An intersectional framework for counterfactual fairness in risk prediction

Area of Science:

• Health Informatics
• Biostatistics
• Health Equity Research

Background:

• Data-driven models are increasingly used in health policy and decision-making.
• Algorithmic fairness methods often fail to address intersecting demographic groups and clinical risk prediction complexities.
• Existing bias correction techniques can exacerbate health inequities.

Purpose of the Study:

• To develop novel unfairness metrics addressing intersecting groups and clinical risk prediction challenges.
• To create a framework for estimating and inferring these new unfairness metrics.
• To evaluate the framework's application in a real-world COVID-19 risk prediction model.

Main Methods:

• Development of new unfairness metrics accounting for intersectionality.
• Introduction of the unfairness value (u-value) for quantifying bias extremity.
• Utilizing an alternative to standard bootstrap for confidence intervals and standard errors.

Main Results:

• The proposed framework successfully addresses limitations of existing algorithmic fairness methods.
• Novel metrics and estimation tools provide a robust approach to measuring and correcting bias.
• Application to a COVID-19 risk model demonstrated practical utility in a health system.

Conclusions:

• The developed framework offers a significant advancement in assessing and mitigating algorithmic unfairness in healthcare.
• Addressing intersectionality and clinical context is crucial for equitable health AI.
• This work provides essential tools for fairer health policy and patient care.