Validation of a breast cancer risk prediction model based on the key risk factors: family history, mammographic density and polygenic risk

Abstract

Purpose: We compared a simple breast cancer risk prediction model, BRISK (which includes mammographic density, polygenic risk and clinical factors), against a similar model with more risk factors (simplified Rosner) and against two commonly used clinical models (Gail and IBIS). Methods: Using nested case–control data from the Nurses’ Health Study, we compared the models’ association, discrimination and calibration. Classification performance was compared between Gail and BRISK for 5-year risks and between IBIS and BRISK for remaining lifetime risk. Results: The odds ratio per standard deviation was 1.43 (95% CI 1.32, 1.55) for BRISK 5-year risk, 1.07 (95% CI 0.99, 1.14) for Gail 5-year risk, 1.72 (95% CI 1.59, 1.87) for simplified Rosner 10-year risk, 1.51 (95% CI 1.41, 1.62) for BRISK remaining lifetime risk and 1.26 (95% CI 1.16, 1.36) for IBIS remaining lifetime risk. The area under the receiver operating characteristic curve (AUC) was improved for BRISK over Gail for 5-year risk (AUC = 0.636 versus 0.511, P < 0.0001) and for BRISK over IBIS for remaining lifetime risk (AUC = 0.647 versus 0.571, P < 0.0001). BRISK was well calibrated for the estimation of both 5-year risk (expected/observed [E/O] = 1.03; 95% CI 0.73, 1.46) and remaining lifetime risk (E/O = 1.01; 95% CI 0.86, 1.17). The Gail 5-year risk (E/O = 0.85; 95% CI 0.58, 1.24) and IBIS remaining lifetime risk (E/O = 0.73; 95% CI 0.60, 0.87) were not well calibrated, with both under-estimating risk. BRISK improves classification of risk compared to Gail 5-year risk (NRI = 0.31; standard error [SE] = 0.031) and IBIS remaining lifetime risk (NRI = 0.287; SE = 0.035). Conclusion: BRISK performs better than two commonly used clinical risk models and no worse compared to a similar model with more risk factors.