This retrospective study included patients with TLE-IAS or TLE-AS who underwent cerebral ¹⁸F-FDG PET/computed tomography (CT) examinations at the Department of Nuclear Medicine, Lanzhou University Second Hospital, between January 2018 and January 2025.

The inclusion criteria were as follows: patients aged between 6 and 65 years; complete clinical data with a documented history of impaired awareness seizures (i.e., reduced responsiveness to verbal commands during seizures or inability to recall the ictal episode) or aware seizures; a confirmed clinical diagnosis of TLE; and no seizures occurring within 24 hours before or during the PET scan. The exclusion criteria included: indeterminate seizure type; MRI evidence of major structural abnormalities, such as encephalitis, cerebral infarction, intracerebral hemorrhage, cerebrovascular malformation, or intracranial tumor (except hippocampal sclerosis); prior intracranial surgery; severe metabolic disorders; or neuropsychiatric illness.

Based on these criteria, 30 patients with TLE-AS and 193 patients with TLE-IAS were enrolled. The control group consisted of 193 individuals, who were matched to the TLE-IAS group for age and sex. All controls underwent PET/CT during the same period and had no history of central nervous system diseases (e.g., healthy subjects or patients with non-neurological disorders). We systematically collected demographic data (age, sex), clinical characteristics (onset age and duration of seizures), and brain MRI findings for all enrolled patients.

All PET/CT scans were performed using a Discovery 690 scanner (GE Healthcare, Waukesha, WI, USA). For patients with epilepsy, the scans were performed during an interictal period, which was confirmed by continuous clinical observation for at least 24 hours before and throughout the PET scanning procedure. All subjects fasted for at least 6 hours before the examination, ensuring fasting blood glucose levels between 3.9 and 6.3 mmol/L. Subsequently, 2.96–5.55 MBq/kg of ¹⁸F-FDG was administered intravenously, followed by a 40–60 minute rest period in a quiet, dimly lit room for all participants. PET/CT imaging was then performed in the supine position. Brain PET acquisition lasted 5 minutes per bed position. A low-dose CT scan was first acquired using the following parameters: slice thickness, 3.75 mm; tube current, 50–220 mA; tube voltage, 120 kV; and matrix size, 512 $\times$ 512. PET data were then acquired in 3D+time-of-flight (TOF) mode with a matrix size of 256 $\times$ 256 and reconstructed using the VUE Point FX algorithm (18 subsets, 4 iterations).

Image preprocessing was conducted using Statistical Parametric Mapping (SPM12, Wellcome Trust Centre for Neuroimaging, London, UK) running on MATLAB R2022a (MathWorks, Natick, MA, USA). After converting the DICOM-format PET images into NIfTI format, the origin coordinates were manually adjusted. All individual PET images were then spatially normalized to the Montreal Neurological Institute (MNI) PET template provided in SPM12 and resampled to a voxel size of 2 $\times$ 2 $\times$ 2 mm³.

The PET metabolic brain networks were constructed and analyzed as follows. Preprocessed PET images from the TLE-IAS, TLE-AS, and control groups were parcellated into regions of interest using the Automated Anatomical Labeling 90 atlas. This divided the whole brain into 90 anatomical regions that served as network nodes. The standardized uptake value (SUV) of each region was calculated. The cerebellum—which exhibits relatively stable glucose metabolism—was used as a reference region to obtain the standardized uptake value ratio (SUVR) for each region [23, 24].

At the group level, Pearson correlation coefficients between all pairs of the 90 brain regions (retaining only positive correlations) were computed separately for the TLE-IAS, TLE-AS, and control groups, generating three 90 $\times$ 90 correlation matrices. To optimize the network topology, each correlation matrix was binarized across a sparsity range of 0.25–0.75 (incremented by 0.01) [24, 25].

Subsequently, global and nodal network metrics were computed using GRETNA 2.0 (https://www.nitrc.org/projects/gretna/), including characteristic path length (Lp), clustering coefficient (Cp), global efficiency (Eg), local efficiency (Eloc), and betweenness centrality (Bc) [26, 27]. Hub nodes were identified as those with a normalized betweenness centrality (Bc/mean Bc) greater than 2 and statistically significant according to permutation testing [28, 29]. Visualization of hub node distributions was performed using BrainNet Viewer (https://www.nitrc.org/projects/bnv/) [30].

Statistical analyses were performed using SPSS 25.0 (IBM Corp., Armonk, NY, USA) and MATLAB R2022a. Continuous variables following a normal distribution were presented as mean $\pm$ standard deviation (SD). For inter-group comparisons, the t-test was used. Categorical data were summarized as proportions and compared using the chi-squared test. Non-normally distributed continuous data are presented as median (interquartile range) [M (IQR)], and comparisons between two groups were performed using the Mann-Whitney U test. Group differences in PET-based metabolic network metrics were assessed using nonparametric permutation tests (1000 iterations) supplemented by bootstrap resampling (1000 iterations) for robustness validation. The false discovery rate (FDR) correction was applied to control for multiple comparisons in hub node identification. A two-sided p 0.05 was considered statistically significant.

No significant differences in demographic or clinical characteristics (age, sex, onset age of seizure, duration of seizure, and MRI findings) were found between the TLE-AS group and the other two groups (all p $>$ 0.05) (Table 1).

Metabolic connectivity matrices for the TLE-IAS and control groups were constructed by calculating Pearson correlation coefficients of glucose metabolism across all brain regions (Fig. 1). The interregional metabolic connections of both groups were visualized, with the color bar representing correlation strength. The TLE-IAS and TLE-AS groups tended to exhibit weaker metabolic connectivity compared to the control group, suggesting reduced interregional coordination of glucose metabolism. Additionally, we generated metabolic connectivity difference matrices comparing each patient group (TLE-IAS and TLE-AS) with the control group to visualize alterations (Supplementary Fig. 1).

Fig. 1.

Metabolic connectivity matrices of brain networks in the three groups. (A) Control group, (B) TLE-AS, and (C) TLE-IAS groups. TLE-AS, temporal lobe epilepsy-aware seizures; TLE-IAS, temporal lobe epilepsy-impaired awareness seizures.

Within the sparsity range of 0.25–0.75, graph-theoretical analysis of the metabolic networks revealed alterations in global and local properties across all groups. Compared with the control group, the TLE-IAS group exhibited significantly increased Lp and decreased Cp, with no significant differences in Eg and Eloc (Fig. 2). After FDR correction for each sparsity level, significant between-group differences were observed for Lp at certain sparsity levels (0.47–0.52, 0.58–0.75) (p_FDR ranging from 0.044 to 0.048, mean $\mathrm{\Delta }$Lp: –0.1436), whereas Cp showed significant intergroup differences at specific sparsity levels (0.34, 0.37, 0.38, 0.42–0.75) (p_FDR ranging from 0.003 to 0.036, mean $\mathrm{\Delta }$Cp: 0.0880). No significant intergroup differences were observed for Eg and Eloc across all sparsity levels (all p_FDR $>$ 0.05).

Fig. 2.

Comparison of graph-theoretical metrics between the control and TLE-IAS groups. (A–D) Differences and confidence intervals of Lp, Cp, Eg, and Eloc values between the control group and the TLE-IAS group within the sparsity range of 0.25–0.75. The green area represents the 95% CI, red triangles represent significant between-group differences at each sparsity level (p_FDR 0.05), while blue circles indicate non-significant between-group differences at each sparsity level. Lp, characteristic path length; Cp, clustering coefficient; Eg, global efficiency; Eloc, local efficiency; CI, confidence intervals.

Compared with the control group, the TLE-AS group exhibited significantly increased Lp and significantly decreased Cp and Eloc values, without any significant difference in Eg (Fig. 3). Sensitivity analysis using bootstrap resampling (1000 iterations) confirmed the robustness of these differences. The 95% confidence intervals (CI) for the area under the difference curve were [0.0061, 0.0350] for Lp, [0.0401, 0.1220] for Cp, and [0.1800, 0.4972] for Eloc, indicating stable overall intergroup differences. Furthermore, after FDR correction across sparsity levels, Lp and Eloc showed significant intergroup differences throughout the entire sparsity range (p_FDR ranges: 0.012–0.021 and 0.032–0.037, respectively; with mean $\mathrm{\Delta }$Lp: –0.3561, mean $\mathrm{\Delta }$Eloc: 0.2618). The Cp exhibited significant differences at specific sparsity (0.32–0.75) (p_FDR: 0.007–0.045; mean $\mathrm{\Delta }$Cp: 0.1514). In contrast, no significant intergroup differences were found for Eg at any sparsity level (all p $>$ 0.05).

Fig. 3.

Comparison of graph-theoretical metrics between the control and TLE-AS groups. (A–D) Differences and confidence intervals of Lp, Cp, Eg, and Eloc values between the control group and the TLE-AS group within the sparsity range of 0.25–0.75. The green area represents the 95% CI, red triangles represent significant between-group differences at each sparsity level (p_FDR 0.05), while blue circles indicate non-significant between-group differences at each sparsity level.

Direct comparison between the TLE-AS and TLE-IAS groups revealed no statistically significant intergroup differences in Lp, Cp, Eg, or Eloc (Fig. 4).

Fig. 4.

Comparison of graph-theoretical metrics between the TLE-AS and TLE-IAS groups. (A–D) Differences and confidence intervals of Lp, Cp, Eg, and Eloc values between the TLE-AS and the TLE-IAS group within the sparsity range of 0.25–0.75. The green area represents the 95% CI, red triangles represent significant between-group differences at each sparsity level (p_FDR 0.05), while blue circles indicate non-significant between-group differences at each sparsity level.

Hub nodes, characterized by high centrality, play a key role in information integration and relay within brain networks. In this study, hub nodes were defined at the group level as nodes with Bc values greater than twice the mean Bc and surviving permutation testing.

To compare the hub nodes, a common sparsity threshold of 0.48 was applied, defined as the minimum value ensuring a global efficiency $>$0.7. The control group contained 13 hubs at this threshold, whereas the number of hubs was reduced to 8 in the TLE-AS and TLE-IAS groups (Fig. 5, Table 2). In the control group, the hub nodes were widely distributed across the default mode network (angular gyrus, left (ANG.L), precuneus, right (PCUN.R), posterior cingulate gyrus, left (PCG.L), etc.), other higher-level association cortices (superior temporal gyrus, left (STG.L), fusiform gyrus, left (FFG.L), cuneus, left (CUN.L)), and additionally spread across the limbic/paralimbic system and subcortical nuclei. In the TLE-IAS group, 6 out of 8 hub nodes were located in the limbic/paralimbic system, followed by subcortical nuclei and limited primary auditory areas (Heschl gyrus, left (HES.L)). Similarly, in the TLE-AS group, 5 out of 8 hub nodes were in the limbic/paralimbic system, including the subcortical nuclei, but the primary auditory cortex (HES.L and Heschl gyrus, right (HES.R)) showed a bilateral distribution.

Fig. 5.

Hub nodes in the control, TLE-IAS, and TLE-AS groups. Blue, red, and yellow spheres represent hub nodes in the control, TLE-IAS, and TLE-AS groups, respectively.

Compared with the control group, both TLE patient groups exhibited a consistent loss of hub nodes in the default mode network and a widespread reduction of hubs in other higher-order association cortices. Concurrently, the overall hub distribution shifted toward limbic and paralimbic regions, accompanied by the emergence of new hub nodes in the primary auditory cortex. Additionally, differences in hub composition were observed between both patient groups: the TLE-IAS group exhibited two limbic hubs compared to one in the TLE-AS group, while the TLE-AS group had two primary auditory hubs compared to one in the TLE-IAS group.

PET-based analysis of the metabolic brain network offers high robustness and reproducibility [25], partly due to its independence from neurovascular coupling and the whole-brain coverage provided by ¹⁸F-FDG PET. This study revealed that compared to controls, both patient groups exhibited widespread disturbances in metabolic connectivity, suggesting impaired large-scale functional coherence in TLE. Graph-theoretic analysis further demonstrated that both patients with TLE-IAS and TLE-AS shared a similar pattern of network impairment, characterized by increased Lp and decreased Cp, reflecting longer communication distances and reduced local integration efficiency. These findings align with previous studies on brain networks in epilepsy [35, 36, 37].

Notably, a direct comparison between the TLE-IAS and TLE-AS groups revealed no significant differences in any of the four graph-theoretical parameters. This suggests that the overall severity of network impairment at the topological level is similar in both patient groups. This pattern differs from the weakening of network connectivity reported in patients with pharmacologically induced unconsciousness. A plausible explanation is that the chronic and widespread network damage caused by TLE may constitute a dominant pathological background, masking the more subtle or state-specific topological changes related to varying degrees of awareness. Additionally, since all imaging data were acquired during the interictal period, dynamic network alterations directly driven by impaired awareness may be less pronounced than those seen in pharmacological models. Therefore, future studies should aim to identify awareness-sensitive topological properties and subnetworks [38], or incorporate multimodal recordings (e.g., ictal EEG) to capture dynamic changes in brain networks.

Hub nodes, which concentrate numerous shortest paths, are critical for efficient information exchange in the brain. First, compared with the control group, the TLE-IAS and TLE-AS groups exhibited a similar pattern of network reorganization, including the loss of hub nodes in the default mode network and the superior parietal cortex, with hub distribution converging toward the limbic and paralimbic systems. This finding is consistent with previous studies reporting weakened connectivity or functional abnormalities in default mode networks across multiple epilepsy subtypes [39, 40]. Furthermore, it corroborates earlier observations that in patients with TLE, the hub nodes predominantly exist in the paralimbic/limbic systems [41], supporting Bonilha et al.’s proposed reorganization [42] of limbic regions in TLE. This study also revealed a reduction in hub nodes within the subcortical basal ganglia region across both patient groups (only 1 retained versus 2 in controls), consistent with previous reports showing diminished node strength in this area among epileptics [43]. Regarding the newly identified hub nodes in the primary auditory cortex across both patient groups, we propose that their emergence may compensate for reduced whole-brain integration efficiency and the loss of hub nodes in higher association cortices or may relate to the frequent auditory aura symptoms observed in patients with TLE. Based on the above discussion of the similar pattern in TLE, we propose a preliminary hypothesis that this pattern may facilitate the propagation of epileptic discharges through the limbic/paralimbic system. Further, compensatory alterations in the primary auditory cortex may indicate abnormal discharges in this region during seizures, potentially contributing to auditory auras and other abnormal perceptions. Future studies should further examine the relationship between auditory cortex hub formation and specific clinical symptoms by including a subgroup of patients with auditory aura or by integrating seizure-period EEG data.

The difference in hub distribution between the TLE-IAS and TLE-AS groups provides preliminary insights into the neural substrates underlying their distinct awareness states. Compared to the TLE-IAS group, which exhibited only one additional hub in the primary auditory cortex, the TLE-AS group showed two new hubs in the bilateral primary auditory cortex (HES.L and HES.R), which may provide the neural network foundation for maintaining basic awareness and sensory processing during epileptic seizures. Moreover, compared to the control and TLE-AS groups, the TLE-IAS group exhibited a higher number of hub nodes in the limbic system. This might be due to an accumulation of pathological states. However, whether this difference is directly related to impaired awareness or attributable to epileptic burden remains unclear. Future studies should validate this finding by controlling for key variables, such as seizure frequency and medication use, or expanding the sample size of patients with TLE-AS. Recently, two major hypotheses have been proposed to explain impaired awareness during seizures: the “network inhibition hypothesis” [18] and the “global workspace theory” [17], which collectively highlight the importance of the frontal cortex in the awareness system. However, in contrast to these theories, we observed that REC.L, a hub node in the prefrontal cortex, was specifically lost in the TLE-AS group, whereas it was present in both the control and TLE-IAS groups. Given the limitations of the current cross-sectional study design, the specific mechanisms underlying this phenomenon are difficult to fully explain. To avoid overinterpretation, we present this finding here as a critical observation, warranting targeted experimental designs to underscore its exact significance.

This study has several limitations. First, the sample size of the TLE-AS group is relatively small, mainly due to the low prevalence of TLE-AS in clinical settings and limited PET/CT utilization in this subgroup. Although bootstrap resampling was applied to improve robustness, future studies should aim to expand this cohort to strengthen the conclusions. Second, the network analysis was conducted at the group level, potentially overlooking individual variability. In the future, individualized network approaches, such as those based on Kullback–Leibler divergence, can be adopted to capture subject-specific topological profiles. Third, given the radiation exposure associated with PET/CT, control subjects were recruited from patients without neuropsychiatric conditions rather than healthy volunteers. Even with careful age- and sex-matching, unmeasured confounders may persist. Fourth, although the broad age range (6–65 years) was similar to that of some related studies and showed no statistically significant intergroup differences, developmental differences in the brain metabolic network between children and adults might introduce confounding effects. Finally, owing to the retrospective nature of the study, several clinically relevant variables (e.g., seizure side, medication regimen, seizure frequency, and surgical pathological or stereo-electroencephalography findings) were not systematically controlled. These factors might preclude precise localization of the seizure onset zone for causal inference and may introduce variability in metabolic network patterns, which should be accounted for in future prospective designs.