Ranking Organ Offers Within the Same Patient to Measure Over-Discard

More than 95% of deceased-donor organ offers in the United States are declined. Kidney acceptance sits near 1% and liver acceptance near 3%. While each offer moves down the waiting list, the organ keeps losing quality because it is held cold and without blood supply, the "cold ischemia clock" (Causal Explanation-Guided Learning for Organ Allocation). To reduce discards, one first has to measure how far clinicians' revealed sense of organ quality departs from how those organs actually perform once transplanted. Existing models learn acceptance one offer at a time. Their labels therefore mix three things: the organ, the patient's unrecorded frailty and the surgeon's view of that patient, and the patient's position on the list. Image quality research hit the same problem when absolute ratings mixed picture content with degradation. It addressed it by comparing several versions of the same scene, so that whatever the versions share drops out (Underwater Ranker: Learn Which Is Better and How to Be Better, Learning to Rank Retargeted Images).

We propose to treat each waitlisted candidate as the "scene." A liver candidate typically declines several organs before accepting one. Comparing the accepted organ against the organs that same candidate's team declined cancels everything constant about that candidate, including unrecorded factors. What remains is a clinician-revealed organ desirability score, which we learn with a neural organ encoder. We then put that score next to post-transplant outcomes. Organs the score ranks low but that did well after being accepted elsewhere are a direct, auditable measure of over-discard. The claim is not that pairwise losses beat pointwise ones in general. It is that within-candidate comparison yields an organ score that is less contaminated by candidate-level confounding and by list-position shortcuts, and that this cleaner score makes the gap between clinician preference and organ outcome visible.

Core Claim

Let $q_\theta(o)$ be a learned desirability score for organ $o$, computed from donor and organ features by a network with parameters $\theta$. Let $\tau_c(t)$ be candidate $c$'s acceptance threshold at offer time $t$. We model acceptance as

$$P(\text{accept } o \mid c, t) = \sigma\big(q_\theta(o) - \tau_c(t)\big), \qquad \tau_c(t) = \alpha_c + g_\phi(x_c(t)),$$

where $\sigma$ is the logistic function and $\alpha_c$ is a per-candidate constant capturing everything stable and unrecorded about the patient. The term $g_\phi$, a small network with parameters $\phi$, describes how the threshold shifts with the candidate's time-varying state $x_c(t)$: urgency score at the offer, time on list, and list position.

We take the candidate's eventual accepted offer $(o^+, t^+)$ and each earlier declined offer $(o^-, t^-)$. In the logit difference for this pair, $\alpha_c$ cancels exactly. Training on these pairs is a stratified Bradley-Terry objective, the same form PieAPP uses for image pairs (PieAPP: Perceptual Image-Error Assessment Through Pairwise Preference):

$$\mathcal{L} = -\sum_{c}\sum_{(o^+,o^-)} \log \sigma\Big(\big[q_\theta(o^+) - q_\theta(o^-)\big] - \big[g_\phi(x_c(t^+)) - g_\phi(x_c(t^-))\big]\Big).$$

The second bracket matters. Candidates deteriorate over time and accept worse organs later. Without the bracket, the model would read rising urgency as falling organ quality.

As in PieAPP, the pairwise structure is used only during training. At inference, $q_\theta$ scores a single organ in one pass. This answers the latency concern raised against comparison-based quality models, since the cold ischemia clock leaves little time.

Relation to Prior Work

Three works define the space this proposal sits in.

Causal Explanation-Guided Learning for Organ Allocation (CLEXNET) models offer acceptance pointwise. It removes allocation-policy context with adversarial balancing and adds synthetic counterfactual offers nudged in the direction implied by stated refusal reasons. It motivates this proposal because it documents that pointwise acceptance models take a shortcut through list rank. We differ by removing candidate-level confounding structurally, through within-candidate conditioning, rather than adversarially. We also do not need refusal codes, although they could later be added as extra directional pairs.

OrganITE ranks recipients for a fixed organ using individual treatment effects and organ rarity. OrganSync and Learning Matching Representations for Individualized Organ Transplantation Allocation group organs into discrete clusters to estimate counterfactual survival and waiting times. These supply the outcome side of our comparison. They do not model clinician revealed preference, so they cannot locate where clinicians and outcomes disagree. They also treat organ quality as coarse types, whereas our score is continuous.

From the quality-ranking literature we reuse two ideas. The first is within-scene comparison, from URanker and retargeting ranking. The second is that comparative judgments are more internally consistent than absolute ones. SCOREQ supports this point: its comparison-based training generalized out of domain better than plain squared-error regression, reaching Pearson correlation 0.86 against 0.75 (SCOREQ: Speech Quality Assessment with Contrastive Regression).

We do not claim the loss itself is new. With linear $q$ and $g$, the objective reduces to a conditional logit, which is a standard econometric tool. The contribution is the measurement design and what it reveals about discard.

Research Program

The primary data are the UNOS liver offer records (UNOS-PTR), the same 2021–2024 roughly 1.1-million-offer source CLEXNET used. These are linked to post-transplant graft and patient survival in the UNOS liver registry that OrganSync used. Candidate identifiers are required to form within-candidate strata.

We build the organ encoder $q_\theta$ over donor demographics, labs, donation type, and cold ischemia time at offer. We build the threshold network $g_\phi$ over candidate urgency, time on list, and sequence number. We train on candidates who eventually received a transplant, splitting by candidate and by calendar year to prevent leakage.

The analysis then has three steps. First, we test whether $q_\theta$ is free of list-position shortcuts. Among offers to different candidates with matched organ features, $q_\theta$ should not vary with sequence number. Second, we test how well $q_\theta$ predicts outcomes among transplanted organs. Third, and centrally, we construct a discard-gap estimate. For organs declined many times and then transplanted, we compare their revealed-preference rank with their realized graft survival relative to outcome-model predictions. We then characterize the donor feature regions where clinicians systematically under-rank organs that perform adequately.

Because real data lack ground-truth organ quality, we follow CLEXNET's semi-synthetic protocol. We plant a true organ quality and an unobserved candidate-level factor that drives both acceptance and outcome, then simulate acceptance decisions. This gives exact recovery targets.

Baselines

The naive reference point is a pointwise logistic regression and a gradient-boosted classifier predicting acceptance from organ and candidate features with no structural treatment of confounding, and with list position included. Both CLEXNET and InsighTx use gradient boosting or similar methods as the representative of this family (A machine-learning approach to human ex vivo lung perfusion predicts transplantation outcomes and promotes organ utilization). Its learned organ coefficient or partial dependence serves as its "desirability score."

The strongest existing competitor is CLEXNET. It targets exactly this confounding problem on exactly this data with a deep model. We will reimplement it on identical splits and features, with an equal hyperparameter search budget and the same encoder width. We will extract its organ-only component, the acceptance logit with candidate features fixed at population means, as its desirability score. No reported result was found on Lacuna for CLEXNET's raw Brier, AUC, or AUPRC values; the file states only that it ranked best against logistic regression, random forests, and causal baselines. On the outcome side, the strongest reported reference is OrganSync's factual survival mean absolute error of 715 days on UNOS, against 2,087 days for clinical baselines (OrganSync). We will use OrganSync's survival model as the outcome reference against which discard gaps are measured, not as a competitor.

Two ablations isolate the key ingredient. A pooled pairwise ablation forms the same accepted-versus-declined pairs but across candidates, so $\alpha_c$ does not cancel. This separates the effect of the pairwise loss from the effect of within-candidate conditioning. A linear conditional logit ablation tests whether the neural encoder adds anything beyond the design.

Beating the competitors means three things. In the semi-synthetic setting, Spearman correlation between $q_\theta$ and the planted organ quality must be higher, with a margin exceeding seed-to-seed variation across five seeds, as the strength of the planted candidate confounder grows. On real data, the score must show weaker dependence on sequence number conditional on organ features. It must also show higher concordance with graft survival among transplanted organs, assessed with paired bootstrap intervals.

Why not just…?

Why not just add candidate or center fixed effects to a pointwise acceptance classifier? With a linear score this is mathematically our method, and we say so plainly. The linear conditional logit ablation is exactly this alternative. If it matches the neural version, the honest conclusion is that the within-candidate design, not the network, does the work. That would still be a useful finding for allocation research, which currently does not use it. Fitting a separate intercept per candidate inside a boosted tree or deep classifier is not equivalent, however. With few offers per candidate, the intercepts are poorly estimated and leak into the organ effect. Conditioning removes them without estimating them, and the semi-synthetic sweep over candidate-level confounding will show whether this matters.

Why not just predict post-transplant outcomes directly and ignore clinician decisions? Outcome models such as OrganSync and OrganITE observe outcomes only for accepted organs. They therefore cannot say whether an organ was discarded because it was bad or because clinicians were cautious. Measuring over-discard requires both a model of what clinicians prefer and a model of what organs do, compared on the same organs. Our score is the first of these; OrganSync's survival model serves as the second.

Why not just use CLEXNET's refusal-reason signals? CLEXNET corrects confounding through adversarial balancing and through synthetic offers that it requires to receive higher predicted acceptance than the originals. It assumes that recorded refusal reasons map onto single feature axes. Within-candidate comparison needs no refusal codes and makes no assumption about what they mean. The two are complementary, and the direct head-to-head on shortcut dependence and recovery will settle which removes more confounding. If CLEXNET wins in the semi-synthetic setting, we will add its directional pairs to our loss rather than argue against them.

Why not just rank organs within the same donor, or within the same organ across centers? Within-organ comparison across the candidates who declined it is a natural complementary design. It isolates candidate thresholds rather than organ quality. We will report it as a secondary analysis. It cannot by itself produce an organ score, because the organ is held fixed.

Decisive Checks

Four results decide the proposal. In the semi-synthetic UNOS-PTR setting, the within-candidate model must keep its correlation with planted quality as candidate-level confounding increases, while the pointwise, CLEXNET, and pooled pairwise baselines degrade. If it degrades alongside them, the core claim fails. Without the threshold-drift term $g_\phi$, the model should wrongly rate organs accepted late in a candidate's course as worse; this confirms the deterioration confound is real and handled. On real data, the learned score should depend less on offer sequence position than the baselines' scores do. Finally, among organs declined at least ten times before transplant, the fraction with realized graft survival at or above outcome-model expectations yields a concrete over-discard estimate tied to identifiable donor feature regions. If that fraction is indistinguishable from the fraction among readily accepted organs, the clinician preference gap is small and the scarcity argument for rescue weakens. Either way, the answer matters for allocation policy.