List of Papers Browse by Subject Areas Author List
Abstract
Multimodal longitudinal records provide critical insights into disease evolution and support disease understanding and early intervention. However, existing methods struggle with irregular and incomplete multimodal data, overlook the population-level progressions, and offer limited interpretability for clinical forecasting. To address these challenges, we propose a unified multimodal longitudinal forecasting framework that jointly predicts future disease status while generating explanations. Our approach employs a multimodal language model (MLM) with a novel longitudinal language tuning scheme that learns from generated hierarchical interpretations, enabling forecasting alongside structured explanatory outputs. Furthermore, a multimodal bipartite graph module is proposed to model missing and heterogeneous records, while a dual-memory module retrieves population-level patterns to support report generation. Experimental results on the ADNI dataset demonstrate significant improvement of the proposed method in both disease forecasting and result interpretation compared with state-of-the-art methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1800_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/yhf42/ENIGMA
Link to the Dataset(s)
ADNI dataset: https://adni.loni.usc.edu/data-samples/adni-data/
BibTex
@InProceedings{FanYuh_Interpretable_MICCAI2026,
author = { Fan, Yuheng AND Qin, Hongyi AND Xie, Jianyang AND Chen, Zhengju AND Luo, Yimin AND Alam, Uazman AND Lip, Gregory Y. H. AND Cheng, Li AND Zheng, Yalin AND Zhao, He},
title = { { Interpretable Longitudinal Disease Forecasting with Graph-Based Multimodal Encoding and Population-Aware Language Generation } },
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 ENIGMA (ENable Interpretable multimodal longitudinal modeling with bipartite Graph encoding and population Memory-Augmented generation), a multimodal language model (MLM) framework for interpretable Alzheimer’s disease forecasting. The core innovations include:
1.First MLM-based Longitudinal Forecasting Framework: To the authors’ knowledge, this is the first multimodal language model specifically designed for longitudinal disease forecasting that jointly predicts future disease status while generating hierarchical structured explanations.
2.Bipartite Graph Encoding (BGE): A novel module that handles incomplete and heterogeneous multimodal longitudinal records (MRI images, time-variant clinical variables, time-invariant demographics) using graph neural networks. The bipartite graph structure represents records as nodes with missingness reflected in connectivity, enabling robust handling of missing data.
3.Hierarchical Interpretation Construction: The framework generates structured explanations at four temporal scales:
- Visit Impression (VI): Current normative status
- Recent Changes (RC): Most recent evolution patterns
- Long-term Changes (LC): Historical trajectory from initial visits
- Progression Status (PG): Expected evolution toward forecasting target These hierarchical interpretations are constructed using normative modeling (FreeSurfer volumetric quantiles) and synthesized via GPT-5.2 to create teaching supervision.
4.Dual Population Memory (DPM): Two memory banks (record memory and impression memory) that capture population-level multimodal disease evolution patterns and textual impressions, enabling population-aware forecasting and clinically relevant explanation generation.
5.Longitudinal Language Tuning: A novel training scheme where the MLM learns from generated hierarchical interpretations, enabling forecasting alongside structured explanatory outputs.
The framework addresses key limitations of existing methods: inability to handle irregular/incomplete multimodal data, lack of population-level pattern modeling, and limited clinical interpretability.
- 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.
Strength 1: Novel Integration of MLMs for Interpretable Clinical Forecasting
- This is pioneering work bringing multimodal language models to longitudinal disease forecasting with structured interpretability, addressing a critical gap in medical AI where most models are “black boxes. “
- The hierarchical explanation structure (VI, RC, LC, PG) is clinically meaningful and mirrors how physicians reason about disease progression - considering current status, recent changes, long-term trends, and expected trajectory.
- The dual-task learning (forecasting + explanation generation) is elegant: the model learns better disease representations by being forced to generate clinically grounded explanations, not just predict labels.
Strength 2: Sophisticated Handling of Missing and Heterogeneous Data
- The bipartite graph encoding (BGE) is technically sound for handling incomplete multimodal longitudinal records - a ubiquitous challenge in real clinical data.
- Graph structure naturally represents variable-time relationships with missingness encoded in connectivity patterns, more principled than simple imputation or loss masking.
- The method handles three types of variables (time-variant tabular, time-invariant tabular, imaging patches) in a unified framework using GNN with multi-head attention.
Strength 3: Strong Empirical Results with Comprehensive Evaluation
- Impressive improvements over strong baselines: 84.2% accuracy vs. 76.4% for best baseline (UniLM), 89.3% AUC vs. 89.0% for GraFITi.
- Outperforms recent LLM/MLM methods (DTGPT, NoteLM, ViT-LM) across all metrics, demonstrating the value of the proposed architectural innovations.
- Multi-faceted evaluation: quantitative forecasting metrics (ACC, SEN, SPE, F1, AUC), NLG quality metrics (BLEU, METEOR, ROUGE-L, BERTScore), LLM-based judge evaluation, AND clinician evaluation across 6 dimensions.
- Comprehensive comparisons: 9 baselines covering single-modality, multi-modality, and VLM approaches.
- Qualitative analysis (Fig. 3) shows the model generates more accurate, structured, and clinically coherent explanations compared to strong baseline (NoteLM).
- 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.
Weakness 1: Questionable Quality of GPT-Generated Teaching Labels
- Issue: The entire hierarchical interpretation structure (VI, RC, LC, PG) is generated by GPT-5.2 and used as teaching supervision. The paper states “To avoid manual summarization, we provide GPT-5.2 with the complete four-visit longitudinal records. . . “ However, there is no validation mechanism to verify the correctness of these GPT-generated annotations.
- This is problematic because: (1) The model is learning from potentially hallucinated or inconsistent explanations, (2) GPT-5.2 may not have medical expertise to accurately interpret normative deviations and generate clinically appropriate explanations, (3) No gold standard from actual clinicians is used for comparison.
- The reviewer asks (from focus points): “What are the mechanisms put in place for verifying the annotation given by ChatGPT considered as reference for the teaching for the different annotation VI, RC, LC, PG?” - This critical question is not addressed in the paper.
- Specific references: For concerns about LLM-generated medical data: Thirunavukarasu et al. , “Large language models in medicine,” Nature Medicine 2023; Singhal et al. , “Towards Expert-Level Medical Question Answering with Large Language Models,” arXiv 2023.
Weakness 2: Missing Statistical Significance Testing
- Issue: No statistical significance tests are reported for the forecasting results in Table 1.While the proposed method shows improvements (e. g. , 84.2% ACC vs. 76.4% for UniLM), there are no p-values, confidence intervals, or significance tests (e. g. , DeLong test for AUC comparisons, McNemar’s test for accuracy).
- Given that results are reported on a test set of only 297 clips, statistical testing is crucial to establish whether observed differences are significant or could occur by chance.
- Reviewer’s focus note: “Need statistical test DeLong for AUC”
- Specific references: DeLong et al. , “Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach,” Biometrics 1988.
Weakness 3: Inconsistent and Unexplained Baseline Comparisons
- Issue: The reviewer notes “Why is LaTIM missing in the Visit missing comparison?” suggesting that some baselines are not consistently evaluated across conditions.
- Some baseline methods (DTGPT, NoteLM) are given an unfair advantage by providing ground-truth volumetric measurements directly as input, while other methods must extract this from imaging. The paper acknowledges this (“DTGPT and NoteLM are unable to process MRI images, instead ground-truth volumetric measurements are directly provided”) but doesn’t fully account for this advantage in interpretation of results.
- The paper claims improvements over NoteLM on explanation quality, but NoteLM has access to privileged information (ground-truth volumes), making the comparison less meaningful.
- 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
Major Comments: 1.CRITICAL: Provide validation mechanism for GPT-generated hierarchical interpretations (VI, RC, LC, PG). Consider having expert clinicians validate a subset of generated explanations against ground truth, or compare GPT-generated vs. human-generated explanations on a validation set. 2.Add statistical significance testing (DeLong test for AUC, McNemar’s test or bootstrap for accuracy) for all comparisons in Table 1. 3.Clarify why LaTIM is missing from visit missing comparisons and ensure all baselines are evaluated consistently. 4.Provide more details on clinician evaluation methodology: How many clinicians? What were their qualifications? How were the 10 cases selected? Inter-rater agreement?
Minor Comments: 1.Consider ablation studies isolating the contribution of: (a) bipartite graph encoding vs. simpler missing data handling, (b) dual population memory vs. no memory, (c) hierarchical interpretations vs. single-level explanations. 2.Provide training time and computational cost comparisons with baselines. 3.The claim “first MLM-based longitudinal forecasting framework” should be carefully verified - provide explicit comparison with any concurrent work. 4.Include failure case analysis: When does the model generate incorrect explanations? When does forecasting fail?
For extension:
1.How sensitive is the method to the quality of GPT-generated teaching labels? What happens if GPT-5.2 generates incorrect interpretations? 2.Can the framework handle variable numbers of visits (not just 4)? 3.How does performance degrade with increasing amounts of missing data?
- 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?
Recommendation: Accept. The paper makes significant methodological contributions with strong empirical validation. Weaknesses can be addressed in camera-ready version (add statistical tests, clarify baselines, provide GPT label validation).
Positive Factors:
- Novel and ambitious work bringing multimodal language models to interpretable longitudinal disease forecasting - addresses a critical need for explainable medical AI
- Strong technical contributions: bipartite graph encoding for missing data, dual population memory, hierarchical interpretation framework
- Comprehensive evaluation with impressive results: 84.2% ACC, 89.3% AUC, outperforming 9 baselines including recent LLM/MLM methods
- Multi-faceted evaluation methodology: forecasting metrics, NLG quality, LLM-based judging, AND clinician evaluation
- Qualitative results (Fig. 3) demonstrate clinically meaningful, structured explanations that align with normative evidence
- Tackles real clinical challenges: missing data, heterogeneous modalities, need for interpretability
- Use of public ADNI dataset enables reproducibility
- Code release commitment
Concerns:
- Primary concern: Lack of validation for GPT-generated teaching labels is a methodological weakness, but doesn’t invalidate the overall contribution since the model still achieves strong empirical results
- Missing statistical significance testing (but this can be added)
- Some baseline comparison inconsistencies
- Missing some implementation details
- Reviewer confidence
Very confident (4)
- [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.
My concerns were tackle and I was already convinced that the paper should be accepted.
Review #2
- Please describe the contribution of the paper
1.The authors aim to develop a model to explain the disease progression via text generation using LLM. 2.The authors proposed Bipartite graph encoding module to encode incomplete multimodal longitudinal records.
- 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 authors introduced a graph-based multimodal encoder tailored to irregular, missing longitudinal data by treating time and variables as a bipartite graph with edge embeddings for observed values and a query-time aggregation mechanism.
- 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 authors mentioned that their main contribution is to design Bipartite graph encoding (BGE). However, the authors have not emphasized the good points of this module. What is the novelty of this module? In addition, what is L_da^i loss in equation (1)? How is it computed? 2.I wonder if there is some data leakage. Why 399 participants but have 2342 clips. If one participant has more than one clip. I wonder if these clips from same participant appear in both training and testing sets. The author should clarify this 3.The author mentioned that this is multimodal longitudinal forecasting framework but looks like they only use MRI scan along with tabular data. I think we should not consider tabular data as a modality. 4.What feature do tabular data contain (age, sex, education level, gene, …)? 5.The author mentioned the limitation of existing method are irregular and incomplete multimodal data. But it’s not clear how the authors solve this problem. How about a subset of MRI scan or tabular data are missing? 6.Figure 1 is also not very clear for readers to understand how the proposed model work. The author should redraw it. 7.In Table 3, using only image without tabular data, the accuracy and F1 score drop ~30% compared to the use of image and tabular data. Could the authors explain this? Does the model dominate tabular data? 8.The authors lack of compare the proposed method with SOTA method in multimodal longitudinal predictions with missing data such as: +https://doi.org/10.1016/j.inffus.2022.11.028
- https://doi.org/10.1109/JBHI.2024.3472462
- https://doi.org/10.1016/j.media.2022.102643
- Please rate the clarity and organization of this paper
Poor
- 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 describtion of the proposed method is not clear at the current version
- 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 #3
- Please describe the contribution of the paper
The authors propose ENIGMA, a framework for forecasting Alzheimer’s disease progression from longitudinal multimodal records (3D MRI + tabular clinical/demographic variables across multiple visits) that produces both a categorical diagnosis at a future visit and a structured natural-language explanation. The system has three main components. First, a Bipartite Graph Encoding (BGE) module represents each subject’s multi-visit records as a graph where variable nodes connect to time nodes, with edges present only where data is observed; this is intended to handle missingness and modality heterogeneity, and is processed with a 4-layer graph neural network using multi-head attention. Second, a Dual Population Memory (DPM) module maintains two memory banks — one over multimodal record embeddings, one over text-derived “impression” embeddings — trained with a SwAV-style swap loss to capture cohort-level progression patterns. Third, a multimodal language model (Qwen3-4B-Instruct with LoRA) consumes the individual embedding, the population memory, and a task instruction, and is trained to generate explanations at multiple temporal scales (Recent Changes, Long-term Changes, Progression) alongside the diagnostic prediction. The training supervision for the explanation task is itself synthetic: GPT-5.2 is given the structured longitudinal data (cognitive scores, brain volume quantiles, w-scores, slope labels) and asked to write the reference explanations, which then serve as teacher-forcing targets. Evaluation is on 2,342 four-visit clips from 399 ADNI participants, with subjects who converted to AD before visit 4 excluded. Forecasting is scored against the actual visit-4 diagnosis (ACC, SEN, SPE, F1, AUC); explanations are scored with NLG metrics (BLEU, METEOR, ROUGE-L, BERTScore), an LLM-as-judge using GPT-5.2 across six dimensions (Consistency, Historical Fidelity, Prognostic Fidelity, Structural Specificity, Helpfulness, Human-likeness), and a small clinician panel. Reported forecasting accuracy of 0.842 substantially exceeds the strongest baseline at 0.764.
- 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 forecasting result itself is the cleanest piece of evidence and looks genuinely strong. The visit-4 diagnostic label is an objective ground truth independent of the language modeling, and a jump from 0.764 to 0.842 accuracy over the best baseline is not trivial. The robustness experiments under missing tabular variables and missing visits (Fig. 4c, 4d) are a sensible addition and the model appears to degrade gracefully. The bipartite graph formulation for handling irregular and incomplete longitudinal records is a reasonable design choice and builds on a defensible line of prior work (GraFITi, GRAPE). The motivation — that visualization-based interpretability via image synthesis demands expert reading, while structured language explanations are more directly useful — is well-taken.
- 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.
Circularity in the explanation evaluation. This is my main substantive concern, and it runs through several parts of the evaluation in a way that is not adequately acknowledged.
The reference explanations used as training targets are written by GPT-5.2 from the structured input data. The same GPT-5.2 is then used as the LLM-judge to score candidate explanations across the six clinical-correctness dimensions in Table 2.The model under evaluation is, to a meaningful extent, being trained to imitate the judge’s writing on inputs the judge has already seen in structured form, and is then graded by that judge. The first row of Table 2 makes this concrete: “GPT (GT)” is itself scored and receives 4.04–4.96 across the six dimensions, i.e., the judge rates its own outputs as near-ceiling. The authors do not comment on this, and the implications for the relative ordering of methods in Table 2 are unclear.
The clinician evaluation in Fig. 2 is the right instrument for breaking this loop, but as reported it cannot do that work. Ten explanations, across five methods, scored on six dimensions by two clinicians, with no inter-rater agreement statistics, no confidence intervals, no significance tests, no rubric beyond the one-sentence definitions in the text, no statement on blinding or order randomization, and no information about the clinicians’ subspecialty or experience reading ADNI-style cases. With those gaps, the radar plot in Fig. 2 is suggestive at best. I would want to see, at minimum: the full prompt and rubric used by the LLM-judge (in supplementary); inter-rater agreement (Cohen’s κ or ICC) for the clinician panel; a substantially larger clinician sample and ideally a blinded pairwise preference protocol rather than absolute Likert scores, which are notoriously poorly calibrated across raters; and explicit discussion of the circularity introduced by using GPT-5.2 as both supervisor and judge.
Asymmetric baselines for explanation quality. DTGPT and NoteLM are given ground-truth volumetric measurements as text input, while image-based methods (including ENIGMA) must extract this from MRI. The authors note this and treat it as a handicap they overcome, but it cuts in both directions: the text-only baselines lack feature-extraction noise, which should favor their explanation quality, but they also lack any mechanism to identify imaging findings the volumetrics don’t capture. It is hard to read Table 2 cleanly given this asymmetry, and a more apples-to-apples baseline — say, NoteLM given the same MRI features ENIGMA extracts — would strengthen the case. Single-cohort evaluation. All results are on ADNI, with one preprocessing pipeline (FreeSurfer + MNI152), one LLM backbone (Qwen3-4B), and an exclusion criterion (no pre-visit-4 converters) that selects for slower progressors. There is no external cohort validation. For a paper whose central claim includes “population-aware” forecasting, the absence of any out-of-distribution test is notable. Minor / Clarification Questions The notation nn.Embedding should be spelled out (it’s a learnable lookup table); calling it by its PyTorch class name will not be transparent to all readers. The phrase “L = 4 layers GNN using multi-head attention blocks” deserves a sentence clarifying that the GNN is attention-based message-passing rather than convolutional, since “GNN” is overloaded in the literature. The w-score concept is used heavily but never defined; one sentence (“residual from a covariate-adjusted normative model, divided by residual SD”) would help readers outside the normative-modeling community. The rationale for selecting α_cl = 2.0, α_ca = 1.0, the 25% router sparsity target, and the memory length of 64 (Fig. 4a is shown but the criterion is unstated) should be made explicit. The footnote deferring “detailed implementation documentation” to a post-camera-ready GitHub release is a familiar dodge, and given how much methodological detail is unspecified in the main text, I would want to see at least a methods supplement at review time.
- Please rate the clarity and organization of this paper
Poor
- 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?
Overall impressive performance and reasonable choices, but the fairness of much of the evaluation remains unclear. However, the clinical labeling evaluation is straightforward, and here the method appears to do very well. So, weak accept.
- Reviewer confidence
Not confident (1)
- [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 authors addressed some of the biggest concerns:
1.Evaluation circularity is reasonably mitigated (also, additional use of Sonnet for evaluation criteria is reassuring(
2.Cohen’s κ = 0.79 for interrater reliability is decent.
Review #4
- Please describe the contribution of the paper
1) The paper proposes ENIGMA, an interpretable multimodal language model framework for longitudinal disease prediction. 2) It introduces a bipartite graph encoding scheme to handle incomplete, multimodal, and longitudinal records. 3) The paper further proposes a dual population memory to support individual-level interpretation with population-level patterns.
- 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) This paper focuses on predicting future status from historical follow-up records while requiring the model to provide interpretable explanations. 2) The three types of explanations enable the model to distinguish recent changes, long-term trends, and future progression.
- 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 explanation supervision heavily relies on GPT-generated content. As a result, the model may learn GPT-style explanations rather than explanations that truly reflect clinical expert reasoning. 2) Section 2.1 states that GPT uses complete four-visit longitudinal records and diagnoses to generate explanations, where the progression status describes the change from the most recent record to the target visit. This suggests that the explanation supervision may include the true diagnosis, cognitive scores, or even future imaging statistics from the fourth visit. While using such information as training targets may be acceptable, it raises concerns when evaluating explanation quality. If the reference explanations themselves depend on future ground-truth information, a closer match between the generated and reference explanations does not necessarily indicate that the model has truly inferred these changes from historical data. 3) The paper states that 2342 four-visit clips were constructed from 399 ADNI participants, with a train/validation/test split of 1764/281/297.However, it does not clearly specify whether the split was performed at the participant level. If different clips from the same participant appear in both the training and test sets, the longitudinal prediction task would suffer from serious data leakage, and the reported performance could be overestimated. For multi-visit datasets such as ADNI, this is a critical issue that must be clarified. 4) The method is complex, and its reproducibility is relatively poor. 5) n Table 1, the AUC of the proposed method is 0.893.However, in the modality ablation study in Table 3, the Image + Tabular setting achieves an AUC of 0.932, while the full MLM component in the last row again reports an AUC of 0.893.6) Figure 1 is difficult to read, and the paper introduces too many variables, which substantially reduces the overall readability.
- Please rate the clarity and organization of this paper
Poor
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The explanation supervision and evaluation rely excessively on GPT, while the paper suffers from poor readability and limited reproducibility.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The authors clearly state that the data split was performed at the participant level, which substantially alleviates the main concern about longitudinal data leakage. However, they still do not sufficiently explain whether GPT-generated explanations truly reflect clinical reasoning, whether the reference explanations rely on future ground-truth information, whether text-matching evaluation is fair, how reproducible the method is, why the AUC results are inconsistent, or how the readability of the figures can be improved. In particular, the AUC inconsistency between Table 1 and Table 3, as well as the potential use of future information in explanation supervision, require direct clarification.
Author Feedback
We thank reviewers/ACs and clarify key concerns below:
Dataset [R2Q2Q4/R4Q3]: We split dataset in participant level (319 train/41 val/39 test), yielding clips with no participants overlap. Tabular variables include 6 demographics (e.g. Sex/Age/Edu/Ethnicity/Race/Marry), APOE4, and 6 cognitive scores. MRI data are complete for all samples, whereas 22% records exhibit tabular missing values.
GT-reliability [R1/3/4Q1]: We construct GT similarly to how junior doctors are taught to trace evidence given target status. We also invite two senior clinicians to recognize GT’s quality across all 6 dims (CS 4.8 HF 4.9 PF 4.6 SS 4.7 HP 4.5 HL 4.8) together with Fig 2, suggesting GT’s reliability. Notably, similar GT-summary setups have been adopted in recent medical reasoning works, such as GEM(NIPS25) and ClinRaGen(ACL25), which supports our design.
LLM-judge circularity [R3Q1]: Using same LLM in GT-summary and LM-judge is adopted by prior works (GEM) since two stages are independent with different contexts and tasks. LM-judge‘s prompt instructs LM to rate candidate hieratically using given data, stats, rubric and output format, e.g. “You are a clinical reasoning evaluator. Input… Evaluate quality on… Output STRICT JSON…”. Rubric emphasizes objective evidence, e.g. “missing a key V1-V3 trend should score 3” in historical fidelity (full prompt and rubric in anonymized link). We further use Sonnet-4.6 as LM-judge across metrics, GPT-summarized GT achieves avg score of 3.86 as upper bound and ENIGMA ranks best among all methods (3.70 vs. suboptimal ViTB-LM 3.51).
Human evaluation [R3Q1/R4Q2]: Beyond GT text matching, the evaluation also includes historical and prognostic fidelity in both the LM-judge and clinician rubrics. Specifically, we include two clinicians’ blinded ratings for 10 cases, providing them with full records, labels, normative stats and shuffled text candidates. Two clinicians achieved high agreement (Cohen’s κ 0.79).
Fairness [R1Q3/R3Q2]: DTGPT/NoteLM cannot read MRI, we provide volumetric features as a realistic upper-bound proxy, and we also include MLM baselines, e.g. ViTB-LM, with MRI encoders and structured tabular for symmetric comparisons.
BGE novelty [R2Q1/5]: Unlike prior pre-visit fusion or missing completion works, BGE models fine-grained cross-time interactions in a unified graph. It is trained directly with missing value and irregularity thanks to edge disconnection design and learnable time encoding. L_da^i is a cross-entropy loss aligning the representation of visit i with disease status.
Reproducibility [R3Q4]: Full settings/code are online (sent to ACs due to Rebuttal Guide). Some important details are shared below. Experiments use 1 L40S GPU and cost 3.4h pretraining (batch size 16/100 epochs) and 2.0h finetuning (batch size 4/6 epochs) with lr=3e-5 and anneal decay. Checkpoints are selected by validation AUC with early stopping. We empirically search αcl from 1 to 3, αda in 0.5 1.0 1.5, sparsity 1/8 1/4 1/2. Statistic significance [R1Q2]: McNemar tests (p<0.05) are conducted to assess statistical significance showing ENIGMA significantly improves accuracy.
Modality ablation justification [R4Q5]: First panel in Table 3 presents an ablation study of the multimodal encoder (BGE) under different modality inputs, rather than the full MLM pipeline. which leads to AUC difference.
Tabular data domination [R2Q7]: Experimental results indicate that tabular data contribute more, likely due to their closer relevance to AD diagnosis (e.g., cognitive scores). However, MRI still provides complementary anatomical cues.
Visit missing comparison [R1Q3]: Fig 4 tests extra input missing at inference stage. As LaTiM utilize the single record for inference, it is not directly comparable in this setting.
Related work [R2Q8]: Direct comparison is limited as W1/W2 lack public code. We compared W3 by adjusting into status prediction without PET, obtaining F1=0.652/AUC=0.796.We will discuss W1–3 in Related Work.
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.
This paper presents an ambitious and potentially impactful multimodal language model framework for interpretable longitudinal AD forecasting, with one reviewer strongly supporting acceptance while others provide mixed assessments (scores: 5 / 4 / 3 / 2) . While reviewers acknowledge strong empirical performance and novelty in combining forecasting with structured explanation generation, several critical concerns remain. In particular, the reliance on GPT-generated supervision without validation, potential data leakage in longitudinal splits, and circularity in explanation evaluation (e.g., using the same LLM for supervision and judging) need to be carefully addressed. Additional issues include unclear methodological details, baseline fairness, and reproducibility; the authors are encouraged to respond to these points in the rebuttal to clarify the validity and robustness of the proposed approach.
- 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 the rebuttal, two reviewers recommended acceptance, one reviewer suggest reject, and one reviewer did not respond (with the original rating 2). Most of the concerns are raised due to clarity and reproducibility, which may not be easy considering the complexity of the method. The authors should try to improve the figures and descriptions, and include the discussions in the rebuttal in the revision.
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 reviewing the authors’ rebuttal, I have revised my initial assessment to Accept. The authors have provided a thorough, evidence-based response that directly addresses the core concerns raised by the reviewers, including participant-level split to prevent data leakage, validation of GPT-generated ground truth via clinician assessment, mitigation of evaluation circularity using Sonnet-4.6 as an independent judge, high clinician inter-rater agreement (Cohen’s κ = 0.79), statistical significance testing with McNemar tests, clarification of AUC inconsistencies, and provision of an anonymized code repository for reproducibility.
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.
The rebuttal substantially alleviates the key concerns on participant-level data leakage, GPT-generated supervision reliability, LLM-judge circularity, statistical testing, and reproducibility, and although some issues about future-information-dependent explanations and table consistency remain, the strong empirical performance and positive post-rebuttal assessments support acceptance.
