|
Activity Number:
|
374
|
|
Type:
|
Topic Contributed
|
|
Date/Time:
|
Wednesday, August 9, 2006 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Statistics in Epidemiology
|
| Abstract - #306102 |
|
Title:
|
Prediction Modeling Using Survival Data for Gene Expression Prognostic Test for Breast Cancer
|
|
Author(s):
|
Kit Lau*+ and Alice Wang and John Sninsky and Trevor Hastie
|
|
Companies:
|
Celera Diagnostics and Celera Diagnostics and Celera Diagnostics and Stanford University
|
|
Address:
|
1401 Harbor Bay Parkway, Alameda, CA, 94502,
|
|
Keywords:
|
gene expression ; prognostic test ; prediction modeling ; survival data ; breast cancer
|
|
Abstract:
|
Gene expression profiles have been shown to be prognostic for cancer. A supervised principal components method (Bair 2004) using survival endpoint was used to build a breast cancer signature for distant metastasis. A metastasis score (MS) from gene expression was derived from the Cox model. Probabilities of distant metastasis at any time can be calculated from the MS. To avoid over-fitting, "pre-validation" technique (Tibshirani 2002) was used to estimate test performance. The MS was calculated with 10-fold cross-validation (MS(CV)). Time-dependent AUC (Heagerty 2000), hazard ratios, and survival rates of stratified groups from MS(CV) were evaluated. Univariate and multivariate Cox regression were performed for the MS(CV) and Adjuvant!, an online prognosticator. An integrated MS that combines Adjuvant! and gene expression data had an improved AUC over both.
|
