Fairness in Predicting Cancer Mortality Across Racial Subgroups.

Area of Science:

• Oncology
• Medical Informatics
• Health Equity

Background:

• Machine learning (ML) models can aid cancer care by identifying patients needing serious illness conversations.
• Evaluating ML models for racial bias is crucial to prevent exacerbating existing health disparities.

Purpose of the Study:

• To assess racial bias in a predictive ML model designed to identify 180-day cancer mortality risk.
• To ensure the model's equitable performance across diverse racial groups.

Main Methods:

• A cohort study utilized a random forest algorithm with retrospective data from cancer registries and EHRs.
• The model predicted 180-day cancer mortality risk for adult patients diagnosed between 2016-2021.
• Performance and fairness metrics (e.g., AUROC, F1 score, equal opportunity, equalized odds) were evaluated across racial categories.

Main Results:

• The validation dataset comprised 43,274 patients with balanced demographics.
• Model performance metrics (AUROC, F1 score) showed reasonable concordance across Asian, Black, Native American, White, and other/unknown race groups.
• Fairness metrics (equal opportunity, equalized odds, disparate impact) indicated no significant variation, suggesting absence of racial bias.

Conclusions:

• The ML model for cancer mortality risk demonstrated fair performance across racial groups, indicating no significant racial bias.
• The findings support the potential for equitable clinical application of this model.
• Continuous monitoring in clinical settings is recommended to ensure sustained equitable patient care.