Profit Analysis and Bootstrapping of Mass Spectrometry Data

• Activity Number: 435
• Type: Topic Contributed
• Date/Time: Wednesday, August 5, 2009 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Brian Daniels+ and Katie Lentz*+
• Companies: The College of New Jersey
• Keywords: binary classification ; profit analysis ; mass spectrometry ; bootstrapping ; neural network ; logistic regression

Abstract

This study investigates prediction and diagnosis of cancer via a number of data mining tools on a set of mass spectrometry data. The original data has 200 patients and 360,000 predictors. Our study uses dynamic binning and feature selection analysis to facilitate the modeling and prediction of the binary target (cancer vs. non-cancer). The data mining tools in this study include decision tree, logistic regression, neural network, and support vector machine. In the modeling process, we impose a profit/cost structure from the insurance view point and then compare the model performance by the profits of the tools. A bootstrapping technique is used to facilitate a fair comparison of the models.