J A. Welge1,2,
Richard A. Komoroski2
1Environmental Health, University of
Cincinnati, Cincinnati, OH, United States; 2Center for Imaging
Research, University of Cincinnati, Cincinnati, OH, United States
Using
prior 31P NMR data for the composition of phospholipid (PL) and PL
metabolites in postmortem schizophrenic and matched control brains, we
searched for multivariate regression models to classify these samples.
Because the number of measurements exceeded the number of samples, variable
selection was required. We employed Akaikes Information Criterion in
conjunction with repeated cross-validation using random splits of the data
into model-building and validation subsets. This procedure addressed the risk
of over-fitting the sample data and generated predictions from data not used
to select the model. Certain metabolites that were not individually
significant produced accurate classification when modeled jointly.