Interpretable Hypergraph Neural Network Improves Glioblastoma Survival Prediction

Summary

The paper proposes a multimodal framework for glioblastoma multiforme (GBM) survival prediction that aims to improve predictive discrimination without sacrificing interpretability. It combines multi-modal MRI with clinical and genomic information, addressing the limitations of approaches that use a single imaging modality or treat accuracy and transparency as competing goals. The model uses a sheaf hypergraph neural network with directional, asymmetric message passing to represent higher-order relationships among tissue patches. A concept bottleneck layer compresses the learned representation into clinically grounded concepts, providing ante-hoc interpretability. An extension sufficiency test regularizer is applied during training to penalize explanations that do not faithfully reflect the model's decision process. Clinical and genomic features are integrated through gated fusion, preserving the prognostic contribution of molecular markers while retaining concept-level traceability. On 593 patients from the UPenn-GBM dataset, evaluated with 5-fold cross-validation, the framework achieved a concordance index of 0.643 and the lowest fold-level variance among the compared models, with a standard deviation of 0.015. The authors describe this as the first combination of sheaf hypergraph convolution, concept bottleneck supervision, and EST regularization for interpretable survival prediction from brain MRI.