Survival After Radical Cystectomy for Bladder Cancer: Development of a Fair Machine Learning Model

Area of Science:

• Oncology
• Biostatistics
• Health Informatics

Background:

• Machine learning (ML) models are increasingly used in healthcare, but can exhibit bias, leading to unfair performance across different population subgroups.
• This bias is a significant concern in bladder cancer, where known disparities exist among sex and racial groups.

Purpose of the Study:

• To develop an ML model for predicting survival after radical cystectomy in bladder cancer patients.
• To assess the developed model for potential bias across sex and racial subgroups.
• To compare different techniques for mitigating algorithmic unfairness and improving model fairness.

Main Methods:

• Trained and compared various ML classification algorithms using the National Cancer Database to predict 5-year survival post-radical cystectomy.
• Evaluated model performance using the F1-score and fairness using the equalized odds ratio (eOR).
• Compared three algorithmic unfairness mitigation techniques to enhance the eOR.

Main Results:

• The best-performing naive model (extreme gradient boosting) achieved an F1-score of 0.860 and an eOR of 0.619.
• All tested mitigation techniques improved the eOR, with the correlation remover technique yielding the highest increase to 0.750.
• The mitigated model maintained strong performance across diverse subgroups, with F1-scores of 0.86 (overall), 0.904 (Black males), and 0.824 (Asian females).

Conclusions:

• The initial ML model for bladder cancer survival prediction demonstrated bias across sex and racial subgroups.
• Algorithmic unfairness mitigation techniques effectively improved model fairness, as indicated by the increased eOR.
• This study underscores the importance of evaluating and actively mitigating bias in ML models to ensure equitable healthcare delivery, culminating in the deployment of the first web-based fair ML model for this purpose.