A0592
Title: Sparse functional linear discriminant analysis for high-dimensional predictors
Authors: Limeng Liu - University of Minnesota (United States) [presenting]
Guannan Wang - College of William & Mary (United States)
Sandra Safo - University of Minnesota (United States)
Abstract: Functional data analysis (FDA) aims to analyze functional datasets where observations are not individual data points but functions. Most FDA approaches are for continuous time domains with a single variable. However, multivariate data measured at dense or sparse time points are collected in biomedical research with a typical goal of finding time-variant profiles to distinguish between classes. Fisher's linear discriminant analysis (LDA) is a popular multivariate dimension reduction method for finding linear combinations of variables that optimally separate classes. Most existing LDA methods for the FDA only apply to a single variable or binary classes and cannot identify variables discriminating between classes over time. Sparse functional linear discriminant analysis (SFLDA) is proposed to find linear combinations of multiple functional predictors that optimally discriminate between two or more classes over time and identify functional predictors. Simulations are used to demonstrate the effectiveness of SFLDA. SFLDA is applied to the inflammatory bowel disease study, and a personal omics profiling dataset is integrated to identify longitudinal biomarkers of disease progression.
