pith. sign in

arxiv: 2606.20757 · v1 · pith:NBIU2A2Xnew · submitted 2026-06-18 · 💻 cs.LG

Evidential Fusion Network for Multimodal Survival Prediction under Missing Modalities

Yucheng Xing , Hailan Mo , Zi Wang , Ling Huang , Mengling Feng This is my paper

Pith reviewed 2026-06-26 18:13 UTC · model grok-4.3

classification 💻 cs.LG
keywords multimodal survival predictionmissing modalitiesevidential fusionDempster-Shafer theoryuncertainty estimationcancer datasetsGaussian Random Fuzzy Numbers
0
0 comments X

The pith

EMMS fuses multimodal survival data by treating missing modalities as vacuous evidence within a Dempster-Shafer framework.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a model that performs survival prediction from multiple clinical data sources even when some sources are absent. It combines predictions from available modalities through an evidential fusion process that accounts for different kinds of uncertainty and the trustworthiness of each source. Missing inputs receive no weight in the combination step, which raises overall uncertainty in a controlled manner. This matters because incomplete records are routine in medical practice, and the method avoids the separate step of creating substitute values for absent data.

Core claim

The EMMS model performs multimodal survival prediction under missing modalities by employing Dempster-Shafer theory and Gaussian Random Fuzzy Numbers for decision fusion. It considers aleatoric and epistemic uncertainty along with modality reliability. Missing modalities are treated as vacuous evidence, which prevents them from interfering with available inputs and leads to increased yet calibrated uncertainty. Experiments on four cancer datasets show state-of-the-art performance with calibrated uncertainty estimates and no extra computational cost.

What carries the argument

The evidential fusion mechanism based on Dempster-Shafer theory and Gaussian Random Fuzzy Numbers that integrates modality predictions while assigning vacuous evidence to absent modalities.

If this is right

  • The approach yields calibrated uncertainty estimates that increase when modalities are absent.
  • No separate generative model or imputation step is required for missing data.
  • Performance reaches state-of-the-art levels on four cancer survival datasets.
  • Computational cost remains comparable to models trained on complete data.
  • Uncertainty estimates are produced alongside the survival predictions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same fusion rule could be applied to other multimodal clinical tasks such as diagnosis or treatment recommendation.
  • Clinical data collection protocols might tolerate more missing entries if this style of fusion is used downstream.
  • The uncertainty signal could be used to trigger additional tests only when missing data meaningfully degrades reliability.

Load-bearing premise

Treating missing modalities as vacuous evidence will prevent interference with available inputs and will automatically produce increased but calibrated uncertainty.

What would settle it

Run the model on the same cancer datasets but with increasing fractions of modalities removed at random and check whether the reported uncertainty rises in proportion to the actual rise in prediction error; persistent under- or over-estimation of uncertainty would falsify the central claim.

Figures

Figures reproduced from arXiv: 2606.20757 by Hailan Mo, Ling Huang, Mengling Feng, Yucheng Xing, Zi Wang.

Figure 1
Figure 1. Figure 1: Overview of the proposed EMMS framework. WSI and gene modalities generate GRFN survival evidence, modeling both aleatoric and epistemic uncertainty. Missing modalities are treated as vacuous evidence and fused via Dempster–Shafer combination to produce calibrated survival predictions. ties [28]. While improving average performance, these methods may hallucinate spurious biological or morphological signals,… view at source ↗
Figure 2
Figure 2. Figure 2: shows that our method remains closer to the diagonal line across all four datasets, indicating better calibration than the compared baselines. This con￾sistent behavior suggests improved confidence estimation and high￾lights the reliability of our uncertainty quantification in survival pre￾diction. In terms of computational cost, our approach is comparable to the imputation-free baseline MUSE, while remain… view at source ↗
read the original abstract

Recent multimodal survival prediction models have demonstrated strong predictive performance by leveraging complementary information across modalities. However, such models generally assume data completeness and exhibit limited robustness toward missing modalities, which are frequently encountered in real-world clinical settings. We propose the Evidential Missing Modality Survival Fusion (EMMS) model for multimodal survival prediction under missing modalities. EMMS offers a straightforward, computationally effective approach to survival analysis without requiring a generative phase for missing data. By employing Dempster-Shafer theory and Gaussian Random Fuzzy Numbers for multimodal decision fusion, it considers both aleatoric and epistemic uncertainty alongside modality reliability for fusion. Moreover, the model treats missing modalities as vacuous evidence, preventing interference with available inputs and naturally reflecting increased uncertainty and calibrated predictions. Extensive experiments on four cancer datasets demonstrate state-of-the-art performance while providing calibrated and interpretable uncertainty estimates under incomplete multimodal observations, without introducing additional computational overhead.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript introduces the Evidential Missing Modality Survival Fusion (EMMS) model for multimodal survival prediction. It fuses per-modality evidence using Dempster-Shafer theory combined with Gaussian Random Fuzzy Numbers, explicitly models both aleatoric and epistemic uncertainty plus modality reliability, and assigns vacuous evidence to missing modalities so that they do not alter the combined mass from observed modalities while increasing total uncertainty. Experiments on four cancer datasets are reported to show state-of-the-art predictive performance together with calibrated uncertainty estimates and no extra computational cost relative to complete-modality baselines.

Significance. If the empirical claims hold, the work supplies a direct, non-generative fusion rule that exploits the algebraic neutrality of vacuous belief under Dempster combination. This yields a computationally lightweight method for incomplete multimodal survival data that simultaneously reports calibrated uncertainty, which is a practical advantage in clinical settings where modality dropout is routine.

major comments (1)
  1. [Abstract] Abstract and experimental section: the central claim of state-of-the-art performance and calibrated uncertainty under missing-modality regimes is asserted without any description of the competing baselines, the statistical tests employed, the train/validation/test splits, or the precise handling of right-censoring; these omissions make the effectiveness claim unverifiable from the available text and are load-bearing for the paper's primary contribution.
minor comments (2)
  1. Clarify the precise definition and parameterization of Gaussian Random Fuzzy Numbers when they are first introduced; the current description leaves the mapping from per-modality survival outputs to mass functions underspecified.
  2. Add a short paragraph contrasting the proposed vacuous-evidence rule with standard imputation or modality-dropout baselines to highlight the claimed computational advantage.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed review and constructive suggestion to improve verifiability of the empirical claims. We address the comment below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract and experimental section: the central claim of state-of-the-art performance and calibrated uncertainty under missing-modality regimes is asserted without any description of the competing baselines, the statistical tests employed, the train/validation/test splits, or the precise handling of right-censoring; these omissions make the effectiveness claim unverifiable from the available text and are load-bearing for the paper's primary contribution.

    Authors: We agree that the abstract, as a concise summary, omits these specifics, and that the experimental section should make the evaluation protocol fully explicit to support the SOTA and calibration claims. In the revised manuscript we will (i) expand the abstract with a brief clause listing the main baselines (e.g., the complete-modality and missing-modality variants of the compared multimodal survival models), (ii) add a dedicated paragraph in the experimental section that enumerates the competing methods, reports the exact train/validation/test splits (including any cross-validation scheme), describes the statistical tests used for significance (e.g., paired Wilcoxon or log-rank tests with p-values), and details the right-censoring handling (standard negative log-likelihood under the Cox partial likelihood with censoring indicators). These additions will render the effectiveness claims directly verifiable while preserving the paper's length and focus. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper's central fusion mechanism applies standard Dempster-Shafer combination rules to per-modality evidence modeled via Gaussian Random Fuzzy Numbers, with missing modalities assigned vacuous belief functions. Vacuous evidence is the neutral element under Dempster's rule by the axioms of the theory itself, so the claimed increase in uncertainty and non-interference with observed modalities follows directly from the external DST properties rather than from any fitted parameter or self-referential definition internal to the paper. No equations or claims in the provided text reduce a prediction to a fitted input by construction, and no load-bearing uniqueness theorem or ansatz is imported via self-citation. The four-dataset experiments are presented as independent empirical validation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Review performed on abstract only; no explicit free parameters, axioms, or invented entities are quantified in the provided text.

axioms (1)
  • domain assumption Dempster-Shafer theory combined with Gaussian Random Fuzzy Numbers can represent both aleatoric and epistemic uncertainty for multimodal fusion when modalities are missing.
    Core modeling choice stated in abstract.

pith-pipeline@v0.9.1-grok · 5688 in / 1212 out tokens · 30377 ms · 2026-06-26T18:13:17.194092+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 4 canonical work pages · 1 internal anchor

  1. [1]

    In: Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining

    Chen, J., Zhang, A.: Hgmf: heterogeneous graph-based fusion for multimodal data with incompleteness. In: Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. pp. 1295–1305 (2020)

    2020

  2. [2]

    In: Proceedings of the IEEE/CVF interna- tional conference on computer vision

    Chen, R.J., Lu, M.Y., Weng, W.H., Chen, T.Y., Williamson, D.F., Manz, T., Shady, M., Mahmood, F.: Multimodal co-attention transformer for survival pre- diction in gigapixel whole slide images. In: Proceedings of the IEEE/CVF interna- tional conference on computer vision. pp. 4015–4025 (2021)

    2021

  3. [3]

    In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion

    Chen, Z., Du, Y., Hu, J., Liu, Y., Li, G., Wan, X., Chang, T.H.: Multi-modal masked autoencoders for medical vision-and-language pre-training. In: Interna- tional Conference on Medical Image Computing and Computer-Assisted Interven- tion. Springer (2022)

    2022

  4. [4]

    Information Fusion51, 259–270 (2019)

    Choi, J.H., Lee, J.S.: Embracenet: A robust deep learning architecture for multi- modal classification. Information Fusion51, 259–270 (2019)

    2019

  5. [5]

    Fuzzy Sets and Systems471, 108679 (2023)

    Denœux,T.:Parametricfamiliesofcontinuousbelieffunctionsbasedongeneralized gaussian random fuzzy numbers. Fuzzy Sets and Systems471, 108679 (2023)

    2023

  6. [6]

    IEEE Transactions on Fuzzy Systems31(10), 3690–3699 (2023)

    Denœux, T.: Quantifying prediction uncertainty in regression using random fuzzy sets: the ennreg model. IEEE Transactions on Fuzzy Systems31(10), 3690–3699 (2023)

    2023

  7. [7]

    Fuzzy Sets and Systems453, 1–36 (2023)

    Denœux, T.: Reasoning with fuzzy and uncertain evidence using epistemic random fuzzy sets: General framework and practical models. Fuzzy Sets and Systems453, 1–36 (2023)

    2023

  8. [8]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Ding, K., Zhou, M., Metaxas, D.N., Zhang, S.: Pathology-and-genomics multi- modal transformer for survival outcome prediction. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 622–631. Springer (2023)

    2023

  9. [9]

    Nature Medicine pp

    Ding, T., Wagner, S.J., Song, A.H., Chen, R.J., Lu, M.Y., Zhang, A., Vaidya, A.J., Jaume,G.,Shaban,M.,Kim,A.,etal.:Amultimodalwhole-slidefoundationmodel for pathology. Nature Medicine pp. 1–13 (2025)

    2025

  10. [10]

    arXiv preprint arXiv:2205.14204 (2022)

    Geng, X., Liu, H., Lee, L., Schuurmans, D., Levine, S., Abbeel, P.: Multi- modal masked autoencoders learn transferable representations. arXiv preprint arXiv:2205.14204 (2022)

    work page arXiv 2022

  11. [11]

    Jack Geraghty, Andrew Hines, and Fatemeh Golpayegani

    Gong, L., Liu, Y., Sun, L., Bi, Y., Liu, J., Zhu, X.: Embracing aleatoric uncer- tainty in medical multimodal learning with missing modalities. arXiv preprint arXiv:2601.21950 (2026)

    work page arXiv 2026

  12. [12]

    Statistics in medicine18(17- 18), 2529–2545 (1999)

    Graf, E., Schmoor, C., Sauerbrei, W., Schumacher, M.: Assessment and comparison of prognostic classification schemes for survival data. Statistics in medicine18(17- 18), 2529–2545 (1999)

    1999

  13. [13]

    Jama247(18), 2543–2546 (1982) 10 Y

    Harrell, F.E., Califf, R.M., Pryor, D.B., Lee, K.L., Rosati, R.A.: Evaluating the yield of medical tests. Jama247(18), 2543–2546 (1982) 10 Y. Xing et al

    1982

  14. [14]

    Nature communications12(1), 4423 (2021)

    Howard, F.M., Dolezal, J., Kochanny, S., Schulte, J., Chen, H., Heij, L., Huo, D., Nanda, R., Olopade, O.I., Kather, J.N., et al.: The impact of site-specific digital histology signatures on deep learning model accuracy and bias. Nature communications12(1), 4423 (2021)

    2021

  15. [15]

    Medical Image Analysis97, 103223 (2024)

    Huang, L., Ruan, S., Xing, Y., Feng, M.: A review of uncertainty quantification in medical image analysis: Probabilistic and non-probabilistic methods. Medical Image Analysis97, 103223 (2024)

    2024

  16. [16]

    In: International Conference on Belief Functions

    Huang, L., Xing, Y., Denoeux, T., Feng, M.: An evidential time-to-event prediction model based on gaussian random fuzzy numbers. In: International Conference on Belief Functions. pp. 49–57. Springer (2024)

    2024

  17. [17]

    IEEE Transactions on Fuzzy Systems (2025)

    Huang, L., Xing, Y., Lin, Q., Duan, J., Ruan, S., Feng, M.: Esurvfusion: An ev- idential multimodal survival fusion model based on epistemic random fuzzy sets. IEEE Transactions on Fuzzy Systems (2025)

    2025

  18. [18]

    International Journal of Ap- proximate Reasoning181, 109403 (2025)

    Huang, L., Xing, Y., Mishra, S., Denœux, T., Feng, M.: Evidential time-to-event prediction with calibrated uncertainty quantification. International Journal of Ap- proximate Reasoning181, 109403 (2025)

    2025

  19. [19]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Jaume, G., Vaidya, A., Chen, R.J., Williamson, D.F., Liang, P.P., Mahmood, F.: Modeling dense multimodal interactions between biological pathways and histology for survival prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11579–11590 (2024)

    2024

  20. [20]

    In: Proceedings of the AAAI conference on artificial intelligence

    Ma, M., Ren, J., Zhao, L., Tulyakov, S., Wu, C., Peng, X.: Smil: Multimodal learning with severely missing modality. In: Proceedings of the AAAI conference on artificial intelligence. vol. 35, pp. 2302–2310 (2021)

    2021

  21. [21]

    Encyclopedia of artificial intelligence1(330- 331), 3 (1992)

    Shafer, G.: Dempster-shafer theory. Encyclopedia of artificial intelligence1(330- 331), 3 (1992)

    1992

  22. [22]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wang, H., Chen, Y., Ma, C., Avery, J., Hull, L., Carneiro, G.: Multi-modal learning with missing modality via shared-specific feature modelling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15878– 15887 (2023)

    2023

  23. [23]

    In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=Je5SHCKpPa

    Wu, Z., Dadu, A., Tustison, N., Avants, B., Nalls, M., Sun, J., Faghri, F.: Multimodal patient representation learning with missing modalities and labels. In: The Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=Je5SHCKpPa

    2024

  24. [24]

    DPsurv: Dual-Prototype Evidential Fusion for Uncertainty-Aware and Interpretable Whole-Slide Image Survival Prediction

    Xing, Y., Huang, L., Ma, J., Hong, R., Qiu, J., Liu, P., He, K., Fu, H., Feng, M.: Dpsurv: Dual-prototype evidential fusion for uncertainty-aware and interpretable whole-slide image survival prediction. arXiv preprint arXiv:2510.00053 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  25. [25]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Xu, Y., Chen, H.: Multimodal optimal transport-based co-attention transformer with global structure consistency for survival prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 21241–21251 (2023)

    2023

  26. [26]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Xu,Y.,Zhou,F.,Zhao,C.,Wang,Y.,Yang,C.,Chen,H.:Distilledpromptlearning for incomplete multimodal survival prediction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5102–5111 (June 2025)

    2025

  27. [27]

    Yun, S., Choi, I., Peng, J., Wu, Y., Bao, J., Zhang, Q., Xin, J., Long, Q., Chen, T.: Flex-moe: Modeling arbitrary modality combination via the flexible mixture- of-experts (2024), https://arxiv.org/abs/2410.08245

    work page arXiv 2024

  28. [28]

    In: Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining

    Zhang, C., Chu, X., Ma, L., Zhu, Y., Wang, Y., Wang, J., Zhao, J.: M3care: Learning with missing modalities in multimodal healthcare data. In: Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining. pp. 2418–2428 (2022)

    2022