Epilepsy is one of the most common neurological diseases, affecting ~50 million people worldwide (1). In total, ~30% of patients with epilepsy are drug-resistant and suffer from seizure recurrence (2). Increased intractable rates leading to premature mortality, the global burden of disease, and social stigma have been public health concerns (1). Temporal lobe epilepsy (TLE) accounts for the largest proportion of drug-resistant epilepsy (DRE) (3). The mechanisms of DRE were proposed as alterations in the antiseizure medication (ASM) targets with the loss of receptor sensitivity and overexpression of multidrug resistance (MDR) proteins (MDR1, MDR1/P-glycoprotein, etc.) that affect drug transport (2). In addition, abnormal neural circuits, such as hippocampal–diencephalic–cingulate paths (4), extratemporal networks (5), and dynamic network evolution (6), were suggested to be relevant to seizure severity in TLE. Clinically recognizing potential severities and deciding on treatment strategies for patients with TLE remains a challenge.

Temporal lobe epilepsy is characterized as a neural network disorder that propagates beyond the regions that are anatomically connected (7, 8). The underlying basis of the neural network in TLE involves structural and functional alterations, including neurogenesis and functional plasticity (9). It has been reported that extratemporal networks, such as the orbitofrontal, insular, medial frontal, bifrontal, and parietal regions, were involved in some patients with TLE (7, 10). Indeed, wider spreading networks were spanning more interconnected regions in TLE patients with drug-resistance and focal to bilateral tonic–clonic seizures (11). Furthermore, rapid spread in the lateral temporal cortex was detected in TLE patients with recurrent seizures (12). Several studies have illustrated the anatomical and electrophysiological connections with the extemporal regions in patients with TLE. A typical pathway from the hippocampus to the prefrontal cortex, which is called the hippocampus–prefrontal pathway, has been uniquely reported as a monosynaptic and unidirectional projection mediated by glutamate (13). The lateral temporal, frontal, and occipital neocortexes develop from the same origin of cortical gradients of laminar elaboration and organize comparable neurogenetic time courses. They are connected via various fibers, such as the arcuate fasciculus (AF), uncinate fasciculus (UF), cingulate, and longitudinal fasciculus (14). Thus, an accessible tool to distinguish the evolution of the extratemporal network may help differentiate outcomes in patients with TLE.

Functional magnetic resonance imaging (fMRI) and electroencephalography (EEG) were utilized to map the functional epileptic network (15). Using fMRI, a seizure extratemporal propagation network of TLE was identified, including the bilateral thalamus, insula, midcingulate, and precuneus (16). However, MRI measures the blood–oxygen signal and indirectly characterizes electrophysiological functions, regardless of its mental contraindications and expense. In contrast, EEG directly reflects neural activity (15). Stereoelectroencephalography (SEEG) is the gold standard for marking epileptogenic and propagation networks (8) but is limited to invasion, spatial undersampling, and heterogeneous topological signatures for individual electrode placement (17). Scalp EEG offers a standard acquisition and non-invasive scheme for determining the network. However, in terms of its anatomical disadvantages (15), a more precise arithmetic tool based on scalp EEG is required to overcome the challenge of mapping the seizure spread and epilepsy characteristics in TLE.

Graph theory provides a topology of the underlying neural architecture (18). Recently, small-worldness, minimum spanning tree, and community structure are the popular network methods to help clinicians better understand how seizures originate and propagate. Based on EEG, the alternations of local segregation (characterized by clustering coefficient, CC) and global integration (characterized by shortest path length, PL) were reported in the epileptic network study (19, 20). Compared to other measures (i.e., magnitude squared coherence, synchronization likelihood, phase locking value, etc.), PLI was proved to be the least affected functional connectivity (FC) metric by spurious influences in a simulated volume conduction environment (21). Phase lag index (PLI), indicating brain synchronization, was decreased with seizure reduction in focal DRE (22). A minimum spanning tree (MST) was introduced into the brain network field to avoid the threshold choice during the network construction (23–25). The clinical application of MST is still lacking because the method is not as popular as the small-world network. In 2016, van Diessen et al. found that MST features are more sensitive to identifying the interictal network alterations than classical weighted network analysis at an early stage of focal epilepsy (26). Thus, the MST was adopted to detect the alternations of the patients with TLE in this study. Community structure quantifies the extent of brain functional network partition and was recently introduced to help classify epileptic EEG events (27). However, the community construction in EEG network analysis always lacks its physiological meaning because of the bad EEG spatial resolution. By using source location techniques, the community analysis may reveal the hidden network mechanisms of the seizure propagation. Distributed EEG source localization using standardized low-resolution brain electromagnetic tomography analysis (sLORETA) is a relatively new method to yield three-dimensional (3D) images of electrical neuronal activity (28). Despite the robustness of indicators that was demonstrated in this approach, it still lacked consistent results and clinical applications. This study aimed to find the brain network indicators based on EEG and provide an accessible approach to indicate seizure severity in patients with TLE and guide the treatment strategy for routine clinical practice.

Temporal lobe epilepsy is a common type of DRE that involves complex network mechanisms. Here, we adopted a graph theory analyzing pipeline based on EEG, including the weighted functional network, MST, and community structure analysis methods. We reported the independent correlations between frontotemporal functional connection and seizure severity in patients with TLE. This study highlighted the crucial role of the frontal lobe in TLE and provided a non-invasive and easy-to-use indicator for drug-resistant patients with TLE.

Compared to negative findings using routine power spectrum analysis, the graph theory network results showed increased frontotemporal connections ipsilaterally during interictal-resting periods. The frontal and temporal cortices are connected via various fibers, such as the AF and UF (14). Whether the frontal lobe participates in seizure propagation in patients with TLE remains unknown. Some studies have reported the involvement of the frontal lobes in patients with TLE, with controversial conclusions. Ipsilateral hyperperfusion in regions was observed in patients with mTLE using single-photon emission computed tomography. At the same time, significant hypoperfusion was found in the bilateral prefrontal areas (37). However, decreased frontal connectivity has also been detected during seizures (38). Other studies reported that frontotemporal, especially orbitofrontal, connections were interictally increased after interictal epileptic discharges (39). In this study, sLORETA images, according to EEG sources, a newer tomographic method to calculate electrical neuronal activity, were consistent with those connected by AF and UF. The pars triangularis (BA 45), posterior frontal gyrus (BA 8, 9), and pars opercularis (BA 44) were the regions in which the largest branch of AF was terminated (40). The other branch connects the posterior part of the superior temporal gyrus (BA 41, 42) to the inferior frontal gyrus (BA 44, 45) (41). The UF connects the anterior temporal lobe (BA 38) with the orbitofrontal (BA 11, 47) and frontal pole cortex (BA 10). The ventral branch of the UF terminates in the orbitofrontal cortex (BA 11, 47). In addition, the anterior and middle branches connect the anterior cingulate gyrus (BA 32) and frontal pole (BA 10) (41). These regions co-activated in R-TLE, L-TLE, and B-TLE may indicate the frontotemporal reconfigurations of the epileptic functional network. Because the presence of structural abnormality affects the epileptic network, the patients with abnormal extratemporal abnormalities on routine MR were excluded from this study to avoid confounding factors such as non-epileptogenic lesions. For the temporal lesions, multivariate analysis showed consistent results in 104 TLE patients with negative MRI.

Compared to other bands, the theta band was the most significant in the functional network in this study. Previous studies reported a unique theta oscillation in the human hippocampus, which was widely associated with attention, execution, and emotion (42). Little attention has been paid to epilepsy research. Theta rhythm was previously reported to regulate the firing frequency and discharge time of individual neurons (43). A study of 10 patients with DRE found that small-world parameters in the theta band were increased before seizures (44). Seizure behavior was related to theta activity in the kainic acid model (45). All the results indicate that the theta band plays an important role in seizure propagation.

The most important clinically relevant finding was that using our analyzing pipeline, frontotemporal PLI values in the theta band with IFCN-standard EEG were independently correlated with VA-2 scores. PLI values between 0 and 1 were used to describe brain synchronization and reported to be much better to avoid volume conduction than other parameters (synchronization likelihood, Fourier coherence coefficient, etc.) (46). Although weighted PLI (wPLI) is slightly better than PLI in avoiding volume conduction, Christodoulakis, et al. (47) addressed that the brain network indicators calculated based on wPLI are not significantly different from those calculated based on PLI. In addition, they emphasized that excessive pursuit of reducing the impact of volume conduction will result in a loss of the ability to identify abnormal epilepsy EEG changes. Therefore, we adopt the method of PLI, which can not only effectively avoid the impact of volume conduction, but also sensitively capture the network significance in patients with TLE compared to the controls. The larger the PLI was, the more vital synchronization was observed in the brain. Synchronized neurons are critical for brain function, whereas excessive synchronization is notably related to epilepsy (48). Meanwhile, increased CC and decreased PL were observed, indicating that local and global network connectivities were abnormal. The changed small-world topologies in patients with TLE were consistent with previous studies (19, 49), except for some controversial results (20, 50). These inconsistent results might be due to the different network types (structural or functional, global, or local network) or neuroimaging modalities (EEG or MRI). MST results showed that the frontotemporal diameter in patients with epilepsy was shorter than that in the controls. We observed that alterations in the frontotemporal MST parameters were more notable than those in the whole brain. However, in this study, we did not detect its sensitivity compared to the weighted-graph indicators, which might have been because of its focus on the network skeleton and ignorance of the weaker connections (51). In terms of prognosis prediction, other methods such as machine learning were preliminarily applied. Croce et al. (52) provide a pipeline using a machine learning approach for predicting the levetiracetam response in 23 patients with epilepsy. Tang et al. (53) develops and validates a machine learning-powered approach in neuroimaging for surgery outcome prediction. However, it remained to be improved in the clinical application, such as excessive feature extraction, computational cost, time consumption, the risk of overfitting in small datasets, etc. The changes in TLE frontotemporal network characteristics that we have discovered so far may guide us to use machine learning methods to predict clinical outcomes in the future.

According to the previous EEG brain network research, the epoch length of the EEG signal will affect the functional connectivity values, which in turn affects the brain network analysis results. The brain functional connectivity tends to be stable when the continuous EEG epoch is longer than 10 s (25, 54). Consistent with our previous study, 90-s EEG epochs were segmented and analyzed (55). At present, the latest electrophysiology network modeling analysis mainly depends on invasive EEG monitoring or high-density EEG (56). However, these were limited by routine clinical use and individualized implantation schemes. In 2017, the IFCN recommended a new electrode positioning system for routine practice. It was a 25-channel system, adding an inferior temporal chain based on the conventional 10–20 system (33). It was reported to have the same effect as high-density recordings (above 128 channels) in diagnostic yield (57). It could compensate for the lack of routine practice and individualization compared to SEEG and high-density EEG. We followed the IFCN guidelines to identify alterations within the temporal and frontal lobes. This study provided a non-invasive indicator with IFCN-system EEG for seizure severity, which was more accessible and standardized in routine practice.

The limitation of the present study is the insufficient analysis of the EEG network to describe the anatomical spatial characteristics. A better network characteristic could be described with a high-density EEG recording. Further work is required for the comparison of our data with SEEG and high-density EEG results. Second, it is a lack of accuracy for the source-level analysis using standard coordinates of electrodes. We will obtain the source location coordinates from each individual MRI and then transform them in Talairach space to precisely determine the distribution of region of source involvement in the following studies. Moreover, the predicted model will be acquired for multi-dimensional parameters, large sample sizes, and multi-center research.

In summary, our network analysis illustrated that ipsilateral frontotemporal regions are activated in the interictal state in patients with TLE. The frontotemporal PLI values in the theta band independently correlated with VA scores in patients with TLE. We suggested that increased frontotemporal theta synchronicity might be correlated with seizure severity. This study introduced a non-invasive method with routine scalp EEG for seizure severity and guided the treatment strategy for drug-resistant epilepsy.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by Zhongshan Hospital, Fudan University. The patients/participants provided their written informed consent to participate in this study.

LM: study concept and design, acquisition of data, study coordination, and drafting. GZ: data analysis, interpretation, and revising the article. YC: data interpretation. WL: acquisition of data. QZ: acquisition of data. WP: acquisition of data. JD: study design, interpretation of data, and revising the article for content. XW: study concept and design and revising the article for content. All authors contributed to the article and approved the submitted version.

This work was supported by Project grants from the Shanghai Municipal Committee of Science and Technology (Codes 17411962500, 16JC1420201, and 2018BR05).

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.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.