An adversarial training framework for mitigating algorithmic biases in clinical machine learning.

Area of Science:

• Healthcare AI
• Machine Learning Ethics
• Biomedical Informatics

Background:

• Machine learning (ML) is increasingly used in healthcare, but can perpetuate existing biases.
• Identifying and mitigating these biases is crucial for equitable healthcare delivery.
• Existing ML tools may inadvertently introduce or amplify disparities in patient outcomes.

Purpose of the Study:

• To introduce an adversarial training framework to mitigate data-driven biases in healthcare ML models.
• To address site-specific (hospital) and demographic (ethnicity) biases in predictive models.
• To evaluate the framework's effectiveness in improving fairness while maintaining clinical utility.

Main Methods:

• Developed an adversarial training framework designed to counteract biases acquired during data collection.
• Applied the framework to a real-world task of rapid COVID-19 prediction.
• Utilized the statistical definition of equalized odds to measure and improve fairness.

Main Results:

• Adversarial training significantly improved outcome fairness, specifically addressing hospital and ethnicity biases.
• The framework achieved clinically effective screening performance, with negative predictive values exceeding 0.98.
• Prospective and external validation across four independent hospital cohorts demonstrated the method's robustness.

Conclusions:

• The proposed adversarial training framework effectively mitigates biases in healthcare ML applications.
• This approach enhances fairness without sacrificing the clinical performance of predictive models.
• The framework is generalizable across various outcomes, models, and fairness definitions, offering broad applicability.