Recalibration of Genomic Risk Prediction Models in Prostate Cancer to Improve Individual-Level Predictions

• Activity Number: 555
• Type: Contributed
• Date/Time: Wednesday, August 12, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Genomics and Genetics
• Author(s): Voleak Choeurng* and Bin Luo and Kasra Yousefi and Zaid Haddad and Heesun Shin and Ashley Ross and Edward Schaeffer and Robert Den and Adam Dicker and Jeffrey Karnes and Elai Davicioni and Darby Thompson
• Companies: GenomeDx Biosciences and University of Western Ontario and GenomeDx Biosciences and GenomeDx Biosciences and GenomeDx Biosciences and Johns Hopkins Medical Institution and Johns Hopkins Medical Institution and Thomas Jefferson University and Thomas Jefferson University and Mayo Clinic and GenomeDx Biosciences and EMMES Canada
• Keywords: genomics ; biomarker ; prostate cancer ; prostate ; calibration

Abstract

Despite the obvious clinical importance, a test's calibration is rarely described in clinical validation papers. While RNA-expression-based tests are quantitative and can be re-calibrated to provide accurate patient-specific predictions, most are reported as qualitative results. This investigation demonstrates the process of re-calibration and construction of clinically meaningful cut-points for a validated genomic test, Decipher, used for predicting post-surgical cancer progression. Decipher is re-calibrated in a case-cohort study (n=216) using a proportional-hazards (PH) model with time-dependent treatment effects, accommodating departures from the PH assumption. Performance is assessed through a number of metrics; calibration in-the-large, calibration slope, goodness-of-fit, modified Hosmer-Lemshow. Cut-points of the re-calibrated score were identified using resampling and maximizing the partial likelihood of a Cox model. Based on this approach, optimized categories of Decipher were < 0.45, 0.45-0.60 and >0.60. We also demonstrate that these methods can be applied to other genomic-based tests regardless of their method of discovery or model.