Abstract

Most biomarkers, including cancer biomarkers, are interpreted using a single threshold. However, programs for the early detection of cancer involve regular testing at set intervals providing longitudinal biomarker values which may be more informative than a value from a single time point. From initial clinical trials of early detection of ovarian cancer (OC) with serum CA125, we developed an algorithm for early detection of OC which interprets the pattern of CA125 over time. Essentially each woman serves as her own control with the algorithm identifying significant rises above the woman’s CA125 baseline as an earlier and more sensitive method while maintaining the same low acceptable false positive rate as a single threshold rule achieves. The building blocks in the retrospective analysis are hierarchical change-point and mixture models for women who develop OC during the trial (cases) and hierarchical flat profiles in women who do not develop OC (controls). We discuss results from UK and US trials in normal risk postmenopausal women and in high risk women, and briefly mention algorithms with multiple longitudinal markers, and biomarker discovery in longitudinal samples.

Speaker

Steven Skates

Research Area

Statistics seminar

Affiliation

Harvard Medical School

Date

Friday, 17 March 2023, 4pm

Venue

Zoom (link below)