Improving the Diagnostic Accuracy of a Stratified Screening Strategy by Identifying the Optimal Risk Cutoff

Abstract

Abstract We give a novel decision-theoretic approach for finding the optimal risk cutoff for additional breast cancer screening for women who have a high model-predicted risk of breast cancer. Methods: The American Cancer Society (ACS) suggests a stratified screening strategy for breast cancer. The strategy includes assessing women's risk of breast cancer using a risk model, screening women at high risk of breast cancer with both contrast-enhanced breast MRI and mammography, while screening women at low risk of breast cancer with mammography alone. The ACS used expert consensus opinion to choose the risk cutoff for additional screening. Instead, we suggest a risk cutoff chosen to maximize the full area under the receiver operating characteristic curve for the strategy. We use three inputs to find the cutoff, including: 1) the distribution of five-year breast cancer risk scores, 2) the probability of breast cancer given the risk score, and 3) the diagnostic accuracies of digital mammography, and contrast-enhanced screening breast MRI. No data were publicly available for testing the performance of stratified screening strategies based on the ACS suggested risk models. Instead, we used publicly available Breast Cancer Screening Consortium (BCSC) risk data to seek an optimal risk cutoff. A previous study suggests that the BCSC risk model has similar predictive accuracy as the Tyrer-Cuzick and Claus models. For comparison, we used a hypothetical risk model with much better predictive accuracy than the BCSC model. Results: For the BCSC risk model, the strategy with the highest diagnostic accuracy for the entire population is to screen almost all women with both digital mammography and contrast- enhanced breast MRI. Under the strategy based on the hypothetical risk model, only women at 18% or higher five-year risk would receive dual modality screening. Conclusion: The results occur because the BCSC risk model does not accurately predict which women will or will not develop breast cancer. However, if there were a more accurate risk model, a stratified screening strategy would offer benefits for the entire population by balancing the harms of breast cancer screening with the potential for improved breast cancer detection.