Genetic risk prediction models such as BRCAPRO play a critical role in identification and management of women who carry mutations of breast cancer genes BRCA1 and BRCA2. Although BRCAPRO is widely used in genetic counseling, an impediment to its use in primary care is the fact that it requires potentially extensive information on counselee and her family history to estimate the carrier probabilities of BRCA1/2 genes. On the other hand, primary care settings such as mammography centers are ideal for identifying women at high risk for breast and ovarian cancers at a large population level. As a big proportion of women who are at high risk typically are unaware of their risk, implementing risk prediction models in primary care can make a huge impact in identification and management of genetically pre-disposed women. To bring BRCAPRO to this level, we need to balance the tradeoff between the amount of information used and accuracy achieved. With this motivation, we propose a two-stage approach. In the first stage, only a limited amount of family history information will be collected and that data will be analyzed using a simpler version of BRCAPRO or other simpler models. If the assessed risk at this stage is sufficiently high, full version of BRCAPRO will be used in the second stage to obtain more accurate estimates. We propose several first stage tools that vary by the amount of information they require. In some of these tools, we augment the collected information by imputing the missing (unasked) information such as the current ages of unaffected relatives or the unaffected relatives themselves, which BRCAPRO can utilize to make potentially more accurate prediction. We will compare the first stage tools in terms of their sensitivities, specificities, and other related statistics at a range of cutoffs, and area under the ROC curve (AUC). Further, we develop a methodology to evaluate the overall performance of the two-stage approach that takes into account the fact that the second stage results are conditional on those of the first stage. In particular, we derive the overall sensitivity, specificity, and AUC of the approach. After developing the approach, we plan to validate it on an independent set of data. Our total sample exceeds 6,000 families and come from a wide range of sources - Cancer Genetics Network, University of Texas (UT) MD Anderson Cancer Center, UT Southwestern Medical Center, Newton-Wellesley Hospital, St. Barnabas Health Care system, Yale University, Middlesex Hospital, and Bermuda Cancer Genetics and Risk Assessment Program. We will also implement the approach in BayesMendel and HughesRiskApps software.
A key to preventing or surviving breast and ovarian cancers is early identification and appropriate management of women at high risk for these cancers, and one such high-risk group comprises those carrying genetic mutations. Our approach is intended to identify these genetically susceptible women at a large-scale population level through primary care settings. Thus, successful implementation of the proposed approach is expected to have a wide impact in prevention and treatment of these cancers in this population.
|Biswas, Swati; Arun, Banu; Parmigiani, Giovanni (2014) Reclassification of predictions for uncovering subgroup specific improvement. Stat Med 33:1914-27|
|Biswas, Swati; Atienza, Philamer; Chipman, Jonathan et al. (2013) Simplifying clinical use of the genetic risk prediction model BRCAPRO. Breast Cancer Res Treat 139:571-9|