|
Activity Number:
|
22
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Sunday, July 29, 2007 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
Section on Bayesian Statistical Science
|
| Abstract - #309427 |
|
Title:
|
Classification with Spectroscopic Data Using Bayesian Variable Selection
|
|
Author(s):
|
Hongxiao Zhu*+ and Dennis D. Cox
|
|
Companies:
|
Rice University and Rice University
|
|
Address:
|
6100 Main St MS 138 , Houston, TX, 77005,
|
|
Keywords:
|
Bayesian Variable Selection ; Classification ; Probit Model ; Latent Variable ; Fluorescence Spectroscopy ; Cervical Cancer
|
|
Abstract:
|
Fluorescence Spectroscopy provides a non-invasive tool for in vivo, real time diagnosis of cervical pre-cancer. An important issue involved is to accurately classify diseased tissue from normal using extremely high dimensional data-the Fluorescence excitation-emission matrices (EEMs). This paper presents a Bayesian probit model for classification purpose based on features extracted from EEMs. Latent variables are included to simplify computation, and variable selection is performed through a mixture normal prior of the regression coefficients. MCMC methods-a Gibbs sampler and a Hybrid Metropolis-Hasting/Gibbs sampler, are used to obtain posterior samples. Simulations show that this model produces accurate variable selection results and good estimates of the regression coefficients. Application to spectroscopic data gives improved classification performance compared with other methods.
|
