Discriminative analysis of resting-state functional connectivity patterns of schizophrenia using low dimensional embedding of fMRI

Abstract

Recently, a functional disconnectivity hypothesis of schizophrenia has been proposed for the physiological explanation of behavioral syndromes of this complex mental disorder. In this paper, we aim at further examining whether syndromes of schizophrenia could be decoded by some special spatiotemporal patterns of resting-state functional connectivity. We designed a data-driven classifier based on machine learning to extract highly discriminative functional connectivity features and to discriminate schizophrenic patients from healthy controls. The proposed classifier consisted of two separate steps. First, we used feature selection based on a correlation coefficient method to extract highly discriminative regions and construct the optimal feature set for classification. Then, an unsupervised-learning classifier combining low-dimensional embedding and self-organized clustering of fMRI was trained to discriminate schizophrenic patients from healthy controls. The performance of the classifier was tested using a leave-one-out cross-validation strategy. The experimental results demonstrated not only high classification accuracy (93.75% for schizophrenic patients, 75.0% for healthy controls), but also good generalization and stability with respect to the number of extracted features. In addition, some functional connectivities between certain brain regions of the cerebellum and frontal cortex were found to exhibit the highest discriminative power, which might provide further evidence for the cognitive dysmetria hypothesis of schizophrenia. This primary study demonstrated that machine learning could extract exciting new information from the resting-state activity of a brain with schizophrenia, which might have potential ability to improve current diagnosis and treatment evaluation of schizophrenia.

Introduction

In recent years, there has been a growing interest in exploring brain variance between schizophrenic patients and healthy controls based on neuroimaging data. One motivation for investigating disease-related structural and functional abnormalities in patient brains is to find stable neuroimaging-based biomarkers, which are expected to provide additional information for current clinic diagnostic systems based solely on clinical manifestations (Kawasaki et al., 2007, Fan et al., 2007). Early research mainly reported structural and functional abnormalities in some special brain regions in schizophrenia using voxel-based analysis methods (Friston and Frith, 1995, DeLisi et al., 1998, Friedman et al., 1999, Gaser et al., 1999). Some recent studies have begun to pay attention to abnormal functional integration associated with the frontal, temporal, parietal and occipital regions in schizophrenia (Heckers et al., 2002, Burns et al., 2003). In particular, some studies on the resting-state functional network in schizophrenia have observed that functional disconnectivity in schizophrenia distributes widely throughout the entire brain rather than being restricted to a few specific brain regions, and have suggested that schizophrenia might arise from abnormalities in a distributed network of brain regions (Boksman et al., 2005, Liang et al., 2006, Welsh et al., 2008). This disconnection hypothesis of schizophrenia has also been supported by increasing evidence from structural neuroimaging data such as diffusion tensor imaging (DTI) and magnetic resonance imaging (MRI) (Loeber et al., 2001, Kubicki et al., 2005, Buchsbaum et al., 2006, Zhou et al., 2008).

Previous studies of discriminative analysis in schizophrenia focus mainly on univariable and group-level statistical methods, which are less helpful to diagnosis, due to complex and widespread dysfunction and disconnectivity in the entire brain in schizophrenia (Lawrie et al., 2002, Honey et al., 2005, Bluhm et al., 2007, Yoon et al., 2008). In the last few years, a growing number of studies have shown that machine learning is capable of extracting stable structural or functional patterns from neuroimaging data, and may potentially be useful for finding significant neuroimaging-based biomarkers (Pereira et al., 2009). For example, some linear classifiers based on linear dimensionality reduction technology, such as principal components analysis (PCA) (Pagani et al., 2009), independent components analysis (ICA) (Jafri et al., 2007), and the multivariate linear model (Kawasaki et al., 2007), have increasing applications in discriminative analysis of mental disorder patients. The common idea underlying these multivariable methods is to map a high-dimensional data space into a linear subspace spanned by some components named eigen-images or eigen-time series. Then, the reconstructed subspace models are used to discriminate or identify patients from healthy controls. However, these methods assume that the measured signal should be a linear combination of eigen-images or eigen-time series. It is argued that this assumption may lack physiological reasons and does not accord with the fact of the complex and intrinsic nonlinear neuro-dynamics of the brain. Another limitation of PCA and ICA is that both of them make a priori assumptions about the components with orthogonality or statistical independence. Such constraints may not be necessary and may have no physiological justification in discriminative analysis.

Recently, some machine learning-based classifiers such as support vector machine (SVM) have been developed to identify mental patients from healthy controls (LaConte et al., 2005, Fan et al., 2007). Derived from statistical learning theory, a SVM classifier aims at finding a hyperplane maximizing the margin between positive and negative samples while simultaneously minimizing misclassification errors in the training set. As a powerful tool for statistical pattern recognition, SVM has been used to analyze fMRI data with the goal of decoding the information represented in the subject's brain at a particular time and obtaining rather high classification accuracy (LaConte et al., 2005, Haynes and Rees, 2006, Kriegeskorte et al., 2006). In these studies, SVM uses labeled data to find statistical properties in the fMRI training data that discriminate between two or more brain states, and then obtains the mapping from a pattern of brain activity represented by fMRI to a subject's cognitive states.

In the present work, we purposefully focus on the discrimination problem of schizophrenia using resting-state fMRI data. Various linear or nonlinear models for extracting morphological patterns of the brain, have successfully classified patients from healthy controls with satisfactory classification accuracy (Kawasaki et al., 2007, Fan et al., 2007). Until now, however, little attention has been paid to classifying schizophrenia using resting-state functional connectivity patterns. Studies have suggested that a rest-based functional analysis can detect a more complete and more accurate connectivity map than a task-driven analysis (Xiong et al., 1999), and the mental activity occurring during rest is relevant to the phenomenology of schizophrenia (Malaspina et al., 2004). The advantages of using resting-state data also include the fact that resting-state neuroimaging is easy to perform without any complicated task design, and thus can be readily accepted by schizophrenic patients. Although resting-state neuroimaging studies have demonstrated that schizophrenia may arise from widespread improper functional integration (Liang et al., 2006), further studies about whether some special spatiotemporal patterns exist as potential biomarkers in these resting-state functional networks are still necessary. A fundamental and interesting problem is how we train an effective classifier to decode behavioral symptoms of disorder or other variables of interest from resting-state fMRI data.

Here, we develop a data-driven method integrating low-dimensional embedding and C-means clustering of fMRI to extract spatiotemporal patterns associated with schizophrenia from resting-state functional connectivity for classification. The main idea underlying this classifier is that resting-state functional networks of the entire brain of all subjects are treated as points distributed in a high-dimensional feature space, and moreover, the spatiotemporal patterns associated with schizophrenic symptoms are hypothesized to lie on a low-dimensional manifold embedded in the feature space. We are interested in learning the organization of these points in high dimensions and extracting useful information for data classification. A nonlinear manifold learning technique named locally linear embedding (LLE) (Roweis and Saul, 2000) is applied to obtaining a low-dimensional embedding of fMRI while preserving the intrinsic structure in the data. Manifold learning has gradually attracted attention recently due to its nonlinear nature, geometric intuition, and computational feasibility, and has already been used to detect activated voxels (Shen and Meyer, 2008) in fMRI. By assuming that the subset of the disorder-related resting-state functional network forms a low-dimensional structure, we partition the dataset and separate subsets of points with low dimensionality. An unsupervised classifier based on C-means clustering (Theodoridis and Koutroumbas, 2006) was used to decide the classification of subjects. We have also conducted several experiments with in-vivo datasets using different classifiers including SVM to demonstrate the performance of this approach.

The remainder of this paper is organized as follows: In the Materials section, we introduce the experimental data acquisition and preprocessing. In the Methodology section, we detail three important steps in our methodology: feature selection, dimensionality reduction based on LLE, and C-means clustering classification. The cross-validation strategy for estimating the performance of the classifier is also described. The experimental results on real datasets along with comparison with other methods are presented in the Results section. The discussions are in the Discussion section followed by the Conclusion section.