List of Papers Browse by Subject Areas Author List
Abstract
Imaging-derived phenotypes (IDPs) summarize multi-organ physiology but provide only static snapshots of diseases that evolve over time. In contrast, longitudinal electronic health records encode disease trajectories through temporal dependencies among past diagnosis events and comorbidity structure. We hypothesize that IDPs and disease trajectories contain partially shared disease-relevant structure. We propose a trajectory-aware distillation framework that transfers structural knowledge from a generative disease trajectory Transformer into an organ-wise IDP encoder. A population-scale trajectory model trained on longitudinal diagnosis sequences produces subject-level embeddings that supervise IDP representation learning via geometry-preserving alignment. During downstream prediction, trajectory and imaging representations can also be fused via cross-attention. Across 159 diseases in the UK Biobank cohort, trajectory-aware pretraining consistently improves both discrimination (AUC) and time-to-onset prediction (MAE), with the largest gains for low-prevalence diseases. Similarity relationships in IDP embedding space also align with those in trajectory space, providing supportive evidence for partially aligned representation geometry. These results suggest that population-scale generative disease models can serve as structural priors for data-limited imaging modalities, improving robustness under realistic cohort constraints.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1438_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WanZia_From_MICCAI2026,
author = { Wang, Zian AND Lan, Lizhen AND Wang, Guangming AND Zhang, Haosen AND Xu, Minxuan AND Li, Qing AND He, Tianxing AND Yang, Mo AND Mao, Wenyue AND Zhang, Yajing AND Li, Yan AND Wang, Chengyan},
title = { { From Trajectories to Phenotypes: Disease Progression as Structural Priors for Multi-organ Imaging Representation Learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a distillation framework that transfers disease progression knowledge from a EHR-based trajectory transformer into a multi-organ imaging phenotype encoder. The pre-trained imaging transformer demonstrated improved classification performance across diseases. The overall idea of leveraging population-scale longitudinal EHR data to enrich imaging-based disease prediction is interesting and practically motivated.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The core idea of borrowing structural knowledge from large-scale longitudinal EHR data to guide imaging predictors is well-motivated and practically useful. The framework is evaluated across a large number of diseases with a carefully partitioned cohort design that attempts to prevent information leakage. The comparison of different alignment objectives provides useful insight into how distillation loss choice affects downstream performance for classification versus time-to-onset prediction tasks.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The major concern is a potential information leakage issue that undermines the fairness of the comparison, particularly for the fusion model. Since the fusion model directly uses each evaluation subject’s own pre-imaging trajectory embedding, the model may already have direct access to prior diagnosis of the disease being predicted or closely related comorbidities. A fair evaluation would require predicting only diseases with first onset strictly after the imaging visit. In addition, I have some concerns regarding the disagreement of IDP similarity and trajectory similarity (Fig. 3). Despite the positive correlation, the IDP similarity is very low while the trajectory similarity is generally above 0.8.Also, the confidence interval in fig 3b is not consistent with Fig. 3a, where IDP similarity spans a much wider range at any given trajectory similarity. This result to some extent challenges the claim that imaging and trajectory representations share a common latent disease manifold. Finally, several methodological details including the specific imaging modalities, organ systems, and IDP selection criteria are never clearly stated.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The major concern is a potential information leakage issue in the fusion model, which directly uses each evaluation subject’s own pre-imaging trajectory embedding at inference time. For chronic and recurring diseases this gives the model direct access to prior diagnosis information, making the comparison against the scratch imaging model unfair. The paper’s central claim of a shared latent disease manifold is also not convincingly supported in Fig 3.The major positive point is in the improved performance pre-trained IDP encoder, which only leveraged the knowledge from population-wide longitudinal EHR.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
Addressing the broad challenge of disease progression modelling, the authors present a method that a) uses ICD-codes based patient trajectories as priors to guide pretraining of an encoder that produces predictive embeddings from Image-Derived Phenotypes (IDP) [which they call their trajectory-distillation framework], and b) fuses information from temporal disease trajectories with static IDPs to show improved predictive performance on disease occurrence and time-to-event (disease) prediction.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Novel Method: Using ICD trajectory embeddings as structural priors is an interesting formulation, and the authors show that it helps learn embeddings from IDPs which boost prediction performance on disease forecasting, especially in low-prevalence disease. This could be a useful paradigm even outside of this context, in the larger scope of learning with asynchronous modalities recorded at different time resolutions. 2.Experimental designs are sound and rigorous, with the baselines of IDPs and trajectory embeddings providing meaningful context to interpret their results. 3.Post-hoc analysis of disease-wise gains in AUC is further informative.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The “shared latent manifold” hypothesis for IDP and disease trajectories is hard to digest. While I understand the intuition, I am not convinced that the experiments and results support the claim. The spearman of 0.166 on the permutation test on pairwise IDP and trajectory embeddings is significant, however the following could be confounders: a. Age, BMI, Sex: Controlling for these in the experiment would give a stronger support for the shared latent manifold theory, since these could be potential confounders that cause similar embeddings in both spaces. This is because, as I understand it, the claim is that the disease progression trajectory and IDP share a common latent manifold (independent of demographics). If this is not the claim, then could the authors please clarify? b. The InfoNCE Alignment objective: The embeddings are learned by aligning projections of the embeddings using contrastive learning. This alignment might be driving the pairwise similarity, in which case the effect is a result of the training strategy not a natural phenomenon. It would be more meaningful to do the same experiment but on the non-pretrained embeddings [as a suggestion for future work]. 2.Unclear Writing: Some parts of the paper could be rewritten for clarity. [See Comments below] 3.Missing Discussion of Limitations, Related Work and Future Work: I understand that the authors show extensive experiments, and there is an 8-page limit, but there is no discussion of related work on predicting disease progression on IDPs, or disease trajectories, or even mention of related technical works [Contrastive Representation Distillation, ICLR 2020]. Most importantly, there is no discussion of limitations or implications for use in future research, which is important for any reader trying to interpret these results.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
1.Writing and Experimental Details: a. The abstract and introduction mention “temporal dependencies and comorbidity structures” : For example, “Disease progression is governed by temporal dependencies and comorbidity relationships that cannot be fully captured by static imaging measurements alone.” It is unclear to the reader (who hasn’t read ahead) what temporal dependencies are being referred to. I think the authors might mean to say “temporal dependencies in past disease diagnosis.” b. The Methods Section mentions: “Training objective. The model is trained with a joint objective consisting of next-event cross entropy and a time-to-next-event likelihood term that models the distribution of inter-event time gaps” – Could the authors please clarify how they modelled the time-to-next-event term (the distribution on the inter-event time gaps). 2.I would be interested in the authors’ comments about the issues mentioned with the Shared Latent Manifold Hypothesis. I understand that this is a conference paper, so extensive experimentation could be out of scope, and I think this paper should be presented at MICCAI either ways, but it is important to discuss the interpretation and limitations of their results. 3.In a related sense, the missing discussion of limitations, related works should be addressed by the authors.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I think this work should be accepted and discussed at MICCAI but the paper could be improved (preferably addressing the points in the comments section) before publication.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes a trajectory-aware distillation framework that transfers disease progression knowledge from a generative EHR-based trajectory model to an imaging-derived phenotype (IDP) encoder via representation alignment. It shows that incorporating trajectory-informed structure improves disease prediction, especially for low-prevalence conditions, and enables effective fusion of longitudinal and imaging data.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
-
Clean experimental design to prevent subject and temporal leakage
-
Intuitive and interesting motivation for using temporal trajectories to enrich static IDPs
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
-
Some central claims are only supported by weak correlations, making the evidence less convincing.
-
Experiments are conducted on a single dataset (UK Biobank) and mainly compare internal variants.
-
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
-
I have concerns about the strength of evidence supporting some key claims. In particular, the prevalence-dependent trend in Fig. 2b appears weak (low correlation and marginal significance), and the representation alignment results in Fig. 3 are based on relatively small effect sizes. I suggest either moderating these claims or providing stronger supporting analyses.
-
I think the experimental evaluation could be strengthened. Table 1 mainly presents internal comparisons (e.g., scratch vs. pretrained, different fusion strategies), but lacks comparison to strong prior baselines such as EHR-based models (e.g., BEHRT, Med-BERT) or multimodal methods without distillation. This makes it difficult to assess how the proposed method compares to existing approaches.
-
The model descriptions are somewhat unclear, particularly in Table 1.The differences between “IDP Baseline,” “IDP Organ (Transformer),” and the fusion models are not fully specified. Providing a clearer taxonomy or summary table of model variants would improve readability and reproducibility.
-
All experiments are conducted on a single dataset (UK Biobank). I think evaluating additional cohorts, if feasible, would strengthen the claims about generalizability.
-
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The contribution appears somewhat incremental, building on existing distillation and contrastive alignment frameworks. In addition, the evaluation lacks rigorous comparison to strong prior work, making it difficult to assess the true advantage of the proposed method.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
Dear Area Chair and Reviewers,
We sincerely thank all reviewers for the constructive feedback and positive assessment of our work. We are encouraged that the reviewers found the proposed trajectory-aware distillation framework interesting and potentially useful for MICCAI. Our primary goal is to evaluate whether longitudinal disease trajectories can provide useful structural priors for multi-organ IDP representation learning under realistic imaging-cohort constraints, rather than to claim definitive evidence for a universal disease manifold.
Leakage and Prediction Setting (R1, Meta-review)
We agree that the evaluation protocol should be clarified more explicitly. Our downstream prediction setting follows an incident-only risk-set design with imaging as baseline. For each target disease, subjects with diagnosis onset at or before imaging are excluded as prevalent cases, and positives therefore correspond only to incident post-imaging cases. Controls are sampled from subjects who had already undergone imaging and remained event-free at the corresponding case time. Thus, the task is future-onset prediction after imaging rather than current-state classification.
To further prevent temporal leakage, trajectory embeddings used for distillation and downstream fusion are computed only from diagnosis records strictly prior to imaging acquisition. In addition, the downstream IDP-only and fusion pipelines explicitly exclude target-derived variables such as
labelanddelta_timefrom the feature set, with runtime leakage checks. We will revise the manuscript to make this protocol explicit and easier to follow.Shared Structure Claim (R1, R2, R3)
We agree that our current evidence should not be framed as definitive proof of a shared latent disease manifold. The observed alignment between IDP and trajectory similarity is statistically significant but modest, and should be interpreted as preliminary empirical evidence consistent with partially aligned disease-relevant structure. We will moderate the relevant wording accordingly.
Baselines and Scope (R3, Meta-review)
Our design isolates the effect of trajectory-informed priors through controlled comparisons across IDP-only, trajectory-only, and fusion settings, rather than through an exhaustive benchmark. We agree that the absence of broader comparisons to EHR and multimodal methods is a limitation of the current study. In our formulation, models such as BEHRT and Med-BERT can also be viewed as alternative trajectory encoders/teachers rather than only downstream baselines, and exploring such alternatives is an important direction for future work.
Methods (R1, R2, R3)
We will clarify the problem formulation, the Delphi trajectory-teacher objective, and the role of trajectory-aware distillation in the Methods more clearly. We will also expand the IDP description: the 301 quantitative phenotypes cover five system-level groups, namely brain MRI grey-matter volumes, skeleton/body-composition measures, cardiac MRI traits, abdominal MRI traits, and pulmonary measures including spirometry and lung volume. The IDP baseline operates on concatenated IDPs, whereas the organ-aware variants group features by organ and aggregate them by concat, weighted pooling, or a transformer.
Overall, we will revise the manuscript to (1) make the leakage-prevention and evaluation protocol fully explicit, (2) reduce overclaiming regarding manifold-sharing interpretations, (3) improve methodological clarity and model taxonomy, and (4) strengthen discussion of limitations, external validation, alternative trajectory models, and future directions.
We thank the reviewers again for their thoughtful comments and suggestions.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Overall all reviewers found this manuscript presents a useful and interesting algorithm for MICCAI. I am recommending this manuscript for acceptance. Authors in the final version of the manuscript, I recommend addressing the information leakage (R1) and limitations of the comparison methods (R3).
