List of Papers Browse by Subject Areas Author List
Abstract
Accurately predicting the temporal evolution of clinical biomarkers is crucial for the early diagnosis and management of neurodegenerative diseases such as Alzheimer’s disease. However, this relies on longitudinal data to capture biomarker changes over time, which is often sparse and irregular due to the high cost, labor-intensive nature, and patient burden. To address these challenges, we propose ENC-ODE, an Event-level Neurodegenerative modeling in Continuous time with neural Ordinary Differential Equations. ENC-ODE predicts future biomarker evolution by modeling clinical events through diagnosis-conditioned continuous dynamics. A target-conditioned attention mechanism weights and aggregates event-level predictions for the target time and modality without history compression. Extensive experiments on Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset demonstrate that ENC-ODE outperforms representative sequence models while offering a scalable and neuroscientifically grounded solution for clinical support. The code is available at https://github.com/JardinDelSol/enc-ode.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1872_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/JardinDelSol/enc-ode
Link to the Dataset(s)
ADNI dataset: https://adni.loni.usc.edu/data-samples/adni-data/
BibTex
@InProceedings{SonYuj_ENCODE_MICCAI2026,
author = { Song, Yujee AND Baek, Seunghun AND Wu, Guorong AND Kim, Won Hwa},
title = { { ENC-ODE: Event-level Neurodegenerative Modeling in Continuous Time with Neural ODEs } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contribution of the paper is the development of a time-series analysis framework based on Neural Ordinary Differential Equations (nODE) for predicting the evolution multimodal data in Alzheimer’s disease accounting for past observations through diagnosis-conditioned ODE dynamics. This system allows to capture stage-specific progression rates across irregular time intervals.
- 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.
- To the best of my knowledge, the decoding based on an attention mechanisms aggregating past hidden states is novel. This is the main contribution of this work.
- Comparison with a large panel of competing methods
- Promising results in terms of prediction metrics
- 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.
- Acknowledgement and comparison with prior work. Modeling hiddend state dynamics with ODEs is not novel in the domain of disease progression modeling of multimodal time-series (e. g. Abi Nader et al, Brain Comm 2021). In this sense, the actual contribution of this work appears to be rather incremental.
- Still concerning the acknowledgement of SOTA, it is not clear whether the other methods in the benchmark are applied in the multi- or uni-modal setting, or how missing data and time-stamps are handled. It seems that all these approaches were implemented ad-hoc by the authors. To provide a relevant assessment, there is a large literature on time-series analysis tailored to multimodal Alzheimer’s disease modeling and ADNI that should be acknowledged and compared with.
- The proposed attention-based aggregation approach seems overkill given the generally low amount of follow-up per subjects (as shown in Table 2). I wonder about the reliability of such method when the number of obsservations per individuals is so low. This approach could make more sense when applied to time series with many more time stamps.
- The ablation study should include simpler decoding methods (e. g. linear) beyond the trivial ones presented in Table 4.
- The proposed approach has an advantage with respect to the benchmark as it is conditioned on information relevant to the trajectory prediction task (e. g. clinical stage). This information is not available to the other methods.
- The magnitude and variability of the metrics across experiments is concerning. The need to go to the fourth decimal to appreciate SUVR prediction errors is unrealistic, it does not seem clinically meaningful, nor within the measurement error for this kind of data.
Minor. The notation || used in equation (1) is not clear and ambiguous.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper does not position well the proposed methods with respect to the large amount of existing contributions on progression modeling in Alzheimer’s disease. The experimental assessment seems also lacking.
- 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
- Propose ENC-ODE, a Neural ODE-based continuous-time modeling framework for predicting the temporal evolution of neurodegeneration-related clinical biomarkers from irregularly sampled multimodal longitudinal data.
- Introduce learnable embeddings for modalities and diagnostic stages, enabling a shared encoder and diagnosis-conditioned ODE dynamics.
- Design a target-conditioned attention mechanism that selectively aggregates event-level predictions from past observations based on their relevance to the target time point and modality, avoiding the information loss caused by conventional history compression.
- 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 proposed ENC-ODE architecture effectively shares a common latent space across modalities while utilizing learnable embeddings for both modalities and diagnostic stages.
- The target-conditioned attention mechanism allows the model to selectively aggregate event-level predictions by considering their relevance to the target time point and modality, which helps preserve fine-grained information that is often lost in traditional history compression. -The attention weight analysis (Fig. 3) provides meaningful insights into cross-modality relationships and complementary roles among biomarkers (AMY, FDG, and TAU). The authors conduct comprehensive experiments with appropriate baselines, including both discrete-time models (RNN, Transformer, Mamba) and continuous-time models (Neural ODE, ODE-RNN, Neural Flows, CRU) that can handle irregularly sampled time-series data.
- The ablation studies, particularly the removal of diagnosis-conditioned dynamics (Table 4), clearly demonstrate the contribution of each proposed component.
- 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.
Weaknesses)
- The learnable diagnosis embeddings may not be sufficiently trained for rare classes such as SMC. This raises questions about the reliability of diagnosis-conditioned dynamics for underrepresented stages.
- The paper does not provide per-diagnosis performance analysis. Although later disease stages are expected to show higher heterogeneity and consequently lower prediction accuracy, a stratified evaluation by diagnostic stage would still be helpful for future works.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper tackles a clinically important problem by introducing a continuous-time Neural ODE framework designed for multimodal and multi-diagnosis biomarker modeling. The proposed method demonstrates clear advantages over existing discrete and continuous baselines through its event-level modeling and target-conditioned attention mechanism. I recommend acceptance, as this work is likely to influence future research in medical time-series modeling and neurodegenerative disease analysis.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I think the paper presents a valuable idea by leveraging prior longitudinal events to predict future brain metabolism trajectories. The event-level formulation could provide meaningful benefits to the field of longitudinal neurodegenerative disease modeling.
Review #3
- Please describe the contribution of the paper
The paper proposes a continuous-time framework for the task of neurodegenerative disease modeling by leveraging Neural ODEs. The authors introduce key ideas of modeling events through diagnosis-conditioned ODE and target-conditioned attention mechanism to avoid compressing the entire history into a single latent state. Experimental results on ADNI have shown that the proposed model outperforms various baseline models.
- 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 paper identifies a critical limitation of the existing models which compress longitudinal multimodal data into a single latent vector so that event-level dependencies are lost; The proposed model is quite relevant for handling irregular clinical time series data, and the problem is well-motivated
-
The paper proposes several methodological novelties including diagnosis-conditioned event-level continuous trajectory modeling and target-conditioned multimodal aggregation, which are suited for the problem and designed with interpretability
-
The results as shown in Table 3 demonstrate that the model consistently outperforms all other baseline models across different metrics in both unimodal and multimodal settings, where the gains are more notable in multimodal setting
-
The paper further includes the ablation studies and provides with the interpretability analysis to showcase the contribution of the components and modalities
-
- 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 current experimental setting lacks external validation dataset as only ADNI has been used, which limits the clinical generalization of the method
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 paper is well-written with novel methodological contributions to tackle clinically relevant problems for irregular time series data and disease modeling. The experimental results also show the consistent strong performance of the proposed model. Although the current dataset used is only ADNI, that should not be considered as a major bottleneck as compared to the overall quality of the work.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal has addressed my concerns, and I will remain my decision for the paper.
Author Feedback
We thank reviewers for their insightful feedback. We focus on main concerns while highlighting recognized strengths such as event-level formulation, target-conditioned aggregation, broad baselines, ablation, and interpretable attention.
R1: Novelty Boundary We do not claim ODE-based hidden-state modeling itself is novel, which is why we included multiple ODE baselines. ENC-ODE differs from existing methods by propagating events independently rather than compressing history into a single latent state. Through target-conditioned aggregation, ENC-ODE handles sparse, irregular data and provides observation-level interpretability via learned weights.
R1: Baseline and problem setting Our work focuses on multimodal longitudinal prediction under irregular timestamps and missing modalities. We compared representative sequence models under a unified protocol, treating each modality-specific measurement as an independent timestamped event. Public implementations were used when available, and standard implementations were matched in model capacity and training settings. Regarding Abi Nader et al. [1], the problem settings differ substantially. Their framework models subject-level latent disease trajectories assuming relatively complete baseline multimodal states, whereas our work focuses on target biomarker forecasting from sparse and irregular event histories with missing modalities.
R1: Reliability with limited follow-up visits Our attention mechanism focuses on relevance-aware aggregation over sparse, heterogeneous events rather than sequence length. As shown in Table 4, selective weighting yields superior accuracy compared to uniform and latest-observation baselines, particularly for the sparsest modality (Tau). This demonstrates that weighted aggregation is beneficial even for short histories.
R1: Simpler Decoding Ablations While the decoder remains a simple fully connected layer (Sec. 3.1), a fixed linear aggregation over event-level predictions is ill-suited to our architecture. Because history length varies, such an approach would force arbitrary padding or truncation. The submitted ablations directly isolate diagnosis-conditioned dynamics and target-conditioned aggregation.
R1: Conditioning on Diagnosis/Stage Diagnosis is an observed attribute of each clinical event, not future information. Table 4 isolates this concern. Even without leveraging diagnosis information, ENC-ODE outperforms the strongest multimodal baselines in Table 3 for all target biomarkers. Therefore, the gains cannot be attributed solely to privileged disease-stage information. Larger degradation from uniform aggregation and latest observation shows target-aware aggregation is a main driver.
R1: Metric Precision and Clinical Significance We do not claim individual-level clinical decisions from such differences. Metrics are used for controlled model comparison on the same normalized SUVR targets. Our conclusion relies on consistent RMSE, MAE, and R2 rankings over 5 runs, not tiny decimals. In Table 3, ENC-ODE reduces RMSE by about 17% for FDG, 31% for Tau, and 11% for AMY over the strongest baselines. Thus, the evidence supports improved longitudinal biomarker forecasting under sparse PET follow-up, not immediate clinical deployment from small SUVR differences.
R2: Stratified Evaluation We appreciate this valuable suggestion. Table 1 reports rare classes such as SMC and the associated imbalance. Since our scope is time irregularity and modality sparsity, we treat subgroup-specific analysis as valuable future work.
R3: Dataset Extension We selected ADNI because it provides sparse irregular longitudinal PET biomarkers, with AMY, FDG, and Tau measured at different visits rather than as complete synchronized modalities. This directly matches our event-level setting. While the formulation is general to such multimodal sequences, validation on other neurodegenerative cohorts remains valuable future work.
[1] Abi Nader et al., Brain Communications, 2021.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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.
Reviewers summarized strengths of your work, novel aspects and noted that your solution may be of potential interest. They all agree on some novelty of this research, namely a time-series analysis framework based on nODE with a novel decoding mechanism and find the comprehensive comparison and promising results positive aspects. However, reviewers also point in detail to some major weaknesses of the current work which include limited acknowledgement of previous work, questions on the ablation study, questions on the metrics used across experiments, and lack of clear assessments of experiments. Reviewers also questioned the reliability of the diagnosis-conditioned dynamics, and as a major critique found that the experimental setting lacks external validation datasets beyond ADNI. In summary, we see that weaknesses and limitations somewhat balance its merits and thus may pose questions for this work to be accepted for MICCAI.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
I think that the idea is interesting, and major concerns have been addressed in the rebuttal. Authors have provided a good discussion on this early-stage but interesting research, and have provided some convincing responses to specific critiques. I strongly encourage authors to include better clarifications and revisions as suggested by reviewers to strengthen the final version. I would argue for accept.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After carefully reading the reviewer’s comments and the author’s rebuttal my opinion of this paper is strongly aligned with Rv2 and 3 that lean for the acceptance of the manuscript. I believe that the improved version of the manuscript coming out from this review process will result into an extremely valuable contribution to the community.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After considering the rebuttal and the updated reviews, I recommend acceptance. The paper addresses a meaningful problem in longitudinal neurodegenerative disease modeling and presents a potentially useful formulation. Overall, the reviews are sufficiently positive to support acceptance.