Pith. sign in

REVIEW 4 major objections 5 minor 26 references

Bipartite Patient-Modality Graph Learning with Event-Conditional Modelling of Censoring for Cancer Survival Prediction

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that CenSurv, a bipartite patient-modality graph model with event-conditional modeling of censoring, outperforms the best prior multimodal method by 3.1% in mean C-index across five TCGA cancer datasets while remaining…

desk verdict Solid architecture, but the SOTA claim ignores the closest prior work—add Centime before believing the 3.1%. read the letter →

arxiv 2507.16363 v1 pith:IFSHICPQ submitted 2025-07-22 cs.LG cs.MM

Hailin Yue , Hulin Kuang , Jin Liu , Junjian Li , Lanlan Wang , Mengshen He , Jianxin Wang This is my paper
classification cs.LGcs.MM
keywords cancersurvivalpredictionmultimodallearningcensoreddatabipartitegraphevent-conditionalmodelingC-indexTCGAplug-and-playmodule
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that two neglected resources—censored patients with unknown true survival times and the pattern of which modalities each patient has—can be converted into measurable gains in cancer survival prediction. It introduces CenSurv, which links patients to their available modalities in a bipartite graph and uses a complete-incomplete alignment loss so the model remains accurate when some modality is missing. A plug-and-play module, ECMC, selects censored patients the model is confident about, assigns them longer survival times, flips their status to death, and retrains on them as uncensored samples. Across five TCGA datasets, CenSurv reports a mean C-index of 0.708 versus 0.677 for the best prior multimodal method, and ECMC improves eight baselines by an average of 1.3%.

What carries the argument

The core machinery is the Event-Conditional Modeling of Censoring (ECMC) module, which uses Dynamic Momentum Accumulation Confidence (DMAC)—an exponentially smoothed measure of ranking stability across epochs—to select reliable censored samples. For each selected sample, the survival time is updated by searching for the value that maximizes the C-index within its K nearest neighbors in risk ordering, and the event status is flipped from alive to dead (δ: 0→1). The second component is the Bipartite Patient-Modality Graph (BPMG): a graph with patient nodes and modality nodes, where available modalities create edges; a siamese GNN encodes the complete graph and an edge-dropped incomplete copy, and a cosine-similarity alignment loss pulls the two patient representations together to learn modality-agnostic features.

What would settle it

A concrete falsifying experiment would be to take a dataset with near-complete follow-up, artificially censor a random subset of patients, run ECMC, and compare the assigned survival times against the true event times; if the assigned times deviate systematically from the true residual lifetimes, or if models trained with ECMC fail to outperform models trained on genuinely uncensored data alone, the central claim would be undercut.

Watch

Extended reading notes

Core claim

The central claim is that reclassifying carefully chosen censored samples as uncensored—using the model's own confidence and a local C-index optimization to update their event times—combined with a bipartite patient-modality graph whose incomplete version is aligned to the complete one, yields state-of-the-art multimodal survival prediction. The paper reports that CenSurv outperforms the best prior multimodal method by 3.1% in mean C-index (0.708 vs 0.677) over five TCGA datasets, with logrank test p-values confirming separation between high- and low-risk groups; in modality-missing settings it maintains higher mean C-index than existing missing-modality methods.

Load-bearing premise

The load-bearing premise is that the survival times ECMC assigns to censored patients are accurate enough to serve as uncensored labels; if the model's own risk ordering is biased, updating censored times to maximize C-index within K neighbors can create a self-reinforcing loop that inflates apparent performance.

Editorial extensions

If this is right

  • Censored patients become a source of training signal instead of being discarded, increasing the effective size of the uncensored training set.
  • Multimodal survival models can be made robust to missing modalities by aligning complete and incomplete graph representations, not just by imputing features.
  • Because ECMC is plug-and-play, existing survival predictors can be upgraded without redesigning their architectures; the paper reports mean C-index gains for all eight baselines tested.
  • A 3.1% improvement in mean C-index on five TCGA datasets corresponds to better ranking of patients by risk, which is the basis for more personalized treatment decisions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to test ECMC on cohorts with long follow-up where true event times are known, to check whether the pseudo-labels it assigns are well calibrated rather than merely rank-consistent.
  • The DMAC confidence signal could be used as a stopping criterion or as a per-sample weight in the loss, both of which the paper leaves implicit.
  • The same complete-incomplete alignment strategy might transfer to other multimodal clinical prediction tasks, such as diagnosis or treatment response, where missing inputs are common.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes CenSurv, a multimodal cancer survival prediction method combining a bipartite patient-modality graph with a complete-incomplete alignment loss and a plug-and-play event-conditional modelling of censoring (ECMC) module. The method extracts pathological, genomic, and clinical features, builds patient-modality graphs, and uses ECMC to pseudo-label selected censored samples as uncensored by updating their survival times and flipping their event status. The authors report a mean C-index of 0.708 across five TCGA datasets, outperforming eight multimodal baselines by 3.1%, and report that adding ECMC to eight baselines improves their mean C-index by 1.3%. They also evaluate robustness under missing-modality scenarios and provide ablation studies. The code is publicly available.

Significance. If the results are reliable, CenSurv would be a practically useful contribution to multimodal survival analysis, particularly for exploiting censored data and handling missing modalities. The paper is strengthened by five-fold cross-validation, ablations, plug-and-play experiments across eight baselines, and public code. However, the central claims currently rest on an evaluation that omits the closest prior work (Centime), and the ECMC label-update mechanism is both under-specified and potentially circular. The paper's significance can only be assessed after these issues are resolved.

major comments (4)
  1. [Section 3.2, Table 1] The state-of-the-art claim in the Abstract and Section 3.2 is established by comparing CenSurv against GSCNN, MultiSurv, OuterP, MetricL, MCAT, HGCN, SurvMamba, and SurvPath, but Centime (reference [17]) is cited and never included in Table 1 or Figure 3. Since Centime's method is literally 'event-conditional modelling of censoring in survival analysis', it is the closest prior work to ECMC, and the 3.1% mean C-index improvement over 'the best state-of-the-art' is not meaningful without this comparison. Please add Centime to the full experimental comparison, or explicitly restrict the SOTA claim and justify why Centime cannot be evaluated on these datasets.
  2. [Section 2.3] The ECMC update procedure is circular and under-specified. Censored samples are assigned new survival times by optimizing the C-index within a K-neighborhood of the model's own risk ordering, after which their event status is flipped to uncensored and they are used to retrain the same model. The manuscript does not specify K, the DMAC selection threshold, the exact time-search procedure, or any safeguard preventing training-fold label updates from being informed by validation/test-fold predictions. Please provide the full algorithm, including the selection criterion and the fold-restriction mechanism, and report sensitivity of the Table 2 results to K and the DMAC threshold.
  3. [Fig. 1 and Eq. (4)] The censoring notation is internally inconsistent. Fig. 1 states that uncensored data have status delta=0 and censored data have delta=1, whereas Eq. (4) uses delta_i as the multiplier in the Cox partial likelihood, which conventionally is 1 for observed events and 0 for censored observations. With the Fig. 1 convention, Eq. (4) would incorrectly sum over censored samples. Please reconcile the definition of delta between Fig. 1, Section 2.3 (where the text says delta: 0 -> 1 for a status update to death), and Eq. (4).
  4. [Table 2 and Section 3.4] The ablation row 'w/o DMAC' is not described anywhere. It is unclear whether removing DMAC means using all censored samples, randomly selecting censored samples, or using some other selection rule. Without this description, the claim that DMAC 'avoids adding noisy samples and thus improves model performance' is not supported. Please specify the exact selection protocol used in the 'w/o DMAC' ablation and, if applicable, report the number of samples selected in each variant.
minor comments (5)
  1. [Fig. 1 and Fig. 2] The captions contain repeated typos: 'Statue' should be 'Status' in Fig. 1 and 'Statue Time' should be 'Status Time' in Fig. 2.
  2. [Section 3.2] The text refers to 'SOAT multimodal methods' and 'SOAT methods'; this should be 'state-of-the-art' for clarity.
  3. [Section 2.1] The paragraph introducing 'hyperedge' says all nodes of a modality are linked by a hyperedge, but the subsequent equation and pooling operation do not explicitly use a hyperedge construction; please clarify whether the hyperedge is a formal part of the graph or just a conceptual description.
  4. [Eq. (3)] The summation index in Eq. (3) is written as N, while the patient set is denoted P elsewhere; please unify the notation to avoid ambiguity.
  5. [Section 2.3 and Experimental Setting] Several hyperparameters are not reported: the temperature phi in Eq. (3), the neighborhood size K in ECMC, and the edge dropout rate used to construct the incomplete graph. Please report these values or state how they were chosen.

Circularity Check

1 steps flagged · score 6.0 of 10

ECMC's pseudo-labeled survival times are generated from CenSurv's own risk ranking, so part of the reported C-index gain is a self-consistency loop.

  1. fitted input called prediction [Section 2.3, 'Selection and Update for Censored Data'; ablation Table 2]
    "Specifically, all samples are first sorted by the predicted survival risk from CenSurv. Then, for each censored sample, with the goal of obtaining the best C-index within its K immediate range, the survival time and survival status of the sample are updated. Finally, traverse and update the survival time and status of all selected censored samples. This process ensures the updated survival time exceeds the original censored time, and the status is updated to death from alive (δ: 0→ 1)."

    The 'updated' survival times for censored samples are not obtained from an independent external source; they are chosen to optimize C-index within K neighbors of CenSurv's own risk ordering, then the status is flipped to death. These model-fitted labels are reinserted as uncensored training targets in the Cox loss (Section 2.2, Eq. 4), and the resulting improvement is presented as ECMC recovering 'more accurate survival times' (Abstract) and as a 0.678→0.708 ablation gain (Table 2). The training labels are therefore constructed from the model being evaluated; the training-set concordance gain is by construction, and only the held-out test evaluation plus the synthetic-pruning MAE check supply independent evidence.

full rationale

The main independent content is the bipartite patient-modality graph and complete-incomplete alignment (Section 2.2), which alone moves the mean C-index from 0.677 (best baseline) to 0.698 (Table 2, w/o ECMC). The remaining 0.678→0.708 gain attributed to ECMC rests on pseudo-labels that are generated by optimizing C-index in a K-neighborhood of CenSurv's own risk ranking (Section 2.3), which is a fitted-input loop rather than an externally grounded label source. This makes the SOTA claim partially circular, though not entirely: the five-fold test evaluation is on true labels, and the synthetic-pruning MAE check (Section 3.4) is an external sanity check. Separately, the SOTA comparison omits Centime (ref [17]), the closest prior work with nearly the same event-conditional censoring formulation; that is a benchmark-completeness and attribution concern rather than a circularity step, but it compounds the risk that the headline 3.1% gain is over-stated.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several fitted hyperparameters (alpha, beta, lambda, phi, K, dropout rate) and on domain assumptions about the sufficiency of pretrained features, the validity of random dropout as a missingness model, and the reliability of model-generated survival labels. No new physical or independent entities are introduced; DMAC, BPMG, and ECMC are methodological constructs, not independently falsifiable entities.

free parameters (6)
  • alpha (Cox loss weight) = 5
    Selected 'after extensive experiments' (Section 3.1); changes the balance between survival loss and alignment loss.
  • beta (alignment loss weight) = 1
    Selected 'after extensive experiments'; controls the strength of complete-incomplete alignment.
  • lambda (DMAC momentum balance) = 0.4
    Chosen through additional experiments; determines how much weight is given to current vs. previous confidence estimates.
  • phi (InfoNCE temperature) = not reported
    Temperature in the complete-incomplete alignment loss (Eq. 3); value not given in the implementation section.
  • K (neighborhood size for censored time update) = not reported
    ECMC updates survival times to optimize C-index within the K immediate range; K is never stated.
  • Edge dropout rate for incomplete graph = not reported
    Random edge dropout generates the incomplete patient-modality graph; the dropout rate is not specified.
assumptions (5)
  • standard math Cox proportional hazards assumption
    The loss in Eq. 4 is a Cox partial likelihood, which assumes proportional hazards and a linear risk score.
  • domain assumption Non-informative censoring
    Standard survival analysis assumption required for Cox loss; not stated explicitly in the paper.
  • domain assumption KimiaNet features capture histology relevant to survival
    Pretrained KimiaNet is used without fine-tuning; the paper assumes its patch features are adequate for survival prediction.
  • ad hoc to paper Random edge dropout simulates modality-missing scenarios
    The incomplete graph is generated by random edge dropout (Section 2.2), with no validation that this matches real missing-data distributions.
  • ad hoc to paper Pseudo-labeled censored samples can be treated as uncensored
    Core assumption of ECMC: updated survival times and event statuses of censored samples are used as ground truth in the Cox loss (Section 2.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bipartite Patient-Modality Graph Learning with Event-Conditional Modelling of Censoring for Cancer Survival Prediction." pith.science (2026). https://pith.science/paper/IFSHICPQ

@misc{pith2026250716363,
  author       = {Pith},
  title        = {Pith review of: Bipartite Patient-Modality Graph Learning with Event-Conditional Modelling of Censoring for Cancer Survival Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IFSHICPQ}},
  note         = {Machine review of arXiv:2507.16363}
}
read the original abstract

Accurately predicting the survival of cancer patients is crucial for personalized treatment. However, existing studies focus solely on the relationships between samples with known survival risks, without fully leveraging the value of censored samples. Furthermore, these studies may suffer performance degradation in modality-missing scenarios and even struggle during the inference process. In this study, we propose a bipartite patient-modality graph learning with event-conditional modelling of censoring for cancer survival prediction (CenSurv). Specifically, we first use graph structure to model multimodal data and obtain representation. Then, to alleviate performance degradation in modality-missing scenarios, we design a bipartite graph to simulate the patient-modality relationship in various modality-missing scenarios and leverage a complete-incomplete alignment strategy to explore modality-agnostic features. Finally, we design a plug-and-play event-conditional modeling of censoring (ECMC) that selects reliable censored data using dynamic momentum accumulation confidences, assigns more accurate survival times to these censored data, and incorporates them as uncensored data into training. Comprehensive evaluations on 5 publicly cancer datasets showcase the superiority of CenSurv over the best state-of-the-art by 3.1% in terms of the mean C-index, while also exhibiting excellent robustness under various modality-missing scenarios. In addition, using the plug-and-play ECMC module, the mean C-index of 8 baselines increased by 1.3% across 5 datasets. Code of CenSurv is available at https://github.com/yuehailin/CenSurv.

Figures

Figures reproduced from arXiv: 2507.16363 by the authors.

Figure 1
Figure 1. (a) The distinction of censored and uncensored data. For uncensored data, the survival status is dead (δ = 0), and the survival time is accurate. For censored data, the survival status is alive (δ=1), and the survival time calculated with the follow-up time as the event endpoint is inaccurate and smaller than the actual survival time. (b) The number of censored and uncensored data on 5 datasets in our study. With ad… view at source ↗
Figure 2
Figure 2. Overview of the bipartite patient-modality graph learning with event￾conditional modelling of censoring (CenSurv). data, improving the mean C-index by 1.3% across 8 baselines evaluated on 5 cancer datasets. 2) We propose a bipartite patient-modality graph that fully ex￾ploits the flexibility of graph learning and enables the model to not only work with different modality-missing scenarios but also achieve better per… view at source ↗
Figure 3
Figure 3. The C-index of compared multimodal methods with and without ECMC. Red represents the methods with ECMC, and blue represents the methods without ECMC. 𝒑 ✓ 𝒄 𝒈 x x 𝒑 ✓ 𝒄 𝒈 x x 𝒑 ✓ 𝒄 𝒈 x x 𝒑 ✓ 𝒄 𝒈 ✓ x 𝒑 ✓ 𝒄 𝒈 x ✓ 𝒑 𝒄 𝒈 x ✓ ✓ ✓ is available x is missing 𝒑: Pathological Sections 𝒄: Clinical Records 𝒈: 𝐺𝑒𝑛𝑜𝑚𝑖𝑐 𝐹𝑒𝑎𝑡𝑢𝑟𝑒𝑠 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The mean C-index of methods under various modality-missing scenarios [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 20 canonical work pages

  1. [17]

    Medical Image Analysis 91, 103016 (2024)

    Shahin, A.H., Zhao, A., Whitehead, A.C., Alexander, D.C., Jacob, J., Barber, D.: Centime: Event-conditional modelling of censoring in survival analysis. Medical Image Analysis 91, 103016 (2024)

    work page 2024

  2. [1]

    Bioinformatics35(14), i446–i454 (2019)

    Cheerla, A., Gevaert, O.: Deep learning with multimodal representation for pan- cancer prognosis prediction. Bioinformatics35(14), i446–i454 (2019)

    2019

  3. [2]

    IEEE Transactions on Medical Imaging 41(4), 757–770 (2020)

    Chen, R.J., Lu, M.Y., Wang, J., Williamson, D.F., Rodig, S.J., Lindeman, N.I., Mahmood, F.: Pathomic fusion: an integrated framework for fusing histopathology and genomic features for cancer diagnosis and prognosis. IEEE Transactions on Medical Imaging 41(4), 757–770 (2020)

    work page 2020

  4. [3]

    In: Proceedings of the IEEE/CVF Inter- national 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 Inter- national Conference on Computer Vision. pp. 4015–4025 (2021)

    work page 2021

  5. [4]

    arXiv preprint arXiv:2404.08027 (2024)

    Chen, Y., Xie, J., Lin, Y., Song, Y., Yang, W., Yu, R.: Survmamba: State space model with multi-grained multi-modal interaction for survival prediction. arXiv preprint arXiv:2404.08027 (2024)

    arXiv 2024

  6. [5]

    Medical Image Analysis99, 103346 (2025)

    Elforaici, M.E.A., Montagnon, E., Romero, F.P., Le, W.T., Azzi, F., Trudel, D., Nguyen, B., Turcotte, S., Tang, A., Kadoury, S.: Semi-supervised vit knowledge distillation network with style transfer normalization for colorectal liver metastases survival prediction. Medical Image Analysis99, 103346 (2025)

    work page 2025

  7. [6]

    Advances in neural information processing systems30 (2017)

    Hamilton, W., Ying, Z., Leskovec, J.: Inductive representation learning on large graphs. Advances in neural information processing systems30 (2017)

    2017

  8. [7]

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

    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 H. Yue et al

    work page 1982

Show all 26 references
  1. [8]

    IEEE Transactions on Medical Imaging42(8), 2462–2473 (2023)

    Hou, W., Lin, C., Yu, L., Qin, J., Yu, R., Wang, L.: Hybrid graph convolutional network with online masked autoencoder for robust multimodal cancer survival prediction. IEEE Transactions on Medical Imaging42(8), 2462–2473 (2023)

    2023

  2. [9]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)

    Jaume, G., Vaidya, A., Chen, R., Williamson, D., Liang, P., Mahmood, F.: Mod- eling dense multimodal interactions between biological pathways and histology for survival prediction. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)

    2024

  3. [10]

    BMC medical research methodology18, 1–12 (2018)

    Katzman, J.L., Shaham, U., Cloninger, A., Bates, J., Jiang, T., Kluger, Y.: Deep- surv:personalizedtreatmentrecommendersystemusingacoxproportionalhazards deep neural network. BMC medical research methodology18, 1–12 (2018)

    2018

  4. [11]

    npj Digital Medicine7(1), 2 (2024)

    Lee, K.H., Choi, G.H., Yun, J., Choi, J., Goh, M.J., Sinn, D.H., Jin, Y.J., Kim, M.A., Yu, S.J., Jang, S., et al.: Machine learning-based clinical decision support system for treatment recommendation and overall survival prediction of hepato- cellular carcinoma: a multi-center...

    2024

  5. [12]

    IEEE Journal of Biomedical and Health Informatics (2024)

    Liu, H., Shi, Y., Xu, Y., Li, A., Wang, M.: Agnostic-specific modality learning for cancer survival prediction from multiple data. IEEE Journal of Biomedical and Health Informatics (2024)

    2024

  6. [13]

    Proceedings of the National Academy of Sciences115(13), E2970–E2979 (2018)

    Mobadersany, P., Yousefi, S., Amgad, M., Gutman, D.A., Barnholtz-Sloan, J.S., Velázquez Vega, J.E., Brat, D.J., Cooper, L.A.: Predicting cancer outcomes from histology and genomics using convolutional networks. Proceedings of the National Academy of Sciences115(13), E2970–E2979 (2018)

    2018

  7. [14]

    European urology81(6), 576–585 (2022)

    Psutka, S.P., Gulati, R., Jewett, M.A., Fadaak, K., Finelli, A., Legere, L., Morgan, T.M., Pierorazio, P.M., Allaf, M.E., Herrin, J., et al.: A clinical decision aid to sup- port personalized treatment selection for patients with clinical t1 renal masses: Re- sults from a mult...

    2022

  8. [15]

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

    Qu, L., Huang, D., Zhang, S., Wang, X.: Multi-modal data binding for survival analysis modeling with incomplete data and annotations. In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention. pp. 501–510. Springer (2024)

    2024

  9. [16]

    Medical image analysis70, 102032 (2021)

    Riasatian, A., Babaie, M., Maleki, D., Kalra, S., Valipour, M., Hemati, S., Za- veri, M., Safarpoor, A., Shafiei, S., Afshari, M., et al.: Fine-tuning and training of densenet for histopathology image representation using tcga diagnostic slides. Medical image analysis70, 102032 (2021)

    2021

  10. [18]

    IEEE Transactions on Medical Imaging42(9), 2552–2565 (2023)

    Shao, W., Liu, J., Zuo, Y., Qi, S., Hong, H., Sheng, J., Zhu, Q., Zhang, D.: Fam3l: Feature-aware multi-modal metric learning for integrative survival analysis of hu- man cancers. IEEE Transactions on Medical Imaging42(9), 2552–2565 (2023)

    2023

  11. [19]

    Advances in neural information processing systems34, 2136–2147 (2021)

    Shao, Z., Bian, H., Chen, Y., Wang, Y., Zhang, J., Ji, X., et al.: Transmil: Trans- former based correlated multiple instance learning for whole slide image classifica- tion. Advances in neural information processing systems34, 2136–2147 (2021)

    2021

  12. [20]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Shao, Z., Chen, Y., Bian, H., Zhang, J., Liu, G., Zhang, Y.: Hvtsurv: Hierarchical vision transformer for patient-level survival prediction from whole slide image. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37, pp. 2209– 2217 (2023)

    2023

  13. [21]

    In: Proceedings of the 28th ACM international con- ference on multimedia

    Shen, G., Wang, X., Duan, X., Li, H., Zhu, W.: Memor: A dataset for multimodal emotion reasoning in videos. In: Proceedings of the 28th ACM international con- ference on multimedia. pp. 493–502 (2020) Bipartite Patient-Modality Graph Learning for Survival Prediction 11

    2020

  14. [22]

    Advances in neural information processing systems20 (2007)

    Steck, H., Krishnapuram, B., Dehing-Oberije, C., Lambin, P., Raykar, V.C.: On ranking in survival analysis: Bounds on the concordance index. Advances in neural information processing systems20 (2007)

    2007

  15. [23]

    Proceedings of the National Academy of Sciences 102(43), 15545–15550 (2005)

    Subramanian, A., Tamayo, P., Mootha, V.K., Mukherjee, S., Ebert, B.L., Gillette, M.A., Paulovich, A., Pomeroy, S.L., Golub, T.R., Lander, E.S., et al.: Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proceedings of the...

    2005

  16. [24]

    CA: a cancer journal for clinicians 71(3), 209–249 (2021)

    Sung, H., Ferlay, J., Siegel, R.L., Laversanne, M., Soerjomataram, I., Jemal, A., Bray, F.: Global cancer statistics 2020: Globocan estimates of incidence and mor- tality worldwide for 36 cancers in 185 countries. CA: a cancer journal for clinicians 71(3), 209–249 (2021)

    2021

  17. [25]

    arXiv preprint arXiv:1806.06176 (2018)

    Tsai, Y.H.H., Liang, P.P., Zadeh, A., Morency, L.P., Salakhutdinov, R.: Learning factorized multimodal representations. arXiv preprint arXiv:1806.06176 (2018)

    2018 arXiv

  18. [26]

    Scientific Reports11(1), 13505 (2021)

    Vale-Silva, L.A., Rohr, K.: Long-term cancer survival prediction using multimodal deep learning. Scientific Reports11(1), 13505 (2021)

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.