The Algorithmic Waiting List
A public hospital uses a software triage system to rank patients for elective surgery. The vendor's model was trained on five years of historical data in which older patients and patients from a rural region received surgery later, partly because of travel distances and partly because of clinician bias in referral letters. The hospital board knows this. Removing age and postcode from the model reduces its predictive accuracy by 4 percentage points but does not eliminate the disparity, because referral letters contain proxy cues. A 71-year-old rural patient, Mr Chen, is ranked 340th; a clinician estimates his untreated condition gives him a 25% chance of serious complication within a year. The board must decide whether to keep the model, patch it, or replace it with a slower manual panel.
1. Analyse the board's options using a Rawlsian framework. What would a party behind a veil of ignorance choose, and why?
Read the worked answer
Behind Rawls's veil of ignorance, parties do not know their age, location, health status or position in the queue. They know the general facts of society - that some will be old, some rural, some sick.
The relevant primary good here is not income but access to healthcare, and Rawls's second principle requires that social and economic inequalities be attached to positions open to all under fair equality of opportunity, and to the greatest benefit of the least advantaged. The difference principle, applied to health institutions, directs attention to the worst-off group.
A veil-of-ignorance chooser would reason:
- If I turn out to be Mr Chen, the current model gives me a 25% annual risk of serious complication while others with equal need are treated sooner. That is a serious loss.
- If I turn out to be a patient who benefits from the model's accuracy, I lose some speed but not the same magnitude of harm.
- Because I might be the worst-off, I will choose the arrangement that maximises the position of the worst-off - which points toward the manual panel or a heavily patched model with an appeals mechanism, even at the cost of overall efficiency.
A strong answer notes the limits: Rawls's framework was designed for the basic structure, not individual clinical decisions, and the veil does not tell us how to weigh a 4-point accuracy loss against a diffuse fairness gain. It also notes that the rural disparity is partly caused by geography, not injustice - distance is a real burden, not a bias, and a just system might address it with transport funding rather than queue-jumping.
2. A utilitarian board member argues that keeping the model maximises total health outcomes. Construct the strongest reply, and state what evidence would settle the dispute.
Read the worked answer
The utilitarian claims the model saves more quality-adjusted life-years overall; the 4-point accuracy loss from removing proxies means some patients are treated later, and aggregate welfare falls.
The strongest reply has three parts.
- The baseline is not neutral. The historical data encodes past injustice. Aggregate welfare calculated from a biased ranking may be high precisely because it concentrates harm on a group with less political voice. A utilitarian should count Mr Chen's 25% risk at full weight, not as a rounding error - and should ask whether the model's gains come partly from systematically under-serving the rural region.
- Rule-level versus act-level. Even if this model maximises this year's QALYs, adopting a rule that permits opaque, historically biased tools erodes trust in the health system. Patients who believe the queue is rigged may delay presentation, which costs lives. Rule-utilitarianism may therefore favour transparency and appeal rights.
- The comparison is unfair. The manual panel is described as 'slower', but slowness is not the same as lower welfare. A panel that is 4 points less accurate but corrects the disparity may produce better outcomes for the worst-off at modest aggregate cost.
Evidence that would settle it: a prospective audit comparing complication rates and time-to-surgery by age and region under each option; a cost-effectiveness estimate of the manual panel; and data on whether the rural disparity is driven by bias or by distance. Without that, both sides are trading intuitions.