Clinical Report: Mitigating Age Bias in Diabetes Risk Prediction through Adversarial Debiasing
Overview
This study evaluates the effectiveness of adversarial debiasing with a gradient reversal layer (GRL) in improving diabetes risk prediction while addressing age-related biases. The findings indicate that the adversarial model enhances recall for older age groups, but its impact on fairness across different demographic partitions is variable.
Background
Diabetes mellitus is a significant global health issue, affecting over 500 million adults and leading to serious complications. Machine learning models have shown promise in diabetes risk prediction but may harbor biases that affect diagnostic accuracy across demographic groups, particularly age. Addressing these biases is crucial to ensure equitable healthcare delivery.
Data Highlights
Model
Recall (>50 years)
ROC-AUC
Recall Parity Gap
Adversarial Model
0.7778
0.7896
0.2153
Logistic Regression
0.5556
0.7852
0.0996
Key Findings
The adversarial model improved recall for the >50 years age group by 22.22 percentage points.
Overall discrimination (ROC-AUC) was slightly improved from 0.7852 to 0.7896.
The recall parity gap increased from 0.0996 to 0.2153 for the adversarial model.
Across five random seeds, the mean recall parity gap showed variability, with three out of five seeds showing a reduction in the fairness gap.
Clinical Implications
Clinicians should be aware of potential biases in machine learning models.
Conclusion
Adversarial debiasing can improve diabetes risk prediction for underrepresented age groups, but its effectiveness in ensuring fairness is variable.
