Leveraging clinical epidemiology concepts to strengthen machine learning fairness evaluations

Area of Science:

• Clinical Epidemiology
• Machine Learning Fairness
• Health Informatics

Background:

• Machine learning (ML) is increasingly used in clinical care, raising critical fairness concerns.
• Fairness, defined as the absence of disparities, shares conceptual links with clinical epidemiology.
• Existing ML fairness evaluations can be strengthened by established epidemiological principles.

Purpose of the Study:

• To apply clinical epidemiology frameworks to evaluate ML fairness.
• To enhance understanding of ML fairness concepts and assessment methodologies.
• To strengthen the rigor of fairness assessments in clinical ML applications.

Main Methods:

• Examined connections between clinical epidemiology bias terms and ML fairness concepts.
• Investigated the relationship between diagnostic testing metrics and ML fairness criteria.
• Explored strategies for fairness considerations by leveraging clinical epidemiology principles.

Main Results:

• Unfairness in ML can stem from selection, measurement, misspecification, cognitive, and implementation biases.
• Fairness criteria (independence, separation, sufficiency, predictive parity) can be assessed using diagnostic testing metrics.
• Fairness evaluations are susceptible to multiple testing issues, particularly in subgroup analyses.

Conclusions:

• Significant parallels exist between ML fairness and clinical epidemiology in conceptualizing unfairness and defining criteria.
• Well-established principles from clinical epidemiology can be leveraged for methodologically sound ML fairness approaches.
• Applying epidemiological frameworks improves the understanding and assessment of fairness in clinical ML.