Trustworthy Multimodal Fusion for Cancer Survival Analysis

The complexity of survival analysis

Predicting cancer survival — formally, time-to-event analysis — is one of the harder tasks in oncology. A prognosis rarely rests on a single data point; clinicians draw on whole-slide images (WSIs), genomic profiles, structured records, and free-text notes. Two problems make this hard for AI:

  1. Heterogeneity: how to fuse a pixel-dense pathology slide with a text-based clinical record.
  2. Censoring and noise: survival data is noisy and frequently censored (a patient leaves a study before an event occurs). Standard black-box models handle this poorly and tend to be overconfident.

Evidential fusion and epistemic random fuzzy sets

My research builds models that represent their own uncertainty. Supported by the Dame Julia Higgins Fund for the project “Revolutionizing Cancer Survival Analysis Through Trustworthy AI and Multimodal Integration”, this work changes how models handle heterogeneous clinical data.

One example is EsurvFusion, an evidential multimodal survival model published in IEEE Transactions on Fuzzy Systems. Rather than forcing the modalities through a standard network, it uses epistemic random fuzzy sets — specifically Gaussian random fuzzy numbers — to quantify both data and model uncertainty. A reliability-discounting step down-weights a modality (say, a low-quality scan) when the model judges it unreliable, leaning instead on the clearer sources.

Dual-prototype architectures for pathology

We also extend these methods to whole-slide pathology. In DPsurv (Dual-Prototype Evidential Fusion for Uncertainty-Aware and Interpretable Whole-Slide Image Survival Prediction), we handle the scale of whole-slide images with a dual-prototype evidential fusion that extracts survival-relevant features while keeping uncertainty explicit. Related work on time-to-event prediction (International Journal of Approximate Reasoning) keeps the quantified uncertainty well calibrated to real outcomes.

Clinical impact

By centring uncertainty quantification and reliability, this work aims to make predictive models safe, transparent partners for oncologists — giving not just a survival estimate, but an indication of why the model reached it and how confident it is.