Persistent Proxy Discrimination in HIV Testing Prediction Models: A National Fairness Audit of 386,775 US Adults

Background: Human immunodeficiency virus (HIV) disproportionately affects marginalized communities in the United States, with Black Americans comprising 42% of new diagnoses despite representing 13% of the population. Machine learning (ML) prediction models are increasingly deployed for HIV-related decision-making, yet whether they perform equitably across demographic groups remains largely unexamined. Methods: We conducted a fairness audit using the Behavioral Risk Factor Surveillance System (BRFSS) 2024 dataset (N=386,775 adults). We trained four ML classifiers to predict HIV testing uptake and evaluated disparities using demographic parity difference (DPD), equalized odds difference (EOD), and calibration metrics across eight racial/ethnic groups. We tested whether excluding race from models ("race-blind" approaches) eliminates disparities, and applied fairness mitigation methods. External validation used Ryan White HIV/AIDS Program data (N=372,220). Results: All baseline models exhibited substantial disparities exceeding fairness thresholds by 5-6 fold. Selection rates ranged from 12.1% (Asian) to 66.0% (Black), with highly significant Black-White differences (z=81.5, p<0.001). Critically, race-blind models still exhibited DPD of 0.154-0.169, demonstrating that proxy discrimination through correlated variables (insurance status, depression, cost barriers, geography) accounts for approximately 70% of baseline disparity. Fairness mitigation reduced disparities by 84-97% (95% CI: 84.8%-97.1%) with 1-3% accuracy loss. Ryan White ground truth confirmed real-world disparities align with model-detected inequities. Conclusions: HIV testing prediction models exhibit persistent proxy discrimination even when race is excluded from feature sets. Bias propagates through socioeconomic variables that are themselves products of structural racism. Fairness-aware ML methods can substantially reduce disparities, but we highlight a critical tension: strict demographic parity may reduce screening recommendations for high-incidence populations such as Black MSM. The choice of fairness criterion requires stakeholder deliberation given unequal HIV burden across populations. We recommend routine fairness auditing prior to clinical deployment.