APPLICATION OF SUPPORT VECTOR MACHINES TO LONGITUDINAL FUNCTIONAL NEUROIMAGING DATA

Abstract

The principal objective of this thesis was to test a novel adaptation of the support vector machine (SVM), called a longitudinal support vector machine(LSVM), on longitudinal functional neuroimaging data. LSVM performance was compared to a traditional SVM and logistic regression (LR) using classification accuracy and interpretability of feature weights. Classification accuracy was measured as the percentage of subjects placed into their correct categories, and feature weights by how closely they matched the known signal. The first study involved purely simulated data, which found the LSVM had higher classification accuracy for data without heteroscedasticity, but performed worse when heteroscedasticity was introduced. The second study used real magnetoencephalography (MEG) resting state readings added to a simulated trend. The LSVM had similar classification accuracy, and only had more interpretable feature weights at the highest SNR dataset. Currently the LSVM is not recommended over the SVM/LR algorithms.