Mitigating Disparities in Prostate Cancer Survival Prediction Through Fairness-Aware Machine Learning Models

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Prediction models can perpetuate healthcare disparities due to unequal performance across demographic groups.
• Fairness-aware methods are established for binary outcomes but less explored in survival analysis.
• Racial disparities exist in predicting survival after prostate cancer treatment.

Purpose of the Study:

• To compare two fairness-aware deep learning survival models for predicting survival after radical prostatectomy.
• To mitigate racial disparities in the performance of these prediction models.
• To evaluate the effectiveness of fairness-aware approaches in survival analysis for prostate cancer.

Main Methods:

• Utilized the National Cancer Database to train deep Cox proportional hazards models for overall survival.
• Compared two fairness-aware models: Fair Deep Cox Proportional Hazards Model (Fair DCPH) and Group Distributionally Robust Optimization Deep Cox Proportional Hazards Model (GroupDRO DCPH).
• Assessed model fairness using cross-group and within-group concordance indices (C-index) across racial groups.

Main Results:

• The baseline Deep Cox Proportional Hazards Model (DCPH) showed performance disparities across racial groups (e.g., lower C-index for Black and Hispanic patients).
• Fairness-aware models (Fair DCPH, GroupDRO DCPH) improved cross-group C-indices for Black, Hispanic, and Asian patients.
• These improvements were achieved with minimal performance loss in the White patient subgroup.

Conclusions:

• Two fairness-aware survival models were benchmarked to address racial disparities in post-prostatectomy survival prediction.
• These methods demonstrate potential for ensuring equitable care through fair prediction models.
• The approaches can be extended to other time-to-event models to promote fairness in healthcare predictions.