acceptodds
Withdrawn from ICLR 2027

MEDUSA: Graph-Conditioned Flow Matching for Missing Modalities in EHRs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Electronic health records (EHRs) are inherently multimodal, yet complete records are often unavailable because entire modalities may be missing. Existing approaches either reconstruct missing information or adapt prediction models to incomplete inputs, while graph based methods exploit relationships among patients but typically rely on deterministic message passing. We introduce MEDUSA, a graph conditioned generative framework that combines patient similarity with flow matching to impute missing modality representations. MEDUSA constructs a patient graph from observed modalities and learns a separate flow model for each maskable modality, allowing information from related patients to guide generation. Training follows an expectation maximization (EM) inspired alternating procedure that fits the generative models and jointly optimizes downstream prediction. Across multiple EHR prediction tasks, MEDUSA achieves strong predictive performance, particularly at higher missingness. Its imputations also better preserve the patient similarity structure of fully observed data than competing methods. These results show that combining relational structure with generative imputation is effective for learning from incomplete multimodal EHRs.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Graph Diffusion Models, Missing Modalities, Electronic Health Records

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_HQTqOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This work aims to address the problem of missing modality information in electronic health record (EHR) tasks. The authors propose MEDUSA, a graph-conditioned generative framework for handling missing modalities in multimodal EHR data. MEDUSA combines patient similarity graphs with generative models and employs an EM-inspired alternating optimization procedure. Experiments are conducted to evaluate the effectiveness of the proposed method.

Critical strengths
  1. The research topic of addressing missing modalities in EHR data is practically relevant and has received considerable attention.
  2. Overall, this work is easy-to-follow.
Critical weaknesses
  1. The novelty of this work appears limited and remains unclear. The proposed method combines patient similarity graphs and generative models to address missing modalities in EHR data. However, both techniques have been widely studied in related domains. The unique insights offered by this combination for the target task appear limited. The authors should clarify the fundamental differences between MEDUSA and existing approaches and highlight its novel contributions.
  2. The motivation should be enhanced. This works mainly combines the patient graphs and generative modeling to handle the missing modality issues in EHR. However, the motivation behind this combination is limited and unclear. The distinct advantages of this technical design are missing. The authors should clarify why these two components are complementary and explain how their integration addresses the limitations of existing methods more effectively than alternative approaches.
  3. The discussion for limitations of current methods should be enhanced. In current version, the discussion is unclear and lack strong supports. In line 42-43, the authors simply mentioned that "they primarily rely on graph aggregation, fusion, or representation alignment, and therefore do not explicitly model a distribution over the missing modality representations." However, the exact limitations for these methods are not clearly explained. Moreover, the rationale for explicitly modeling a distribution over missing modality representations remains unclear. The authors should provide a more rigorous analysis of the limitations of existing approaches and explain the unique advantages of distribution modeling in addressing missing modalities.
  4. The technical rationales should be enhanced. The authors did not clearly discuss their motivation for technical designs. For example, why are generative models suitable for addressing missing modalities in EHR data, and what advantages do they offer over alternative approaches? Furthermore, why is patient relational structure incorporated into the generative models, and what unique benefits does this integration provide? The authors should provide a clearer justification for these design choices and explain how each component contributes to the overall effectiveness of MEDUSA.
  5. The performance improvements achieved by MEDUSA over existing baselines appear limited and inconsistent. For example, Table 1 shows that MEDUSA achieves only marginal improvements over the strongest baselines in some settings, including an improvement of just 0.03 in AUROC for eICU-Readmission. More importantly, MEDUSA is substantially outperformed by baseline methods on eICU-Mortality. These results raise concerns about the effectiveness and robustness of the proposed method. In addition, statistical significance tests are missing, making it difficult to determine whether the observed improvements are statistically meaningful.
Questions for authors

Please refer to the Weakness seciton.

Reviewer_8j4TOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MEDUSA, a graph-conditioned flow matching framework for imputing missing modality representations in EHRs. In particular, it combines patient similarity graphs with modality-specific flow matching to generate missing modality representations and employs an EM-inspired alternating optimization procedure for downstream prediction.

Critical strengths
  1. The paper addresses an important practical challenge in multimodal EHR learning, where missing modalities can significantly affect downstream prediction performance.

  2. The proposed framework integrates patient similarity information with flow matching to generate missing modality representations.

Critical weaknesses
  1. MEDUSA combines several existing techniques, including patient similarity graphs, flow matching, and alternating optimization. In particular, DiffPuter already employs an EM-based alternating optimization framework for generative imputation. Also, M3Care leverages patient similarity to recover missing modality representations, and SADI incorporates patient similarity into diffusion-based imputation for temporal EHRs. Although MEDUSA extends these ideas to graph-conditioned flow matching, the methodological contributions beyond these existing approaches are not sufficiently established. Overall, the novelty of the proposed framework is limited.

  2. The motivation for adopting flow matching is not sufficiently clear. Although the paper motivates generative imputation as an alternative to deterministic approaches, it does not adequately explain why flow matching is particularly suitable for missing modality recovery compared to other generative methods, such as diffusion models. Moreover, MEDUSA averages multiple generated samples into a single embedding for downstream prediction, without analyzing whether modeling a conditional distribution provides additional benefits over directly predicting the missing embedding. The experiments also lack controlled comparisons that isolate the contribution of flow matching from other components.

  3. The paper does not compare the proposed approach with alternative conditioning strategies. In particular, directly conditioning the generative model on observed modality embeddings may provide comparable performance without requiring a patient similarity graph. A controlled comparison would help establish whether graph-based conditioning offers meaningful advantages over simpler alternatives.

Questions for authors

Please see the concerns detailed in the Weaknesses.

Reviewer_d78pOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed a framework to handle missing modalities in EHR. The research question itself is indeed an important gap in EHR research; however the method design, experiments and evaluations are not convincing given some major technical issues therefore the current results are not sufficient to justify a meaningful and novel method.

Critical strengths

The overall writing is not difficult to follow.

Critical weaknesses

CW1. The assumption that demographic are always observed is not true and the Missingness simulation in Section 4 is unrealistic.

On page 6 "Missingness simulation" the authors simulated missing modalities in MCAR , which is usually not the case in real-world heath data, and this is based on selected complete modalities (".... retain only admissions with all modalities available, so that missingness can be simulated in a controlled way. ") The experiments then built up on this cherry picking ideal scenarios and never analyze MAR/MNAR scenario.

In addition, the assumption that demographic are always observed is not true in real world health data and this assumption can affect the methods significantly because apparently equation (4) depends on this.

In summary both the method and experiment depend on heavy assumptions that are difficult to hold in EHR data

CW2. The reported evaluation is not sufficient and properly evaluated.

The paper evaluated the proposed method by downstream prediction performance (Table 1) but not the imputation itself. What is more meaningful in MVI problems are whether the method could recover the original data distribution if possible. Since this is a simulated missingness this could be of course be computed (by using h^\hat{h} for instance).

In addition, the hyperparameter-tuning details are somewhat ambiguous. It appears that several fixed hyperparameters are shared across all models and both prediction tasks (Section 4 "Baselines and model selection". However, the method-specific hyperparameters for each baseline should be tuned independently for each prediction task; the current experimental design raises my concerns about the fairness of the comparisons.

Questions for authors

See Above.

Reviewer_HXtWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MEDUSA, a graph-conditioned flow matching framework for imputing missing EHR modalities. It combines patient similarity graphs with modality-specific generative models and EM-inspired alternating optimization. Experiments on three EHR datasets demonstrate competitive predictive performance and improved preservation of patient similarity structures.

Critical strengths

The paper evaluates the proposed method across multiple EHR datasets and missingness settings, with extensive baseline comparisons and ablation studies.

Critical weaknesses
  1. Limited Evidence for Graph Conditioning. The contribution of graph conditioning is not convincingly demonstrated. In Table 2, removing the graph even improves performance at 10% missingness, while improvements at higher missingness are relatively small. This raises questions about the necessity of the central graph-conditioning mechanism.
  2. Insufficient Validation of Generative Imputation. The flow models use median-filled embeddings as training targets for missing modalities, potentially introducing bias into the learned distribution. Despite having access to complete patient records, the paper lacks direct evaluations of imputation fidelity, relying primarily on downstream performance and patient similarity preservation.
Questions for authors

Graph Conditioning: Could the authors further clarify the contribution of graph conditioning, given its limited or negative impact in some ablation settings? Generative Imputation: Could the authors provide direct evaluations of imputation quality and investigate the potential bias introduced by median-filled training targets?

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

Related papers

Discussion (0)

Sign in to comment.