Adversarial debiasing for age-equitable diabetes prediction: performance-fairness trade-offs and partition dependency in machine learning.

Background: Machine learning models used for diabetes risk prediction may encode age-related biases that reduce diagnostic accuracy for specific demographic groups. Adversarial debiasing with a gradient reversal layer (GRL) offers a theoretically principled approach to learning representations that are invariant to a protected attribute; however, its practical effectiveness under realistic conditions of subgroup imbalance in healthcare datasets has not been fully characterised. Research question: Does adversarial debiasing with a GRL improve age-equitable diabetes prediction, and how do its fairness effects vary across different data partitions? Methods: Adversarial debiasing was evaluated for age-bias mitigation in diabetes prediction using the publicly available Pima Indians Diabetes Database ( Results: On the primary test partition, the adversarial model improved recall for the smallest age group [>50 years: 0.5556 → 0.7778, +22.22 percentage points (pp)] while maintaining comparable overall discrimination (ROC-AUC: 0.7852 → 0.7896, +0.45 pp). However, the recall parity gap increased from 0.0996 to 0.2153 (+11.57 pp), reflecting a concurrent decline in recall for the <30-year group (-6.25 pp). Across five random seeds, the mean recall parity gap showed a modest mean reduction (0.3282 → 0.3033, -2.49 pp), but with high variability (SD > 0.27) exceeding the mean difference. The adversarial model reduced the fairness gap in three of five seeds, increased it in one, and produced no change in one. Conclusion: Adversarial debiasing can improve predictive recall for underrepresented demographic subgroups but does not guarantee consistent fairness improvements across data partitions, particularly when subgroup sample sizes are small. Multi-seed evaluation is essential for reliable fairness assessment; single train-test splits are insufficient.