Single-Axis Fairness Interventions Produce Asymmetric Cross-Axis Effects in Clinical Prediction

Objective: To systematically evaluate whether single-axis demographic balancing introduces cross-axis fairness trade-offs in clinical prediction, and to characterize the directional asymmetry of these effects across model architectures and balancing strategies. Materials and Methods: We evaluated cross-axis fairness effects of demographic balancing in one-year all-cause mortality prediction using the MIMIC-IV database (N=64,427). Seven machine learning architectures were trained under six balancing strategies targeting either gender or race, with performance assessed via outcome-stratified 80-20 train-test splits repeated 30 times across both targeted and non-targeted axes using AUC, TPR, and Brier score. Results: Gender-targeting interventions largely preserved race fairness, while race-targeting consistently disrupted gender fairness across all methods and the majority of architectures. This asymmetry was invisible to same-axis evaluation alone. Race-targeting also incurred greater performance costs and calibration loss, with observed fairness gains potentially reflecting leveling down rather than genuine improvement. The same intervention could appear successful under TPR but fail under AUC evaluation. Discussion: The asymmetry likely reflects differential category complexity: binary gender balancing requires modest distributional shifts, whereas multi-category race balancing necessitates aggressive reweighting that propagates to correlated axes. Cross-axis fairness effects are directionally dependent and metric-sensitive, indicating that single-metric, single-axis evaluation is insufficient. Conclusion: Single-axis fairness optimization cannot guarantee cross-dimensional equity. Cross-axis, multi-metric fairness evaluation should be integrated into pre-deployment auditing of healthcare artificial intelligence (AI) models.