An Accurate and Interpretable Hyper Graph Neural Network for GBM Survival Prediction

Mushahid Intesum

cs.AISep 19, 2026

An Accurate and Interpretable Hyper Graph Neural Network for GBM Survival Prediction

Authors: Mushahid Intesum

Abstract

Survival prediction for glioblastoma multiforme (GBM) demands models that are both accurate and interpretable, yet existing approaches treat these objectives as com- peting, where performant models sacrifice transparency, while interpretable models accept degraded predictive power. We argue that this trade-off is not inherent. Graph neural net- works offer a structural foundation for extracting interpretable, explainable representations without compromising discriminative ability. Furthermore, current methods typically rely on a single imaging modality, underutilizing the complementary information available across multi-modal MRI and clinical metadata. We propose a multi-modal framework that inte- grates three components to address both objectives simultaneously: (1) a sheaf hypergraph neural network that captures higher-order relationships among tissue patches through direc- tional, asymmetric message passing; (2) a concept bottleneck layer that compresses learned representations into clinically grounded concepts, enforcing ante-hoc interpretability; and (3) an extension sufficiency test (EST) regularizer that penalizes unfaithful explanations during training, ensuring that model explanations genuinely reflect the internal decision process. Clinical and genomic features are incorporated through gated fusion, preserving the dominant prognostic signal of molecular markers while retaining concept-level traceabil- ity. Evaluated on 593 patients from the UPenn-GBM dataset under 5-fold cross-validation, our framework achieves a concordance index of 0.643 with the lowest fold-level variance among all compared models (std = 0.015). To our knowledge, this is the first work to unify sheaf hypergraph convolution, concept bottleneck supervision, and EST regularization for interpretable survival prediction from brain MRI

Figures & tables

Figure 1. Architecture overview. Multi-modal MRI patches are encoded into a dual-space sheaf hypergraph, processed by SheafHGNN, compressed through a concept bottleneck (8 concepts), refined by HECRL, and fused with clinical features via gated attention for survival prediction. EST regularization (training only) penalizes prediction shifts when non-explanation nodes are added.

Exp	Components
E1	kNN graph + GNN + SurvivalHead
E2	Dual-space hypergraph + SheafHGNN
E3	E2 + CBM + HECRL
E4	E3 + Multi-modal fusion
E5	E4 + Multi-granular tree pooling
E6	E4 + EST training-time regularization

Table 1. Progressive ablation configurations. E1 through E4 build the core pipeline, while E5, E6, and E7 evaluate auxiliary modules.

Exp	C-Index	Std	Params
E1	0.5519	0.012	1,026,992
E2	0.5601	0.016	1,026,992
E3	0.5391	0.019	1,027,281
E4	0.6373	0.027	1,074,963
E5	0.6392	0.041	1,167,131
E6	0.6431	0.015	1,074,963

Table 2. 5-fold cross-validation results on 593 UPenn-GBM patients.

Figure 2. Progressive ablation study (E1–E7). Phase 1 builds the core pipeline; Phase 2 evaluates auxiliary modules atop E4. Error bars denote standard deviation across 5 folds. Clinical fusion (E3

\rightarrow

E4) contributes the largest single gain (+18.2%). E6 (red) achieves the best accuracy–stability trade-off.

Step	$\Delta$ C-Index	Interpretation
E1 $\rightarrow$ E2 (+Hypergraph)	+0.008 (+1.5%)	Sheaf structure captures higher-order relations
E2 $\rightarrow$ E3 (+CBM)	$-$ 0.021 ( $-$ 3.7%)	Interpretability bottleneck tax
E3 $\rightarrow$ E4 (+Fusion)	+0.098 (+18.2%)	Clinical data is the dominant prognostic signal
E4 $\rightarrow$ E5 (+Tree)	+0.002 (+0.3%)	Tree pooling benefits from wider embedding space
E4 $\rightarrow$ E6 (+EST)	+0.006 (+0.9%)	EST regularization improves stability
E6 $\rightarrow$ E7 (+Tree)	+0.000 (+0.1%)	Negligible gain when EST already active

Table 3. Marginal contribution of each component. Clinical fusion dominates all other deltas.

Exp	Fold 1	Fold 2	Fold 3	Fold 4	Fold 5	Min	Max
E1	0.540	0.574	0.543	0.553	0.549	0.540	0.574
E4	0.612	0.622	0.681	0.615	0.657	0.612	0.681
E5	0.606	0.639	0.681	0.589	0.681	0.589	0.681
E6	0.624	0.639	0.661	0.633	0.659	0.624	0.661
E7	0.626	0.620	0.682	0.629	0.658	0.620	0.682

Table 4. Per-fold C-Index for selected experiments. E6 shows the tightest spread with no fold below 0.62.

Model	C-Index	Std	Params	Input Modalities
DeepSurv [ 15 ]	0.6502	0.028	11K	Clinical only
HyperCBM [ 30 ]	0.5346	0.021	1.00M	Imaging + concepts
MRePath [ 26 ]	0.6419	0.021	1.05M	Imaging + clinical
E6 (Ours)	0.6431	0.015	1.07M	Imaging + clinical + concepts
E7 (Ours)	0.6435	0.023	1.17M	Imaging + clinical + concepts

Table 5. Cross-model comparison on 593 UPenn-GBM patients with 5-fold CV. All models use the same data splits.

Model	C-Index	td-AUC 6mo	td-AUC 12mo	td-AUC 18mo
DeepSurv [ 15 ]	0.6502	0.732	0.748	0.703
HyperCBM [ 30 ]	0.5346	0.601	0.554	0.519
MRePath [ 26 ]	0.6419	0.708	0.725	0.682
E6 (Ours)	0.6431	0.726	0.737	0.697
E7 (Ours)	0.6435	0.713	0.724	0.702

Table 6. Time-dependent AUC at 6, 12, and 18 months post-diagnosis (4-fold average). C-Index values are reproduced from Table 5 for reference.

Figure 3. Kaplan-Meier risk stratification. (a) E6 produces well-separated survival trajectories across risk tertiles (log-rank

p<0.001

). (b) HyperCBM fails to achieve meaningful separation (

p=0.516

), confirming that concept supervision alone is insufficient for clinically actionable risk grouping.

Model	Fold 1	Fold 2	Fold 3	Fold 4
DeepSurv	$<0.001$	$<0.001$	$<0.001$	0.002
HyperCBM	0.516	0.697	0.774	0.128
MRePath	0.004	0.002	$<0.001$	$<0.001$
E6 (Ours)	$<\textbf{0.001}$	0.001	$<\textbf{0.001}$	0.002
E7 (Ours)	0.002	0.003	$<0.001$	$<0.001$

Table 7. Log-rank

p

-values (high vs low risk tertile) per fold. Values

<0.05

indicate statistically significant survival separation.

Comparison	Mean $\Delta$	Paired $t$ -test $p$	Wilcoxon $p$
E6 vs DeepSurv	$-$ 0.007	0.562	0.625
E6 vs HyperCBM	+0.109	0.002	0.063
E6 vs MRePath	+0.001	0.885	1.000
E6 vs E7	$-$ 0.001	0.941	1.000

Table 8. Paired statistical tests between E6 and each comparison model across 5 folds. Bold indicates

p<0.01

Exp	Sufficiency Rejection	Mean Suf. Shift	Mean EST Shift
E2	0%	0.65%	0.80%
E4	0%	2.52%	0.55%
E5	12%	4.10%	0.40%
E6	0%	2.75%	0.69%
E7	0%	2.75%	0.74%

Table 9. Faithfulness audit results. Explanation subgraph is the top 20% of nodes by concept activation magnitude. Rejection threshold is 0.10.

Figure 4. Concept analysis. (a) Spearman correlation between each learned concept and patient survival time. DTI MD ( $\rho=+0.142$ , $p=0.046$ ) and Boundary ( $\rho=-0.242$ , $p=0.001$ ) reach statistical significance (red bars). Enhancement approaches significance ( $\dagger$ , $p=0.061$ ). (b) Inter-concept Pearson correlation matrix reveals expected biological groupings: DTI features (MD, FA) cluster with Heterogeneity ( $r>0.70$ ), while Enhancement and Boundary are positively correlated ( $r=0.52$ ).

Concept	Mean	Std	Min	Max
c1: Enhancement	0.652	0.103	0.059	0.877
c2: FLAIR/Edema	0.664	0.158	$-$ 0.028	1.787
c3: T2 Abnormality	$-$ 0.001	0.006	$-$ 0.024	0.042
c4: DTI MD	0.209	0.062	$-$ 0.036	0.361
c5: DTI FA	0.972	0.192	0.089	1.189
c6: Heterogeneity	0.146	0.022	0.043	0.198

Table 10. Learned concept distributions across 200 patients (E6 model). Concepts with higher variance (c1, c2, c5) capture patient-level heterogeneity, while low-variance concepts (c3, c7, c8) encode more uniform tissue properties.

Concept	Spearman $\rho$	$p$ -value
c1: Enhancement	$-$ 0.133	0.061
c2: FLAIR/Edema	$-$ 0.060	0.398
c3: T2 Abnormality	+0.060	0.401
c4: DTI MD	+0.142	0.046
c5: DTI FA	+0.017	0.810
c6: Heterogeneity	+0.118	0.096

Table 11. Spearman correlation between learned concepts and survival time (200 patients, E6 model). * denotes

p<0.05

Cancer Survival Prediction

Cancer survival prediction estimates patient outcomes and time to events from clinical, molecular, and medical-imaging data.

See all 92 papers

Quantifying Cross-Modal Interactions in Multimodal Glioma Survival Prediction via InterSHAP: Evidence for Additive Signal Integration
Mar 31, 2026Iain Swift, JingHua Ye, Ruairi O'ReillyMultimodal Fusion Multimodal Survival Analysis
Department of Computer Science, Munster Technological University, Cork, Ireland
Bridging the gap between Performance and Interpretability: An Explainable Disentangled Multimodal Framework for Cancer Survival Prediction
Mar 2, 2026Aniek Eijpe, Soufyan Lakbir, Melis Erdal Cesur +4Cancer Survival Prediction Multimodal Survival Analysis
AI Technology for Life, Department of Information and Computing Sciences, Department of Biology, Utrecht University, Utrecht, The Netherlands · Department of Metabolic Diseases, Wilhelmina Children’s Hospital, University Medical Center Utrecht, Utrecht, The Netherlands · Regenerative Medicine Center Utrecht, Utrecht, The Netherlands +3
Attention-Based Multimodal Survival Prediction with Cross-Modal Bilinear Fusion
May 12, 2026Hassan Keshvarikhojasteh, Josien P. W. Pluim, Mitko VetaRepresentation Learning Cancer Survival Prediction
Department of Biomedical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands

Mar 31, 2026J/K move · Enter open · S save

Quantifying Cross-Modal Interactions in Multimodal Glioma Survival Prediction via InterSHAP: Evidence for Additive Signal Integration

Iain Swift, JingHua Ye, Ruairi O'Reilly

Department of Computer Science, Munster Technological University, Cork, Ireland

Multimodal Fusion Multimodal Survival Analysis Cancer Survival Prediction Multimodal Model Interpretability Shapley Value Attribution

Open paper AskPDF arXiv

Multimodal deep learning for cancer prognosis is commonly assumed to benefit from synergistic cross-modal interactions, yet this assumption has not been directly tested in survival prediction settings. This work adapts InterSHAP, a Shapley interaction index-based metric, from classification to Cox proportional hazards models and applies it to quantify cross-modal interactions in glioma survival prediction. Using TCGA-GBM and TCGA-LGG data (n=575), we evaluate four fusion architectures combining whole-slide image (WSI) and RNA-seq features. Our central finding is an inverse relationship between predictive performance and measured interaction: architectures achieving superior discrimination (C-index 0.64

\to

0.82) exhibit equivalent or lower cross-modal interaction (4.8%

\to

3.0%). Variance decomposition reveals stable additive contributions across all architectures (WSI

{\approx}

40%, RNA

{\approx}

55%, Interaction

{\approx}

4%), indicating that performance gains arise from complementary signal aggregation rather than learned synergy. These findings provide a practical model auditing tool for comparing fusion strategies, reframe the role of architectural complexity in multimodal fusion, and have implications for privacy-preserving federated deployment.

Abstract

Cancer Survival Prediction

Explore similar work

Quantifying Cross-Modal Interactions in Multimodal Glioma Survival Prediction via InterSHAP: Evidence for Additive Signal Integration

Bridging the gap between Performance and Interpretability: An Explainable Disentangled Multimodal Framework for Cancer Survival Prediction

Attention-Based Multimodal Survival Prediction with Cross-Modal Bilinear Fusion