Transplant Center Acceptance as a Response to the Allocation Rule

Machine learning proposals for organ allocation are judged in simulators built from historical registry data. In those simulators, the decision that matters most is whether a transplant center accepts an offered heart, and it is modeled as a fixed probability learned from past offers. That modeling choice is the weak point. Centers compete for a scarce supply of organs. When they decide whether to accept, they weigh the organ in front of them against what they expect to be offered later, and against how a risky transplant will look on their next performance report. Both of those depend on the allocation rule itself. A simulator that replays historical acceptance behavior under a new rule is therefore replaying the equilibrium of the old rule. This is the retrospective-study version of Goodhart's law, and it is the gap that the incentive-focused position paper on heart allocation names without resolving (Position: Machine Learning for Heart Transplant Allocation Policy Optimization Should Account for Incentives).

We propose to model each center's accept-or-decline choice as a best response: the center compares the value of transplanting now, minus an audit penalty tied to its report-card standing, against the continuation value of waiting for a later offer. That continuation value is computed from the allocation rule, so changing the rule changes predicted acceptance. We fit the model on UNOS heart offer data from before the October 2018 policy change and test whether it predicts post-2018 acceptance better than a stationary acceptance classifier. We then re-rank published allocation policies under the re-equilibrated behavior.

Core Claim

Acceptance behavior in the registry is policy-dependent in a way that can be estimated. A structural acceptance model, one whose parameters describe the center's preferences rather than its observed accept rates, should transfer across the 2018 rule change where a reduced-form classifier cannot.

If this holds, the reported gains of non-myopic allocation policies must be re-measured. Those gains were computed either under the assumption that centers always accept (Learning Potentials for Dynamic Matching and Application to Heart Transplantation) or with a fixed acceptance predictor (Policy Optimization for Dynamic Heart Transplant Allocation). The re-measured values may shrink, and the ranking among policies may change.

Relation to Prior Work

Four lines of work motivate or constrain this proposal, and one supplies the underlying simulator.

The heart-allocation program of Anagnostides, Zilberstein, Sandholm and colleagues supplies both the simulator and the motivation. Policy Optimization for Dynamic Heart Transplant Allocation builds a UNOS simulator with a LightGBM acceptance classifier and Cox survival models. It finds that "more permissive" centers would raise efficiency by about 23%, but it treats permissiveness as an exogenous dial. Learning Potentials learns opportunity-cost potentials by imitating an omniscient matching oracle, and it explicitly assumes every offer is accepted.

The position paper is the empirical motivation. It documents three behaviors that no stationary model captures:

  • a jump in intra-aortic balloon pump use from 7% to nearly 25% after the 2018 tier change;
  • a May spike in transplants just after the semi-annual audit cutoff;
  • small centers accepting offers at a rate nearly 50% higher than large centers.

It proposes no estimator. We keep their simulator and replace its acceptance component.

Causal matching methods constrain the design. Improving Causal Transplant Outcomes through Dynamic Organ Offer Estimation (DynamITE) models how sequential rejections degrade the organ through cold ischemic time, the hours an organ spends preserved outside the body. It treats this as a violation of the no-interference assumption, but its acceptance model is a stationary probability evaluated on simulated decisions. OrganITE removes historical assignment bias with adversarial representations under an assumption of no unmeasured confounding. We borrow DynamITE's treatment of rejection-induced degradation. We differ from both in making the acceptance probability a function of the rule and of the center's competitive position.

Retrospective-bias studies set the ceiling on what we can claim. The graft survival model in the UNOS simulator reaches a concordance index (C-index, the probability that a model ranks two patients' survival times correctly) of only 0.600. Propensity methods balance only recorded covariates (Laparoscopic versus open pancreaticoduodenectomy). So our model must absorb unrecorded organ quality rather than ignore it.

The budget-constrained policy evaluator of Improved Policy Evaluation for Randomized Trials of Algorithmic Resource Allocation is the closest statistical tool. It rules itself out here because it assumes no interference between units, it requires trial arms, and it treats recipients as passive.

Research Program

The acceptance model has three parts. A center $c$ offered organ $o$ for patient $p$ at time $t$ accepts when

$$ u_c(p,o) - \lambda_c \, g_c(t) \, r(p,o) + \eta_o + \epsilon > V_c^{\pi}(p,t). $$

The symbols are:

  • $u_c(p,o)$ is the center's expected graft benefit, taken from the survival models in the existing simulator.
  • $r(p,o)$ is the predicted one-year graft-failure risk.
  • $g_c(t)$ measures report-card pressure: the center's current margin against its expected-outcome threshold, together with the days remaining to the next audit cutoff.
  • $\lambda_c$ is a center-specific weight on that pressure, pooled hierarchically by center size.
  • $\eta_o$ is a latent organ-quality term.
  • $\epsilon$ is a logistic shock.
  • $V_c^{\pi}(p,t)$ is the continuation value of declining. It is the expected value of the next acceptable offer for patient $p$ before the patient dies or deteriorates, computed by simulating future arrivals under allocation policy $\pi$ and the patient's waitlist hazard.

Two design choices carry the proposal.

The first is how we handle unmeasured organ quality. The same heart is offered to many candidates in sequence, so the shared term $\eta_o$ is identified from the pattern of rejections across centers. A string of refusals is itself evidence of quality the registry does not record. This is our targeted answer to the unmeasured-confounding problem that caps the C-index at 0.600. It explains away only the part of confounding that is shared across all decision-makers seeing the same organ. Patient-level confounding remains, and we will report that residual honestly.

The second is that $V^{\pi}$ depends on $\pi$. This dependence is what lets the model predict that a rule change reshapes acceptance. Center competition enters through $V^{\pi}$: a large center that expects neighbors to absorb marginal organs sees a higher value of waiting.

Estimation is by maximum likelihood with an inner loop that recomputes $V^{\pi}$ by simulation. We use a fixed-point iteration so that each center's continuation value reflects other centers' fitted acceptance rules. Counterfactual evaluation of a new policy repeats this fixed point under the new rule and then runs the existing simulator.

We assume access to offer-level match-run records of the kind the UNOS simulator already uses. Out-of-sequence allocations are about 19% in 2023 per the position paper. We flag them and model them as a separate bypass channel rather than as ordinary accept decisions.

Baselines

The naive reference point is the always-accept assumption of Learning Potentials. Under it, reported mean population life years gained per month (PLYG) are:

Policy Mean PLYG per month
Status quo 1,422
Proposed continuous distribution 2,353
Myopic maximizer 3,280
Potential-based neural network 3,302
Omniscient oracle 3,475

We will rerun all five policies under each acceptance model. A strategic acceptance model matters for policy choice if it changes the gap between the potential-based network and the myopic policy by more than simulation noise, or if it reorders any pair of policies.

The strongest existing competitor is the stationary LightGBM acceptance predictor of Policy Optimization for Dynamic Heart Transplant Allocation. It is the component actually used to evaluate these policies. That paper reports a graft-survival C-index of 0.600. For acceptance-prediction accuracy we found no reported result on Lacuna for the LightGBM acceptance predictor, so we will retrain it ourselves. It gets the same pre-2018 training offers, the same patient and donor features, and the same hyperparameter search budget as our model, and is scored on the same post-2018 held-out offers. Beating it means better log-loss and calibration on post-2018 offers, especially within strata of center size. It also means reproducing the direction and size of the post-2018 shift in each center's acceptance rate.

A second competitor is DynamITE's PatientNet-style multi-task acceptance model (DynamITE). We refit it on the same heart offer data, since its published results are on semi-synthetic liver data.

The ablations isolate each ingredient. One removes the audit term by setting $\lambda_c = 0$. One removes policy dependence by freezing $V$ at its pre-2018 values. One removes the shared organ-quality term $\eta_o$. The frozen-$V$ ablation is the decisive one. If it transfers across 2018 as well as the full model, our core claim fails.

Why not just…?

Why not just add incentive features to the LightGBM acceptance model? This is the real threat. Features such as days to audit cutoff, report-card margin and neighbor-center load may well reproduce the May spike and the gap between small and large centers within a single policy regime. What such a reduced-form model cannot do is respond to a rule it has never seen. The allocation rule is not one of its inputs, so the value of waiting under a new rule has no way to enter.

We will give LightGBM exactly these features and compare it with our model on the post-2018 test. If augmented LightGBM matches us out of regime, we will conclude the structural layer is unnecessary for this change and report that plainly. In that case the remaining value of our model is limited to rules that alter the audit itself, where no historical variation exists to fit.

Why not just run sensitivity analysis on a uniform permissiveness dial, as in the 23% counterfactual? A single dial assumes all centers move together. The position paper's finding of a nearly 50% higher acceptance rate at small centers suggests heterogeneous, possibly opposite-signed responses. We will test this directly: we check whether the observed post-2018 changes in center acceptance rates are well explained by one shared shift, measured by the variance explained across centers.

Why not just use a causal policy evaluator such as assignment permutation or inverse propensity weighting? Improved Policy Evaluation shows that inverse propensity weighting has exploding variance under budgeted sequential allocation and is undefined for deterministic ranking policies. Its permutation fix requires trial arms and assumes no interference. Organ allocation has no randomized arms, and one center's refusal changes both the organ and the queue for the next candidate. So these tools do not apply without the kind of behavioral model we propose.

Decisive Checks

The first test is out-of-regime transfer. We train on offers from 2010 to October 2018 and evaluate log-loss, calibration and per-center acceptance-rate shifts on offers from late 2018 to 2019. The full model, the frozen-$V$ ablation, augmented LightGBM and the PatientNet variant are all scored on this split.

The second test is behavioral replication. The fitted $\lambda_c$ should reproduce the May transplant spike in held-out years. For the status incentives, we attach to the model a simple choice of patient status in which clinicians trade intervention cost against tier priority. We ask whether it predicts a rise in balloon-pump use of roughly the documented size, compared with the 7% to nearly 25% jump the position paper reports.

The third test is aggregate validation. Simulated post-2018 transplant volumes and waitlist deaths under each acceptance model should be compared with realized values. The Kaplan–Meier one-year survival of 89% reported by Luo et al. serves as a sanity anchor.

Only after these tests pass do we report re-ranked PLYG for the five policies. The headline result would be a measured change in how much the potential-based policy gains over the myopic one under re-equilibrated acceptance.