Title: Position Paper: Why We Must Abandon the C-Index for Evaluating Survival Analysis Models

Abstract:

The landscape of survival analysis evaluation is currently compromised by a widespread disconnect between the metrics employed and the actual objectives of the models. Furthermore, numerous assessments rely on censoring assumptions that remain either unstated or unsupported, rendering reported performance figures potentially deceptive and incapable of addressing the core scientific or modeling inquiries they were designed to answer. This position paper offers a critical review of existing evaluation methodologies in survival analysis, emphasizing how censoring introduces fundamental distinctions between survival tasks and traditional regression or classification problems. We scrutinize the excessive reliance on concordance-based metrics, specifically the C-index, which dominates current literature. To assist in selecting suitable metrics, we outline essential desiderata and introduce the "double-helix ladder" framework, which posits that rigorous evaluation demands strict alignment between the chosen metric and the underlying modeling assumptions. Our controlled experiments demonstrate that failing to maintain this alignment results in skewed and misleading model comparisons. The paper concludes with actionable recommendations for the proper evaluation of survival models.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC