Technological advances have yielded a wealth of
biomarkers that have the potential to detect
chronic diseases such as
cancer. However, most
biomarkers considered for further validation turn out not to have strong enough performance to be used in clinical practice. Group sequential designs that allow early termination for futility may be cost-effective for
biomarker studies based on biobanks of stored specimens. Previous studies proposed a group sequential design for the validation of a single
biomarker. In this article, we adapt a 2-stage design to the setting where a panel of candidate
biomarkers are under investigation. Conditional estimators of the clinical performance are proposed under an updated risk model that uses all accrued data, and can be computed through resampling procedures. Under a special case where a multivariate binormal distribution applies for
biomarkers following a suitable transformation, these estimators have analytical forms, alleviating the computational burden while retaining statistical efficiency. Performance of the proposed 2-stage design and estimators are compared with a traditional fixed-sample design and an existing 2-stage design that allows early termination but does not update the risk model with accrued information. Our proposed design and estimators show an ability to reduce sample size when the
biomarker panel is not promising, while controlling rejection rate and gaining efficiency when the panel is promising. We apply the proposed methods to a
biomarker panel development for the detection of high-grade
prostate cancer in a study conducted within the National Cancer Institute's Early Detection Research Network.