acceptodds
Withdrawn from ICLR 2027

TaskEvi-Fuse: Task-Adaptive Clinical Evidence Fusion for Multimodal ICU Outcome Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

ICU outcome prediction must integrate structured EHR histories, chest radiographs, and radiology reports, although the contribution of each modality may vary by outcome. We present TaskEvi-Fuse, a multimodal clinical prediction framework that combines Gated Residual Modality Alignment (GRMA), which refines representations through gated residual updates and modality-subset experts, with Validation-Controlled Risk Construction (VCRC) for outcome-specific risk-policy selection and calibration. We evaluate the frozen endpoint-specific primary system separately from a controlled common-bank analysis. On a subject-disjoint MIMIC-IV/MIMIC-CXR cohort, using three initializations on one fixed split, the primary system achieves AUROC/AUPRC values of 0.8438/0.4432 for in-hospital mortality, 0.7419/0.7797 for long-stay, and 0.5895/0.3072 for readmission. Among the evaluated protocol-adapted multimodal comparators, it ranks best across the six reported metrics for both long-stay and readmission, and on four of six mortality metrics. Controlled analyses reveal outcome-dependent patterns: different representation variants and risk policies are favored for different endpoints. These findings highlight the value of evaluating multimodal representation refinement and risk construction across clinical outcomes rather than assuming a uniform effect.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multimodal learning, mixture of experts, representation alignment, risk calibration, clinical outcome prediction

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_ZRW6OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the problem of how to construct task dependent risks while aligning them into a same representation space to help fusion. And they proposed a fusion method for multimodal ICU outcome prediction. The core idea is that fusion benefits from task-adaptive representations rather than enforcing an identical representation across different tasks. Specifically, the approach integrates two main components: Gated Residual Modality Alignment (GRMA), which generates alignment-aware multimodal representations, and Validation-Controlled Risk Construction (VCRC), which applies development-data risk construction to each individual task. Consequently, the framework accommodates distinct execution paths for different clinical endpoints, avoiding an uniform pathway that may also introduce task-irrelevant signals.

Critical strengths
  1. Adaptive fusion is an important problem for multimodal clinical outcome prediction.
  2. It is useful to guide representations by specific tasks, allowing for flexible clinical evidence, risks, and configurations tailored to different endpoints, thereby enhancing prediction adaptability.
Critical weaknesses
  1. Originality: The authors tried to clarify their main contributions, which however are still not clear to me.

  2. Presentation: The paper is hard to read in general.

  • Section 1 advocates that the paper is addressing system-level questions, and the concepts like "endpoint", "executable configuration", etc. are used. It is important to include sufficient background to provide the context to improve readability.
  • Section 2 is hard to read. Better to provide an overview of the different components and the rationale behind the design first before providing details in different subsections.
  • Section 2.2 The key rationale of using GRMA is not clearly presented (even though that is not the key contribution of the paper). ada_alignmoe_v2 is suddenly introduced without much explanation.
  • Section 2.3 refers to Table 1 for more details, but the table is not clearly explaining how to read it. Also, many acronyms like OURS_AlignMOE_ExpertStack_v4, 3.0_bal1, are used without too much background information provided.
  • The other sections share similar problems.
  1. Experiments:
  • The experiments lack significance tests and standard deviations for the reported statistics, making it difficult to determine whether the performance improvements are statistically meaningful.
  • In Table 2, there are some notations highlighting protocol-adopted implementation and simplied protocol-adapted implementation. More explanation is needed. In particular, how to make sure that the performance comparison is fairly conducted.
  • In the controlled representation ablation study (Table 3), the performance margins between variants are too small, especially given the lack of standard deviation. It is hard to conclusively identify the major contributor to the improvement.
  • The analysis fails to adequately explain why the method performs well for long-stay and readmission predictions while exhibiting varying performance for mortality prediction.
  • Computational costs are neither analyzed theoretically nor demonstrated empirically in comparison with baseline methods.
Questions for authors

Please refer to the "Critical Weaknesses" section.

Reviewer_b4UvOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary
  • TaskEvi-Fuse combines structured EHR data, chest radiographs, and radiology reports for mortality, long-stay, and readmission prediction.

  • introduces a gated residual representation module, GRMA, and an outcome-specific procedure, VCRC, for selecting or combining candidate predictions and calibrating the resulting risks.

  • reports two distinct evaluations: an archived primary system with endpoint-specific candidate banks, and a controlled analysis using a common ten-candidate bank.

Critical strengths
  • Separating representation quality, risk aggregation, and calibration helps identify which stage contributes to performance so the evaluation framing addresses a useful practical question

  • it also distinguishes primary results from controlled analyses, acknowledges inconclusive matched comparisons, and avoids interpreting router weights as causal explanations. These disclosures make it possible to assess the actual scope of the evidence.

Critical weaknesses
  • Section 3.1 states that test predictions were viewed during earlier development, leaving unclear which subsequent decisions may have been informed by them. Section 2.1 also states that CXR acquisition and report-signing times are not fully audited against the 48-hour prediction landmark. This leaves potential use of information unavailable at prediction time unresolved.

  • Section 2.2 states that the exact checkpoint/configuration linkage between the specified GRMA implementation and the archived primary alignment candidate is not established.

Questions for authors
  • What development decisions occurred after test predictions were inspected?

  • Can you verify that every included image and report was available by the 48-hour landmark?

  • Can you provide the exact mapping from the GRMA specification to the primary candidate checkpoints and their training partitions? If that mapping cannot be recovered, which results from the controlled implementation alone establish the proposed method’s advantage?

Reviewer_8UQTOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TaskEvi-Fuse combines Gated Residual Modality Alignment (GRMA), which refines EHR, chest X-ray, and report representations and uses modality-subset experts, with Validation-Controlled Risk Construction (VCRC), which handles risk-policy selection, calibration, and thresholding. On MIMIC-IV/MIMIC-CXR, the authors predict in-hospital mortality, long stay, and 30-day readmission. They report the best results among protocol-adapted comparators on all six metrics for long-stay and readmission and on four of six for mortality. They also conclude that controlled analyses reveal outcome-dependent representation and policy choices.

Critical strengths
  • The paper is unusually transparent about test exposure, validation reuse, and missing provenance.
  • It separates the primary system from a controlled analysis.
  • It reports calibration alongside discrimination, uses a subject-disjoint split, and includes ablations.
  • It avoids claims of clinical utility or universal superiority.
Critical weaknesses

The paper’s main message is that the best representation and risk policy may vary by clinical outcome. Methodologically, however, the proposed approach combines several established components, and the authors acknowledge that these ingredients are not new. Although they offer a high-level rationale, namely that aggressive alignment can dilute modality-specific evidence and that useful evidence varies by outcome, they do not explain why this particular design is needed. Nor do their results show that it delivers the intended benefit over simpler alternatives or existing multimodal methods. Without that justification, it is difficult to identify the new insight this work contributes.

Other issues include:

  • The headline results do not evaluate the method as specified.
  • The primary evidence is exploratory and cannot be fully verified.
  • The motivating premise is stated but not demonstrated.
  • The proposed components show no reliable gain over simpler alternatives.
  • Outcome dependence cannot be separated from seed variation.
Questions for authors
  1. Can you evaluate one fully specified GRMA and VCRC system on a sequestered test set, using multiple subject-level splits?
  2. What does each modality add for each endpoint compared with EHR-only prediction, with intervals?

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

Related papers

Discussion (0)

Sign in to comment.