Accurate risk stratification of cancer patients is essential to propose a relevant treatment for each patient. In the clinical field, risk stratification is usually based on models or scoring systems relying on known biomarkers for the pathology. Yet, learning-based methods have been proven to be able to improve stratification compared to such empirically-defined indicators. These methods usually either learn directly to classify the patients into binary or multiclass risk groups [21, 22], or they perform a survival analysis from which they derive such risk groups, an approach strongly privileged by clinicians (as in [2]). However, these methods only use the individual information of each patient to make a prediction, without taking advantage of the information of individuals who have a profile similar to that of the patient considered.