Assessing Phenotype Definitions for Algorithmic Fairness.

Area of Science:

• Health Informatics
• Epidemiology
• Health Equity Research

Background:

• Phenotyping is essential for observational health research, influencing disease characterization, risk stratification, and treatment studies.
• Ensuring cohort representativeness across demographics and social determinants of health is critical for unbiased research outcomes.
• Existing phenotype definitions may inadvertently introduce bias, affecting downstream analyses and health equity.

Purpose of the Study:

• To propose best practices for assessing the fairness of phenotype definitions in observational health research.
• To evaluate the performance of various phenotype definitions using established fairness and epidemiological metrics.
• To highlight disparities in phenotype definition performance across different demographic subgroups.

Main Methods:

• Leveraged established fairness metrics from predictive modeling and related them to epidemiological metrics.
• Conducted an empirical study using phenotype definitions for Crohn's disease and type 2 diabetes from existing literature.
• Analyzed phenotype definition performance across racial and gender subgroups.

Main Results:

• Different phenotype definitions exhibited significant variations in performance when assessed for fairness.
• Disparities in performance were observed across various fairness metrics and demographic subgroups.
• Empirical studies demonstrated that phenotype definitions can lead to non-representative patient cohorts.

Conclusions:

• Proposed best practices can aid in the development of fair and inclusive phenotype definitions.
• Fairness assessment of phenotype definitions is critical for mitigating bias in health research.
• Adopting these practices can enhance the representativeness of patient cohorts and promote health equity.