Coverage-First Audits for Proxy Label Bias in Clinical Risk Scores

Many clinical risk models are trained on a stand-in for what they are meant to predict. Healthcare spending stands in for health need. Readmissions or emergency visits stand in for poor health. The stand-in is called a proxy label. When access to care depends on income, transport and neighborhood, the proxy drifts away from the real target, and it drifts most for the patients the record knows least about. Dissecting racial bias in an algorithm used to manage the health of populations showed the cost of this: at the same risk score, Black patients carried 26.3% more chronic illness than White patients. Finding such gaps requires a gold-standard measurement of need, for example a clinician's chart review, and that is expensive. A hospital can usually afford a few hundred reviews. The practical question is which few hundred patients to review so that a proxy gap gets caught wherever it sits in the social and clinical landscape, including in small groups that nobody thought to name ahead of time.

We propose to answer this with a tool from seed banks. Gene banks build "core collections," small subsets that keep a huge collection's diversity, by clustering accessions and drawing a guaranteed share from every cluster, sparse ones included. A sharper version from phylogenetics picks representatives that minimize the average distance from every item to its nearest representative. That objective covers the whole population while avoiding noisy outliers. Our claim is that an audit design built this way, then randomized inside each cell so standard debiasing inference stays valid, finds proxy gaps in rare social-clinical subgroups with fewer chart reviews than uniform sampling or variance-targeted adaptive sampling. It does this without giving up valid confidence intervals for the population-level gap.

Core Claim

Write $Y$ for the gold-standard need (chart-reviewed), $P$ for the cheap proxy-trained score, and $m(P)=E[Y\mid P]$ for the population calibration curve of need against score. The audit target is the residual gap for a subgroup $G$:

$$\Delta_G = E[\,Y - m(P)\mid G\,],$$

where a positive $\Delta_G$ means members of $G$ are sicker than their score implies. This is the Obermeyer finding written as a per-group quantity.

Existing label-efficient methods are built to estimate one population parameter precisely. M-estimation under Two-Phase Multiwave Sampling with Applications to Prediction-Powered Inference sends labels to regions where the proxy model is uncertain or the debiasing correction is large. It reports up to 1.8x effective sample size over uniform sampling at nominal 90% coverage. It was tested on satellite, protein and forestry data, not on health-disparity audits. Towards the Efficient Inference by Incorporating Automated Computational Phenotypes under Covariate Shift treats the choice of who gets reviewed as given and corrects for it afterward. It reports up to 60% shorter confidence intervals than prediction-powered inference on a 297-label pediatric diabetes cohort.

Neither method asks whether every region of social-clinical space has enough audited patients nearby for $\Delta_G$ to be estimable for groups defined after the fact. Our design does ask this. It makes three changes:

  1. Allocation is driven by geometric coverage of a joint space of community SDoH (social determinants of health) and clinical features, not by proxy uncertainty.
  2. Every cell receives a minimum number of reviews, the analogue of a core collection preserving rare alleles.
  3. Reviews are drawn at random inside each cell with known inclusion probabilities, so inverse-probability-weighted and prediction-powered estimators remain valid.

Because the auditor controls selection and it depends only on observed covariates, the covariate-shift assumption that the automated-phenotype estimator needs holds by construction. This removes one of the tensions in the literature for audit settings. It does not remove the case where the gold standard itself is distorted.

Research Program

The build has three parts.

A joint patient space. We combine clinical features with county-level indices from County Health Rankings, the Social Vulnerability Index and the AHRQ SDOH Database, as linked in MIMIC-IV-SDOH. We will compare two representations: a learned embedding, and a simple standardized-feature distance.

Partitioning by the closest-leaf objective. Following Minimizing the average distance to a closest leaf in a phylogenetic tree, we build a hierarchical clustering tree (UPGMA, as in The Genetic Diversity Assessment of Broomcorn Millet). We then choose $k$ representatives that minimize the average distance from each patient to the nearest one, and assign each patient to the cell of its nearest representative. That paper's exact dynamic program scales only to a few hundred leaves. We therefore collapse patients into a few hundred micro-clusters first and run the exact program on that tree. This keeps the optimization exact where it matters and avoids the local minima of medoid swapping.

The choice of this objective over diversity maximization is deliberate. The same paper found that maximizing phylogenetic diversity over-selects chimeric sequences on long isolated branches, while the closest-leaf objective selects them less often than random sampling. In an electronic health record, the analogue of a chimera is a miscoded or linkage-error patient. An audit that spends its budget on those patients learns nothing about bias.

Allocation and inference. Each cell receives a floor of reviews. The remaining budget goes out in waves using the multiwave predict-then-debias rule restricted within cells. The gap is estimated with wave-specific inverse probability weighting, plus the doubly robust cross-fitted estimator that uses the cheap phenotype on unreviewed patients.

The analysis question is the trade-off between the floor size and the efficiency of the overall population estimate. Floors cost precision on the population average. We will measure how much, and characterize when the trade is worth it.

Baselines

Uniform random sampling. This is the naive reference: chart review drawn uniformly from the cohort, analyzed with the same prediction-powered estimator. It is unbiased, and it is the default in fairness audits such as Evaluating the Impact of Social Determinants on Health Prediction in the Intensive Care Unit, which used random splits. It shares our label budget, phenotype model and estimator. Beating it means a lower worst-subgroup error in $\Delta_G$ at equal budget.

Multiwave predict-then-debias, unmodified. This is the strongest existing method. Its paper reports up to 1.8x effective sample size over non-adaptive sampling with 90% coverage, on housing, AlphaFold and forest-cover tasks. No reported result was found on Lacuna for it on clinical or disparity-audit data. We will run it with the authors' greedy rule, the same budget, the same number of waves (about six, where its reported gains plateau), and the same proxy model. Beating it means matching its population-level confidence interval width within 15% while cutting worst-subgroup gap error or the labels needed to detect a planted gap.

Stratified sampling on prespecified demographic strata. This uses race by income tertile, with Neyman allocation (labels split in proportion to stratum size times outcome spread). It is the textbook design and a serious competitor whenever the at-risk groups are known in advance. We tune it with the same stratum-floor search we give our method.

Ablations isolating the key ingredient. We keep the floors and randomization but replace the closest-leaf cells with k-means cells. We drop SDoH features from the space. We drop the floors. Finally, we use a diversity-maximizing selector to test whether the artifact trap transfers to health records.

Why not just…?

Why not just sample uniformly? Uniform sampling is valid, but for a rare cell it delivers only about budget times prevalence reviews. Prior work shows rare classes are badly served by it: SHAC active learning retrieved 151% more drug-use and 109% more homelessness mentions than random sampling, and core-collection work (Prunus sibirica) needed allele-preferred sampling to retain rare variants. Our experiments quantify the same effect for gap estimation, as the detection rate of planted gaps in cells holding under 2% of patients.

Why not just stratify on race and income? This is the real threat. If the groups to be audited are fixed and few, prespecified stratification may match us. We settle it with a held-out-subgroup protocol. Each method is designed without seeing a set of evaluation subgroups, defined for example by transit reliance crossed with diabetes, or low education crossed with age band. All methods are then scored on those subgroups. If prespecified stratification wins there too, we will report that the geometric layer adds nothing. That would itself be a useful negative result.

Why not just use multiwave predict-then-debias? Its rule chases proxy uncertainty. A systematically biased proxy can be confidently wrong in exactly the underserved regions. In Obermeyer et al., the cost model was well calibrated for cost in both groups. We test this directly by planting gaps in low-uncertainty regions.

Why not just retrain on a better label, or add fairness constraints? Retraining on chronic-condition counts cut Obermeyer's disparity by 84%, but someone first has to discover which label is wrong and for whom. That discovery is the audit. Fairness constraints evaluated against the proxy only reproduce it, as Fogliato et al. and Li argue. Adding SDoH features does not fix the label either: Yang et al. found such features accuracy-neutral for discrimination, even though they lowered calibration error from 0.11 to 0.03.

Proposed Experiments

Semi-synthetic audits on MIMIC-IV-SDOH. The proxy is a model trained on utilization (30-day readmission). The gold standard is a need score built from chronic-condition burden and biomarkers, mirroring Obermeyer's need label. It is computed for everyone but revealed only for audited patients, which simulates chart review. We plant known gaps of varying size and prevalence in social-clinical regions, both inside and outside the prespecified strata. Budgets are 100, 300 and 1,000 reviews. The 300 level matches the size of the pediatric diabetes study.

Metrics. We report:

  • error in the population gap and its confidence interval coverage;
  • mean and worst error of $\Delta_G$ over held-out subgroups;
  • planted-gap detection power at a 5% false discovery rate;
  • the fraction of the budget spent on injected corrupt records.

Replication. We repeat the design on the MC-BEC emergency cohort, with ED revisit as the proxy. Its baseline AUPRC of 0.11 to 0.24 marks revisit as a weak stand-in for need. We also run All of Us, where survey-based individual SDoH can serve as a gold-standard check on whether county-level cells capture person-level gaps, the ecological-fallacy question.

Decisive checks. The proposal succeeds if, on held-out subgroups, the coverage design beats both multiwave predict-then-debias and prespecified stratification on worst-subgroup error at equal budget, while keeping population coverage near nominal. It also requires the k-means ablation to lose, which would show the closest-leaf geometry, not the floors alone, does the work. It fails if prespecified stratification ties it on held-out subgroups, or if the floors widen population-level intervals by more than they gain at the subgroup level.