Explainable AI for Pan-Cancer Risk Stratification Accurate risk stratification remains central to precision oncology, yet conventional survival models often fail to translate complex analytics into clinically actionable insights. This study introduces a novel, model-agnostic artificial intelligence framework that enables unsupervised identification of prognostically distinct patient groups by directly optimizing survival heterogeneity across clusters. The approach was validated across multiple cancer types and data modalities, including laboratory datasets from multiple myeloma patients (CoMMpass dataset) and imaging data (CT scans) from non-small cell lung cancer (Lung1 dataset).
The model demonstrated robust performance in identifying clinically meaningful risk clusters without predefined labels. Importantly, post-hoc explainability analyses revealed key features driving patient…
