Influence of Race and Ethnicity on Advanced Breast Cancer Risk Model Performance
Overview
Removing race and ethnicity from the Breast Cancer Surveillance Consortium advanced breast cancer risk model led to miscalibration, overestimating risk in Asian women and underestimating risk in Black women. This exclusion reduced the model's ability to accurately identify Black women at intermediate or high risk, potentially impacting tailored screening strategies.
Background
Breast cancer incidence and mortality vary significantly across racial and ethnic groups, prompting inclusion of these factors in risk prediction models to guide screening and prevention. The Breast Cancer Surveillance Consortium developed an advanced breast cancer risk model incorporating race and ethnicity to inform screening intervals and supplemental imaging. Concerns exist that including race and ethnicity may introduce algorithmic bias, while excluding them might reduce model accuracy and equity. Evaluating the impact of race and ethnicity inclusion is critical for optimizing risk-targeted breast cancer screening.
Data Highlights
Race/Ethnicity
Expected/Observed Ratio (Original Model)
Expected/Observed Ratio (Race-Excluded Model)
95% CI (Original)
95% CI (Race-Excluded)
Asian (Annual Screening)
1.00
1.28
0.82–1.29
1.05–1.66
Black (Annual Screening)
1.00
0.61
0.88–1.15
0.53–0.70
Other/Multiple Race (Annual Screening)
1.00
0.82
0.85–1.21
0.70–0.99
Hispanic (Biennial Screening)
1.00
1.14
0.80–1.33
0.92–1.52
Model Discrimination (AUC)
0.682
0.677
0.670–0.694
0.665–0.689
Key Findings
Excluding race and ethnicity from the model overestimated advanced breast cancer risk in Asian women (expected/observed ratio 1.28 vs. 1.00).
Risk was underestimated in Black women when race and ethnicity were excluded (expected/observed ratio 0.61 vs. 1.00), leading to fewer Black women identified at intermediate/high risk.
The model including race and ethnicity showed slightly better overall discrimination (AUC 0.682) compared to the race-excluded model (AUC 0.677).
Advanced breast cancer among Asian women classified as intermediate/high risk increased from 6.1% to 16.7% when race and ethnicity were excluded.
Among Black women, intermediate/high risk classification decreased from 75.3% to 47.5% without race and ethnicity in the model.
Biennial screeners showed similar calibration issues, with overestimation of risk in Hispanic women when race and ethnicity were excluded.
Clinical Implications
Inclusion of race and ethnicity in advanced breast cancer risk models improves calibration and discrimination, particularly for Black and Asian women. Removing these factors may lead to under-identification of high-risk Black women and over-identification of risk in Asian women, potentially resulting in suboptimal screening recommendations. Clinicians should consider the impact of race and ethnicity on risk prediction to optimize personalized screening strategies and reduce disparities in breast cancer outcomes.
Conclusion
Race and ethnicity are important components of advanced breast cancer risk prediction models that enhance accuracy and equity. Excluding these variables may introduce bias, impairing the model's clinical utility in guiding screening and supplemental imaging decisions.
