Localized activity alternations in periventricular nodular heterotopia‐related epilepsy

Abstract Objective Periventricular nodular heterotopia (PNH) is a common type of heterotopia usually characterized by epilepsy. Previous studies have identified alterations in structural and functional connectivity related to this disorder, but its local functional neural basis has received less attention. The purpose of this study was to combine univariate analysis and a Gaussian process classifier (GPC) to assess local activity and further explore neuropathological mechanisms in PNH‐related epilepsy. Methods We used a 3.0‐T scanner to acquire resting‐state data and measure local regional homogeneity (ReHo) alterations in 38 patients with PNH‐related epilepsy and 38 healthy controls (HCs). We first assessed ReHo alterations by comparing the PNH group to the HC group using traditional univariate analysis. Next, we applied a GPC to explore whether ReHo could be used to differentiate PNH patients from healthy patients at an individual level. Results Compared to HCs, PNH‐related epilepsy patients exhibited lower ReHo in the left insula extending to the putamen as well as in the subgenual anterior cingulate cortex (sgACC) extending to the orbitofrontal cortex (OFC) [p < 0.05, family‐wise error corrected]. Both of these regions were also correlated with epilepsy duration. Furthermore, the ReHo GPC classification yielded a 76.32% accuracy (sensitivity = 71.05% and specificity = 81.58%) with p < 0.001 after permutation testing. Interpretation Using the resting‐state approach, we identified localized activity alterations in the left insula extending to the putamen and the sgACC extending to the OFC, providing pathophysiological evidence of PNH. These local connectivity patterns may provide a means to differentiate PNH patients from HCs.

visible, while the majority of sporadic cases with unilateral or single nodules do not have Filamin A mutations. 4 Current clinical PNH diagnoses are mostly based on radiologists' interpretations. Although most PNH can be detected by neuroradiologists using moderate-to-high-quality magnetic resonance imaging (MRI), the process is time-consuming and requires considerable expertise. Especially in resource-limited regions, there are still some problems with PNH diagnostic report as the neuroimaging techniques and personal experience of radiologists are key points to diagnose PNH. Machine learning has become increasingly popular over the past decade and has served as a supplementary diagnostic tool for diseases such as glioma 5 and malignant lung nodule, 6 as well as PNH and MCDs. 7 Furthermore, multivariate pattern analysis (MVPA) has been used to develop brain signatures for clinical diagnoses that are more effective than traditional linear models. 8 A machine learning approach to PNH diagnosis could potentially accelerate and improve conventional neuroradiological interpretation.
The neuroanatomical alterations related to PNH have been well investigated. 9 However, the functional neural correlates of PNH remain elusive, though previous work demonstrated brain abnormalities in PNH using a seed-based functional connectivity approach. 10,11 Local functional analysis could provide an opportunity to discover localized functional disruptions in patient groups without a priori constraints, thus enabling the discovery of previously unconsidered regional brain abnormalities. Regional homogeneity (ReHo) is a voxel-based physiological metric that reflects local signal similarity and is derived from a blood oxygen level-dependent (BOLD) functional MRI (fMRI) time course. Recently, ReHo has been shown to be a valid metabolic proxy. 12 Since brain functional metabolism is significant to both epilepsy diagnosis and epileptogenic foci localization, local functional analyses may further our knowledge of PNH-related epilepsy's underlying neuropathological mechanisms and may assist in clinical outcome predictions.
Here, we combined both univariate analysis and a Gaussian process classifier (GPC) approach to explore ReHo in PNH. A training stage and a testing stage were both included in our analysis comparing PNH patients and healthy controls (HCs). We hypothesized that the GPC approach to the ReHo maps would (1) be able to discriminate individual patients with PNH from HCs and (2) provide information on neurobiological changes that could help to elucidate the mechanisms that cause PNH.

| Participants
Thirty-eight PNH-related epilepsy patients were included from neurology department, West China Hospital, Sichuan University. All methods were complying with guidelines and regulations. All subjects provided their informed consent before being enrolled, and this current research was approved by the local ethical committee of West China Hospital, Sichuan University. All of them presented with seizures based on the definition of International League Against Epilepsy. 13 Thirty-eight healthy subjects were also enrolled through advertisement in local region. They were age-and sex-matched. The clinical information was collected through standard questionnaire.

| MRI data acquisition and preprocessing
3.0-T MRI scanner (Siemens Trio) was applied to collect fMRI data. We adopted the DPABI toolbox 14 to preprocess the resting-state fMRI data. Specifically, the first ten time points were excluded in order to confirm the stability of the MRI signals. The remaining 190 volumes were used to perform the slice timing and head motion correction. We used the Friston 24-parameter model for head motion correction in the current study, because previous publication had reported that the Friston 24-parameter model is superior to the traditional 6-parameter model. 15 The translational and rotational parameters for all of the image data did not exceed ±1.5 mm and ± 1.5 degree, while the mean framewise displacement (FD) value did not exceed 0.2. Afterward, we spatially normalized these image data to the EPI template with the resolution of 3 × 3 × 3 mm 3 . The covariates including the signal from the cerebrospinal fluid, white matter, and global mean signal intensity were regressed out. Finally, we removed the linear trend of the fMRI images and performed the band-pass filtering (0.01-0.08 Hz).

| Regional Homogeneity Calculation
We used the REST software (http://resti ng-fmri.sourc eforge.net) to conduct the ReHo analysis. 16 Specifically, we calculated the Kendall coefficient of concordance (KCC) value of the time series between one single voxel with adjacent 26 voxels of its neighbors for creating the ReHo map of each participant. Afterward, a whole-brain mask was adopted to remove all the non-brain tissues for standardization.

| Univariate Analysis
We used a standard, univariate approach to investigate ReHo alterations between patients with PNH and controls by means of twosample t-tests in SPM12 software. The results were thresholded at p < 0.001 uncorrected at the voxel level, and a minimum cluster extent of 100 and the statistical threshold of cluster level was set at p < 0.05 using family-wise error (FWE) correction. Brain areas with significant ReHo alterations between two groups were extracted as region of interest for Pearson's correlation analyses with clinical characteristics.

| Multivariate Pattern Analysis Approach
We adopted GPC approach to discriminate patients with PNH from controls based on the whole-brain individual ReHo maps obtained in the previous univariate analysis. The theoretical rationale and implementation details of GPC have been described in previous publications. [17][18][19] Here, we just provided a brief introduction of this multivariate pattern classification method. The GPC is a probabilistic model on the basis of the Bayesian extension of logistic regression.
The main strength of GPC model, compared with other alternative methods such as support vector machines, is that the predicted class is augmented by an estimation of the certainty of the prediction.
Based on Bayesian principle, the posterior distribution of functions that represent the training data is identified in an optimal way. This posterior distribution is applied for classifying new examples according to the rules of probability. 20 Thus, the GPC classification could provide probabilistic class predictions and accurately quantify the predictive confidence assigned to each data point.
In the current study, we used a receiver operating characteristic curve (ROC) for evaluating the classification performance of GPC model. The ROC curve could plot the classifier's true positive rate (sensitivity) against its false-positive rate (1-specificity) as the decision threshold is varied. Besides, the leave-one-out crossvalidation (LOOCV) strategy 21 was adopted to assess the stability of the GPC classifier. This LOOCV approach discarded each subject once and used the rest of the participants to train the GPC classifier. Subsequently, the discarded participant pair was used to test the differentiating ability of GPC to reliably distinguish between two groups. Finally, we used a permutation test 22 to determine the statistical significance of the discrimination accuracy yielded by the GPC classifier. Specifically, the permutation test repeated the discrimination process 1000 times using a different random permutation of the training group labels and counting the number of permutations that achieved the same or higher sensitivity and specificity relative to the true labels. All these MVPA procedures were performed using the PROBID toolbox (http://www.brain map.co.uk/probid.htm) as implemented in the MATLAB software.
HCs were excluded if MRI scanning showed any structural lesions. We identified no significant differences between the PNH group and the HC group concerning age, sex, handedness, years of education, or minimental state examination (MMSE) performance (p > 0.05) ( Table 1).
We found that the PNH group had lower ReHo values in the left insula extending to the left putamen as well as in the subgenual anterior cingulate cortex (sgACC) extending to the orbitofrontal cortex (OFC) (p < 0.05, FWE correction at the cluster level) ( Table 2 and Figure 1A,B). We identified no regions with higher ReHo values.

| DISCUSS ION
Here, for the first time, we applied both univariate analysis and MVPA to explore ReHo alterations in patients with PNH. We achieved an  33 Additionally, altered prefrontal cortex local activity has been widely reported in subtypes such as generalized epilepsy 34,35 and temporal lobe epilepsy. 36 A previous task-based  Abbreviations: MNI, Montreal Neurological Institute; OFC, orbitofrontal cortex; PNH, periventricular nodular heterotopia; ReHo, regional homogeneity; sgACC, subgenual anterior cingulate cortex.
*p < 0.05 with whole-brain family wise error correction at cluster level.

TA B L E 2 Decreased ReHo in patients
with PNH compared with healthy controls.
fMRI study revealed that PNH patients had significantly modulated heterotopia as well as widespread cortical activation, 37 and during a task-free resting state, unique directionality was observed within an interconnected network. 11 Another study used seed-defined nodules that combined structural and functional connectivity and showed disrupted FC in multiple brain areas including the cerebellum, 10 while an additional study used seeds defined by altered ALFF areas and showed higher FC in the fronto-parieto-limbic neurocircuitry and lower FC in the DMN. 28 Methodological differences are one possible reason for the inconsistencies among PNH studies. We evaluated functional changes in PNH by assessing local functional disruptions in PNH-related epilepsy groups without a priori constraints rather than focusing on nodule dysregulation as in seed-based functional connectivity studies. ROI definitions for seed-based analysis also differed. Our previous study 28

F I G U R E 2
The plotting for Gaussian process classifier (GPC) classification using ReHo maps are presented in left. The receiver operating characteristic (ROC) curve evaluating the GPC classification performance based on ReHo maps are shown in right.

CO N FLI C T O F I NTE R E S T S TATE M E NT
All authors have completed the ICMJE uniform disclosure form. The authors all have no conflicts of interest to declare.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings are available upon reasonable request.