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

Aniek Eijpe, Soufyan Lakbir, Melis Erdal Cesur, Sara P. Oliveira, Angelos Chatzimparmpas, Sanne Abeln, Wilson Silva

Keywords: Cancer survival prediction, Disentangled representation learning, Multimodal deep learning, Explainable AI, Histopathology whole-slide images, Transcriptomics data

Registration Requirement: Yes

Abstract:

While multimodal survival prediction models are increasingly accurate, their complexity limits interpretability, reducing insight into how different data sources drive predictions. We introduce DIMAFx, an explainable multimodal framework that learns disentangled modality-specific and modality-shared representations from histopathology whole-slide images and transcriptomics data. Across multiple cancer cohorts, DIMAFx\footnote{This short paper is a shortened version of our preprint~\cite{eijpe2026bridging}.}
achieves state-of-the-art performance while improving disentanglement. Through a case-study in breast cancer, we find that the most predictive features are modality-shared, linking solid tumor morphology with late estrogen response---consistent with known biology---while key modality-specific features capture microenvironmental signals from the WSI. These results show that disentangled multimodal representations provide biologically meaningful insights without sacrificing predictive performance, supporting their use in precision medicine.

Reproducibility: https://github.com/Trustworthy-AI-UU-NKI/DIMAFx

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 68