CenSurv: Multimodal Cancer Survival Prediction

• CenSurv is a multimodal cancer survival prediction framework that integrates pathology, genomics, and clinical data to estimate time-to-event outcomes robustly.
• It constructs a bipartite patient-modality graph with complete-incomplete alignment using GraphSAGE encoders to ensure robustness even when some modalities are missing.
• Utilizing a plug-and-play ECMC module with dynamic momentum accumulation confidence, CenSurv actively reassigns censored samples to achieve notable improvements in concordance index over leading baselines.

CenSurv is a multimodal cancer survival prediction framework introduced in “Bipartite Patient-Modality Graph Learning with Event-Conditional Modelling of Censoring for Cancer Survival Prediction” (Yue et al., 22 Jul 2025). It is designed for time-to-event prediction under two conditions that are common in clinical datasets: heavy censoring and missing modalities. The framework combines graph-based multimodal representation learning, a bipartite patient-modality graph with complete-incomplete alignment for modality-missing robustness, and a plug-and-play Event-Conditional Modelling of Censoring (ECMC) module that selects reliable censored samples, updates their survival times, and reincorporates them into training as uncensored data.

1. Problem setting and conceptual scope

CenSurv addresses cancer survival prediction as a time-to-event problem in which each sample is associated with a survival or follow-up time $t_i$ and a status indicator $\delta_i$. The framework is motivated by two limitations identified in prior work. First, many methods use censored samples only through conventional survival losses and do not attempt to recover more informative event-time supervision from them. Second, many multimodal survival methods assume that pathology, genomics, and clinical data are all available, which leads to performance degradation under modality-missing scenarios and can create difficulty at inference.

The paper distinguishes uncensored or event samples, for which the death event and survival time are observed, from censored samples, for which the patient is alive at last follow-up and the observed time is smaller than the actual survival time. It also notes some inconsistency in the verbal description of $\delta$, while stating that the operational loss uses the usual Cox-style event-indicator behavior in which only event samples contribute as anchors in the Cox loss.

The framework is positioned against unimodal and multimodal baselines including Cox$(c)$, Cox$(g)$, Cox$(p)$, TranMIL$(p)$, HvtSurv$(p)$, GSCNN, MultiSurv, OuterP, MetricL, MCAT, HGCN, SurvMamba, SurvPath, ZeroP, MFM, and MMDB. Its central thesis is that multimodal robustness and more active exploitation of censored data should be addressed jointly rather than as separate design problems (Yue et al., 22 Jul 2025).

2. Multimodal representation learning and bipartite graph construction

CenSurv uses three modalities: pathological sections, genomic profile, and clinical records. Each modality is first converted into a graph representation. For pathology, original pathology images are cropped at $10\times$ magnification into non-overlapping $512 \times 512$ patches, and patch features are extracted using pretrained KimiaNet. Each patch becomes a node, and each patch is connected to its 8 spatially adjacent patches. This produces a pathology graph intended to capture both local and global tumor information.

For genomics, the method generates 5 genomic embeddings using Gene Set Enrichment Analysis (GSEA), and these embeddings are connected to form a graph representation of the genomic profile. For clinical data, each clinical feature is one-hot encoded, the vectors are fully concatenated, and the result is used to create a graph representation. To unify the three modalities, all node features are padded to 1024 dimensions via zero padding, and three independent GraphSAGE encoders are applied, one per modality.

A hyperedge-based aggregation mechanism is then used to obtain a modality-level representation. For modality $\delta_i$0 with $\delta_i$1 nodes, the paper defines

$\delta_i$2

This yields up to three patient-specific modality embeddings: $\delta_i$3, $\delta_i$4, and $\delta_i$5.

The central structural element is the bipartite patient-modality graph

$\delta_i$6

with patient nodes

$\delta_i$7

and modality nodes

$\delta_i$8

In the reported implementation, $\delta_i$9. An edge $\delta$0 exists if patient $\delta$1 has modality $\delta$2; otherwise it is absent. Available edges are initialized with the corresponding patient-specific modality representation, so missing modalities are represented naturally as missing edges rather than by forcing a fixed fusion vector.

Training uses two graph views. The complete graph $\delta$3 preserves all available modalities, while the incomplete graph $\delta$4 is obtained by random edge dropout to simulate missing-modality scenarios. A siamese GNN encodes both graphs: $\delta$5 The complete and incomplete embeddings are later aligned so that the same patient remains close in representation space under varying modality availability (Yue et al., 22 Jul 2025).

3. Complete-incomplete alignment and survival optimization

The complete-incomplete alignment strategy is intended to learn modality-agnostic patient features. The paper’s motivation is that a patient represented with all modalities and the same patient represented with some modalities dropped should preserve what is essential for risk prediction. CenSurv therefore aligns the complete representation $\delta$6 and incomplete representation $\delta$7 of the same patient while contrasting them against incomplete representations of other patients. The alignment loss $\delta$8 is described as an InfoNCE-like contrastive objective using cosine similarity and a temperature hyperparameter $\delta$9.

Survival supervision is provided through a Cox loss. The paper presents the framework as optimizing Cox-style ranking behavior over the at-risk set, with event samples serving as anchors. The total objective is

$(c)$0

with

$(c)$1

This weighting gives the survival objective greater emphasis than the alignment term.

The reported implementation uses Adam with learning rate $(c)$2, five-fold cross-validation, and selection of the best validation model in each fold for testing. A plausible implication is that the method treats missing-modality robustness not as a post hoc correction but as a directly optimized property of the learned patient representation. The paper’s formulation also makes clear that the alignment objective can be applied to both uncensored and censored samples because it depends on patient identity consistency across complete and incomplete graph views rather than on exact event times (Yue et al., 22 Jul 2025).

4. Event-Conditional Modelling of Censoring (ECMC)

ECMC is the second main component of CenSurv and is explicitly described as a plug-and-play module. Its purpose is to exploit censored samples more aggressively than conventional survival losses by estimating confidence in sample rankings over training, selecting reliable censored samples, updating their survival times and statuses, and treating them as uncensored samples in later training.

The key confidence mechanism is Dynamic Momentum Accumulation Confidence (DMAC): $(c)$3 with

$(c)$4

Here $(c)$5 is described as the reciprocal of the absolute value of the difference between the current epoch and the previous epoch in ranking. Stable rankings therefore produce higher confidence, while unstable rankings produce lower confidence. The paper interprets DMAC as a reliability filter for censored samples.

After sorting all samples by predicted survival risk, ECMC examines censored samples within a $(c)$6 immediate range. The paper does not provide a closed-form selection equation, but it states that for each censored sample the method searches within this local neighborhood and updates survival time and status with the goal of achieving the best C-index in that region. Two constraints are explicit: the updated time must exceed the original censored time, and the status is updated from alive to death, $(c)$7. The updated censored samples are then incorporated back into training as event samples.

Training is divided into two stages. In the preheating stage,

$(c)$8

the original censored and uncensored statuses are preserved. In the iterative update stage,

$(c)$9

ECMC updates selected censored samples and reincorporates them into training. The paper reports a synthetic censoring study in which the initial MAE for uncensored and censored data is 14.5 months and drops to 5.7 months after ECMC. This is used as evidence that the reassigned survival times are substantially more accurate than the original censored observations.

A common misconception is that ECMC merely duplicates what a Cox loss already does. The paper rejects that reading. Standard Cox training handles censoring statistically, but ECMC adds confidence estimation over training dynamics, selective conversion of censored samples into event-supervised samples, and local event-conditional reassignment of survival times. In this sense, ECMC changes the effective supervision structure rather than only reweighting existing samples (Yue et al., 22 Jul 2025).

5. Experimental evaluation and reported performance

CenSurv is evaluated on five publicly available TCGA cancer datasets: KIRC, LIHC, ESCA, LUSC, and UCEC. The experiments use pathology images, genomics, and clinical records. Performance is measured primarily by Concordance index (C-index), with logrank-test $(g)$0-value used to assess separation between high- and low-risk groups.

Under complete-modality conditions, the paper reports that CenSurv attains a mean C-index of 0.708. The best unimodal baseline is HvtSurv with mean 0.651, so the reported improvement is 5.7 percentage points. The strongest multimodal baseline is SurvPath with mean 0.677, so the reported gain is 0.031, described as a 3.1% mean C-index improvement over the best state-of-the-art. Dataset-specific CenSurv results are reported as $(g)$1 on KIRC, $(g)$2 on LIHC, $(g)$3 on ESCA, $(g)$4 on LUSC, and $(g)$5 on UCEC.

For survival-group stratification, the reported logrank $(g)$6-values are $(g)$7 for KIRC, $(g)$8 for LIHC, $(g)$9 for ESCA, $(p)$0 for LUSC, and $(p)$1 for UCEC. The paper states that these values are better than those of the comparison methods.

The paper also emphasizes robustness under various modality-missing scenarios. Exact numeric values for each missingness condition are not printed in the textual summary, but the reported conclusion is that CenSurv remains ahead of ZeroP, MFM, HGCN, and MMDB under incomplete-modality settings. ECMC is additionally tested in plug-and-play mode on 8 multimodal survival methods across 5 datasets, where it yields an average gain of 1.3% mean C-index; one explicit example is MultiSurv + ECMC, which improves mean C-index by 2.1% over MultiSurv. Reported model complexity is 2,308.91K parameters and 2.26G FLOPs (Yue et al., 22 Jul 2025).

6. Ablations, interpretation, and limitations

The ablation study separates the contributions of ECMC, the bipartite patient-modality graph, and DMAC.

Variant	Mean C-index	Interpretation
w/o ECMC	0.678	Lower than full model
w/o BPMG	0.698	Lower than full model
w/o DMAC	0.667	Largest drop among removals
CenSurv	0.708	Best overall mean

The full per-dataset ablation results are also reported. Without ECMC, the model scores 0.745 on KIRC, 0.691 on LIHC, 0.643 on ESCA, 0.567 on LUSC, and 0.747 on UCEC. Without the bipartite patient-modality graph, it scores 0.747, 0.699, 0.694, 0.585, and 0.765. Without DMAC, it scores 0.733, 0.662, 0.637, 0.557, and 0.745. The paper highlights that removing DMAC causes the largest mean decline, which it interprets as evidence that confidence-based censor filtering is critical. One exception is ESCA, where the variant without BPMG slightly exceeds full CenSurv, 0.694 versus 0.691, but the full model remains best in mean performance and on most datasets.

The paper attributes CenSurv’s robustness to four design choices: structural modeling of patient-modality availability through the bipartite graph, explicit training for representation stability through complete-incomplete alignment, active exploitation of censored samples through ECMC rather than passive use in a survival loss, and DMAC-based filtering to reduce harmful pseudo-labeling. At the same time, it identifies several limitations and open directions. The discussion mentions future work on interpretability across multiple modalities, model complexity, and additional evaluation metrics. The summary also notes that exact details of some components, including the neighborhood size $(p)$2 and the exact pseudo-time search rule, are not fully spelled out in the manuscript, and that evaluation is limited to 5 TCGA datasets and 3 modalities. Code availability is reported at https://github.com/yuehailin/CenSurv (Yue et al., 22 Jul 2025).