PaperlayerView exact revision on arXiv

Source-linked AI summary

When Should AI Decide, Defer, and Deviate? Responsible Automation for Deceased Donor Kidney Allocation

Madeleine Pollack, Thomas Daillak, Mehmet Toner, Korkut Uygun, Heidi Yeh, Swati Gupta

arXiv:2610.39045v1cs.AI

TL;DR

Historical kidney-allocation decisions may not be worth replicating, while OOS placement trades reduced nonuse against fairness concerns. The paper develops an uncertainty-aware, interpretable Arbitrator and audits its rules against clinical literature; it resolves most placement decisions and flags some historically discarded kidneys for expedited placement.

  • Problem

    The paper asks when AI trained on historical kidney-allocation decisions should decide, defer, or deviate, given that historical actions may not reflect what should have been done and uncertain predictions may be unsafe to act on.

    Source: §1 · p. 2

  • Method

    The Arbitrator combines interpretable early-matchability models, uncertainty-based certification and deferral, and a clinical-literature audit of resolved rules.

    Source: §1 · p. 3

  • Results

    93.5% of organs were resolved with interpretable rules, and 70.1% of historically discarded low-KDPI test kidneys were flagged for expedited placement; if performance generalizes, 43.9% of historically discarded kidneys could be flagged.

    Source: §4 · p. 13

  • Takeaways & Limitations

    The policy certifies interpretable rules for possible automation only when recommendations are stable under uncertainty and explicitly approved after clinical-literature review.

    Source: §4 · p. 13

    Early matchability uses acceptance by original PTR sequence 100 as a proxy for difficult placement, and the model predicts placement likelihood rather than transplant success.

    Source: §4 · p. 13

Abstract

from arXiv · show

We ask when an AI trained on historical decisions should be allowed to recommend decisions, defer to human experts, or in other cases, flag concerning patterns in data and deviate from past practice. This can be challenging because past decisions record what was done rather than what should have been done, and where data is sparse, a model's predictions may be too uncertain to act on. We provide a framework for addressing this problem using a case study in deceased-donor kidney allocation. Nearly 30% of kidneys recovered from deceased donors in the United States are never transplanted, largely because they are hard to place through the standard waitlist sequence. One can use AI to flag kidneys at risk of delayed or failed placement for out-of-sequence (OOS) allocation, but this comes at the cost of bypassing candidates higher on the waitlist. Thus, OOS allocation should only be used when the organ is demonstrably at high risk of discard on the standard sequence. Using Organ Procurement and Transplantation Network match-run data from 2016-2019 and 2023-2025 (excluding COVID-19-disrupted years), we build the Arbitrator framework, which combines various interpretable models and acts only when the uncertainty in their predictions cannot change the preferred action. Our Arbitrator resolves 93.5% of organ placement decisions using explicit medical criteria learned from data, and defers the remaining 6.5% to a human or a more complex model. We show that the Arbitrator correctly identifies 70.1% of historically discarded low-KDPI kidneys as OOS candidates, projecting that up to 43.9% of historically discarded kidneys could be flagged for expedited placement and potentially transplanted. We then exploit the interpretability of the Arbitrator to create an agentic audit comparing the learned rules to published clinical literature in order to flag decision-making patterns that are contentious.

1 Introduction

Kidney allocation must balance the risk of nonuse against the fairness cost of bypassing higher-priority candidates. The paper proposes an interpretable, uncertainty-aware framework that can decide, defer, or flag recommendations for review against clinical literature.

1 Introduction

  • OOS placement can reduce nonuse of hard-to-place kidneys, but it bypasses higher-priority candidates and may worsen inequities, so it should be reserved for organs unlikely to place through the standard sequence.The match run is designed to make offers systematic, fair, and auditable.Source: §1 · p. 2
  • The Arbitrator predicts early-matchability and considers OOS allocation when a kidney is unlikely to be accepted within 100 match-run offers.The preferred action depends on β, the relative preference for maintaining the match run versus routing high-risk organs OOS.Source: §1 · p. 3
  • Models certify an action only when its confidence interval lies entirely on one side of the decision threshold; otherwise, the Arbitrator defers the case for additional review.The paper uses XGBoost for deferred organs.Source: §1 · p. 3
  • An auditor compares each resolved rule with clinical literature and can provisionally retain the action, recommend the opposite action, or request further review.This is intended to avoid blindly replicating historical decisions that may reflect bias, noise, or outdated guidance.Source: §1 · pp. 2–3

2 Materials and Methods

The methods construct a standard-sequence kidney cohort, fit interpretable predictors with validation-based uncertainty intervals, and use cost-sensitive certification to resolve or defer recommendations. A literature-assisted audit then assesses the certified rules.

2.1 Data sources and cohort construction

  • The cohort uses OPTN match-run data linked to donor and waitlist registries, with donors split across training, validation, and test sets to keep each donor’s kidneys together.The analysis includes 100,674 training, 21,573 validation, and 21,574 test organs; test outcomes are reserved for final evaluation.Source: §2.1 · p. 4

2.2 Prediction models and calibration

  • The study evaluates six predictors: thresholded and calibrated KDPI rules, logistic regression, CART, Random Forest, and XGBoost; isotonic calibration is fitted for four models, while raw KDPI remains uncalibrated.Source: §2.2 · pp. 4–5

2.3 Local uncertainty

  • CART and SparseRF estimate early-matchability from validation organs and use 99% Wilson intervals; unsupported organs receive the uninformative interval [0, 1].SparseRF extracts a compact rule set from a trained Random Forest and combines intervals when multiple supported rules apply.Source: §2.3 · p. 5

2.4 Cost-aware certification and arbitration

  • With β setting the relative cost of unfair OOS routing, the optimal threshold is τβ = 1/(1 + β), and a model certifies only when its full uncertainty interval supports one action.The Arbitrator uses a certified action when available and defers when both models abstain or certify conflicting actions.Source: §2.4 · p. 5

2.5 LLM-assisted clinical rule audit

  • The exploratory audit evaluates 113 certified rules using literature selected by an AI agent and criteria it developed to inform decide, defer, or deviate assessments.The audit prioritized creatinine-driven OOS recommendations for reconsideration related to acute kidney injury.Source: §2.5 · p. 6

3 Results

The results evaluate donor-only models and an uncertainty-aware Arbitrator for routing deceased-donor kidneys, then audit its learned rules against clinical literature. The Arbitrator resolves most decisions with interpretable rules, while deferring uncertain or conflicting cases.

3.1 Comparative Performance of ML Models

  • At β = 0.5, XGBoost identifies 66.7% of low-KDPI non-EM kidneys for potential OOS placement while unnecessarily routing 13.7% of low-KDPI EM kidneys OOS.Compared with XGBoost, calibrated KDPI identifies 485 fewer low-KDPI non-EM kidneys and unnecessarily routes 1,111 more EM kidneys OOS.Source: §3.1 · p. 7, Figure 2 · p. 7

3.2 Interpretable Resolution with the Arbitrator

  • At β = 0.5, the Arbitrator resolves 93.5% of organs with interpretable rules and defers the remaining 6.5% for expert review.CART and SparseRF jointly certify 79.4% of the test set, while single certification covers another 14.2%.Source: §3.2 · p. 8, Figure 3 · p. 8

3.3 Modest Cost of Interpretability

  • Across test kidneys, the Arbitrator raises average decision cost by 9.1% versus XGBoost; dual-certified cases cost only 4.8% more, while single-certified cases cost 28.0% or 20.0% more.The 28.0% and 20.0% gaps apply to CART-only and SparseRF-only certification, respectively.Source: §3.3 · pp. 9–10, Table 2 · p. 10
  • Single-certified kidneys are especially difficult: they comprise 14.2% of the test set but account for about two-thirds of the Arbitrator’s additional cost.On deferred kidneys, XGBoost has higher accuracy and lower average cost than on either single-certified group, suggesting these pathways identify different kinds of difficult cases.Source: §3.3 · pp. 10–11
  • Certification and deferral concentrate at intermediate KDPI: in the 60–70% range, deferral rises from 7.1% to 16.0% and single certification from 18.4% to 36.5%.Dual certification is 55.8% for KDPI 70–85%, but 93.7% for KDPI at most 20% and 94.9% at least 85%.Source: Figure 4 · p. 11, §3.3 · p. 11

3.4 Auditing the Arbitrator Pathways

  • The audit assessed 113 certified rules and recommended retaining 62 actions, referring 43 for physician interpretation, and reconsidering eight opposite actions.If its recommendations are well-founded, 81.0% of all test kidneys remain eligible for automation; the eight deviations propose OS instead of learned OOS placement.Source: §3.4 · p. 12, Table 3 · p. 12
  • For one low-KDPI rule involving high terminal creatinine and no recorded anoxia, the audit proposed OS placement, citing research suggesting AKI need not imply worse graft survival without other contributing factors.The authors say more extensive clinical review is needed to verify the audit’s utility and limitations.Source: §3.4 · p. 12

4 Discussion

The Arbitrator permits automation only when interpretable recommendations are stable under uncertainty and audited against clinical literature. In kidney allocation, it resolves many decisions and flags historically discarded low-KDPI kidneys, while routing remains a separate challenge.

4 Discussion

  • The Arbitrator automates a rule only when its recommendation is sufficiently stable under uncertainty and the rule has been audited against clinical literature and explicitly approved.The framework aims to automate easier cases so limited human experts can focus on challenging ones.Source: §4 · p. 13
  • 93.5% of organs were resolved with interpretable rules, and 70.1% of historically discarded low-KDPI test kidneys were flagged for expedited placement.If test-set performance generalizes, 43.9% of historically discarded kidneys could be flagged; applying that share to 2023 nonuse yields approximately 3,700 kidneys potentially targeted and saved.Source: §4 · p. 13
  • The current agentic auditor is a proof of concept, and more powerful literature-review pipelines remain future work.The authors also note that interpretability comes at some cost to predictive performance in their setup.Source: §4 · p. 13
  • The framework decides whether a kidney should leave the standard sequence, but not where it should go; routing must account for compatibility, ischemia time, and fairness.The authors identify construction of such a routing system as future work.Source: §4 · p. 13

Competing interests

S1.1 Data sources and the match run

The study links OPTN offer histories to donor and kidney registry data, defines early matchability from final acceptance within the first 100 sequence positions, and constructs a filtered cohort for donor-grouped model evaluation.

S1.1 Data sources and the match run

  • PTR records capture offer histories and responses, while STAR registry data supply donor and kidney characteristics; matching donor identifiers link these sources.The match run ranks biologically compatible candidates under OPTN policy, and PTR sequence numbers indicate rank rather than deliberate rejections because some offers are bypassed automatically.Source: S1.1 Data sources and the match run · p. 21

S1.2 Early-matchable outcome

  • Early-matchability labels a kidney positive only for eligible final acceptance by original PTR sequence 100; provisional acceptance does not qualify.For donors with two kidneys and exactly one early final acceptance, registry transplantation status determines kidney labels; this assignment convention applies to 22.79% of donors in the primary union.Source: S1.2 Early-matchable outcome · p. 22

S1.3 Cohort construction

  • The cohort excludes donors with OOS indicators and removes date-specific operational-refusal offers without renumbering the original sequence.The refusal-code vocabulary changed on December 2, 2021; exclusions also remove kidneys without eligible offer histories or retained disposition categories.Source: S1.3 Cohort construction · pp. 22–23

S1.4 Donor-grouped splits

  • Donors, rather than individual kidneys, were randomly assigned to approximately 70/15/15% training, validation, and test sets, stratified by KDPI decile and donor EM count.Keeping each donor’s kidneys together prevents shared characteristics and offer history from crossing splits; the eligible cohort contains 143,821 organs from 71,913 donors.Source: S1.4 Donor-grouped splits · p. 24

S1.5 Operational populations and the 100-offer cutoff

The study treats KDPI below 85% as a policy proxy for acceptable kidneys and uses acceptance by sequence 100 to identify early matchability. Historical acceptance and discard patterns support that cutoff as a marker of elevated discard risk, not certain discard.

S1.5 Operational populations and the 100-offer cutoff

  • KDPI < 85% is the operational low-KDPI proxy for whether an organ should be accepted on the match run absent further deterioration; higher-KDPI kidneys can still be transplanted.Programs must obtain written, informed consent before offering kidneys above 85% KDPI because of increased graft-failure risk.Source: S1.5 Operational populations and the… · p. 24
  • The intervention analyses examine low-KDPI non-EM organs and low-KDPI kidneys with explicit historical discard; the test set contains 4,944 and 2,738 such kidneys, respectively.Across the pooled cohort, the corresponding populations contain 33,068 and 18,302 kidneys.Source: S1.5 Operational populations and the… · pp. 24–25
  • At sequence 100, 91.76% of eventually accepted kidneys with KDPI < 85% had been accepted; among kidneys still unplaced, explicit discard probability was 58.46% versus 20.35% at match-run start.The latter is a 2.87-fold increase; the authors note that non-placement by sequence 100 does not mean a kidney is unusable or destined for discard.Source: S1.5 Operational populations and the… · p. 25, Figure S2 · p. 25

S1.6 Baseline outcomes and year effects

Acceptance and early-matchability rates declined across the recovery-year epochs in the pooled analysis, while the pandemic years were kept as a descriptive comparison outside the primary cohort. The paper also describes six prediction models with differing predictor sets.

S1.6 Baseline outcomes and year effects

  • Acceptance rates declined from 79.12% in 2016–2019 to 64.36% in 2023–2025, while EM prevalence fell from 71.82% to 57.77%.The 2020–2022 descriptive cohort had 73.21% acceptance and 65.78% EM prevalence; the authors do not quantify causes of the difference between the first and last epochs.Source: S1.6 Baseline outcomes and year effects · p. 26, Table S3 · p. 26

S2.1 Predictors and missingness

  • The study trains raw and calibrated KDPI cutoffs, logistic regression, CART, Random Forest, and XGBoost; logistic regression uses five predictors, while tree models use the full preprocessed donor feature set.The five logistic-regression predictors are creatinine, age, diabetes history, KDPI, and pediatric-candidate status; the full tree feature set was curated from 188 source inputs.Source: S2.1 Predictors and missingness · pp. 26–27

S2.2 Training and calibration

The section describes calibration and operating-point evaluation, then explains how validation-supported uncertainty is attached to CART and SparseRF rules. SparseRF makes a forest auditable through selected rules, while conservative intervals govern when its predictions can support action.

S2.2 Training and calibration

  • 0.005–0.011 was the kidney-weighted mean absolute gap between predicted and observed early-acceptance rates across models on the held-out test set.No bin with at least 200 kidneys differed by more than 0.06 for CART and XGBoost or 0.03 for the other listed models.Source: S2.2 Training and calibration · p. 27, Figure S3 · p. 27

S2.3 Operating points and performance

  • At β = 0.5, all calibrated models use τ0.5 = 2/3, and held-out performance ranks XGBoost, Random Forest, CART, logistic regression, then KDPI baselines.The fixed operating point is used for accuracy, F1, and average cost; log loss does not depend on the threshold.Source: S2.3 Operating points and performance · p. 28, Table S4 · p. 28

S3 Validation-supported uncertainty and SparseRF

  • CART and SparseRF attach validation-set 99% Wilson intervals to rules learned from training data, using the interval to represent uncertainty in each rule’s early-acceptance rate.Kidneys not covered by a rule with validation support receive the uninformative interval [0, 1].Source: S3 Validation-supported uncertainty… · p. 28

S3.3 SparseRF uncertainty: the consensus envelope

S3.1 CART Wilson intervals

  • CART uses disjoint terminal leaves, so each kidney matches exactly one rule; no CART leaf in this study lacked validation support.Wilson intervals quantify group-level event-rate uncertainty, not individual outcomes, and within-donor dependence may mildly violate independence.Source: S3.1 CART Wilson intervals · p. 29

S3.2 SparseRF: selecting rules with FIRE

  • FIRE selects a sparse subset of Random Forest leaves using training labels, trading fewer rules against broader kidney coverage; the 128-rule cap yielded 80 selected rules.Neither validation nor test outcomes select the rules; each selected rule’s uncertainty comes from its validation Wilson interval.Source: S3.2 SparseRF: selecting rules with FIRE · p. 29

S3.4 SparseRF rule-budget trade-offs

The section examines how SparseRF’s rule cap trades compactness for coverage, alongside the cost-based threshold used to choose between on-sequence and out-of-sequence placement.

S3.4 SparseRF rule-budget trade-offs

  • Raising the maximum rule budget increased Arbitrator resolution from 89.2% to 93.5% at β = 0.5 and substantially reduced deferred kidneys, while average cost and its gap to fixed XGBoost changed only slightly.The rule set saturates at 80 rules from nine trees; K is a choice between compactness and broader certification, not a single-best parameter.Source: S3.4 SparseRF rule-budget trade-offs · p. 30, Table S5 · p. 30

S4.1 Decision threshold

  • The cost-minimizing threshold is τβ = 1/(1 + β), reflecting the relative cost β of routing an early-accepted kidney out-of-sequence versus keeping a non-early-accepted kidney on-sequence.The evaluated preference grid includes β values from 0.1 to 9; the reported operating point uses β = 0.5 and τ0.5 = 2/3.Source: S4.1 Decision threshold · pp. 31–32

S4.2 Decision certification under uncertainty

The Arbitrator resolves model certifications when uncertainty supports an action and defers uncertain or conflicting cases. The section reports resolution and cost trade-offs, then describes an exploratory audit of resolved rules against clinical literature.

S4.2 Decision certification under uncertainty

  • 93.54% of test kidneys received rule-based resolution under single certification, leaving 6.46% for review, compared with 79.36% resolution under dual certification.Single certification resolved 3,060 additional kidneys; average cost was 0.1136 for dual-certified organs, while dual-or-single-certified organs cost 9.1% more than applying XGBoost to every kidney.Source: S4.2 Decision certification under… · pp. 32–33, Table S6 · p. 33
  • Example profiles illustrate the pathway distinctions: dual-certified kidneys have both intervals on the same side of τβ, while single-certified profiles have one model abstain.In the examples, SparseRF lacked supported rule coverage for profiles C and D, while CART intervals crossed the threshold for E and F.Source: S4.2 Decision certification under… · p. 33, Table S7 · p. 34

S4.3 Potentially saved organs

  • The authors use low-KDPI kidneys identified for OOS placement as a proxy for potentially saved transplantable organs, comparing the Arbitrator with XGBoost.The supplied table caption reports 4,944 low-KDPI non-EM kidneys and 2,738 low-KDPI historically discarded kidneys in the test data.Source: S4.3 Potentially saved organs · p. 33, Table S8 · p. 35

S5.1 Audit procedure

  • The audit covered 113 resolved contributing rules, using an exploratory agent-selected literature set without prespecified topics, search terms, eligibility criteria, or evidence-rating method.The authors identify prompt refinement and clinical validation as future work.Source: S5.1 Audit procedure · p. 35
  • The auditor generally retained on-sequence profiles and OOS profiles with KDPI lower bounds of at least 78% or sufficiently severe fibrosis or vascular changes, while generally deferring other OOS profiles.Source: S5.1 Audit procedure · p. 36

S5.2 Audit recommendations

  • The audit provisionally retained 62 rule actions, deferred 43 for physician review, and proposed deviation for eight learned OOS actions, all toward considering on-sequence placement.Six deviation candidates involved creatinine or possible AKI and two involved a sclerosis threshold.Source: S5.2 Audit recommendations · p. 36, Table S9 · p. 36

S5.3 Hypothetical audit policy

The hypothetical audit policy retains a resolved kidney for automation when at least one model rule is clean under the audit; otherwise, it defers the kidney. Under this policy, 86.6% of originally resolved kidneys remain eligible, while single-model resolutions are deferred more often than dual resolutions.

S5.3 Hypothetical audit policy

  • 86.6% (17,471 of 20,181) of originally resolved kidneys remain eligible for automation, including 13,508 on-sequence and 3,963 off-sequence actions.The remaining 2,710 defer: 13 have conflicting component actions and 2,697 have no usable action.Source: S5.3 Hypothetical audit policy · p. 37, Table S10 · p. 37
  • A resolved kidney remains eligible if at least one model rule is clean under the audit; otherwise, it is deferred.Decide retains a rule’s learned action, deviate supplies the opposite action, and defer supplies no action; only certified-model rules count toward automation.Source: S5.3 Hypothetical audit policy · pp. 36–37
  • 32.35% of single-model resolutions defer (990 of 3,060), versus 10.05% of dual resolutions (1,720 of 17,121).CART-only and SparseRF-only deferral rates are 27.18% and 38.73%, respectively.Source: S5.3 Hypothetical audit policy · p. 37, Table S10 · p. 37

S5.4 Supported deviation candidates and comparison profiles

The audit examines certified rules against clinical literature, identifying cases where learned actions warrant reconsideration and where rule specificity affects whether a decision should be made or deferred. The examples highlight creatinine, glomerulosclerosis, and KDPI as factors in these assessments.

S5.4 Supported deviation candidates and comparison profiles

  • F25_70 proposes on-sequence placement despite a validation early-acceptance interval of [0.071, 0.317] from 57 kidneys; reversible AKI could motivate early acceptance if research is validated and clinician behavior aligns.The other listed features are relatively low age, low KDPI, and moderately low sclerosis; the AKI rationale cites.Source: S5.4 Supported deviation candidates… · p. 37
  • C36 and C35 differ chiefly in creatinine, splitting at 1.675 mg/dL and receiving opposite certified actions; AKI evidence motivates reviewing whether an isolated increase warrants expedited placement.Their profiles otherwise share moderate KDPI, limited recorded fibrosis/sclerosis, no coronary angiogram, and recovery after 2019.Source: S5.4 Supported deviation candidates… · pp. 37–38
  • C20 and C37 propose on-sequence reconsideration for a learned GS > 10% branch, but prior studies restricted to KDPI above 85% cannot validate that threshold in lower-KDPI kidneys.The GS scale uses categories from 0–5% through > 20%, making the threshold appear too general for immediate OOS action.Source: S5.4 Supported deviation candidates… · p. 38
  • F25_109 is too general for immediate OOS despite an upper confidence bound of 0.298, while C133’s KDPI above 90.5% may warrant OOS despite an upper bound of 0.600.F25_109 lacks GS findings and an upper bound on KDPI; C133 has moderately low sclerosis.Source: S5.4 Supported deviation candidates… · pp. 38–39
Loading 2610.39045v1…