The Illusion of Fairness: When Bias Mitigation Harms Clinical Performance in Imbalanced Healthcare Prediction

Algorithmic fairness in healthcare remains a pressing challenge, particularly when deploying machine learning models on real-world clinical datasets characterized by severe class imbalance and demographic heterogeneity. This study introduces and empirically characterizes the 'illusion of fairness'—a phenomenon wherein standard bias mitigation techniques appear to optimize statistical fairness metrics while simultaneously degrading overall predictive performance, destabilizing decision thresholds, and redistributing errors in clinically harmful ways across subgroups.Using the NHANES dataset (N=5,812), we developed a diabetes prediction task with an optimized XGBoost baseline (Accuracy = 0.819 [95% CI: 0.810–0.828], ROC-AUC = 0.742 [95% CI: 0.731–0.753]). We evaluated two widely used interventions from the Fairlearn library under Equal Opportunity constraints across race/ethnicity groups: post-processing via Threshold Optimizer and in-processing via Exponentiated Gradient Reduction.Results revealed highly counterintuitive outcomes. The Threshold Optimizer worsened the Equal Opportunity Difference (0.312 → 0.478) and reduced global AUC to 0.701. The Exponentiated Gradient method achieved a slightly lower EOD (0.293) but at a substantial clinical cost, collapsing AUC to 0.651 and markedly increasing false negatives in high-prevalence minority cohorts. These findings demonstrate that mathematical parity objectives can act as a proxy for harmful error redistribution rather than meaningful clinical equity.We challenge the plug-and-play application of generic fairness toolkits in medicine and propose a clinician-informed Domain-Aware Fairness Framework that integrates epidemiological priors and cost-sensitive clinical utilities. This work underscores the critical need for closer collaboration between clinicians and AI developers to ensure AI systems deliver genuine patient benefit in real-world healthcare settings.