Fair Survival Time Prediction via Mutual Information Minimization

Area of Science:

• Computer Science
• Statistics
• Machine Learning

Background:

• Survival analysis is crucial for predicting event occurrence over time, especially with censored data.
• Algorithmic fairness has advanced significantly, yet fairness in survival analysis remains under-explored.
• Existing methods lack robust approaches for ensuring fairness in time-to-event predictions.

Purpose of the Study:

• To propose a novel framework for achieving demographic parity in survival analysis.
• To minimize the mutual information between predicted time-to-event and sensitive attributes.
• To develop new disparity assessment metrics for survival predictions.

Main Methods:

• Developed a framework to minimize mutual information between survival predictions and sensitive attributes.
• Implemented techniques to ensure statistical independence of time-to-event predictions.
• Proposed four novel disparity assessment metrics tailored for survival analysis.

Main Results:

• The proposed method effectively minimizes mutual information, promoting fairness.
• Experiments demonstrate systematic improvement in the fairness of survival predictions.
• The approach is robust and performs well even with censored data across benchmark datasets.

Conclusions:

• The introduced framework successfully enhances fairness in survival analysis models.
• Minimizing dependence on sensitive attributes leads to more equitable time-to-event predictions.
• This work provides a valuable tool for developing fair and reliable survival analysis systems.