EMMS model predicts cancer survival with missing data using uncertainty fusion

Researchers from the National University of Singapore and other institutions have introduced the Evidential Missing Modality Survival Fusion (EMMS) model, a new approach to survival prediction that handles missing patient data without requiring generative imputation. Published on arXiv, the model leverages Dempster-Shafer theory combined with Gaussian Random Fuzzy Numbers to fuse multimodal clinical data—such as imaging, genomics, and pathology—even when some modalities are absent. By treating missing inputs as vacuous evidence, EMMS prevents corrupted signals from interfering with available data while naturally increasing uncertainty in predictions. This design yields calibrated and interpretable uncertainty estimates, a critical feature for high-stakes medical decisions.

Extensive experiments on four cancer datasets (including TCGA cohorts) demonstrate that EMMS achieves state-of-the-art survival prediction performance while maintaining robustness under various missing-modality scenarios. Unlike existing methods that either assume complete data or rely on computationally expensive generative models for missing data, EMMS offers a straightforward, computationally efficient solution. It quantifies both aleatoric uncertainty (inherent noise in data) and epistemic uncertainty (model ignorance) for each modality, allowing clinicians to gauge prediction reliability. The work represents a significant step toward deploying AI-driven survival models in real clinical workflows where incomplete patient records are the norm.