A framework for evaluating biomarkers for early detection: validation of biomarker panels for ovarian cancer

Abstract

A panel of biomarkers may improve predictive performance over individual markers. Although many biomarker panels have been described for ovarian cancer, few studies used prediagnostic samples to assess the potential of the panels for early detection. We conducted a multisite systematic evaluation of biomarker panels using prediagnostic serum samples from the Prostate, Lung, Colorectal, and Ovarian Cancer (PLCO) screening trial. Using a nested case-control design, levels of 28 biomarkers were measured laboratory-blinded in 118 serum samples obtained before cancer diagnosis and 951 serum samples from matched controls. Five predictive models, each containing 6 to 8 biomarkers, were evaluated according to a predetermined analysis plan. Three sequential analyses were conducted: blinded validation of previously established models (step 1); simultaneous split-sample discovery and validation of models (step 2); and exploratory discovery of new models (step 3). Sensitivity, specificity, sensitivity at 98% specificity, and AUC were computed for the models and CA125 alone among 67 cases diagnosed within one year of blood draw and 476 matched controls. In step 1, one model showed comparable performance to CA125, with sensitivity, specificity, and AUC at 69.2%, 96.6%, and 0.892, respectively. Remaining models had poorer performance than CA125 alone. In step 2, we observed a similar pattern. In step 3, a model derived from all 28 markers failed to show improvement over CA125. Thus, biomarker panels discovered in diagnostic samples may not validate in prediagnostic samples; utilizing prediagnostic samples for discovery may be helpful in developing validated early detection panels.

Figure 1

Sampling scheme. Note: 119 cases and 952 controls were originally selected following the sampling scheme. One case was excluded due to a coding error, and one control was excluded due to a consent issue, leaving 118 cases and 951 controls in the final sample set. LMP denotes tumors of low malignant potential.

Figure 2

Receiver operating characteristic (ROC) curves for Step 1 models (a) and Step 2 models (b), compared to that of CA125 and the Step 3 (pan-site) model in the within-one-year cases. Black solid: CA125 alone; blue solid: Step 3 model; blue dotted: Panel D; blue dashed: Panel C; red solid: Panel B; red dotted: Panel A; red dashed: Panel E. Note: Figure 2b curves are based on validation set only.