Detecting, Characterizing, and Mitigating Implicit and Explicit Racial Biases in Health Care Datasets With Subgroup...

Area of Science:

• Artificial Intelligence in Healthcare
• Algorithmic Fairness
• Data Science

Background:

• Growing adoption of healthcare algorithms raises concerns about perpetuating bias against disadvantaged groups.
• Existing bias mitigation methods focus on model modification, with limited efforts on data-level interventions.
• Healthcare datasets are prone to biases that can impact diagnostic and prognostic algorithm performance.

Purpose of the Study:

• Introduce AEquity, a novel metric using learning curve approximation to identify and mitigate bias in healthcare data.
• Demonstrate AEquity's effectiveness in guided dataset collection and relabeling for improved algorithmic fairness.
• Evaluate AEquity's robustness across diverse datasets, algorithms, and fairness metrics.

Main Methods:

• Developed AEquity metric based on learning curve approximation for bias detection and mitigation.
• Applied AEquity to chest X-ray datasets, healthcare cost utilization data, and the National Health and Nutrition Examination Survey (NHANES).
• Benchmarked AEquity against state-of-the-art methods like balanced empirical risk minimization and calibration.

Main Results:

• AEquity guided data collection reduced bias in chest radiographs by 29%-96.5% (AUC).
• Significant bias reduction observed for intersectional populations (Black patients on Medicaid) across multiple fairness metrics (e.g., 33.3% FNR reduction).
• AEquity outperformed balanced empirical risk minimization and calibration, showing robust performance across various AI models (CNNs, Transformers, etc.).

Conclusions:

• AEquity is a robust and effective tool for mitigating algorithmic bias at the data level in healthcare.
• The metric demonstrates broad applicability across different datasets, demographic groups, and machine learning architectures.
• AEquity offers a promising data-centric approach to enhance fairness in healthcare AI.