Abstract PR05: Does a comprehensive family history of colorectal cancer improve risk prediction?

Abstract

Abstract Background: Family history of colorectal cancer (CRC) is a strong and well-established risk factor for CRC. To date, however, family history (FH) of the disease is generally only broadly categorized (usually as present or absent) in most risk prediction models (Freedman et al. 2009). These approaches fail to fully utilize information on family history and may lead to suboptimal predictive performance of CRC risk. We investigated the utility of a CRC risk model that incorporates a comprehensive family history of CRC as well as information on known genetic and environmental risk factors and personal characteristics. Methods: We used data from the Colon Cancer Family Registry (CCFR), a large, international consortium of six study centers. Prediction models were developed based on incident invasive CRC cases (N = 4445) and population-based controls (N = 3967) that were recruited from three study sites (Seattle, USA; Ontario, Canada; and Melbourne, Australia). A familial risk profile (FRP) score, a probability index of absolute risk for lifetime CRC was estimated based on family structure, age of onset for affected relatives and the polygenic effect of MLH1, MSH2, MSH6, PMS2 and MUTYH using modified segregation analysis, an approach adapted from Antoniou et al (2002)). Two sets of gender-specific logistic regression models were built: (I) the FRP models, which included FRP and other known risk factors (e.g., BMI, consumption of red meat, calcium and NSAID use duration, smoking amount (pack-years), a history of polyps, and history of FOBT, sigmoidoscopy, colonoscopy, fruit intake, and use of hormone replacement therapy for female); and (II) binary FH models, which replaced FRP with a binary indicator (yes/no) for any self-reported first-degree family member with CRC. 5-year absolute risks were calculated based on the estimated odds ratios (OR), country-, sex- and age-specific CRC incidence rate and mortality due to causes other than CRC. Model validation was conducted with unaffected relatives (N=12,120) and population-based controls (N=1,899) from five study sites based on the follow-up information on incident CRC and death status. The primary endpoint was CRC diagnosis within 5-year after baseline. We used calibration plots to compare the predicted 5-year absolute risks with the observed cumulative incidence rates. Receiver Operating Characteristic (ROC) curve analyses were conducted and areas under the ROC curve (AUC) were used to assess the discriminatory capacity for separating subjects with and without a CRC diagnosis within 5 years, accounting for censoring and competing risk. Results: The ORs (95% confidence interval [CI]) using the FRP per 10% increase were 1.16 (1.11-1.20) for males, and 1.09 (1.06-1.12) for females in the FRP models, while the ORs for the binary FH model were 2.32 (1.88- 2.85) for men and 1.70 (1.38-2.09) for women. The FRP models provided slightly better calibration, with average predicted risks falling within the 95% CIs of the empirical cumulative rates. The binary FH models, by comparison, tended to yield higher estimated CRC risks compared with the observed risks among individuals whose risks were above the top10% of the risk distribution. Both models yielded comparable AUCs using the full validation set. Among individuals with at least one first-degree family member affected with CRC, the FRP model performed significantly better (AUC = 0.71) than the FH model (AUC = 0.63) for male participants; difference equaled 0.09 (95% CI: 0.02, 0.16). The models were comparable for females. Conclusion: Our CRC prediction model that incorporates more comprehensive family history of CRC can provide improved calibration and discrimination of risks compared with the simple FH model, especially in populations with higher underlying risk. The models developed may potentially further improve screening decision making among subgroups with elevated CRC risk.