Adapting a Kidney Exchange Algorithm to Align with Human Values

1 Introduction

• Automated moral decisions may be needed when decisions are too fast for humans, individually too minor to warrant human attention, or inseparable from computational problems beyond human capabilities.Source: §1 · p. 1
• Developing patient priorities is a moral choice, so AI researchers should help determine the weights and understand their consequences, not merely alter an objective function.The paper frames this as a responsibility alongside technical input into the weighting process.Source: §1 · p. 2

1.1 Our Contributions

• The study elicits acceptable prioritization attributes, compares patient profiles that differ on those attributes, and estimates weights for tiebreaking among maximum-sized solutions.The authors describe the work as a proof-of-concept rather than a basis for direct real-world adoption.Source: §1.1 · p. 2
• In simulations, the precise weight values mattered little beyond the profile ordering they induced, while underdemanded blood-type pairs saw chances rise or fall significantly relative to no prioritization.The effects depended on the tiebreaking weights assigned to those pairs.Source: §1.1 · p. 2

2.1 Graph Formulation

• Each patient-donor pair is a graph vertex, compatible donations form directed edges, and cycles represent simultaneous transplants; edge weights can also prioritize patient or donor classes.Figure 1 depicts three pairs and two possible 2-cycles.Source: §2.1 · pp. 2–3, Figure 1 · p. 3
• Fielded exchanges often set edge weights through opaque committee policies, although those weights roughly represent the recipient's utility from a donor kidney.The cited example policy includes calculated reactive panel antibody, failed-exchange history, hospital, and prior donation.Source: §2.1 · p. 2

2.2 Clearing Kidney Exchanges

• With unit edge weights, the baseline seeks a maximum-cardinality matching, then uses the paper's priority weights to break ties among solutions of that size.The matching consists of disjoint cycles and chains, and the IP selects which legal structures are included.Source: §2.2 · p. 3, §2.1 · p. 3

2.3 Tiebreaking and Prioritization: Examples

• In Figure 2's example, the algorithm chooses the 3-cycle over the 2-cycle because it matches more vertices, regardless of additional patient attributes.The example distinguishes maximal solutions from the larger-cardinality solution the algorithm selects.Source: §2.3 · pp. 3–4, Figure 2 · p. 4

3.1 Selecting Attributes

• In an open-ended survey, 100 participants proposed attributes they thought kidney-allocation policy ought or ought not to consider; categories favored by more participants were retained.Attributes already considered by the UNOS algorithm, including medical compatibility, were discarded.Source: §3.1 · pp. 4–5
• The three most-supported attribute categories were Age, Health -– Behavioral, and Health -– General; only these advanced to the next study stage.Health -– General was reported 44 times, versus 18 for the next category, Dependents.Source: §3.1 · p. 5

3.1.1 Participant Demographics

• The attribute survey’s participants ranged from 22 to 64 years old, with an average age of 40, and varied in reported political views and education.Completion averaged 9 minutes, with a range of 2 to 26 minutes.Source: §3.1.1 · p. 5

3.2 Evaluating Pairwise Comparisons

• The pairwise survey presented participants with all 28 possible pairs of eight profiles, asking which fictional patient should receive one kidney.The profiles combined binary alternatives for age, alcohol consumption before diagnosis, and general health; 289 MTurk participants completed the survey.Source: §3.2 · pp. 5–6

3.2.1 Summary of Responses

• Participants preferred profile 3 over 2 and profile 7 over 6, suggesting greater weight for health-behavioral than health-general attributes.The authors note this pattern may not generalize to other attributes in those categories.Source: §3.2.1 · pp. 6–7

3.2.2 Participant Demographics

• The pairwise-survey participants ranged from 19 to 70 years old, averaged 37, and varied in reported political views and education.Completion averaged 7 minutes, with a range of 1.5 to 38.5 minutes.Source: §3.2.2 · p. 7

3.3 Estimating Profile Scores

• Bradley-Terry modeling converts pairwise choices into profile scores, estimated either separately for each profile or as a linear function of attribute values.The scores represent relative preference: a higher score means a randomly selected participant is more likely to choose that profile over another. The authors used direct profile scores in simulations.Source: §3.3 · pp. 7–8

3.4 Adapting the Algorithm

• The adapted algorithm first finds the maximum-cardinality matching, then uses profile weights to choose among solutions matching the same number of patients.The authors restrict weights to tie-breaking because their preference data do not compare allocations involving different quantities of patients.Source: §3.4 · p. 8

• The experiments use simulations to test the effects of the prioritizing algorithm after describing how profile weights were obtained and integrated into the IP-based algorithm.Source: §4 · p. 9

• Each day, incompatible patient-donor pairs enter or depart the pool, an algorithm matches compatible pairs, and remaining pairs stay for later consideration.Matches formed the previous day are executed with probability 0.5 to model medical or logistical failures.Source: §4.1 · p. 9
• Simulated pool demographics were designed to reflect the UNOS kidney-exchange pool where possible and otherwise the general US population.Source: §4.1 · p. 9

4.2.1 Experiment

• The experiment compared STANDARD, which breaks ties arbitrarily, with PRIORITIZED, which breaks ties toward higher patient-profile scores, across 20 simulations of five simulated years.Source: §4.2.1 · p. 9

4.2.2 Hypothesis

• The hypothesis was that STANDARD would match profiles in similar proportions, while PRIORITIZED would favor higher-scoring profiles and disfavor lower-scoring profiles.Source: §4.2.2 · p. 9

4.2.3 Results

• PRIORITIZED matched profiles 1, 3, and 2 substantially more often than STANDARD, and profiles 7, 6, and 8 substantially less often.These differences show that profile scores affected which profiles were matched.Source: §4.2.3 · p. 9
• 61.7% of pairs were matched overall by both algorithms, while PRIORITIZED matched profile 1 nearly twice as often as profile 8.Figure 3 plots proportions matched by profile and algorithm type.Source: §4.2.3 · p. 9, Figure 3 · p. 10

4.3.1 Experiment

• The experiment classified pairs as underdemanded, overdemanded, self-demanded, or reciprocally demanded according to patient and donor blood types.Underdemanded pairs contain a patient with blood type O, a donor with blood type AB, or both; the other three classes were described as substantially easier to match.Source: §4.3.1 · p. 10

4.3.2 Hypothesis

• The hypothesis was that prioritization would affect underdemanded pairs most, favoring higher-scoring profiles at the expense of lower-scoring ones, while other classes would retain similarly high match rates.The stated rationale was that matching opportunities are generally scarce for underdemanded pairs but not for the other types.Source: §4.3.2 · p. 10

4.3.3 Results

• PRIORITIZED matched underdemanded pairs with high profile scores more often and low scores less often than STANDARD, while other blood-type classes had roughly equal match rates.The results suggest the primary difference between the algorithms lies in their treatment of underdemanded pairs; Figure 4 plots underdemanded-pair match proportions.Source: §4.3.3 · pp. 10–11, Figure 4 · p. 11

4.4.1 Experiment

• Because Bradley-Terry score gaps may not reflect proportional importance, the authors tested rank-linear weights with the same profile ordering and equal successive gaps.The alternative vector is linear in profile rank, unlike the original Bradley-Terry scores.Source: §4.4.1 · p. 11, Table 6 · p. 11
• Figure 5 tracks proportions of overdemanded, self-demanded, and reciprocally demanded pairs matched over the simulation, grouped by profile and algorithm.Each box represents 60 simulation runs.Source: Figure 5 · p. 12
• The comparison used 20 simulations of five years of daily matching, contrasting the original Bradley-Terry weights with the alternative weights.Source: §4.4.1 · p. 12

4.4.2 Hypothesis

• The authors hypothesized that profile ranking drives matching differences, so vectors with the same ordering should match profiles in similar proportions.Source: §4.4.2 · p. 12

4.4.3 Results

• Preserving profile order while changing weight gaps produced very little difference in matching; the authors attribute remaining differences to slightly different pairs entering the pool.Other tested weight vectors with the same profile ordering also produced similarly little difference.Source: §4.4.3 · p. 12

5.1 Application in Real Kidney Exchanges

• Direct real-world use is not advocated: the study used limited attribute values and MTurk judgments, while deployment would need expert advice and stakeholder input.The authors also identify structuring that process as future research.Source: §5.1 · p. 13
• Simulations suggest underdemanded pairs would be significantly affected while little would change for others; the authors make no judgment on whether this supports prioritization.They expect this qualitative conclusion to hold with more thorough prioritization and real rather than simulated data, while noting that real data distributions differ.Source: §5.1 · p. 13

5.2 Artificial Morality

• The authors frame their approach as a hybrid of top-down principles and bottom-up learning, aiming to combine comprehensible principles with attention to complex case facts.Source: §5.2 · pp. 13–14
• Crowdsourcing morally relevant features constrains the data and gives predictions an interpretable basis, while attempting to reduce top-down arbitrariness.Source: §5.2 · p. 14
• The hybrid may suit constrained domains with clear morally relevant features; the discussion points to autonomous-vehicle decisions as a related application.Source: §5.2 · p. 14
• Inconsistent judgments within and across people limit prediction to a subset, perhaps a majority, and leave open how social decisions should handle disagreement.Source: §5.2 · p. 14
• Omitting race can avoid explicit racial weighting, but the authors note that subtler hidden bias remains possible.Source: §5.2 · p. 14

5.3 Future Research

• The authors call for further research to generalize from human judgments toward best practices for automated moral decision-making across domains.Source: §5.3 · p. 15