FAIM: Fairness-aware interpretable modeling for trustworthy machine learning in healthcare

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• Machine learning integration in healthcare raises fairness concerns.
• Existing bias mitigation methods may compromise model performance.

Purpose of the Study:

• To introduce a fairness-aware interpretable modeling (FAIM) framework.
• To improve model fairness and performance in high-stakes healthcare applications.
• To reduce intersectional biases related to race and sex.

Main Methods:

• Developed an interpretable framework (FAIM) with an interactive interface.
• Integrated data-driven evidence with clinical expertise for contextualized fairness.
• Validated FAIM on two real-world hospital admission datasets (MIMIC-IV-ED, SGH-ED).

Main Results:

• FAIM models demonstrated satisfactory discriminatory performance.
• FAIM significantly mitigated intersectional biases (race and sex).
• FAIM outperformed commonly used bias mitigation techniques.

Conclusions:

• FAIM effectively improves fairness without compromising performance.
• The framework facilitates multidisciplinary collaboration for tailored AI fairness.
• FAIM offers a practical approach for fair and high-performing AI in healthcare.