain whether future clinical measurements are used to determine target DPI during testing.
  5. The clinical results need more discussion. The full model has higher MMSE error than both persistence and the ODE-only model. Treatment and patient context provide small, inconsistent improvements across metrics. Without repeated runs or confidence intervals, it is difficult to judge these differences. The discussion mostly describes the results rather than explaining what they reveal about the model.
  6. The figures need revision. Figure 1 uses inconsistent fonts and mixes decorative icons, scan images, and latent representations, making the workflow difficult to follow. Training-only supervision and inference inputs should be clearly separated. Figures 2 and 3 need larger views, clearer case descriptions, and quantitative error or change maps. The apparent increase in FDG-SUVR with increasing DPI in Figure 2 also needs an explanation.
  7. The manuscript needs a careful reporting check. Decimal precision varies between the text and tables, some reported values disagree, and cohort counts are not clearly reconciled. Appendix A.1 contains an unresolved “(?)” citation for cLDM-ODE, which the authors identify as their closest prior work.
  8. The submitted manuscript PDF is missing the AI-use statement required by ICLR 2027. Describing the LLM teacher in the methods does not replace the required disclosure section. A reproducibility statement is also absent, although that statement is recommended rather than mandatory.
Questions for authors
  1. How is target DPI obtained at test time? Can ADWM predict a future visit using only information available at the source visit? Please specify and evaluate the progression-speed function.
  2. Can the main comparisons be repeated with the same patients, inputs, targets, and forecast horizons? Does ADWM still improve over a matched Transformer and a generator-only model?
  3. How accurate and stable are the teacher-generated DPI labels? What does the LLM add over established clinical scores or a simple staging model?
  4. Why does adding flow matching improve imaging metrics while worsening MMSE prediction? Is this caused by the representation, the readout, or the objective?
  5. What explains the uptake changes in Figure 2, and do they agree with the observed regional changes in held-out patients?
1 comment
AuthorsOpenReview ↗

To Reviewer zR6H, we believe that addressing these points will help clarify the contributions of our work

Response to Question 1: For test pairs with an observed follow-up visit, we use the DPI assigned to that visit as the target DPI. The teacher LLM assigns visit-level DPIs using tightly constrained prompts and clinical information from the corresponding visit, including cognitive assessments and clinical reports. The resulting image prediction is therefore conditioned on the observed follow-up stage. Forecasting a visit at a specified time using only source-visit information would also require estimating the DPI the patient will reach at that time. We integrate the ODE along the DPI axis and index the reported images by DPI. The implemented framework includes no nonnegative progression-speed function linking DPI to chronological time, so we report no evaluation of such a function. Pairing observations across patients by disease stage allows us to learn from sparse, asynchronous follow-up data over a broader range of stages. In our dataset, the within-patient pairs available for the chronological-time formulation we considered number less than 1% of the cross-patient pairs available for DPI-based training. We will describe the implemented DPI formulation in the Abstract, Introduction, and Conclusion and remove the descriptions of chronological-time integration from the Appendix.

Response to Question 2: ADWM, a matched Transformer, and a generator-only model have been compared using the same held-out patients, source-visit inputs, target observations, forecast horizons, and normalization procedure. All three models were evaluated in the same 120-region SUVR space and, for DPI-conditioned predictions, they receive the same target DPI. AlzheimerODE is the closest published method we identified. As explained in the first paragraph of Section 4.3, the published AlzheimerODE results use a different prediction space, test set, and normalization procedure. We include its metrics in Table 2 as an indirect reference. Our primary experimental evidence comes from internal ablations, which quantify the contributions of individual ADWM components under a common protocol. These results support the feasibility of learning AD progression dynamics with a brain imaging world model.

Response to Question 3: Table 5 shows that the teacher-generated DPI labels are broadly consistent with the distribution of the observed clinical data and that all assignments were accepted. These results describe the labels at the distribution level. The evaluation gives no estimate of accuracy for individual visits against an independent clinical reference or of repeatability across scoring runs. We use the LLM’s prior knowledge of AD progression to combine information from clinical reports and cognitive scales into a single coarse-grained coordinate designed to preserve a clinically plausible ordering of disease stages. The LLM thus brings information from the clinical reports into the stage estimate alongside the scale scores.

Response to Question 4: Flow matching (FM) is trained to fit latent-state residuals without direct MMSE supervision. We decode regional SUVR from the PET embedding and estimate MMSE using a frozen ridge regressor fitted on embeddings from observed training scans. If Δz is the FM correction and w is the ridge regression weight vector, the change in predicted MMSE is wᵀΔz. The same correction can improve regional SUVR by correcting features used by the PET decoder, while its projection onto w moves the ODE-only MMSE estimate farther from the observed score. This explains how the latent residual objective and the fixed MMSE readout can produce different effects on the two metrics through the same change in representation.

Response to Question 5: The changes in Figure 2 come from evolving the PET embedding along the DPI axis and decoding it at different target DPIs. They show how the model predicts relative tracer uptake to change across disease stages. For held-out patients, our reported metrics compare predicted and observed regional SUVR at the target visit.

Reviewer_Qny6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ADWM, a latent world model designed to simulate Alzheimer's disease (AD) progression from sparse longitudinal FDG-PET imaging and regional SUVR profiles. The framework encodes paired PET and SUVR inputs into a frozen, aligned latent space (via ReMAP-PET) and organizes progression along a continuous Disease Progression Index (DPI) supervised by pseudo-labels from an LLM clinical teacher. A shared Neural ODE models population-level latent trajectories along the DPI axis, while a conditional flow-matching (FM) module predicts individual-specific latent residuals conditioned on the initial state, ODE endpoint, patient context, and visit-level medication summaries. The method is evaluated on longitudinal ADNI data via internal ablations and baseline comparisons.

Critical strengths
  1. Decoupling longitudinal disease evolution into a shared, stage-indexed population prior (Neural ODE) and individualized stochastic refinements (conditional flow matching) is conceptually well-suited for heterogeneous and irregularly sampled clinical neuroimaging data.

  2. The manuscript is commendably candid about the model's statistical boundaries (e.g., clarifying that conditioning on observational treatment does not identify causal effects and that sample variance does not guarantee calibrated uncertainty). Furthermore, the appendices provide thorough architectural specifications and the exact prompt contract used for LLM pseudo-labeling.

Critical weaknesses
  1. In the Abstract, Introduction (claiming three main contributions), Related Work, and Conclusion, the paper repeatedly claims to introduce "a separate nonnegative progression speed function that connects latent disease stage with chronological time under longitudinal supervision." However, this component is completely absent from the Methodology (Section 3) and Appendices. Instead, inference requires the target stage pqp_q to be manually supplied as an external input (Section 3.1), and Appendix C.2 treats time-axis and DPI-axis integration as separate experimental modes rather than linking them via a learned speed function.

  2. In Section 4.3 (Table 2), the authors explicitly state that the prediction spaces, test samples, and normalization rules differ between ADWM and baselines such as AlzheimerODE, concluding that "these values describe separate evaluations and do not constitute a controlled comparison." Without a unified evaluation protocol on shared observable targets and identical test splits, the empirical superiority of ADWM cannot be established.

    Additionally, there are notable discrepancies between the main text and appendices: Appendix A.1 claims all baselines are trained and evaluated on the exact same cached latent states and patient splits (contradicting Section 4.3) and describes RSSM/Dreamer and cLDM-ODE as baseline families, which are absent from Table 2.

  3. Table 1 lacks an FM-only variant (without the DPI-ODE prior) and a chronological-time ODE variant, making it difficult to isolate the necessity of the LLM-derived DPI coordinate and the ODE prior.

  4. Adding flow matching (FM) improves metabolic state metrics (SUVR MAE) but degrades downstream clinical cognition prediction (MMSE MAE increases from 1.687 to 1.991, performing worse than the trivial Persistence baseline at 1.902). Moreover, removing treatment (w/o Treatment) or patient context (w/o Context) yields nearly identical or slightly better state prediction errors, weakening the empirical justification for context/treatment conditioning.

Questions for authors
  1. Could the authors clarify whether the "nonnegative progression speed function" linking DPI to chronological time is implemented in the current framework? If so, please provide its mathematical formulation and evaluation results; if not, the Abstract, Introduction, and Conclusion must be revised to accurately reflect the actual model.

  2. Can the authors provide a controlled comparison for Table 2 where all baselines (and an FM-only ablation variant) are evaluated on the exact same held-out test pairs and decoded into the shared 120-region SUVR space?

  3. What causes the conditional flow-matching residual module to substantially increase MMSE MAE (1.687 vs. 1.991) while improving SUVR reconstruction?

2 comments
AuthorsOpenReview ↗

We believe that addressing these points will help clarify the intended scope and contributions of our work

We sincerely appreciate the reviewer's careful and insightful assessment of our work. We particularly value the reviewer's deep understanding of our core motivation and methodological design, as reflected in the precise and constructive questions addressing key aspects of our framework. These questions provide a valuable opportunity to clarify our methodological choices and contributions.

For Question 1, we clarify the role of the Disease Progression Index (DPI) in our framework.

First, our framework does not implement a "nonnegative progression speed function" that explicitly maps DPI to chronological time. DPI and chronological time do not necessarily exhibit a strong correlation, making a reliable mapping challenging. Accordingly, all predicted images are indexed by DPI rather than chronological time.

DPI serves as a coarse-grained indicator of Alzheimer's disease (AD) progression, rather than elapsed chronological time. We leverage large language model (LLM) prior knowledge, strongly constrained prompts, and real clinical information, including cognitive assessments and clinical reports, to construct an approximate disease progression axis reflecting the overall trajectory of AD. Each clinical observation is subsequently assigned a corresponding DPI value. This enables our ODE-based model to capture broader and more comprehensive disease progression patterns from sparsely sampled longitudinal clinical data. Importantly, DPI provides a shared progression coordinate across patients, allowing observations collected at different chronological times to contribute to learning a common disease progression trajectory.

The primary motivation for adopting DPI lies in the limitations of real-world longitudinal AD datasets, including short follow-up durations, irregular and asynchronous observations, and limited sample sizes. ODE integration along the DPI axis enables training pairs to be constructed across patients according to their estimated progression states. This substantially increases available training pairs and expands the range of disease progression patterns the model can learn, allowing it to better capture population-level AD progression dynamics.

In contrast, chronological-time modeling faces significant challenges due to temporal misalignment across patients. Training pairs would generally need to be constructed from longitudinal observations within individual patients. In our dataset, within-patient training pairs account for less than 1% of the cross-patient pairs available under DPI. Moreover, these pairs typically cover only short segments of individual disease trajectories, making it difficult to capture broader population-level AD progression patterns. Thus, chronological-time modeling would substantially restrict both training data availability and disease progression coverage.

We considered chronological-time ODE integration during framework development. However, given sparse and asynchronous follow-up observations and our objective of learning population-level progression, we ultimately adopted DPI. We believe this design better accommodates the characteristics of the available clinical data and our intended modeling objective. Furthermore, a direct comparison between chronological-time-based and DPI-based integration would not be strictly controlled, as they involve substantially different training-pair construction strategies, data availability, and prediction difficulties. These differences also make it difficult to isolate the effect of the integration axis from changes in the underlying training data.

Regarding DPI reliability, Table 5 shows that estimated DPI values broadly align with real clinical data distributions, with no rejected assignments. DPI provides an approximate representation of AD progression based on clinical reports, cognitive assessments, and LLM prior knowledge, rather than exact temporal measurements. Its primary purpose is to preserve a clinically plausible ordering of disease progression states. Compared with conventional clinical scores or simple staging schemes, DPI offers a more comprehensive yet coarse-grained characterization of disease progression. This representation also avoids exposing detailed future clinical information during progression modeling, thereby reducing the risk of information leakage from prediction targets.

Finally, we acknowledge that references to a nonnegative progression speed function and chronological-time mapping may have caused confusion. We will carefully revise the Abstract, Introduction, and Conclusion to accurately describe our DPI-based formulation and remove misleading chronological-time descriptions from the Appendix, ensuring consistency between the mathematical formulation, experimental implementation, and stated contributions.

AuthorsOpenReview ↗

For question2 and 3

For the question 2, we thank the reviewer for this constructive suggestion regarding the fairness and consistency of the baseline comparisons. We would first like to clarify that the existing related AlzheimerODE method operates in a different evaluation space from our framework. Consequently, its reported results cannot be directly interpreted as a controlled comparison in the shared 120-region SUVR space. The first paragraph of Section 4.3 explicitly acknowledges that ADWM and baselines such as AlzheimerODE differ in their prediction spaces, test samples, and normalization procedures, stating that “these numbers describe separate evaluations and do not constitute a controlled comparison.” Without a unified evaluation protocol that evaluates predictions of the same observable targets on identical test splits, the empirical superiority of ADWM over these baselines cannot be established.

A fundamental challenge is the limited availability of directly comparable methods for AD brain imaging world modeling. To the best of our knowledge, we have not identified an existing framework that addresses the same modeling objective under a compatible prediction and evaluation setting. The AD-related ODE method represents the closest relevant work we could identify, and we therefore included its reported metrics as an indirect reference rather than a strictly controlled head-to-head comparison.

Given this limitation, our primary experimental validation focuses on internal ablation studies, which allow us to systematically examine the effectiveness of individual components within a consistent modeling framework. We believe that such controlled evaluations provide meaningful evidence of the proposed method's effectiveness, particularly in an emerging research direction where directly comparable benchmarks and established baselines remain limited.

More broadly, our work aims to explore a relatively underinvestigated direction: learning disease progression dynamics through a brain imaging world model for AD. At this early stage, our primary contribution is to establish a feasible modeling framework and investigate its effectiveness, rather than to claim superiority over existing methods under fully standardized evaluation conditions. We believe that demonstrating the feasibility and value of this modeling paradigm is itself an important step toward further research in this direction.

We fully appreciate the reviewer's emphasis on rigorous and fair comparisons. We have carefully revisited the relevant literature and potential evaluation strategies in response to this concern, but have not yet identified a directly compatible baseline that would enable the requested comparison without substantial methodological modifications. If the reviewer is aware of a suitable baseline or a feasible evaluation protocol, we would be very willing to investigate it. We hope this clarification helps explain our experimental design choices and the current limitations of baseline availability in this emerging research setting.

For the question 3, thank the reviewer for highlighting this observation. The increase in MMSE MAE (from 1.687 to 1.991), despite improved SUVR reconstruction, can be explained by the different objectives and sensitivities of the two prediction pathways.

Specifically, the conditional flow-matching (FM) module is trained to correct residual errors in the predicted latent states without direct supervision from MMSE scores. Regional SUVR values are reconstructed by decoding the PET embeddings, whereas MMSE scores are predicted using a frozen ridge regression model fitted on ground-truth training embeddings.

Consequently, the effect of FM on MMSE predictions depends on the projection of the corrected latent residuals onto the regression weight vector. While FM can reduce latent-state errors that affect regional SUVR reconstruction, it may simultaneously introduce additional shifts along directions to which the MMSE readout is sensitive. If these shifts move the original ODE-based MMSE predictions further away from the ground-truth scores, the MMSE prediction error increases, even when the corrected latent states are overall closer to the true follow-up states.

Therefore, the observed discrepancy reflects a potential misalignment between latent-state reconstruction and downstream cognitive prediction objectives, rather than a contradiction in the effectiveness of FM for SUVR reconstruction.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.