Author's School

School of Engineering & Applied Science

Author's Department/Program

Computer Science and Engineering

Language

English (en)

Date of Award

1-1-2009

Degree Type

Thesis

Degree Name

Master of Arts (MA)

Chair and Committee

Yixin Chen

Abstract

Patients undergoing radiation therapy can develop a potentially fatal inflammation of the lungs known as radiation pneumonitis: RP). In practice, modeling RP factors is difficult because existing data are under-sampled and imbalanced. Support vector machines: SVMs), a class of statistical learning methods that implicitly maps data into a higher dimensional space, is one machine learning method that recently has been applied to the RP problem with encouraging results. In this thesis, we present and evaluate an ensemble SVM method of modeling radiation pneumonitis. The method internalizes kernel/model parameter selection into model building and enables feature scaling via Olivier Chapelle's method. We show that the ensemble method provides statistically significant increases to the cross-folded area under the receiver operating characteristic curve while maintaining model parsimony. Finally, we extend our model with John C. Platt's method to support non-binary outcomes in order to augment clinical relevancy.

DOI

https://doi.org/10.7936/K7XK8CKP

Comments

Permanent URL: http://dx.doi.org/10.7936/K7XK8CKP

Share

COinS