Cognitive impairment is increasingly recognized as a critical component in schizophrenia and therefore an important treatment target (Huffaker et al., 2009; Dickinson et al., 2010a; Wykes, 2010). While a variety of cognitive remediation strategies have been developed, it is unclear which, if any, is effective (McGurk et al., 2009; Dickinson et al., 2010b; Keefe et al., 2010). One particularly promising strategy in the context of schizophrenia begins with training on simple auditory discrimination tasks and builds from this training to improve higher order verbal skills (Adcock et al., 2009; Fisher et al., 2009). A rationale for this approach is that patients with schizophrenia exhibit deficits in early stages of auditory processing that weaken the foundation for more complex cognitive processing (Javitt, 2009). However, these deficits may be ameliorated by practice (Fisher et al., 2009).

How practice influences brain function is as yet unclear. In healthy subjects, practice and improved performance on an auditory discrimination task correlate with an increase in the amplitude of auditory evoked responses (van Wassenhove and Nagarajan, 2007; Tong et al., 2009) likely due to synaptic plasticity (Buonomano and Merzenich, 1998; Tallal, 2004). However, the specific functional brain networks involved in auditory learning have yet to be fully elucidated and it is unclear how they are modified by practice. Current evidence from PET (Zatorre et al., 1994), fMRI (Ohnishi et al., 2001; Gaab et al., 2003, 2006; Temple et al., 2003; Warren and Griffiths, 2003; Rinne et al., 2009; Schulze et al., 2009), and magnetoencephalography (MEG) studies (van Wassenhove and Nagarajan, 2007) suggest that diverse temporo-parietal regions, as well as the left inferior frontal gyrus, and dorsal–lateral prefrontal cortex (DLPFC) contribute to auditory discrimination and working memory.

To characterize the functional brain networks involved in performance and practice of an auditory cognitive remediation task, we used MEG and measured mutual information (MI), a non-linear measure of coherence (David et al., 2004), between signals in the delta through high-gamma frequency bands between synthetic planar sensors. Coherence in fields recorded either intra- or extra-cranially is thought to be important for communication between brain areas (Bullock et al., 1995) and for cognitive performance (Bressler et al., 1993; Varela et al., 2001; Buschman and Miller, 2007; Cavanagh et al., 2009). Furthermore, coherence measurements provide information about the structure and plasticity of brain networks (Wang et al., 2010), that can be described by graph theoretical measures (Bullmore and Sporns, 2009). One such measure – cost efficiency, which quantifies the balance between network efficiency and the number of edges utilized in a graph (Figure 1) – is unique in that it reaches a maximum at a particular cost, thereby providing a non-arbitrary threshold for constructing unweighted graphs (Achard and Bullmore, 2007). This measure is particularly appropriate in our study because previous work has indicated that cost efficiency is positively correlated with performance in the visual N-back task in both controls and people with schizophrenia (Bassett et al., 2009).

In this work, we sought to identify the temporal and spatial characteristics of networks involved in the performance and practice of an auditory pitch discrimination task (Figure 2A) in samples of healthy individuals and patients with schizophrenia. Specifically, we tested whether accuracy in the task at baseline, after practice, and the difference (Δ) between the two, correlated with changes in univariate (signal power), bivariate (MI), and multivariate (cost efficiency) measurements of brain activity in delta through high-gamma bands estimated from MEG data acquired both before and after task practice. Signal power was measured in both sensor space and source space using sliding window synthetic aperture magnetometry (SAM; see Materials and Methods). Comparisons between the two recording sessions were used to determine practice-related changes in functional brain network organization.

We used MEG to investigate whether practicing a computer-based auditory cognitive remediation task would influence functional brain networks operating in the low delta through high-gamma frequency bands in healthy volunteers and patients with schizophrenia. We hypothesized that cost efficiency, a graph theoretical measure based on MI between MEG sensors, would correlate with task accuracy and thereby characterize the functional network involved in pitch discrimination before and after practice. Indeed, cost efficiency of the beta-band network was negatively correlated with task accuracy for probands and controls before practice, and a similar trend was also found among control subjects when the experiment was repeated following practice. Since our sample size was small and we did not observe task performance differences between probands and controls, we did not focus our analysis on characterizing differences between groups. When we compared the before and after practice data sets for all subjects, we found that improved accuracy after practice correlated with increased MI in the beta-band between sensors in temporo-parietal regions. Also, among control subjects improved accuracy correlated with decreased local cost efficiency in the beta band. Therefore, our results support the notion that practice-induced improvements in accuracy in pitch discrimination correlate with changes in the MI between brain signals recorded by temporal–parietal sensors in the beta band, but not changes in global network structure.

To further understand the significance of functional brain network structure, we supplemented our MI-based analysis with SAM analysis, which provides greater spatial accuracy. SAM analysis of beta power showed increases in brain regions important in auditory processing (Haenschel et al., 2000), and decreases in region important for sensorimotor integration (Engel and Fries, 2010). One possible interpretation of these findings is that network structure, which correlates with accuracy, regulates the integration of segregated beta-band mediated processes in different brain areas (Bassett et al., 2009). It is important to be aware in utilizing this methodological analysis, that a field produced by a single source spreads to multiple sensors, and therefore it is possible for coherence (i.e., MI data) to be contaminated by artifact. Yet, artifact is unlikely to completely explain the reported changes in connectivity that correlate with accuracy because sources associated with field spread may be constant across variations in accuracy (Schoffelen and Gross, 2009).

A role for the beta band in a pitch discrimination task is not surprising in light of the diverse functional roles of the beta band, including maintaining the status quo for sensorimotor integration (Engel and Fries, 2010), top-down attentional processing (Buschman and Miller, 2007), and working memory processing (Bassett et al., 2009; Palva et al., 2010). Increased working memory load has been shown to increase clustering of functional networks in the beta band (Palva et al., 2010), and performance on the visual N-back task was shown to correlate with increased global cost efficiency (Bassett et al., 2009), potentially due to the global network maintaining a more metabolically efficient state. We found that pitch discrimination task accuracy negatively correlated with beta-band cost efficiency for the combined subjects before practice. The direction of this relationship is in contrast to Bassett et al. (2009), who reported that increased cost efficiency correlated with optimal performance on the N-back task. However, task demands in the two studies differed dramatically. First, N-back engages visuo-spatial working memory systems while the auditory discrimination task engages the phonological system (Baddeley, 2003). Second, the N-back paradigm requires maintenance of information and resistance to distraction across items, while each auditory discrimination task item is discrete.

Our results suggest that prior to practice, pitch discrimination accuracy is optimized with dense short-range connections within bilateral temporo-parietal regions associated with auditory processing, but not between discrete components of the complete network. These short-range connections may be based on neuronal synchrony originating in local cytoarchitecture and monosynaptic circuits (Varela et al., 2001). Such circuits may be at least in part localized to the planum temporale and execute computations critical to pitch discrimination and memory (Griffiths and Warren, 2002). Also, complementary long-range connections between the left frontal gyrus and frontal–temporal regions correlated negatively with performance in the beta band. Importantly, such long-range connections likely rely on cortico–cortico and cortico-thalamic polysynaptic pathways, which may become less active when phonological rehearsal (Baddeley, 2003; Schulze et al., 2009) becomes less essential for task performance.

While our MI-based analysis of pitch discrimination is unique, SAM may prove advantageous in future studies as its spatial accuracy approaches that of fMRI (Rinne et al., 2009), yet its temporal resolution is much greater. Furthermore by combining sensor-space-based correlations between power and task accuracy with SAM analysis we can speculate that before practice there is a positive correlation between accuracy and beta-band power in sensors closest to the left sensorimotor cortex. One possible explanation for this correlation is that accuracy is related to response uncertainty, which in a prior study that utilized a visually cued reaching task was found to correlate with the magnitude of beta-band desynchronization (Tzagarakis et al., 2010).

Also, in the high-gamma band we found that power changes closely correlated with the informational demands of the task (Jerbi et al., 2009; Hinkley et al., 2010a; Uhlhaas et al., 2010), for example the power increased in auditory regions following each stimulus (Edwards et al., 2005; Steinschneider et al., 2008; Griffiths et al., 2010), and in the left motor cortex during the button press (Crone et al., 1998). However, in the case of the power increase in the auditory cortex, it is unclear if it is exclusively event related. Furthermore, the high correlation between pre-practice task performance and high-gamma activity in a region encompassing the right motor cortex, as well as the middle, and the inferior frontal gyri suggests that these regions may play an integrative role in pitch processing.

After practice we observed a scarcity of correlations between accuracy and parameters. We did not identify a correlation between accuracy and cost efficiency across any bands. One possible reason for the marked difference with respect to the pre-practice dataset may be that a majority of the practice-related reconfigurations of the functional brain networks were independent of task performance. Our methodology was not able to directly detect such reconfigurations. Another potential reason for a loss of correlations may be that a number of subjects had near-perfect performance across all blocks of trials after practice. However, in control subjects, we did still observe a trend between accuracy and beta-band cost efficiency. In addition, the MI network structure that was correlated with post-practice accuracy in control subjects exhibited hubs in temporo-parietal regions similar to the pre-practice dataset.

With regards to changes in correlations between accuracy and power after practice among control subjects, correlations with beta occurred in a similar spatial distribution in sensor space as in the pre-practice dataset. However correlations in sensors overlying temporal regions were increasingly evident after practice.

In comparing the post- and pre-practice data sets, broadband power was increased in the left angular gyri. This region has been previously implicated in auditory memory in pitch processing (Rinne et al., 2009; Schulze et al., 2009). It has been previously reported that the auditory evoked potential magnitude increases following hours of practice (van Wassenhove and Nagarajan, 2007; Tong et al., 2009). However, our erSAM analysis indicates that in our experiment, the auditory evoked potential did not involve the angular gyrus and did not change following practice. One possible reason we did not observe an increase in the evoked potential in our study is that the delay between practice and testing was variable and in some cases substantial. In sensor space we found that increased power in gamma- and beta-bands in left temporo-parietal sensors correlated with improved performance suggesting that increased power in the angular gyrus may be behaviorally relevant to practice and improvement.

Improved performance after practice correlated with increased connectivity in the beta-band between sensors primarily in temporo-parietal regions. We did not observe a correlation between improved performance and decreased global beta-band cost efficiency. Perhaps more extensive practice would have extended functional reconfigurations beyond the local networks, to the global network architecture.

In prior studies, auditory (Penolazzi et al., 2010) and somatosensory (Liu and Ioannides, 2004) practice have been previously shown to produce changes in power in the beta and gamma band. Also, increased beta coherence was found to be associated with cognitive recovery in acquired brain injury (Castellanos et al., 2010). It is reasonable to hypothesize that practice-induced synaptic plasticity underlies these meso-scale changes. Further work will seek additional evidence that changes in power and coherence in the beta and gamma band are associated with improved performance after practice and could provide a foundation for the development of cognitive remediation strategies that specifically target and counteract defined behavioral and neurophysiological deficits (Edwards et al., 2010; Hinkley et al., 2010b). However, in the present work, we did not identify behavioral deficits in our probands.

Analyzing MEG data across dimensions by examining power in sensor, and source space, as well as MI between sensors, and graph metrics effectively demonstrated how global and local networks contribute to accurate auditory discrimination. With the use of SAM, we were able to obtain an improved spatial and temporal understanding of this network. A possible future application of our integrative approach is identifying differences in functional brain networks during auditory task performance among large samples of healthy volunteers and patients with schizophrenia. With regards to this study, our data lends additional evidence to the notion that coherence-based graph metrics such as cost efficiency are relevant to behavior (Bassett et al., 2009), and that local functional brain networks based on MI between brain regions might adapt with practice.

Ten healthy volunteers and seven people with schizophrenia (diagnosed according to the Diagnostic and Statistical Manual of Mental Disorders IV criteria) were recruited as part of the Clinical Brain Disorders Branch/National Institute of Mental Health Genetic Study of Schizophrenia. Healthy volunteers had no structural MRI abnormalities or history of psychiatric illness, depression, or loss of consciousness. All patients were receiving antipsychotic drugs and other medication at the time of the study; none of the healthy volunteers were taking psychoactive medication. Three patients had changes in medication during the study. One patient was prescribed lithium, another patient was prescribed a reduced clozapine dose, and the third patient was prescribed a reduced venlafaxine dose. The mean WAIS estimated Full Scale IQ was 81.4 ± 3.82 for the probands (n = 5), and 108.43 ± 2.89 for the control subjects (n = 7).

The participants gave informed consent in writing, pursuant to protocols approved by the National Institute of Mental Health Institutional Review Board. We included patients specifically because we hoped to investigate cognitive remediation in this target population. MEG data were acquired at the National Institute of Mental Health using a 275-channel CTF MEG system (VSM MedTech) with a sampling rate of 600 Hz and third order gradient active noise cancelation.

The experimental paradigm was derived from the “High or Low” task, a component of the Brain Fitness Program (Posit Science, San Francisco, CA, USA), consisting of two 125 ms frequency modulated tones with a base frequency of approximately 500, 1000, or 2000 Hz, with a peak intensity of 87–93 dB that either increased or decreased in frequency (Fisher et al., 2009). The ISI between the stimuli was varied by experimental block. Following stimulus presentation the subject had 4 s to make two selections using a hand held controller with two buttons corresponding to increasing and decreasing frequency, respectively. A response was only considered correct if the subject classified both stimuli. Nine blocks of trials were performed for each subject. Each block consisted of stimuli with a consistent base frequency. For example, blocks 1–3 consisted of stimuli with base frequencies of 500, 1000, and 2000 Hz, respectively, and this was also the case for blocks 4–6, and 7–9. Furthermore, blocks 1–3, 4–6, and 7–9 had ISIs of 800, 400, and 200 ms, respectively. The purpose of altering the stimuli and ISIs by block was to match the stimulus set each subject would encounter during practice. Each block consisted of nine trials. Across the nine blocks, which varied with respect to the ISI and the frequency of the stimuli, we found no differences in accuracy (1D ANOVA, F = 0.280, p = 0.972, df = 8). Measurements were therefore pooled across blocks for subsequent analysis.

Following the first MEG recording, patients and controls were instructed on how to practice the “High or Low” task using the Brain Fitness Program. The sounds in the program were identical in frequency to those used during the MEG recording but in some cases differed in duration. The “High or Low” task uses an algorithm in which task difficulty is adaptively modulated following an initial diagnostic session and then on the basis of the subject's performance. The program provides arbitrary progress bars to indicate improvements through practice: more progress bars indicated a shorter duration of both the stimuli and the ISIs. Inpatients were supervised and monitored during practice sessions lasting for 30 min to 1-h. In contrast, volunteers practiced the task unsupervised and self-reported both practice time and progress. All subjects were asked to train for four to six consecutive days between MEG recording sessions. Six of the 7 patients completed practice, whereas only 7 out of 10 control subjects completed practice.

For coherence and graph theoretical analysis, eye blinks were removed from the raw data recorded by an automated process involving demarcation using a threshold detection function, followed by a principal components analysis based signal space separation. The data was then filtered to attenuate background low-frequency artifact and line noise at 60 Hz by using a 0.3-Hz-width filter, a low-pass filter with a cutoff of 150 Hz, and a high pass filter with a cutoff of 0.15 Hz. The MEG data preprocessing was performed using CTF (VSM MedTech) and FieldTrip software (F. C. Donders Centre for Cognitive Neuroimaging, Nijmegen, The Netherlands). To improve localization specificity, data was transformed into planar space by using FieldTrip software. The time series were then resampled to 120 Hz, and truncated to 34.133 s to constrain the frequency bands of the wavelet transform to conform to the classical EEG frequency bands of interest. A discrete wavelet transform using the Daubechies (4) wavelet was applied to each time series, with further analysis being performed only on wavelet coefficients at levels 1–6, corresponding approximately to classical EEG frequency bands from gamma to low delta (see Table S1 in Supplementary Material). For the high-gamma band, the raw time series was passed through a low-pass filter with a cutoff of 150 Hz before applying the planar transform and was not filtered further. Next, we estimated the MI of the wavelet coefficients of each pair of sensors at each frequency interval in each of the nine blocks of trials and normalized the pairwise MI according to Strehl and Joydeep, 2002. This resulted in a set of normalized MI (association) matrices representing the frequency-band specific functional connectivities between all pairs of sensors for each subject and recording block. The average MI of a network was defined as the average MI over all possible pairs of sensors, N. By applying a range of binary thresholds, each MI association matrix was then converted to a series of adjacency matrices, or graphs, G, with variable connection density or cost, C, (Bassett et al., 2006, 2009; Deuker et al., 2009)  where:

and N is the number of nodes. For each cost interval of size 0.01 in the range 0.01 < C < 1.0, the regional efficiency E(i) was computed for each node in the graph.

where N indicates the index region, j ≠ i denotes a node connected to i, and L_i,j is the minimum path length between nodes i and j. In other words, regional efficiency is inversely related to minimum path length: a region with high efficiency will have short minimum path length to all other nodes in the graph. The regional cost efficiency was calculated as the maximum of the function [E(i) − k], where k is the degree or number of edges connecting the ith region to the rest of the network (Latora and Marchiori, 2001; Achard and Bullmore, 2007). The global cost efficiency was calculated as the mean regional cost efficiency across all nodes.

Small worldness σ was calculated as (σ = γ/λ) where γ = E(i)/E(i)_random and λ = E(G)/E(G)_random, where the global efficiency E(G) (Latora and Marchiori, 2001) was computed for each graph as:

For the computation of E(i)_random and E(G)_random, we calculated 100 random networks with the same degree distribution as the real cortical networks using the Brain Connectivity Toolbox (Rubinov and Sporns, 2010).

All statistical comparisons and computations were performed in Matlab (MathWorks Inc., www.mathworks.com/). Statistical tests performed on all sensors and on the MI association matrix (Hemmelmann et al., 2005) were reported for three levels of stringency: FDR as defined by Benjamini and Hochberg (1995), a more stringent false positive correction p < 1/N, and the Holm–Bonferroni correction.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The Movies 1–8 for this article can be found online at http://www.frontiersin.org/human_neuroscience/10.3389/fnhum.2011.00081/abstract/