Searching Solution Space for SELDI-TOF Cancer Classifiers

• Activity Number: 226
• Type: Contributed
• Date/Time: Tuesday, August 9, 2005 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Eric Siegel*+
• Companies: University of Arkansas for Medical Sciences
• Keywords: SELDI ; proteomics ; classification ; cancer ; mass spectrometry

Abstract

To develop a classification model, we typically divide the data in half to train the model on one half and test it on the other. But when the training phase employs automated variable selection on a large number of variables, as occurs in cancer proteomics, the result is a model whose "solution" (predictors plus parameter estimates) may depend on the half to which the samples are allocated. To see how big a problem this is, I recently iterated the division, training, and testing process 10,000 times to get a distribution of solutions for classifying pancreatic cancer via logistic regression with Forward Selection on intensities of 37 peaks from the SELDI-TOF mass spectra of 103 serum samples. Although classifiers contained from three to seven peaks, every peak was selected at least once (min=12), and no peak was always selected (max=9279), implying that the peaks chosen by logistic regression with Forward Selection constitute a nonrobust and unreliable panel for classification. From each sample's Prob(Cancer|iteration), I developed two distribution-wide measures of classification performance and used them to identify a robust panel of SELDI peaks for cancer classification.