Fisher discriminative analysis of resting-state brain function for attention-deficit/hyperactivity disorder

Abstract

In this study, a resting-state fMRI based classifier, for the first time, was proposed and applied to discriminate children with attention-deficit/hyperactivity disorder (ADHD) from normal controls. On the basis of regional homogeneity (ReHo), a mapping of brain function at resting state, PCA-based Fisher discriminative analysis (PC-FDA) was trained to build a linear classifier. Permutation test was then conducted to identify the brain areas with the most significant contribution to the final discrimination. Experimental results showed a correct classification rate of 85% using a leave-one-out cross-validation. Moreover, some highly discriminative brain regions, like the prefrontal cortex and anterior cingulate cortex, well confirmed the previous findings on ADHD. Interestingly, some important but less reported regions such as the thalamus were also identified. We conclude that the classifier, using resting-state brain function as classification feature, has potential ability to improve current diagnosis and treatment evaluation of ADHD.

Introduction

Attention-deficit/hyperactivity disorder (ADHD) is one of the most commonly diagnosed childhood behavioral disorders which affected approximately 5% of school-age children and characterized by the symptoms of inappropriate inattention, impulsivity, and hyperactivity. Children with ADHD have difficulties in controlling their behaviors or focusing their attentions which result in an adverse effect on academic performance and social function. Moreover, 30%–60% of individuals diagnosed with ADHD in youth have symptoms that persist into adulthood (Biederman, 1998, Biederman et al., 2000). Currently available diagnosis and treatment evaluation of ADHD are mainly made according to the levels of the symptoms listed in the diagnostic criteria from DSM-IV (American Psychiatric Association, 1994). Ranking of the symptoms is usually conducted by the parents or teachers of the children, which is unfortunately subjective. Therefore more objective approaches are highly desired.

Structural and functional magnetic resonance imaging (MRI) techniques have been widely used in the quantitative analysis of the brain for ADHD, and various abnormalities have been reported as the objective evidences for some theoretical hypotheses on the disorder. Structural MRI studies have shown abnormalities of the whole brain and several specific brain areas, such as the frontal lobes, the basal ganglia, the parietal lobe, the occipital lobe, and the cerebellum in ADHD, in comparison to normal controls (Castellanos et al., 1996, Overmeyer et al., 2001, Sowell et al., 2003, Seidman et al., 2006). Using various experimental designs, task-related functional MRI (fMRI) studies found abnormal brain activation of ADHD in the dorsal anterior cingulate cortex (dACC), the ventrolateral prefrontal cortex (VLPFC), and the putamen (Bush et al., 1999, Durston et al., 2003, Teicher et al., 2000). Resting-state fMRI has also been used in the studies of ADHD and abnormalities were found in ACC, prefrontal cortex, putamen, temporal cortex, and cerebellum (Tian et al., 2006, Cao et al., 2006).

Although these studies have indicated that the pathophysiology of ADHD can be associated with the various brain regions, it has been argued that the analysis approaches based on the group-level statistics are less helpful to diagnosis (Seidman et al., 2004). Recently, increasing attention has been directed to the applications of pattern recognition techniques in brain image analysis. Compared with the traditional group-level analysis, such techniques can distinguish normal from abnormal at individual subject level. Hence they are potentially useful procedures for clinical diagnostic purposes. In such studies, various structural characteristics or functional properties of the brain derived from neuroimaging data were used as the feature for classification. For structural MRI, shape of brain structures of interest, deformation field for registration, map of gray matter membership, map of cortical thickness, and so on have been employed to discriminate patients (e.g., schizophrenia and Alzheimer's disease) from healthy controls (Golland et al., 2002, Fan et al., 2007, Kawasaki et al., 2007, Yoon et al., 2007). For task-related fMRI, the original time series and activation maps have been used for discrimination of mental disorders (Kontos et al., 2004, Shinkareva et al., 2006).

Though the promising studies on some psychiatric disorders were reported using classification techniques, few were conducted on ADHD. In addition, though the task-induced brain activities have been used as the classification feature, the brain activities revealed by the resting-state fMRI have not been considered for. In the resting state, low-frequency (0.08 Hz) fluctuations (LFF) of the fMRI signal are considered to be related to spontaneous neuronal activity and the synchrony of LFF was first used to identify functional connectivity among motor cortices (Biswal et al., 1995) and then extended to other functional systems, e.g., between bilateral visual cortices, bilateral auditory cortices, bilateral amygdala, bilateral thalamus, and within the language system (Lowe et al., 1998, Cordes et al., 2000, Stein et al., 2000, Hampson et al., 2002). Abnormal LFF has been reported for ADHD (Tian et al., 2006, Cao et al., 2006) and other neuropsychiatric disorders (Li et al., 2002, Greicius et al., 2004, Liu et al., 2006). These findings inspired us to use the brain activity revealed by the resting-state fMRI as a classification feature to differentiate boys with ADHD from their normal controls in this study.

Besides classification features, learning algorithm is also an important aspect of a classification system and has a great impact on the performance of a classifier. Fisher discriminative analysis (FDA), with the advantages of simplicity, sound theoretical foundation, and ease of interpretation, has been widely used in the domain of pattern recognition (Duda et al., 2001). The traditional FDA, however, cannot be used directly when the within-scatter matrix is singular in the case of small sample size. In order to solve the problem, principal component analysis (PCA) is usually first employed to reduce the dimension of feature space and FDA was then performed in the transformed feature space. Such a PC-FDA approach was first proposed for face recognition (Swets and Weng, 1996) and extended into the field of task-related fMRI data analysis (Mørch et al., 1997). Recently, PC-FDA and its variants (e.g., canonical variate analysis, CVA) have been widely used in the analysis of task-related fMRI data (Carlson et al., 2003, LaConte et al., 2003, Strother et al., 2004, Mourao-Miranda et al., 2005). Carlson et al. (2003) used PC-FDA to investigate patterns of activity in the categorical representation of objects. LaConte et al. (2003) applied CVA on PCA basis to evaluate the impact of preprocessing choices and the number of principal components passed to the CVA on within-subject prediction and reliability. Strother et al. (2004) combined PCA and CVA to optimize preprocessing choices in fMRI data analysis. Mourao-Miranda et al., 2005 applied FDA and support vector machines (SVM) on PCA components for classification of different brain cognitive states using the whole-brain fMRI data. In this study, the PC-FDA was further extended into the resting-state fMRI analysis for the discrimination of mental disorders and applied to ADHD. Our initial trials were presented elsewhere (Zhu et al., 2005).

The rest of the article is organized as follows: the resting-state classification feature, learning algorithm, classifier performance, and discriminative pattern are detailed in the Methodology section. Materials and experimental results are presented in the Materials and Results sections, respectively. The discussions are in the Discussion section followed by the Conclusion section.