Use of detailed family history data to improve risk prediction,with application to breast cancer screening.

As breast cancer represents a major morbidity and mortality burden in the U.S., with about one in eight women developing invasive breast cancer over her lifetime, accurate low-cost screening is an important public health issue. First-degree family history, often simplified as a dichotomous or three-level categorical variable (0/1/>1) based on number of affected relatives, is an important risk factor for many conditions. However, detailed family structure information such as the total number of first-degree relatives, and for each, their current or death age, and age at diagnosis are also important for risk prediction.We develop a family history score under a Bayesian framework, based on first-degree family structure. We tested performance of the proposed score using data from a large prospective cohort study of women with a first-degree breast cancer family history. We used likelihood ratio tests to evaluate whether the proposed score added additional information to a Cox model with known breast cancer risk factors and the three-level family history variable. We also compared prediction performance through Receiver Operating Characteristic (ROC) curves and goodness-of-fit testing.Our proposed Bayesian family history score improved fit compared to the commonly used three-level family history score, both without and with adjustment for other risk factors (likelihood ratio tests p = 0.003 without adjustment for other risk factors, and p = 0.007 and 0.009 under adjustment with two candidate sets of risk factors). AUCs of ROC curves for the two models were similar, though in all cases were higher after addition of the BFHS.Capturing detailed family history data through the proposed family history score can improve risk assessment and prediction. Such approaches could enable better-targeted personalized screening schedules and prevention strategies.

Duke Scholars

Author Yue Jiang Statistical Science