Risk prediction models for colorectal cancer: Evaluating the discrimination due to added biomarkers

Most risk prediction models for colorectal cancer (CRC) are based on questionnaires and show a modest discriminatory ability. Therefore, we aim to develop risk prediction models incorporating plasma biomarkers for CRC to improve discrimination. We assessed the predictivity of 11 biomarkers in 736 men in the Health Professionals Follow-up Study and 639 women in the Nurses' Health Study. We used stepwise logistic regression to examine whether a set of biomarkers improved the predictivity on the basis of predictors in the National Cancer Institute's (NCI) Colorectal Cancer Risk Assessment Tool. Model discrimination was assessed using C-statistics. Bootstrap with 500 randomly sampled replicates was used for internal validation. The models containing each biomarker generated a C-statistic ranging from 0.50 to 0.59 in men and 0.50 to 0.54 in women. The NCI model demonstrated a C-statistic (95% CI) of 0.67 (0.62-0.71) in men and 0.58 (0.54-0.63) in women. Through stepwise selection of biomarkers, the C-statistic increased to 0.70 (0.66-0.74) in men after adding growth/differentiation factor 15, total adiponectin, sex hormone binding globulin and tumor necrosis factor receptor superfamily member 1B (P for difference = 0.008); and increased to 0.62 (0.57-0.66) in women after further including insulin-like growth factor 1 and insulin-like growth factor-binding protein 3 (P for difference = .06). The NCI + selected biomarkers model was internally validated with a C-statistic (95% CI) of 0.73 (0.70-0.77) in men and 0.66 (0.61-0.70) in women. Circulating plasma biomarkers may improve the performance of risk factor-based prediction model for CRC.