Multi-task learning with dynamic re-weighting to achieve fairness in healthcare predictive modeling

Area of Science:

• Healthcare AI
• Machine Learning Fairness
• Predictive Analytics

Background:

• Algorithmic bias in healthcare can lead to disparities in automated decision-making.
• Ensuring fairness in predictive models is crucial to prevent prejudice against minority groups.
• Existing methods often alter data or optimize constraints, which can be complex.

Purpose of the Study:

• To propose a novel fairness-achieving scheme using multitask learning.
• To address the challenge of bias in predictive healthcare modeling.
• To develop a method that ensures prediction performance is equitable across sensitive subgroups.

Main Methods:

• Framing fairness as a task-balancing problem by dividing predictions for sub-populations into separate tasks.
• Implementing a dynamic re-weighting approach during model training.
• Modifying gradients of prediction tasks during neural network back-propagation for fairness.

Main Results:

• The proposed multitask learning scheme significantly reduces disparity between subgroups by 98%.
• The approach achieves fairness with a minimal loss of less than 4% in prediction accuracy.
• Demonstrated effectiveness on a real-world use case for predicting sepsis patient mortality risk.

Conclusions:

• Multitask learning offers a novel and effective framework for achieving fairness in predictive healthcare models.
• The dynamic re-weighting method provides a robust way to mitigate bias during model training.
• This approach balances fairness and accuracy, showing promise for equitable healthcare AI.