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

Area of Science:

• Machine Learning in Healthcare
• Algorithmic Fairness
• Diabetes Risk Prediction

Background:

• Machine learning models for diabetes risk prediction can exhibit age-related biases, impacting accuracy across demographic groups.
• Adversarial debiasing with a gradient reversal layer (GRL) is a method to create bias-invariant representations, but its real-world effectiveness with imbalanced healthcare data is unclear.

Purpose of the Study:

• To evaluate if adversarial debiasing using GRL enhances age-equitable diabetes prediction.
• To determine how fairness effects of adversarial debiasing vary across different data partitions.

Main Methods:

• Adversarial debiasing with GRL was applied to the Pima Indians Diabetes Database (n=768) for age-bias mitigation.
• An adversarial neural model was compared against a logistic regression baseline using all eight predictors and three age groups (<30, 30-50, >50 years).
• Performance metrics (accuracy, recall, ROC-AUC) and recall parity gap were computed, with robustness assessed across five random seeds.

Main Results:

• The adversarial model improved recall for the >50 years group (+22.22 pp) with comparable overall discrimination (ROC-AUC +0.45 pp) on the primary test partition.
• However, the recall parity gap increased (+11.57 pp) due to decreased recall in the <30 years group (-6.25 pp).
• Across five seeds, the mean recall parity gap reduction was modest (-2.49 pp) with high variability; fairness improvements were inconsistent, occurring in 3/5 seeds.

Conclusions:

• Adversarial debiasing can enhance predictive recall for underrepresented subgroups but doesn't guarantee consistent fairness across partitions, especially with small subgroup sizes.
• Reliable fairness assessment necessitates multi-seed evaluation, as single train-test splits are insufficient to capture performance variability.