Algorithmic fairness and bias mitigation for clinical machine learning with deep reinforcement learning.

Area of Science:

• Artificial Intelligence in Healthcare
• Machine Learning for Clinical Applications
• Health Equity and Bias Mitigation

Background:

• Machine learning models in healthcare risk perpetuating data biases.
• Ensuring fairness and equity in AI-driven healthcare is crucial.
• Existing methods struggle to address biases from data collection.

Purpose of the Study:

• To develop a reinforcement learning framework to mitigate biases in healthcare machine learning models.
• To evaluate the model's effectiveness in predicting COVID-19 and improving fairness.
• To demonstrate generalizability across different healthcare settings and tasks.

Main Methods:

• Implemented a reinforcement learning framework with a specialized reward function and training procedure.
• Evaluated the model on predicting COVID-19 in emergency department patients.
• Assessed mitigation of hospital-specific and ethnicity-based biases.
• Performed external validation across three independent hospitals.

Main Results:

• The model achieved clinically effective COVID-19 screening performance.
• Significantly improved outcome fairness compared to benchmarks and state-of-the-art methods.
• Demonstrated generalizability on a patient intensive care unit discharge status task.

Conclusions:

• The proposed reinforcement learning framework effectively mitigates biases in healthcare AI.
• The method offers a pathway to more equitable and reliable AI tools in medicine.
• The approach shows promise for broader applications in clinical decision support.