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

Itai Zilberstein, Ioannis Anagnostides, Tuomas SandholmarXiv 2026 · arXiv

TL;DR

Standard survival analysis models optimize for aggregate statistical metrics like the Concordance Index (C-index), which can lead to disastrously poor decisions in high-stakes resource allocation tasks like organ transplantation. This paper bridges the gap by introducing a decision-focused learning approach that optimizes for Normalized Discounted Cumulative Gain (NDCG), providing theoretical utility guarantees and a bootstrapping framework that doubles top-tier allocation accuracy—potentially saving tens of thousands of life-years annually.

What Problem Does This Paper Address?

When a donor organ becomes available, matching algorithms rely on machine learning predictors to estimate patient survival times and allocate the organ to the candidate who will benefit the most. However, a critical disconnect exists: these predictive models are traditionally trained and evaluated using aggregate metrics like the C-index, which measures pairwise ranking accuracy across an entire dataset.

The authors demonstrate that a model can achieve a near-perfect C-index (e.g., 0.98) while consistently picking the absolute worst candidate for the top spot, yielding an allocation policy no better than random guessing. Because resource allocation is a top-heavy problem—where only the single best candidate receives the item—maximizing aggregate accuracy across the entire population is misaligned with the ultimate decision-making goal.

Figure 1: Illustration of heart transplant allocation with predicted outcomes. The leftmost value shows the unknown, ground-truth patient survival, the middle green value shows the predictions with NDCG@1 = 0.9, and the rightmost red value shows the predictions with C-index = 0.8.

Methods

To align prediction with allocation, the authors borrow a concept from information retrieval: Normalized Discounted Cumulative Gain (NDCG), a metric designed to evaluate the quality of the top-$k$ ranked items.

Implementing NDCG in survival analysis introduces a unique challenge: right censorship. Because some patients are still alive or leave a study early, their true survival times are unobservable, meaning standard NDCG cannot be computed directly. The authors overcome this by introducing two novel, mathematically unbiased estimators for Discounted Cumulative Gain (DCG) under right censorship:

  1. Expected Years (EY) Estimator: Imputes unobserved survival times using conditional expectations derived from a baseline survival function.
  2. Inverse Probability of Censoring Weighting (IPCW) Estimator: Discards censored patients but re-weights observed instances based on the probability of remaining uncensored.

Using these estimators, they propose a two-stage bootstrapping framework:

  • Stage 1: Train a standard baseline survival model to generate conditional survival functions and compute pseudo-labels that blend observed outcomes with expectations.
  • Stage 2: Train a second model (a Gradient Boosted Decision Tree) using a hybrid loss function ($L_{hybrid}$). This loss combines Mean Squared Error (for stability) with a LambdaRank-inspired pairwise ranking penalty scaled by the batch-wide change in NDCG ($|\Delta NDCG_{i,j}|$), forcing the model to prioritize the correct ordering of top-ranked candidates.

Datasets

The framework was evaluated on the United Network for Organ Sharing (UNOS) patient registry, containing clinical data for adult heart transplants in the US from 1987 to the present.

  • Full Dataset: 60,055 total patients, with a 49.47% right-censorship rate and 92 clinical features.
  • Artificial Censoring Setup: To validate the accuracy of their NDCG estimators against a known ground truth, the authors also created a fully observed subset of patients (where exact survival times were known) and introduced simulated random censoring.

Key Insights and Ideas

  • The Illusion of C-Index: A high C-index provides a false sense of security. It is entirely non-informative for top-1 selection because a model can correctly rank 99% of arbitrary patient pairs while failing catastrophically at identifying the single best candidate.
  • NDCG as a Policy Guarantee: The authors mathematically prove that a predictor's NDCG@1 directly translates to a lower bound on the expected utility of a greedy allocation policy.
  • Preserving Relative Rankings: While sorting noisy survival estimates introduces a negative bias in the absolute NDCG value, the authors show that this bias does not affect relative model comparison. If $\widehat{NDCG}_A > \widehat{NDCG}_B$, Model A is definitively superior for allocation.

Key Results

  • Massive NDCG Boosts: On the real-world UNOS dataset, the bootstrapping framework dramatically outperformed standard survival models (like Cox regression, AFT, DeepSurv, and DeepHit). It roughly doubled the NDCG@1 scores, yielding improvements of 50% to 100% across all baselines.
  • Life-Years Gained: This statistical boost translates to monumental real-world impact. When deployed for heart transplant allocation in the US, the framework is projected to gain nearly 50,000 additional life-years annually without sacrificing overall model calibration or C-index.

Limitations and Future Work

The authors note several limitations inherent to historical medical registries, including missing data, systemic clinical biases, and unobserved confounders. Furthermore, the unbiasedness of their DCG estimators relies heavily on the correct specification of "nuisance models" (the baseline survival or censoring distributions); if these underlying models are miscalibrated, the EY estimator inherits that bias, and the IPCW estimator suffers from high variance.

Future work includes exploring how model calibration techniques can mitigate these biases and extending this decision-focused framework to other domains of discrete resource allocation under uncertainty.

Summary updated

Abstract

Machine learning predictors have become essential tools for guiding automated decision making. However, a major misalignment persists: predictive models are typically optimized in terms of standard statistical metrics in isolation from the algorithmic tasks they inform. We highlight this incongruity in the high-stakes domain of organ allocation by demonstrating that any algorithm relying on (even highly accurate) survival predictors optimized for standard metrics -- such as the Concordance index (C-index) -- can yield arbitrarily poor outcomes when used for allocation, failing to guarantee utility better than a uniform random selection. To bridge the gap between survival analysis and policy optimization, we introduce a decision-focused learning approach based on optimizing normalized discounted cumulative gain (NDCG), a mainstay metric in information retrieval. We establish the utility of NDCG in survival analysis by proving that it translates to guarantees on the performance of allocation. Empirically, we propose a bootstrapping approach to optimize the NDCG of existing survival models. Unlike prior work, we also address the challenge of right censorship when evaluating ranking. On historical heart transplant data from the US, our method dramatically boosts the NDCG of baseline models by 50-100%, which translates to tens of thousands of additional life years gained annually when deployed for transplant allocation. We anticipate that our framework will find broader applications in decision making with predictions.
BibTeX
@misc{zilberstein2026aligning,
  author = {Itai Zilberstein and Ioannis Anagnostides and Tuomas Sandholm},
  title = {{Aligning Data-Driven Predictors with Allocation: A Decision-Focused Approach to Survival Analysis}},
  year = {2026},
  eprint = {2606.02671},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2606.02671},
}

Loading related content...

Figure