MADSurv: An Uncertainty-Aware Framework for Multimodal Cancer Survival Analysis.

Area of Science:

• Oncology
• Biomedical Informatics
• Machine Learning

Background:

• Multimodal learning enhances cancer survival prediction using clinical, imaging, and genomic data.
• Existing methods often fail to account for data conflicts or unreliability across modalities.
• Current models typically provide relative risk scores, lacking precise survival probability estimates.

Purpose of the Study:

• To develop a novel framework, MADSurv, for improved cancer survival prediction.
• To address limitations in data fusion and the inability of current models to quantify survival likelihood.
• To enable more personalized and clinically relevant cancer risk assessment.

Main Methods:

• Proposed the Modality-Aware Discrete-Time Survival (MADSurv) framework.
• Implemented an uncertainty-aware attention mechanism for intelligent, confidence-based data fusion.
• Enabled predictions of survival probabilities at discrete yearly intervals, not just relative risk.

Main Results:

• MADSurv demonstrated superior and competitive performance across five cancer datasets.
• The uncertainty-aware attention mechanism improved robustness and personalization by prioritizing reliable modalities.
• The model accurately estimated survival probabilities at yearly milestones, validated by Brier scores.

Conclusions:

• MADSurv offers a more robust and personalized approach to cancer survival prediction.
• The framework enhances clinical utility by providing quantifiable survival probabilities over time.
• This work advances multimodal learning for improved cancer outcome assessment.