AI Insight
This study examined how algorithmic design choices affect equity in healthcare resource allocation by analyzing data from 201,216 adults in the NIH All of Us Research Program. Researchers found that algorithms trained to predict healthcare utilization systematically underidentify patients from disadvantaged groups who face structural barriers to care, with 33% of Black participants reporting poor health but only 22% being classified as high utilizers. Among three modifiable design choices tested, changing the training objective from predicting utilization to predicting self-reported health status had the greatest impact on reducing disparities, achieving near-zero inequality between Black and White patients.
Why it matters
Healthcare systems increasingly use prediction algorithms to allocate limited care management resources, but these tools may inadvertently perpetuate health inequities if designed without equity considerations. This study demonstrates that algorithmic fairness is largely determined by modifiable design choices made during development, suggesting that equity can be engineered into healthcare algorithms rather than addressed only after deployment.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Background. Care management programs increasingly rely on prediction algorithms to decide which patients receive limited healthcare resources. Because these algorithms are typically trained to predict future healthcare utilization or cost, they may systematically underidentify patients whose care-seeking behavior is suppressed by structural barriers. Objective. To evaluate how three modifiable design choices (selection threshold, feature set, and training objective) affect allocation equity in care management targeting. Methods. We constructed an analytical cohort of 201,216 adults from linked electronic health record and participant survey data in the NIH All of Us Research Program (2017-2019). Three prediction models (logistic regression, random forest, and XGBoost) were trained and evaluated under simulation conditions that varied one design choice at a time while holding the others fixed. Equity was assessed using false negative rate (FNR), equal opportunity difference (EOD), and allocation rate conditional on need (an access-independent benchmark based on self-reported general health status). We also applied Oaxaca-Blinder decomposition to identify which predictors contributed most to the Black-White allocation gap. Results. Prior utilization understated self-reported health burden among disadvantaged groups: 33% of Black and 32% of Hispanic participants reported fair or poor health, yet only 22% and 19% were classified as high utilizers. White participants showed the opposite pattern: 25% were classified as high utilizers, while only 19% reported fair or poor general health status. Under baseline design choices, all three models produced similar disparities, with uninsured participants showing the largest gap (EOD: -0.148). Oaxaca-Blinder decomposition identified prior outpatient visits as the dominant contributor to the Black-White gap in predicted allocation, with higher average emergency department utilization among Black participants partially offsetting the effect. Varying the selection threshold had little effect; disparities persisted across all thresholds. Feature set variation produced mixed and group-specific effects; removal of prior outpatient visits achieved near-zero Black-White EOD (0.001, AUC: 0.77), consistent with the Oaxaca-Blinder decomposition finding. Training objective variation produced the most substantial and consistent equity effects, with the health status objective achieving near-zero Black-White EOD (0.008, AUC: 0.69) and the survey-enriched health status condition reversing the direction of disparity for most groups. Conclusions. Algorithmic equity in care management targeting is largely a consequence of modifiable design choices, with training objective emerging as the most impactful of the three design choices evaluated. The causal simulation framework and access-independent equity benchmark developed here provide a replicable approach for evaluating equity as a design consideration during algorithm development rather than a post-deployment correction.