Performance of general-population breast cancer risk prediction models in an international consortium

Background

Several breast cancer (BC) risk prediction models have been developed to provide personal risk assessments. Though individually validated, their performance has not been systematically evaluated across a wide range of populations or ages.

Methods

We harmonized individual-level baseline questionnaire data and incident BC diagnoses from 21 cohorts from North America, Europe, and Australia participating in the Breast Cancer Risk Prediction Project. Five-year absolute risk of invasive BC was estimated for five established risk prediction models using classical risk factors only. Discrimination was evaluated by area under the curve (AUC). Calibration was assessed using average and risk-decile specific expected to observed (E/O) ratios. Performance metrics were meta-analyzed across cohorts and models. Metaregression tested associations between cohort characteristics and performance metrics.

Results

This analysis included 1,595,977 women aged 20-75 years, enrolled in studies between 1976-2015, with 19,062 (1.2%) invasive BC cases ascertained within 5 years from exposure assessment. Age-adjusted AUCs were similar across models and cohorts (pooled AUCs by model: 0.57-0.58), while E/O ratios varied substantially (pooled E/O ratios by model: 0.83-1.25). Overestimation was common among predicted high-risk individuals (>3%). No appreciable differences in model performance by cohort age, birth year, race, and variable missingness emerged. Calibration improved after assigning race-specific incidence rates.

Conclusion

Existing BC risk prediction models provided similar risk discrimination across multiple cohorts, although there was overestimation of risk for high-risk individuals. Performance variation across cohorts was not driven by specific characteristics, which supports development of a unified risk model for diverse populations that leverages appropriate incidence rates.

Key messages

• When using classical risk factor components of existing risk prediction models, we found similar discriminatory ability of models across diverse cohorts.

• Aside from underlying cancer incidence rate, which heavily influenced calibration, no cohort-specific characteristics were consistently associated with model performance.

• Risk was underestimated at lower predicted risk deciles and overestimated at higher predicted risk deciles, indicating a need to improve model fit by integrating more complex risk-factor relationships.