Student or educator? Sign up with your school email and get 6 months of Pro free.

The DOI matches its Crossref record; 77 of the 77 references it cites were found in a registry or an index; 43 of the 65 sources its claims cite were checked for support.

MoMA: a mixture-of-multimodal-agents architecture for enhancing clinical prediction modelling

Open access

Abstract

Multimodal electronic health record (EHR) data provide richer, complementary insights into patient health compared to single-modality data. However, effectively integrating diverse data modalities for clinical prediction modeling remains challenging due to the substantial data requirements. We introduce a novel architecture, Mixture-of-Multimodal-Agents (MoMA), designed to leverage multiple large language model (LLM) agents for clinical prediction tasks using multimodal EHR data. MoMA employs specialized LLM agents ("specialist agents") to convert non-textual modalities, such as medical images and laboratory results, into structured textual summaries. These summaries, together with clinical notes, are combined by another LLM ("aggregator agent") to generate a unified multimodal summary, which is then used by a third LLM ("predictor agent") to produce clinical predictions. Evaluating MoMA with different modality combinations and prediction settings, MoMA outperforms existing methods on three prediction tasks using private datasets, highlighting its enhanced accuracy and flexibility across various tasks.

Overview

CoChat Summary

## TL;DR

MoMA is a modular mixture-of-multimodal-agents architecture that converts non-text EHR modalities into text summaries and uses an LLM pipeline for clinical prediction. On three private UW Health tasks, it outperformed baseline methods, with reported gains including macro-F1 near 0.85 and micro-F1 above 0.90 for chest trauma severity, macro-F1 above 0.75 and micro-F1 close to 0.90 for spine trauma in the multitask setting, and AUROC around 0.75 with AUPR near 0.50 for unhealthy alcohol use screening.

## Key findings

  • MoMA was evaluated on three private UW Health tasks: chest trauma severity stratification, multitask chest and spine trauma severity stratification, and unhealthy alcohol use screening.
  • For chest trauma severity stratification, MoMA achieved macro-F1 scores near 0.85 and micro-F1 scores above 0.90.
  • For spine trauma stratification in the multitask setting, MoMA achieved macro-F1 scores above 0.75 and micro-F1 scores close to 0.90.
  • For unhealthy alcohol use screening, MoMA achieved AUROC of around 0.75 and AUPR near 0.50; a retrained 1D-CNN on the same small cohort underperformed at AUROC 0.641 and AUPR 0.325.
  • In ablation studies, removing the non-text modality reduced performance, indicating the gains were not due only to text understanding but also to use of radiographs or lab measurements.

## Methods

  • Design: MoMA uses specialist LLM agents to convert each non-text modality into text, an aggregator LLM to fuse summaries with the original clinical note, and a predictor LLM to produce the final classification; it is framed as a plug-and-play extension of Mixture-of-Experts and Mixture-of-Agents ideas.
  • Data/sample: Private UW Health datasets with temporal train/test splits; chest trauma tasks used 2,722 unique patients from January 2015 to December 2019, and unhealthy alcohol use screening used 2,096 consented patients from September 2021 to February 2024.
  • Analysis: Compared MoMA against ClinicalBERT, LLaVA-Med, cross-attention fusion, and sparse MoE baselines; evaluated multiclass tasks with macro-/micro-F1 and binary task with AUROC/AUPR; also conducted subgroup and ablation analyses.

## Limitations

  • The paper states that agent interactions in MoMA are “relatively simple,” and better communication/coordination among agents may improve performance.
  • The authors note inherent LLM issues such as hallucination and omitting necessary signals may still affect classification performance; this is stated in the paper.
  • The paper says MoMA may increase inference computational cost because multiple agents must run; this is stated in the paper.
  • Evident from the text: the main validation used private datasets, so external researchers cannot directly access the original data.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41746-025-02219-4

Cites

77 of 77 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Architecture of MoMA. Non-plain-text modalities, including medical images and tabular EHR data, are transformed into text summaries by specialist agents. These agent-generated summaries are then appended to the original clinical notes and passed to the aggregator agent, which produces a comprehensive and concise summary from the combined text. This aggregated summary is subsequently fed into a predictor agent to generate output predictions. Given the well-known capabilities of LLMs in clinical summarization, both the specialist and aggregator agents operate in a zero-shot setting, with only the predictor agent requiring training or fine-tuning, which reduces computational costs during the training phase.
Fig. 1 Architecture of MoMA. Non-plain-text modalities, including medical images and tabular EHR data, are transformed into text summaries by specialist agents. These agent-generated summaries are then appended to the original clinical notes and passed to the aggregator agent, which produces a comprehensive and concise summary from the combined text. This aggregated summary is subsequently fed into a predictor agent to generate output predictions. Given the well-known capabilities of LLMs in clinical summarization, both the specialist and aggregator agents operate in a zero-shot setting, with only the predictor agent requiring training or fine-tuning, which reduces computational costs during the training phase.CC BY-NC-ND · PubMed Central
Fig. 2 Comparison of discriminative performance and 95% confidence intervals between MoMA, the published SOTA baseline, and the cross-attention, MoE-based fusion, and fine-tuned LLaVA-Med. Dotted lines mark the strongest baselines, LLaVA-Med for chest trauma and multitask chest and spine trauma stratification, and the published baseline for the unhealthy alcohol use screening. Confidence intervals were calculated using bootstrap methods. On the unhealthy alcohol use screening task, we converted tabular data to text and passed it together with free text to LLaVA-Med without any image modalities. MoMA consistently outperforms all applicable baselines across these clinical tasks.
Fig. 2 Comparison of discriminative performance and 95% confidence intervals between MoMA, the published SOTA baseline, and the cross-attention, MoE-based fusion, and fine-tuned LLaVA-Med. Dotted lines mark the strongest baselines, LLaVA-Med for chest trauma and multitask chest and spine trauma stratification, and the published baseline for the unhealthy alcohol use screening. Confidence intervals were calculated using bootstrap methods. On the unhealthy alcohol use screening task, we converted tabular data to text and passed it together with free text to LLaVA-Med without any image modalities. MoMA consistently outperforms all applicable baselines across these clinical tasks.CC BY-NC-ND · PubMed Central
Fig. 3 Comparison of model performance across racial and sex subgroups. We excluded the comparison for the white vs. non-white subgroups in the unhealthy alcohol use screening task due to the limited number of cases in the non-white subgroup. Statistically significant differences in performance within each group are marked with asterisks. MoMA achieves the highest and consistent performance across all subgroups in these clinical tasks.
Fig. 3 Comparison of model performance across racial and sex subgroups. We excluded the comparison for the white vs. non-white subgroups in the unhealthy alcohol use screening task due to the limited number of cases in the non-white subgroup. Statistically significant differences in performance within each group are marked with asterisks. MoMA achieves the highest and consistent performance across all subgroups in these clinical tasks.CC BY-NC-ND · PubMed Central
Fig. 4 Ablation study evaluating the impact of removing non-text modalities from MoMA’s input. MoMA consistently achieves superior performance compared to its text-only counterparts, demonstrating that performance improvements arise from the effective utilization of non-text modalities.
Fig. 4 Ablation study evaluating the impact of removing non-text modalities from MoMA’s input. MoMA consistently achieves superior performance compared to its text-only counterparts, demonstrating that performance improvements arise from the effective utilization of non-text modalities.CC BY-NC-ND · PubMed Central
Fig. 5 The upper section showcases the output of the chest X-ray specialist agent, which excludes the presence of severe conditions, enabling MoMA to accurately classify the case as a moderate injury. The lower section provides an example where clinical text and lab results are condensed into a concise summary by the specialist and aggregator agents. This summarized information allows the predictor agent to make predictions with improved interpretability compared to directly using all input text for classification.
Fig. 5 The upper section showcases the output of the chest X-ray specialist agent, which excludes the presence of severe conditions, enabling MoMA to accurately classify the case as a moderate injury. The lower section provides an example where clinical text and lab results are condensed into a concise summary by the specialist and aggregator agents. This summarized information allows the predictor agent to make predictions with improved interpretability compared to directly using all input text for classification.CC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

20 Valid · 7 Not valid · 9 Not checked · 97 Own results and interpretations

42 more claims · 4 with their cited sources checked

Create a free account to see all 133 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 40
  • Probably not a substantial claim. 64

Create a free account to see these claims.

Citations

Cites

Show all 52 references

Research works printed but not linked: 13

  1. Li, J., Li, D., Xiong, C. & Hoi, S. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, 12888–12900 (PMLR, 2022).
  2. Radford, A. et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 8748–8763 (2021).
  3. Hayat, N., Geras, K. J. & Shamout, F. E. Medfuse: multi-modal fusion with clinical time-series data and chest x-ray images. In Machine Learning for Healthcare Conference, 479–503 (PMLR, 2022).
  4. Li, J., Li, D., Savarese, S. & Hoi, S. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, 19730–19742 (PMLR, 2023).
  5. Peng, Z. et al. Kosmos-2: Grounding multimodal large language models to the world. In International Conference on Learning Representations (2024).
  6. Driess, D. et al. Palm-e: an embodied multimodal language model. In International Conference on Machine Learning (2023).
  7. Wang, J., Wang, J., Athiwaratkun, B., Zhang, C. & Zou, J. Mixture-of-agents enhances large language model capabilities. In International Conference on Learning Representations (2025).
  8. Jiang, B. et al. Multi-modal and multi-agent systems meet rationality: A survey. In ICML 2024 Workshop on LLMs and Cognition (2024).
  9. Du, Y., Li, S., Torralba, A., Tenenbaum, J. B. & Mordatch, I. Improving factuality and reasoning in language models through multiagent debate. In International Conference on Machine Learning (2023).
  10. Li, J., Zhang, Q., Yu, Y., Fu, Q. & Ye, D. More agents is all you need. Trans. Mach. Learn. Res. (2024).
  11. Ben Abacha, A., Hasan, S. A., Datla, V. V., Demner-Fushman, D. & Müller, H. Vqa-med: Overview of the medical visual question answering task at imageclef 2019. In Proceedings of CLEF (Conference and Labs of the Evaluation Forum) 2019 Working Notes (9-12 September 2019, 2019).
  12. Shazeer, N. et al. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations (2017)
  13. Xia, P. et al. Rule: Reliable multimodal rag for factuality in medical vision language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 1081–1093 (2024). · DOI · Paper page

Cited by

See all 2

Metadata

Authors
Gao, J., Rahman, M., Caskey, J., Oguss, M., O’Rourke, A., Brown, R., Stey, A., Mayampurath, A., Churpek, M. M., Chen, G., & Afshar, M.
Affiliations
  • University of Wisconsin–Madison (US)
  • Northwestern University (US)
Corresponding author
  • Chen, Guanhua
  • Afshar, Majid
Venue
npj Digital Medicine
Published
2025-12-09
Volume
9
Issue
1
Pages
46-46
ISSN
2398-6352 (ISSN-L)
OpenAlex
W4417156540
Cited by
2
References
52
Citations by year
2026: 2
FWCI
0.80
Type
journal article
Language
en
Open access
Licence
Locations
  • npj Digital Medicine (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • arXiv (Cornell University) (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University) (cc-by)
Funders
  • National Science Foundation, United States (DMS-2054346)
  • National Institutes of Health (R01DA051464)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
Keywords
mixture-of-agents, electronic health records, multimodal data integration, clinical notes, large language models, multimodal summarization

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-09, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-09 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: MoMA: a mixture-of-multimodal-agents architecture for enhancing clinical prediction modelling. https://app.cochat.ai/papers/10.1038/s41746-025-02219-4
The paper
Gao, J., Rahman, M., Caskey, J., Oguss, M., O’Rourke, A., Brown, R., Stey, A., Mayampurath, A., Churpek, M. M., Chen, G., & Afshar, M. (2025). MoMA: a mixture-of-multimodal-agents architecture for enhancing clinical prediction modelling. npj Digital Medicine. https://doi.org/10.1038/s41746-025-02219-4

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.