Twenty rTLE patients (10 females, 10 males, age 26.35 ± 5.97) were recruited consecutively from Epilepsy Clinic, the First Affiliated Hospital of Guangxi Medical University according to the diagnostic manual of the International League Against Epilepsy classification (23). They were recruited from July 2015 to May 2016. All patients underwent standard clinical assessments, including a detailed seizure history, neurological examination, neuropsychological assessment, standard and video-EEG evaluation, and brain MRI. Particularly, all rTLE patients met at least two of the following criteria (24): (1) the typical symptoms of TLE indicated that the epileptogenic lesion was located in the temporal lobe; (2) MRI showed right hippocampus atrophy, sclerosis, or other abnormality of the right temporal lobe. All of the images were assessed by a neuroradiologist. (3) Electroencephalogram (EEG) revealed ictal or interictal discharges in the right temporal lobe, as evaluated by an epilepsy specialist. Our patients had taken regular antiepileptic drugs (AEDs) and had no epileptic seizures in the last 3 months. Additionally, to avoid any confounding effects on cognition, we excluded patients with a Mini-Mental State Examination (MMSE) score <24 as well as any history of neurological or psychiatric disorder other than TLE, traumatic brain injury, or other serious disease.

The control group consisted of 19 age-, gender-, and mean educational years-matched healthy volunteers (9 females, 10 males, age 26.47 ± 3.78). All had no history of neurological or psychiatric disorders. This research was approved by the Ethics Committee of the First Affiliated Hospital of Guangxi Medical University. All participants were right handed and provided signed informed consent prior to the study.

All MR images were performed on a 3-T Achieva MRI scanner (Philips, Netherlands) with a 12-channel phased array head coil. A 3D high-resolution structural image was acquired for each subject with a T1-weighted spin-echo sequence (TR/TE = 3,000/10 ms, slice thickness = 5 mm, slice gap = 1 mm) for spatial brain normalization.

Diffusion tensor imaging images were acquired using a single-shot echo-planar imaging-based sequence with the following parameters: TR/TE = 6,100/93 ms, flip angle = 90°; FOV = 240 mm × 240 mm, slice thickness = 2 mm and no gap, number of signals acquired = 4; data matrix = 256 × 256, flip angle = 90°, voxel size = 0.94 mm × 0.94 mm × 3 mm; resulting in a total of 30 volumes with diffusion gradients applied along 30 non-linear directions (b = 1,000 s/mm²) and 1 volume without diffusion weighting (b = 0 s/mm²). Each volume consisted of 45 contiguous axial slices.

The rsfMRI data were obtained using a gradient-echo echo-planar imaging sequence with parameters of: TR/TE = 2,000/30 ms, flip angle = 90°, FOV = 220 mm × 220 mm, data matrix = 64 × 64, slice thickness = 5 mm, slice gap = 1 mm, and voxel size = 3.44 mm × 3.44 mm × 6.00 mm; 31 slices and 180 volumes were acquired. All participants were instructed to lie still while resting with their eyes open and were forbidden to think of anything in particular.

The DTI data were preprocessed using PANDA¹ (25) in Matlab including the following steps: converting DICOM files into NIfTI images, estimating the brain mask, cropping the raw images, correcting for the eddy current effect, correcting for head motions, estimating the diffusion tensor models by using the linear least-squares fitting method on each voxel, tracking whole-brain fiber in the native diffusion space via Fiber Assignment by using the Continuous Tracking algorithm, and averaging multiple acquisitions and calculating diffusion tensor metrics.

The fMRI data were preprocessed using SPM8² and the GRETNA toolbox³ (26). The preprocessing steps included removal of volumes, slice timing correction, realignment, spatial normalization, and temporal filtering as follows. The first 10 volumes of each subject were removed to ensure magnetization equilibrium. The remaining volumes were then executed for slice timing correction based on the middle slice and then realigned for head motion correction. Two patients were excluded from further calculations due to head motion >2 mm or head rotations <2°. For group average and group comparison purposes, the data were spatially normalized to the standard Montreal Neurological Institute space and resampled with a resolution of 3 mm × 3 mm × 3 mm. Subsequently, signals were typically band-pass (0.01–0.08 Hz) filtered to reduce the effects of low-frequency drift and high-frequency physiological noise (27). Finally, confounding variables, including six head motion parameters, averaged global and white matter signals, and cerebrospinal fluid regressed out.

The nodes of the structural and FC networks were delimited according to an automated anatomical labeling (AAL) algorithm (AAL) algorithm (28). This algorithm scheme parcellated the entire cerebral cortex, except the cerebellum, into 90 anatomical regions (AAL-90), which resulted in 90 nodes covering the non-cerebellar brain and 45 nodes in each hemisphere.

Graph theoretical analyses of the weighted structural and functional networks of rTLE patients and controls were calculated with routines from the GRETNA toolbox. The network topological properties at the global levels were collected, including (1) properties that imply network segregation of brain, such as the weighted clustering coefficient (γ), local efficiency (E_loc), and modularity; (2) properties that indicate network integration of the brain, such as the characteristic path length (λ) and global efficiency (E_glob). E_glob is defined as the average inverse shortest path length; E_loc is defined as the mean of the global efficiencies of subgraphs consisting of the immediate neighbors of a particular node (30). (3) Small-worldness (σ) which evaluates the balance of segregation and integration.

Network topological properties rely on the density of network. So the difference of connectivity strength may affect networks comparisons. When brain graphs constructed, each graph of subjects was thresholded to create an equal number of nodes and edges across subjects (31). We operated network parameters over a range of threshold values to guarantee high correlation coefficients of the remaining connections. As in DTI, we used FA as the value for threshold. And in fMRI, we used the concept of sparsity to analyze the network. The sparsity was defined as a density range of 0.2–0.42, since these densities provide a reasonable trade-off between sparse, but not fully connected networks and highly linked networks, which do not show small-world properties any more (32, 33).

Each participants’ alertness was assessed by attention network test (ANT) (34) based on the E-Prime software platform. The ANT is a common neuropsychological examination that is used to test for attentional deficits and is composed of the Flanker task and cued response time (RTs) task. Alertness is one of the three basic components of attentional function. Through this examination, alertness can be detected by changing the cue prompt and recording both correct and incorrect reactions as well as the reaction time. The correct reaction and its RTs were applied to evaluate alertness according to the formula: mean RT_{no cue} − mean RT_{double cue}. Based on the classification of alertness (12), the no cue condition expresses intrinsic alertness and double cue condition represents phasic alertness. To reduce the effect of executive control function on the intrinsic and phasic alertness RTs, we excluded in-congruent trials during the calculation (35).

Statistical analyses were performed by using IBM SPSS statistics (version 22). A two-sample t-test was performed to analyze group differences in age, years of education, MMSE scores, and ANT scores of alertness between rTLE patients and controls. The Chi square test was employed to compare gender distributions between groups, and p-value less than 0.05 was considered statistically significant.

The graph measurements, such as γ, λ, σ, E_loc, and E_glob were analyzed with ANCOVA to detect differences between rTLE patients and controls over a wide range of threshold, with FDR correction. Age and gender were included in ANCOVA as nuisance covariates.

Pearson correlation analyses were performed to assess the correlations between alertness and graph theoretical measures of the structural and functional networks at each FA or the sparsity threshold value. Considering that the disease duration of epilepsy might be a confounding factor, we performed partial correlation analysis to remove the interference. As these analyses were exploratory in nature, we used a statistical significance level of p < 0.05, uncorrected.

The resultant group level structural and functional networks were displayed using Pajek.⁴

There were no significant difference in age, sex, and educational level between the rTLE group and normal control group. The demographic details of the participants are summarized in Table 1. rTLE patients were treated with drugs without surgery; both groups of subjects were right handed.

The graph measurements of the DTI network were computed over a series of thresholds on FA values (FA = 0.2~0.42) with a step of 0.02. Both rTLE patients and healthy controls showed a small-world organization (σ > 1, with λ close to 1 and γ higher than 1) (Figures 1A,B). Compared with healthy controls, rTLE patients had a higher clustering coefficient (γ) (Figure 2A, FA ≤ 0.28, p < 0.05) and the same characteristic path length (λ) (Figure 2B), which led to significantly elevated small-worldness (σ) (Figure 2C, FA ≤ 0.26 and FA = 0.30). Since σ reflects an optimal balance between fragmentation and coalescence, our results indicate that there is a disturbance in the normal balance of network function. Therefore, network construction was more inclined to regular network characteristics. In addition, rTLE showed unchanged mean local efficiency (E_loc) (Figure 2D) and global efficiency (E_glob) (Figure 2E).

For the rsfMRI datasets, the correlation matrix was thresholded into different sparsities to create the adjacency matrix. Since there is currently no definitive method of selecting a sole threshold, we calculated graph measures of the rsfMRI network over a series of thresholds in a wide range of network sparsity (5~40%) with a step of 1% (Figures 1C,D). Compared with healthy controls, rTLE patients had s lower γ (Figure 3A, sparsity <11%) and higher λ (Figure 3B, 5% < sparsity < 8%), which led to s significantly reduced σ (Figure 3C, sparsity <11%). As a result of these analyses, the FC network architecture tended to a more random organization. After controlling for the FC strength difference using the sparsity threshold, rTLE showed lower Eloc (Figure 3D, sparsity <9%), and lower Eglob (Figure 3E, 5% < sparsity < 8%).

We also analyzed the modularity of the structural and functional networks of the two groups, and there was no significant difference (Figures 2F and 3F).

Alertness did not differ significantly between the rTLE patients and controls. However, the RTs of no cue and double cue of rTLE patients were significantly longer than those of controls. Table 2 provides more details.

We analyzed the correlations between alertness and altered global topological parameters of the structural and functional networks in rTLE patients at each threshold value. For their structural networks, we found that σ and γ were significantly negatively correlated with alertness in the FA value range (0.34 < FA < 0.42) (Table 3). Similarly, after controlling for disease duration as a confounding variable, partial correlation analysis showed the same trend in the FA value range (0.36 < FA < 0.42) (Table 3). No correlation was found in Pearson correlation or partial correlation analyses of alertness with any other parameters.

For the functional networks, σ and γ were also found to be significantly negatively correlated with alertness over a range of sparsity thresholds (for σ, 0.28 < sparsity < 0.40; for γ, 0.26 < sparsity < 0.40; p < 0.05), as analyzed by the Pearson correlation and partial correlation analyses. E_loc had no correlation with the ANT scores by Pearson correlation analysis, but was negatively correlated with alertness at several different sparsity thresholds (sparsity = 0.17, 0.24, and 0.25; p < 0.05) when calculated by partial correlation analysis and controlling for the influence of the disease duration. However, no correlation was found between λ, E_glob and alertness when the above two methods were used (Table 4). In the control group, there was no association between ANT scores and any network parameter.

At this time, DTI is the only non-invasive technique that provides tissue microstructural information in vivo. The DTI-based structural linkage provides a relatively intuitive method of describing the true structural network between brain regions (36). Many neuroimaging studies have shown that TLE is associated with structural abnormalities in specific brain areas. Subsequent findings of white matter damage in or outside the temporal lobe were reported in TLE, including white matter damage in the external capsule, corpus callosum, cingulate gyrus, and hook beam, as well as the amount of the next pillow beam (37). These findings have ignited interest in white matter network structure studies in TLE that may identify experimental targets of pathophysiology mechanisms. In this study, we found that rTLE patients, compared to healthy controls, exhibited an altered topological organization of the white matter structure network and, based on DTI tractography, these alterations included highly increased clustering coefficients and small-worldness.

The clustering coefficient represents the small range of connections between adjacent brain areas, which is used to describe the capability to effectively interchange information and information reprocessing in a closed feedback loop and expresses the constitute forms of cliquishness within network (38, 39). A higher clustering coefficient often indicates modularized information processing and an enhancement of the local specialization and separation of network functionally (38). At present, the increased clustering coefficient and small-worldness could be interpreted as alterations of an efficiently organized network. Our results suggest that microstructure damage might occur in white matter structures because of recurrent, uncontrolled seizures. For instance, subtle alterations in white matter tract volumes were determined to be due to transneuronal degeneration in a diffuse underlying pathology, such as microdysgenesis (2). Aberrant local nerve fibers may be reconstructed as a compensatory mechanism in response to a decrease in long-range connections, such that the relatively high reentrant connectivity within the local clusters (“cliquishness”) increased. In this way, the efficiency of the entire brain network is preserved.

Unlike structural networks, we found significantly decreased clustering coefficients and enhanced shortest path lengths, as well as lower local and global efficiencies in brain functional networks. Our findings are consistent with those of Vlooswijk et al.’s (40). In studies of functional networks in TLE, the clustering coefficients have been widely reported to be either increased (3, 41, 42) or decreased (40, 43, 44). The reasons for these inconsistent findings are unclear, although it is important to account for the sample size, age, epilepsy phenotype, measurement of the connection form, and AED use. A lower clustering coefficient tends to indicate a weakening interconnection between local brain regions, suggesting a possible reduction in the separation function of brain processing information associated with a particular pathological condition. Some authors have observed that the clustering coefficient also changes during the progression of disease. Evidence from a graph theory study of neuron sclerosis in the dentate gyrus showed that the clustering coefficients increased during the sclerosis process and decreased at the final stage (45). In combination with the enhanced shortest path lengths, which indicated that the global integration of the brain was decreased, that is, information transmissions or interactions between remote brain regions were less effective and slow. Therefore, the global information progressing efficiency of the brain functional network of rTLE patients was significantly diminished. Epileptic seizures are due to the synchronous excitability of abnormal neurons in the brain, such as a high synchronization between thalamus and remote cortical regions (46), and increased EEG connection between relevant brain area to the epileptic foci (7), etc. Changes of the above properties may be related to enhanced neural synchronization phenomena. Consequently, high synchronization within a group of neurons in the brain often leads to a decline in global brain function.

The inconsistency of the alteration of the network topology characteristics between structural and functional networks may result from the following reasons. In general, functional networks are considered to be more resilient, while structured networks are considered to be relatively stable (47). We presumed that functional networks are probably more sensitive than structural networks and undergo dynamic and architecture changes at earlier stages of the disease. Yet, structural networks are affected during later stages of the disease. White matter organization is rebuilt to ensure proper brain function, and this reorganization is a compensatory mechanism of brain plasticity. It is also possible that reconstructions of the structure network are under restrictions (48). The compensatory ability of functional networks is limited by the white matter axon plexus structure, and because of these limits, the network efficiency decreases.

Mental and behavioral disorders caused by TLE have recently become a prominent research topic. Alertness is one of the three sub-networks of attention, but it is often overlooked and rarely reported. Alertness consists of two components: intrinsic alertness and phasic alertness. Intrinsic alertness is associated with the body’s internal alert and arousal states. It can respond to the target stimulus without external stimulate signals. Phasic alertness represents a short-term improvement in response to external stimuli. Both of these are measured by the RTs in the ANT test. The alertness that we obtained is the difference between the RTs of two parts (49). Although no significant differences in alertness were found between rTLE patients and the controls, the RTs of no cue and double cue of patients were longer than those of controls. Furthermore, one of our other researchers discovered longer RT_no-cue, RT_double-cue and lower alertness as well (50). Together, these data demonstrate a potential alertness impairment in TLE patients.

According to the association analysis, there was a negative correlation between alertness and the clustering coefficient and the small-worldness property, both in white matter structural networks and in functional networks. For structural networks, these data suggest that the potential decline in alertness might be related to the increase of these two properties. It is likely that repeated seizures injured the microstructure of white matter fibers. Consequently, white matter structure remodeling occurred as a consequence of brain plasticity. Thus, structural networks enhance localized processing to maintain effective interactions in information flow. Similarly, Bonilha et al.’s study indicated that neuronal loss caused by seizures theoretically leads to reorganization of the limbic system (3). Therefore, limbic system reorganization could lead to neuropsychological abnormalities in TLE. As to the correlation in functional networks described above, we presumed that alertness is associated with the degree of global integration of brain functional networks. Further, it means that alertness may be a function for which multiple brain regions process information in parallel and concurrently, rather than as a single group of neurons or a single neuronal circuit. As part of the compensation mechanism of functional networks, reduced local specificity and enhanced functional collaborations maintain the patient’s alertness level. Therefore, differences in clinical behavioral deficits would not be obvious.

In addition, taking into account the possible effects of the disease duration of epilepsy, we used partial correlation analysis to correct for the impact of the length of disease. Most of the results were consistent. However, a negative correlation between alertness and the local efficiency of functional networks was observed, although the negative correlation only occurred over a narrow range of sparsity thresholds. Combined with the non-correlation between alertness and the λ and E_glob of functional networks, the possible explanation would be that, although the network efficiency decreased, both the decline in functional separation and increase in global integration maintained alertness. Therefore, the rTLE patients in our study had no temporarily clinically detected significant alertness deterioration. The disease duration has little involvement in the relationship between alertness and network topology.

In this study, there are several limitations that are worth noting. Graphical analysis of complex neural networks at the macro-scale level is a fast-growing area of research, but there is still controversy regarding the optimal analytic strategies. Node definition and selection are very important issues in building structural or functional networks (51). In addition, the cross-sectional design of this study limited our ability to expose the mechanism of the causal relationship between the network analysis results and clinical behavioral testing. By choosing nodes according to different brain anatomical templates in further studies, we will be able to find more complex information processing mechanisms of the brain.