Understanding Algorithmic Fairness for Clinical Prediction in Terms of Subgroup Net Benefit and Health Equity

Area of Science:

• Health Informatics
• Biostatistics
• Medical Ethics

Background:

• Concerns exist regarding the fairness of clinical prediction models, particularly their performance across protected attributes like ethnicity and gender.
• Current algorithmic fairness approaches may inadvertently lower model performance in well-served subgroups rather than improving it in underserved ones.

Purpose of the Study:

• To propose and demonstrate an expanded net benefit framework for assessing clinical prediction model fairness.
• To evaluate how models distribute benefits, impact health inequalities, and contribute to health equity.

Main Methods:

• Expanded the concept of net benefit to quantify and compare the clinical impact of prediction models across different population subgroups.
• Applied the approach to two case studies: a type 2 diabetes prognostic model and a lung cancer screening algorithm.

Main Results:

• The proposed method allows for a nuanced understanding of how prediction models affect different subgroups.
• Demonstrated potential trade-offs between health equity and other healthcare system objectives due to resource constraints.

Conclusions:

• Assessing fairness through the lens of net benefit distribution provides crucial insights into a model's clinical and social context.
• This approach aids developers in creating models that better uphold health equity.