An empirical comparison of different LDA methods in fMRI-based brain states decoding.

Abstract

Decoding brain states from response patterns with multivariate pattern recognition techniques is a popular method for detecting multivoxel patterns of brain activation. These patterns are informative with respect to a subject's perceptual or cognitive states. Linear discriminant analysis (LDA) cannot be directly applied to fMRI data analysis because of the "few samples and large features" nature of functional magnetic resonance imaging (fMRI) data. Although several improved LDA methods have been used in fMRI-based decoding, little is known regarding the relative performance of different LDA classifiers on fMRI data. In this study, we compared five LDA classifiers using both simulated data with varied noise levels and real fMRI data. The compared LDA classifiers include LDA combined with PCA (LDA-PCA), LDA with three types of regularizations (identity matrix, diagonal matrix and scaled identity matrix) and LDA with optimal-shrinkage covariance estimator using Ledoit and Wolf lemma (LDA-LW). The results indicated that LDA-LW was the most robust to noises. Moreover, LDA-LW and LDA with scaled identity matrix showed better stability and classification accuracy than the other methods. LDA-LW demonstrated the best overall performance.