What is Fair? Defining Fairness in Machine Learning for Health

Ensuring that machine learning (ML) models are safe, effective, and equitable across all patient groups is essential for clinical decision-making and for preventing the reinforcement of existing health disparities. This review examines notions of fairness used in ML for health, including a review of why ML models can be unfair and how fairness has been quantified in a wide range of real-world examples. We provide an overview of commonly used fairness metrics and supplement our discussion with a case-study of an openly available electronic health record (EHR) dataset. We also discuss the outlook for future research, highlighting current challenges and opportunities in defining fairness in health.