Decoding brain states from fMRI connectivity graphs

Abstract

Functional connectivity analysis of fMRI data can reveal synchronised activity between anatomically distinct brain regions. Here, we extract the characteristic connectivity signatures of different brain states to perform classification, allowing us to decode the different states based on the functional connectivity patterns. Our approach is based on polythetic decision trees, which combine powerful discriminative ability with interpretability of results. We also propose to use ensemble of classifiers within specific frequency subbands, and show that they bring systematic improvement in classification accuracy. Exploiting multi-band classification of connectivity graphs is also proposed, and we explain theoretical reasons why the technique could bring further improvement in classification performance. The choice of decision trees as classifier is shown to provide a practical way to identify a subset of connections that distinguishes best between the conditions, permitting the extraction of very compact representations for differences between brain states, which we call discriminative graphs. Our experimental results based on strict train/test separation at all stages of processing show that the method is applicable to inter-subject brain decoding with relatively low error rates for the task considered.

Introduction

Traditional fMRI analysis consists of univariate statistical hypothesis testing to assess changes in the activity of each brain voxel induced by the stimulation paradigm (Frackowiak et al., 1997). More recently, approaches derived from supervised machine learning – commonly termed “brain decoding” in the field of neuroimaging – have shown that it is possible to exploit more subtle relationships in voxels' intensity patterns (Haxby et al., 2001, Haynes & Rees, 2006, Norman et al., 2006). These methods rely on a classifier to predict the subject's brain state from the BOLD responses in a set of selected voxels, such as visual (Cox & Savoy, 2003, Haynes & Rees, 2005, Kamitani & Tong, 2005, Thirion et al., 2006, Kay et al., 2008, Miyawaki et al., 2008) or auditory cortices (Ethofer et al., 2009). Results from brain decoding are often remarkable since they clearly reach beyond the possibilities of univariate techniques, but also because they are able to uncover information from fine-grained cortical activity despite the relatively low spatial resolution of fMRI. Furthermore, instead of a fixed spatial window, one can also apply classification to a so-called “searchlight” that slides over the whole-brain data, such that the classification success for each position of the spotlight can then be mapped to show brain regions that carry discriminative information between different conditions (Kriegeskorte et al., 2006). Another interesting approach is to use spatiotemporal observations as an input to the classifier (Mitchell et al., 2004, Mourao-Miranda et al., 2007).

The study of functional connectivity is concerned with the temporal coherence between neurophysiological events observed in spatially remote brain regions. In early work, correlation with a seed voxel was investigated and revealed bilateral coactivation between sensory cortices (Biswal et al., 1995, Lowe et al., 1998). Further advances have been driven by unsupervised methods such as source separation – mainly principal components (Friston et al., 1993) and independent components analysis (McKeown et al., 1998, Calhoun et al., 2002, Beckmann & Smith, 2004) which allow to identify large-scale cortical networks – and by other methods such as dynamic causal modelling, which tries to establish effective connectivity and requires prior information about the neurological network to investigate Friston et al. 2003. A recent method related to our work proposed to use resting-state correlations between regions of interest as features for an SVM classifier (Craddock et al., 2009). Another attractive methodology to investigate functional networks is to rely on mathematical graph theory; i.e., constructing the (undirected) graph from temporal correlation matrices and computing related measures; e.g., node degree, hubbiness, and so on (Sporns et al., 2000, Salvador et al., 2005). This methodology has brought new insights in functional connectivity at resting state, such as the small-world organisation of cortical networks at low temporal frequencies (Achard et al., 2006).

Here, we bring together brain decoding and graph representations based on functional connectivity measures. These measures, such as temporal correlation, are performed over a given period of time and reflect the timecourse resemblances between different regions. Moreover, we estimate connectivity at different temporal scales using the wavelet transform as a preprocessing step. Then, we build a classifier trained on functional connectivity graphs of a group of subjects to distinguish between different brain states of an unseen subject. The aim of our approach is to identify the connections that are most discriminative between brain states, and to obtain relevant visual representation of the data for neuroscience studies.

Previous brain decoding techniques primarily rely on linear support vector machines (SVMs), which use a soft-margin hyperplane to separate classes. In the present work, we propose polythetic decision trees that fit a hyperplane using the most discriminative features (i.e., connections) at each level, such that potentially complex and non-linear class boundaries can be obtained by multilevel trees. Efficient and effective learning of this type of decision trees relies on recent advances in pattern recognition (Friedman et al., 2000, Gama, 2004, Landwehr et al., 2005) and provides embedded feature selection which yields a compact discriminant function whose parameters are amenable to interpretation. Their variance properties make them good candidates for ensembling, steering the classification strategy towards slightly weaker but simpler (lower capacity) classifiers, which is a desirable behaviour in high-dimensional learning, a classical situation in fMRI where the number of dimensions is much higher than the number of training examples.

Our paper is organised as follows. In the Methods section, we describe our data processing pipeline together with the details of the proposed methodology. Next, we illustrate the feasibility of our approach by a proof-of-concept; i.e., an fMRI experiment with block-based stimulation paradigm (watching short movies) with long resting periods. We extract the set of connections that is the most discriminative between rest and stimulation at different temporal scales. The method is able to correctly classify the conditions in a leave-one-subject-out cross-validation setting. Interestingly, we find that the low-frequency correlations of the BOLD signal (below 0.11 Hz) are the most informative. The discriminative network also confirms a differential modulation of sensory areas, in particular within the visual system, and midline brain areas during movie and rest conditions.