Imputing Missing Values for Neuroimaging Data Based on Principal Component Analysis

• Activity Number: 659
• Type: Contributed
• Date/Time: Thursday, August 4, 2016 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Imaging
• Author(s): Lan Kong and Menghan Li*
• Companies: Penn State University College of Medicine and Penn State University College of Medicine
• Keywords: imputation method ; missing data ; neuroimaging ; principle component analysis ; prediction model

Abstract

Neuroimaging biomarkers have shown great promise in improving diagnosis and prognosis of diseases related to brain disorders. Missing data in neuroimaging measurements present a unique challenge in developing a prediction model. High dimensionality and high correlations in the measurements make many commonly used imputation methods infeasible. The PCA based imputation method is an appealing approach in this case since the principal components are used to estimate the missing values. We will demonstrate the utility of PCA imputation method in a study of relapse risk in alcohol-dependent patients. The neuroimaging data were collected from about 50 patients via functional near-infrared spectroscopy (fNIRS), and summarized to more than 60 variables. There were small to moderate missing data for 40% of the subjects. We will apply PCA based imputation method and regularized logistic regression to evaluate the prediction performance of fNIR data. We will also conduct simulation studies to examine the sensitivity of PCA imputation method to various factors such as sample size, proportion of missing data, correlations among variables, and the number of variables.