Measuring and Reducing Racial Bias in a Pediatric Urinary Tract Infection Model

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Health Equity

Background:

• Clinical predictive models incorporating race can worsen healthcare disparities.
• Respecification or optimization can mitigate racial bias in these models.
• The UTICalc model aims to reduce urinary tract catheterizations in children.

Purpose of the Study:

• To investigate the impact of respecifying the UTICalc model to reduce racial bias.
• To compare fairness and predictive performance between original and respecified UTICalc versions.
• To develop new models for enhanced racial fairness.

Main Methods:

• Removed race from the UTICalc logistic regression model and introduced two new features.
• Compared original and respecified UTICalc models using fairness and predictive performance metrics.
• Derived three new models specifically to improve racial fairness.

Main Results:

• Model respecification showed potential for improving racial fairness.
• Simultaneous improvement across all fairness metrics was not achieved, aligning with impossibility results.
• Enhancing racial fairness sometimes led to a decrease in overall predictive performance.

Conclusions:

• Respecifying clinical predictive models can impact both fairness and performance.
• Achieving perfect fairness across all metrics is challenging.
• Careful consideration of trade-offs is necessary when modifying models for health equity.