Title: Aligning Data-Driven Predictors with Allocation: A Decision-Focused Approach to Survival Analysis

Abstract:

Machine learning models have emerged as critical instruments for facilitating automated decision-making processes. Despite their widespread adoption, a significant disconnect remains: predictive algorithms are generally fine-tuned using conventional statistical benchmarks, often without considering the specific algorithmic tasks they are intended to support. This study underscores the dangers of this disconnect within the critical field of organ allocation. We demonstrate that algorithms dependent on survival predictors—regardless of their high accuracy when measured by standard metrics like the Concordance index (C-index)—can produce arbitrarily suboptimal results when applied to allocation tasks. In fact, such models do not even ensure utility superior to that of a purely random selection strategy.

To reconcile the divide between survival analysis and policy optimization, we propose a decision-focused learning framework centered on the optimization of Normalized Discounted Cumulative Gain (NDCG), a metric widely used in information retrieval. We validate the relevance of NDCG in the context of survival analysis by proving that it provides concrete guarantees regarding allocation performance. Empirically, we introduce a bootstrapping technique designed to enhance the NDCG scores of existing survival models. Distinct from previous studies, our methodology specifically accounts for the complexities of right-censoring during ranking evaluation. Analysis of historical heart transplant data in the United States reveals that our approach increases the NDCG of baseline models by 50% to 100%. When implemented for transplant allocation, this improvement is projected to yield tens of thousands of additional life years each year. We expect that this framework will extend to broader decision-making scenarios involving predictive analytics.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC