High energy consumption characterizes abnormal brain state transitions in temporal lobe epilepsy

Abstract

Temporal lobe epilepsy (TLE) patients experience shifts between non-seizing and seizing brain states, but the structural networks underlying these transitions remain undefined and poorly characterized. We detected dynamic brain states in resting-state fMRI and constructed linked structural networks utilizing multi-shell diffusion-weighted MR data. Leveraging network control theory, we interrogated the structural data for all possible brain state transitions, identifying those requiring abnormal levels of transition energy (low or high) in TLE compared to matched healthy participants (n’s = 25). Results revealed three transitions requiring significantly higher energy in TLE; no abnormally low-energy transitions were observed. In HPs, transitions relied on mediator regions that did not belong to the initial source or final target brain areas. TLE transitions involved a more restricted set of source/target regions, predominantly outside the epileptogenic temporal lobe. Our findings highlight the abnormal and inefficient network mechanisms that accrue from the network entrainment inherent to TLE seizure activity. We argue these findings clarify the pathologic effects and help explain the well-known cognitive inefficiencies and other deficits found in the TLE disorder.

1. Introduction

Epilepsy is a complex neurological disorder characterized by intermittent, often chronic seizure activity afflicting millions of people worldwide (Lariviere et al., 2020; Lariviere et al., 2022). Understanding the underlying functional and structural mechanisms, and in particular the regions, that drive the brain to transition from a state of non-seizure to seizure activity is critical for developing precise brain targets for medical or surgical interventions. A recent shift in scientific paradigms has allowed us to see that epilepsy brings changes to the macroscale organization of the brain (Weng et al., 2020) and that even focal epilepsies constitute a network communication disorder (Bernhardt et al., 2015).

The current study focused specifically on temporal lobe epilepsy (TLE), the most common form of focal epilepsy. We explored macroscale brain organization through the lens of structural network connectivity, specifically neuronal energy resources (i.e., consumption), which allowed us to map out the brain regions that foster and drive both the energy demanding and low resource transitions that occur between brain states (Betzel et al., 2016; Karrer et al., 2020; Parkes et al., 2024). In this project we made no a priori assumptions about the regional membership or organizational topology (Zhang et al., 2024; Hinds et al., 2023) of the brain networks involved in state transitions. Accordingly, we carried out an open, whole brain data-driven analytic approach that allowed any combination of regions or networks to emerge as the source and target topology(ies) defining the brain states. This approach enabled us to interrogate a wide range of state transitions when trying to isolate those that required abnormal levels of energy (high or low) compared to age- and sex-matched healthy participants (HPs).

Transition energy is the amount of energy required by a brain system to move from one state to another (Kim et al., 2018) (e.g., non-seizing to seizing; shift between two cognitive states, and etc.). Several studies (He et al., 2022; Chari et al., 2022; Janson et al., 2024; Scheid et al., 2021) indicated that TLE is generally associated with reductions in transition energy related to the underlying functional or structural burden imposed directly by TLE regional pathology. There is, however, a conundrum that has limited our understanding of the transitions that drive the brain into epileptiform activity. Namely, it has never been established whether the transition to neuronal hypersynchrony among brain regions during the ictal state requires abnormal levels of energy consumption (either high or low). Thus, in our approach, we emphasized any extreme difference in transition energy between our TLE and HP groups, all toward the goal of isolating the brain state transitions at the neurobiologic core of TLE-induced brain reorganization.

Despite advances using MRI, PET, scalp, intracranial and stereo EEG, and neuropsychological testing, there is still no definitive, flawless method for identifying the epileptogenic network in TLE, and some cases remain a challenge to localize. As well, there is no empirical or clinical evidence to suggest that even in focal TLE only one epileptogenic network is at work. In line with this, our method provided a means of identifying not just a single abnormal network, but an array of regions and networks, capable of driving the abnormal brain states evident in TLE. We hypothesized our method would yield two possible scenarios in transition energy, each potentially capturing key mechanisms of epileptogenic pathology. One, compared to HPs, certain transitions in TLE would require abnormally increased energy, a scenario implying that overcoming a homeostatic baseline such as the non-seizure state is consistently, metabolically demanding. Second, decreased transition energy would be evident in TLE compared to HPs, a scenario implying that TLE pathology entrains a neuronal circuit (i.e., presumptive seizure network) that makes traversing between two given states easy, even for transitions present in healthy controls, highlighting that it is the easy facilitation (low energy) feature of the transition that is the key mechanism of seizure generation. Importantly, while both scenarios point to potential structural mechanisms for the construction of epileptiform networks, the dissimilarity in energy requirements will distinguish the roles played by particular regions in fostering the transition to epileptogenic activity. That is, transition energy provides information about inherent energy resources (n.b., either high or low), revealing the capacity of a region to shift the brain toward pathologic seizure activity. By assessing this capacity, transition energy captures a feature goes beyond the functional or structural connectivity of a region (hubness, connection density, small world topology, etc.) as measured by standard graph theory metrics.

We chose to not impose any network topology or organization when isolating the transitions that require either low or high energy for the TLE brain to traverse. We measured the minimum energy required by individual networks constructed from multi-shell diffusion-weighted MRI (dMRI) data to transition between brain states, with these states extracted from dynamic resting-state functional MRI (rfMRI). Next, we utilized network control theory metrics to capture the ability of a specific region to be a driver of brain state transitions (average and modal controllability (Parkes et al., 2024), indexing, respectively, the regional capacity to drive easy versus hard state transitions). We also examined the structural connectivity of the regions involved in the observed abnormal state transitions through standard graph theory measures of hubness (eigen vector centrality, degree centrality). Lastly, we assessed the association of our regional energy, control and graph theory measures to clinical variables such as age of epilepsy onset, duration of epilepsy, and the frequency of seizures, potentially providing clues to the clinical history or trajectory of a region as it developed specific state energy, control, or connectivity characteristics.

2. Materials and methods

2.1. Participants

A total of 25 focal temporal lobe epilepsy (TLE) patients (age = 42.07 ± 12.0) were recruited from the Thomas Jefferson University Comprehensive Epilepsy Center (TJUCEC) (19 left TLE; 6 right TLE; see Table1 for clinical and demographic details). All patients were candidates approved for thermal ablation or resection of the ictal temporal lobe. The seizure focus was determined by multi-modal data (surface video EEG, MRI, PET, neuropsychological assessment) and a consensus decision of the TJUCEC surgical conference committee (Sperling et al., 1992). Diagnosis was further adjudicated through clinical history, clinical interview, and chart review from all providers. A total of 25 age and sex matched healthy participants (HP: age = 39.62 ± 10.54) were recruited from the Thomas Jefferson University community. Exclusion criteria consisted of medical illness with central nervous system impact other than epilepsy; prior or current alcohol or illicit drug abuse; evidence of extra-temporal seizures or MRI/PET pathology; contraindications to MRI; psychiatric diagnosis other than an Axis-I Depressive Disorder; or hospitalization for any Axis I disorder listed in the Diagnostic and Statistical Manual of Mental Disorders, IV. Note, given the substantial co-morbidity of depression and epilepsy, Depressive Disorders were permitted in the patient sample (Tracy et al., 2007). We provided a detailed explanation of the study’s aim and potential consequences to all patients and healthy participants involved, and all participants provided their signed informed consent.

Table 1.

Demographic and clinical characteristics of study participants.^*

	HP (25)	TLE (25)	t/χ2	P
Chorological Age	39.62 ± 10.54	42.07 ± 12.00	−0.75	0.46
Sex (Male/Female)	14/11	16/9	0.72	0.77
TLE Laterality (R/L)	N.A.	6/19
Handedness (R/L)	22/3	21/4	1.40	1.0
Education	17.36 ± 1.64	13.16 ± 2.29	7.29	0.00
Age at Epilepsy Onset (Year)		30.86 ± 12.76
Illness Duration (Year) [Q1,Q2, Q3]		[4.75,7,17.25]
Frequency Of Seizures (Per Month)[Q1,Q2,Q3]		[0.46,1,2.62]

^*Q stands for quartile.

2.2. General imaging protocol

All images were acquired using a 32-channel head coil in a SIEMENS MAGNETOM Prisma 3 T scanner located at Jefferson Neuroscience Hospital. MP-RAGE, Rest fMRI (rfMRI), and multi-shell diffusion-weighted -NODDI- sequences were acquired during the same scanning session. The scanning schedule or time of day was not different for TLE/HP groups, which precluded bias due to scanner calibration that can vary over time. The MP-RAGE volume was acquired in the sagittal plane with a resolution of 320 × 320 mm and a slice thickness of 0.8 mm (208 slices, isotropic voxels of 0.8 mm3, TR = 2500 ms, TE = 2.2 ms, FOV 256 mm, flip angle 80°). The diffusion-weighted images were collected using a single-shot spin-echo EPI pulse sequence (TE = 68 ms, TR = 3000 ms, ~10 min acquisition) with 64 orientations (b-values =1000, 2000, 3000 s mm⁻², anterior-posterior fold-over direction) and 5 non-diffusion volumes (b-value = 0 s mm⁻²). Each brain volume was acquired in the transverse plane with 50 slices (thickness = 2.5 mm, gap = 0 mm) and a matrix size of 88 × 88 (FOV 220× 220 × 125 mm³, voxel size 2.5 × 2.5 × 2.5 mm³). A standard SPIR (spectral pre-saturation with inversion recovery) technique was used to suppress fat. The rfMRI data was obtained using a gradient-echo echo-planar imaging (GE-EPI) sequence in two sequential 8-min periods. Imaging parameters included: TR = 800 ms, TE = 37 ms, 72 slices acquired with a FOV of 208 mm, a slice thickness of 2 mm, and a flip angle of 52 degrees.

2.3. Image preprocessing and reconstruction

All images were preprocessed using standard QSIprep (Cieslak et al., 2021) and fMRIprep (Esteban et al., 2019; Esteban et al., 2020) pipelines (details in supplemental section). Network reconstruction was performed using QSIRecon 0.23.2 (Cieslak et al., 2021), which is based on Nipype 1.8.6 (Gorgolewski et al., 2011). Diffusion orientation distribution functions (ODFs) were reconstructed using generalized q-sampling imaging (GQI) (Yeh et al., 2010), with a ratio of mean diffusion distance of 1.25 in DSI Studio (version 94b9c79). Automatic Tractography was run in DSI Studio (version 94b9c79) (Yeh, 2022) and the Brainnetome atlas (Fan et al., 2016) was used for parcellation and construction of the structural connectivity matrices (all 246 parcels utilized; each parcel/region of interest [ROI] normalized by its gray matter volume). The connectivity matrix of the right TLE (RTLE) patients was flipped, yielding hemispheric data ipsilateral or contralateral to the seizure focus.

2.4. Network control theory measures

We applied network control theory (Parkes et al., 2024) to identify the minimum energy required for transitioning between the initial source and the final target structural network. Network control theory is a framework used to understand how certain regions in a network (in this case, brain regions) can influence the dynamics of an entire system (brain). Minimum energy refers to the least amount of effort or neuronal resources needed to shift the brain from an initial state (n.b., in cognitive studies an example would be the network associated with spatial attention) to a target state (n.b., the network associated with an emotional or decision-making state). Additionally, we assessed average and modal controllability, which reflected the ability of a particular brain region to steer and drive the brain into a wide range of states with minimal versus maximal effort, respectively. As these controllability measures were strongly, inversely related (n.b., showing correlations of r ≈ −1 for all brain parcel pairs), we will only be reporting average controllability.

To determine the brain regions submitted to the above energy transitions and control theory calculations, we utilized our resting state data to identify the brain regions with high levels of intrinsic activation. More specifically, dynamic amplitude of low-frequency fluctuations (dALFF) was calculated using the DynamicBC toolbox (Liao et al., 2014) (n.b., ALFF measures spontaneous fluctuations in the BOLD signal intensity in a given region). For all TLE patients and HPs, we applied 30- and 40-s windows to the rfMRI data, with overlapping steps of 0 (no overlap), 0.25 (25 % overlap) and 0.5 (50 % overlap). ALFF was computed at each time window for each ROI in the Brainnetome atlas. We then binarized the ROI activity for each region in each window. That is, if the ROIs’ ALFF was greater than two standard deviations above the mean of whole-brain activity for a given window, the region was considered to be in an active (state = 1) and otherwise taken to be in an inactive (state = 0). Each intrinsic binarized dALFF constituted the final set of regions/network that defined a distinct brain state. Next, we calculated the minimum energy needed for transitioning between all consecutive pairs of binarized dALFF-defined states, then conducted permutation t-tests between our HP/TLE groups, thereby creating group-difference energy maps for all transitions (10,150 total transitions). Regions present in the initial state are referred to as source regions. Those present in the final terminal state are referred to as target regions. Those ROI’s present in a transition but absent from the list of specific source or target regions are referred to as bulk regions.

2.5. Graph theory (hubness) measures

We computed graph theory hubness measures, including degree centrality (DC), which counts the number of connections for each node, and eigenvector centrality (EC), which assesses the connections among hub nodes (Hagberg et al., 2008).

2.6. Statistical analyses

Whole brain minimum energy measures associated with each possible transition among the dAlff-defined brain states were submitted to independent t-tests (Benjamini–Yekutieli false discovery rate [FDR] multiple test correction, permutation p-value<0.05) to determine which transitions reliably differed between the HP/TLE groups. For each transition showing a significant HP/TLE difference in whole brain minimum energy, we identified the specific brain regions present in the source and/or target states (i.e., energy maps). Group differences in the minimum energy for each individual ROI associated with the three transitions, including bulk regions, are reported. For the individual ROIs we also interrogated for HP/TLE differences in average controllability and hubness (DC and EC). Throughout all t-tests analyses a threshold of permutation p-value (p < 0.05) was utilized to determine statistical significance. Lastly, regional minimum energy in transitions, controllability, and graph theory measures that showed significant HP/TLE differences were examined for their correlation to clinical variables such as age of seizure onset, frequency of seizures (per month over the past year), and illness duration.

3. Results

3.1. Demographical and clinical comparisons

The experimental groups (HP, TLE) did not differ in age, biological sex, or handedness (Oldfield, 1971) (see Table 1). The groups did, however, differ in the years of education attained, with HPs achieving higher levels. Accordingly, to address the potential impact of education on the comparisons, education was regressed out from the metric values using a linear regression model. The resulting residuals then were taken for subsequent statistical analyses and inference. (See Fig. 1.)

Fig. 1.

Network control theory representation of brain connectivity. Tractography-derived structural connections formed the network edges, while dynamic ALFF fMRI activations defined the network states. Minimum transition energy was used for comparing the TLE/HP groups.

3.2. HP/TLE differences in specific state-to-state transitions

As noted, the dALFF algorithm extracted 10,150 brain state transitions from the rfMRI data. Out of these transitions, three demonstrated a reliable HP/TLE difference in whole brain transition energy (to be referred to as transitions one, two, and three; see source [“S”] and target [“T”] region for each transition in Table 2). In all three transitions, the energy consumption required to traverse the transition from source to target was significantly higher in the TLE group. Importantly, no transitions were found to require lower energy in TLE patients compared to HPs. Fig. 2 shows the regional distribution of the activity as given by dALFF for the three transitions. The overall mean and distribution of energy consumption for these transitions, broken down by groups, is shown in Fig. 3.

Table 2.

HP/TLE differences in specific state-to-state transitions. Regions included in the source (S) and target (T) states for the three transitions showing an HP/TLE difference in whole brain minimum energy consumption. Subscript numbers indicate the transition number.

Source/Target	ROI
S1	Superior Parietal Lobule; rostral area 7; Contralateral
S12	Superior Parietal Lobule; caudal area 7; Ipsilateral
S1	Superior Parietal Lobule; intraparietal area 7(hIP3); Contralateral
S1	Precuneus; medial area 7(PEp); Ipsilateral
S123T123	MedioVentral Occipital Cortex; caudal lingual gyrus; Ipsilateral
S123T123	MedioVentral Occipital Cortex; caudal cuneus gyrus; Ipsilateral
S123T123	lateral Occipital Cortex; medial superior occipital gyrus; Ipsilateral
S12T123	lateral Occipital Cortex; medial superior occipital gyrus; Contralateral
S23T123	Parahippocampal Gyrus; caudal area 35/36; Ipsilateral
S3T123	Parahippocampal Gyrus; rostral area 35/36; Ipsilateral
S23T123	Insular Gyrus; ventral agranular insula; Ipsilateral
S23T123	Insular Gyrus; ventral agranular insula; Contralateral
S23T123	Insular Gyrus; ventral dysgranular and granular insula; Ipsilateral

Fig. 2.

Transition source and target maps. Distribution of activity on brain maps (given by dALFF) for the three state transitions showing a significant difference between the TLE/HP groups in whole brain transition energy. The green color indicates the initial state, the red color shows the target state, and the regions in yellow are common to both the source and target states.

Fig. 3.

Whole brain minimum energy distribution in groups. Mean value and distribution of whole brain transition energy for the three state transitions showing significant TLE/HP differences (Benjamini–Yekutieli false discovery rate, multiple test correction).

The regional transition energy data revealed that transition one initiated from superior parietal lobule (SPL, including ipsilateral caudal, contralateral rostral and intraparietal area) to ipsilateral medial precuneus to ipsilateral orbital gyrus (OrG), ipsilateral rostral-caudal parahippocampal gyrus (PhG), contralateral medial amygdala, and bilateral ventral granular and dysgranular insular gyrus (vIg and vId). The ipsilateral caudal lingual gyrus (cLinG), ipsilateral caudal cuneus gyrus (cCung), and the bilateral medial superior occipital gyrus (msOccg) were active and remained unchanged in both the source and target states.

Transition two initiated from the ipsilateral caudal SPL to the ipsilateral OrG and ipsilateral rostral PhG and the bilateral vIg and vId. The ipsilateral caudal PhG, ipsilateral cLing, ipsilateral cCung, along with the bilateral msOccg, were active and remained unchanged in both the source and target states. Accordingly, transition two involved ipsilateral parietal and ipsilateral inferior and medial sources to a target state including ipsilateral medial, subcortical, and mostly posterior brain areas. The PhG played a prominent role in this transition in which ipsilateral structures were almost solely featured.

Transition three also initiated from the ipsilateral caudal SPL, with other regions (ipsilateral OrG, ipsilateral rostral-caudal PhG, bilateral vIg and vId, ipsilateral cCunG, ipsilateral cLinG, bilateral msOccg), active and unchanged in both the source and target states. Thus, transition three was characterized by stability in activation across mostly ipsilateral, posterior regions, and an isolated ipsilateral parietal region initiating but then shifting out of the target state. Again, mesial structures played a limited role in this transition.

Taken together, the regional transition energy patterns were primarily ipsilateral and medial in nature. Activity in all these regions generally remained stable between source and target states. Notably, some contralateral structures were also involved. Mesial structures had a limited role overall. Common to all three transitions was a very limited presence of temporal lobe involvement (only select amygdala and PhG findings) and very limited frontal involvement (e.g., only ipsilateral OrG). Thus, these regions were among the most frequently classified as bulk regions with regard to each transition.

3.3. High energy consumption ROIs in TLE and HPs

Turning our focus to the individual ROIs that required high levels of energy consumption to traverse brain states, the topologic pattern formed by these nodes was quite distinct in the TLE and HP groups. Table 3 displays the specific ROIs showing a group difference in individual ROI consumption (n.b., the number[s] after source [“S”] and target [“T”] denote the transition in which a given a ROI participated). These ROIs can be viewed as the nodes that, in general, served as the facilitator of transitioning from the source (start) to target (termination) nodes of the brain state. Among the source and target regions, those requiring the highest levels of energy consumption in the TLE patients relative to HPs were ipsilateral orbital gyrus (OrG) area 13, ipsilateral parahippocampal gyrus (PhG) rostral area 35/36, ipsilateral medioventral occipital (cortex caudal lingual gyrus and caudal cuneus gyrus [MVOcC; cLinG and cCung]). Table 3 displays the source, target, and bulk nodes by transition (↑ denotes significantly higher energy in TLE than HP and ↓ denotes the inverse direction). Among the bulk regions, those maintaining high levels of energy consumption in the TLE patients relative to HPs were ipsilateral dorsomedial parietooccipital sulcus (dmPOS), precuneus (Pcun; dmPOS), and contralateral lateroventral fusiform gyrus (FuG, area 37lv). Bulk regions requiring higher energy in HPs compared to TLE patients were: ipsilateral rostral middle temporal gyrus (MTG; area 21r), ipsilateral dorsal granular insular gyrus (INS; dIg), and ipsilateral occipital thalamus (Otha).

Table 3.

High energy consumption ROIs in TLE compared to HPs. Displays significant TLE/HP group differences in minimum energy for each individual ROI associated with the three transitions. Bulk (B) regions are included as well as Source (S), Target (T) ROI’s. Subscript number indicates the relevant transition).

Transition		ROI	t-statistic
S3T123	↑	Orbital Gyrus; area 13; Ipsilateral	−3.04, −3.31, −3.28
S123T123	↑	MedioVentral Occipital Cortex; caudal cuneus gyrus; Ipsilateral	−3.42, −3.47, −3.68
S3T123	↑	Parahippocampal Gyrus; rostral area 35/36; Ipsilateral	−3.05, −3.47, −3.43
S123T123	↑	MedioVentral Occipital Cortex; caudal lingual gyrus; Ipsilateral	−2.94, −3.44, −3.10
B13	↑	Precuneus; dorsomedial parietooccipital sulcus (PEr); Ipsilateral	−2.7, −3.09
B23	↑	Fusiform Gyrus; lateroventral area37; Contralateral	−3.13, −3.10
B123	↓	Insular Gyrus; dorsal granular insula; Ipsilateral	2.11, 2.11, 2.11
B123	↓	Thalamus; occipital thalamus; Ipsilateral	3.82, 3.56, 3.89
B123	↓	Middle Temporal Gyrus; rostral area 21; Ipsilateral	2.6, 2.58, 2.58

In summary, the nodes with significantly different energy consumption between the groups were largely ipsilateral to the seizure onset zone of the patients (n.b., the left hemisphere in HPs). These regions generally indicated excessive energy use in TLE, with these high-energy nodes primarily involving source or target regions. In contrast, in nodes where the HPs displayed higher-energy, these nodes involved bulk regions (“B”).

3.4. Hubness (degree and eigenvector centrality) analyses

Several regions displayed significantly higher hubness (regional whole brain connectivity) in TLE patients compared to HPs (see Table 4). Degree centrality (DC) of the bilateral globus pallidus was increased in TLE. Eigenvector centrality of ipsilateral pre-motor thalamus and ipsilateral OrG (area 13) was also higher in TLEs, while EC of contralateral ventrolateral middle frontal gyrus was lower in TLE patients compared to HPs.

Table 4.

Hubness and average controllability analyses. Source (S), Target (T), and Bulk (B) ROI’s displaying significant TLE/HP group differences in Degree Centrality (DC), Eigenvector Centrality (EC), and Average Controllablity (AC).

		ROI	p-value	T
B	↑	Basal Ganglia; globus pallidus; contralateral (DC)	0.003	−3.13
B	↑	Basal Ganglia; globus pallidus ipsilateral (DC)	0.005	−2.96
B	↑	Thalamus; pre-motor thalamus; ipsilateral (EC)	0.007	−2.84
B	↓	Middle Frontal Gyrus; ventrolateral area 6; contralateral (EC)	0.001	3.37
ST	↑	Orbital Gyrus; area 13; ipsilateral (EC)	0.009	−2.74
B	↓	Paracentral Lobule 3 (lower limb region); contralateral (AC)	0.008	2.8
B	↓	Postcentral Gyrus (tongue and larynx region); contralateral (AC)	0.006	2.7
B	↑	Parahippocampal gyrus (entorhinal cortex); contralateral (AC)	0.004	−2.9
B	↑	Basal Ganglia; globus pallidus; contralateral (AC)	0.000	−3.4
B	↑	Basal Ganglia; globus pallidus; ipsilateral (AC)	0.001	−3.3

3.5. Average controllability analyses

The average controllability analyses revealed a select number of significant alterations in ROI control capabilities in TLE patients compared to HPs. Specifically, the contralateral lower limb paracentral lobule and contralateral postcentral gyrus (tongue and larynx) exhibited a diminished ability to drive state transitions in TLE patients. In contrast, the contralateral PhG gyrus (entorhinal) and bilateral globus pallidus (basal ganglia; BG) displayed significantly elevated average controllability in patients with TLE, (see Table 4).

3.6. Associations with clinical features of TLE

Our findings showed notable associations between the clinical factors and the regional minimum energy measures (see Table 5). Late age of onset was associated with higher transition energy in the ipsilateral dorsal area of middle frontal gyrus (MFG) and ipsilateral rostrodorsal area of inferior parietal lobule. Earlier age of onset was associated with higher transition energy in the contralateral dorsolateral anterior superior MTG, contralateral ventrolateral inferior temporal gyrus (ITG), and ventromedial putamen. Higher seizure frequency (per month) was observed in association with higher energy consumption in contralateral superior frontal gyrus, contralateral lateral MFG, contralateral trunk region of postcentral lobule, ipsilateral lower limb region of paracentral lobule, and ipsilateral subgenual cingulate gyrus (CG). Lower energy consumption in ipsilateral dorsal dysgranular INS and contralateral lateral occipital cortex was related to higher seizure frequency. Longer duration of illness was associated with higher energy consumption in the ipsilateral lateral amygdala, contralateral lateral MFG, contralateral ventrolateral ITG, and contralateral pregenual CG.

Table 5.

Control theory metrics association with clinical features of TLEs. Correlation Analyses between controllability (average), eigen vector centrality, minimum energy (n.b., significant findings from Tables 3 and 4) and clinical measures (seizure frequency, age of onset, duration of illness). Degree centrality not shown as all correlations were not significant.

Avg. Controllability	ROI	r	p-value
frequency	Paracentral Lobule; area1/2/3; Ipsilateral	0.57	0.003
frequency	Insular Gyrus; dorsal dysgranular; Ipsilateral	−0.53	0.007
Age of onset	Middle Temporal Gyrus; caudal; Ipsilateral	−0.53	0.007
Duration	Parahippocampal Gyrus; TL; Contralateral	0.57	0.003
Duration	Precuneus; area 31 (Lc1); Contralateral	0.64	0.000
Eigenvector Centrality	ROI	r	p-value
frequency	Paracentral Lobule; area1/2/3; Ipsilateral	0.76	0.000
Duration	Orbital Gyrus; orbital area 12/47; Contralateral	0.54	0.005
Duration	Superior Temporal Gyrus; rostral area 22; Ipsilateral	0.55	0.004
Duration	Precuneus; area 31 (Lc1); Contralateral	0.52	0.008
Duration	Thalamus; sensory thalamus; Contralateral	0.53	0.007
Minimum Energy	ROI	r	p
Age of onset	Middle Frontal Gyrus; dorsal area 9/46; Ipsilateral	0.52	0.009
Age of onset	Middle Temporal Gyrus; dorsolateral area37; Contralateral	−0.54	0.005
Age of onset	Middle Temporal Gyrus; anterior superior temporal sulcus; Contralateral	−0.53	0.007
Age of onset	Inferior Parietal Lobule; rostrodorsal area 39(Hip3); Ipsilateral	0.52	0.008
Age of onset	Inferior Temporal Gyrus; ventrolateral area 37; Contralateral	−0.53	0.007
Age of onset	Basal Ganglia; ventromedial putamen; Contralateral	−0.51	0.009
Frequency	Superior Frontal Gyrus; medial area 9; Contralateral	0.59	0.001
Frequency	Middle Frontal Gyrus; lateral area10; Contralateral	0.58	0.002
Frequency	Paracentral Lobule; area1/2/3 (lower limb region); Ipsilateral	0.51	0.009
Frequency	Postcentral Gyrus; area1/2/3 (trunk region); Contralateral	0.54	0.005
Frequency	Insular Gyrus; dorsal dysgranular insula; Ipsilateral	−0.64	0.000
Frequency	Cingulate Gyrus; subgenual area 32; Ipsilateral	0.54	0.005
Frequency	lateral Occipital Cortex; area V5/MT+; Contralateral	−0.52	0.008
Duration	Middle Frontal Gyrus; lateral area10; Contralateral	0.51	0.009
Duration	Inferior Temporal Gyrus; ventrolateral area 37; Contralateral	0.67	0.000
Duration	Cingulate Gyrus; pregenual area 32; Contralateral	0.53	0.006
Duration	Amygdala; lateral amygdala; Ipsilateral	0.56	0.004

Among the controllability, and hubness measures showing reliable HP/TLE differences, relatively few measures displayed significant associations with the TLE clinical features. Specifically, the average controllability of the ipsilateral caudal MTG decreased as the age of epilepsy onset increased. Similarly, the average controllability of the ipsilateral dorsal dysgranular INS decreased with higher monthly seizure frequency. In contrast, the average controllability of the ipsilateral PCL (areas 1/2/3) increased with an increase in seizure frequency. Additionally, a longer duration of illness was associated with higher average controllability of the contralateral posterior PhG and contralateral precuneus (area 31). Lastly, prolonged illness duration was linked to increased eigenvector centrality in the contralateral OrG (A12/47o), contralateral precuneus (area 31), contralateral sensory thalamus, and ipsilateral rostral superior temporal gyrus (STG; area 22). These results highlight the highly selective relationships between clinical variables and the capacity of a region to serve as a specific driver of state transitions.

4. Discussion

Utilizing a regionally unconstrained method for identifying brain state transitions, we discovered three transitions requiring abnormally high levels of energy in TLE. Notably, compared to HPs, no low energy transitions were unique to TLE. In these high energy transitions, the TLE patients primarily utilized and consumed energy from regions located inside the source and target networks. In contrast, healthy individuals, who exhibited significantly lower energy consumption during these transitions, recruited mediator regions outside the source and target networks, such as the ipsilateral thalamus, insula, and middle temporal regions to bridge the source and target networks. This pattern suggested that the TLE patients relied on more regionally restricted and less efficient networks during these select transitions. These data also indicated that the HPs managed these transitions through incremental steps that led to more energy efficiency, whereas TLE patients relied on more abrupt and larger discrete jumps between the source and target networks. The absence of normally mediated and facilitated network reconfigurations in TLE may reflect diminished regional integration capacity, likely resulting from chronic epileptiform activity. It is important to note that the source and target nodes active in TLE during these high energy transitions were mainly ipsilateral to the seizure onset zone, increasing the likelihood that these nodes were impacted by seizures. These group differences in the topological mechanics of the neuronal networks managing these select transitions may well have functional consequences (i.e., cognitive inefficiencies) for the cognitive processes dependent on these pathways.

Our data on average controllability highlighted a few regions in TLE (e.g., bilateral basal ganglia, contralateral paracentral lobule and postcentral gyrus) that showed abnormal capabilities to drive state transitions. To be more specific, average controllability values for the bilateral basal ganglia in TLE indicated this structure possessed a higher than normal capability to initiate “easy” (i.e., low energy) state transitions. Evidence from our lab (Javidi et al., 2024; He et al., 2020) and others (Deransart et al., 1998; Rossi et al., 2022) have shown BG involvement in seizure activity even in the setting of focal TLE. In contrast, in TLE the contralateral paracentral lobule and the postcentral gyrus appeared to be structures with a high capacity to drive “hard” (high energy transitions). Note, none of these regions with abnormal AC values appeared as source or target regions among three transitions we discovered. This discrepancy suggested these regions either did not possess sufficient intrinsic activation (dALff values) to be included in our transition energy analyses, or the full network transitions these structures tend to drive did not overall, at a network level, require different levels of energy consumption in our HP and TLE groups. Our data on hubness indicated that the bilateral BG and ipsilateral (pre-motor) thalamus possessed an abnormally high level of structural connectivity, consistent with the strong epileptogenic potential of these regions (Zhang et al., 2024; He et al., 2020; Lindquist et al., 2023; Piper et al., 2021; Gonzalez et al., 2019). The ipsilateral orbital gyrus also showed high abnormal density, with one structure (i.e., the contralateral middle frontal gyrus) showing reduced hubness compared to HPs. Of these regions the ipsilateral orbital gyrus appeared in all three of our transitions as a target. Thus, our data pointed to a region whose abnormal connection density played a role in forming inefficient, costly transitions. As both abnormal connection density and transition energy represent characteristics that reflect the imprint of seizures on structural pathways, the observed relationship between these features is not surprising. Overall, however, our data revealed numerous dissociations between TLE related abnormalities in regional driving capacity, node density, and the network transitions most costly to TLE patients. Why such dissociations?

At the most basic level, given our method, these dissociations indicated that in the setting of temporal lobe disease, structural abnormalities in regional node density and controllability are not necessarily expressed as abnormalities in intrinsic (functional) activation. These dissociations also suggested that inefficient, energy-demanding state transitions do not necessarily depend on node density or node control capacity. Thus, these dissociations appeared to represent different yet complementary abnormalities. To clarify the nature of these distinct abnormalities we ran correlations with our clinical history measures (age of epilepsy onset, illness duration, and seizure frequency) and observed relatively few significant associations. Higher average controllability (AC) in the ipsilateral middle temporal gyrus was associated with an earlier age of onset, perhaps indicating that in TLE abnormal specific regional drivers of brain states can emerge in the ictal temporal lobe, with these more likely to develop at an early age of seizure onset.

Another noteworthy finding was the differential association between age of seizure onset and energy consumption across regions. In ipsilateral regions, a positive correlation was observed, while contralateral regions exhibited a negative correlation. This pattern may suggest that age of seizure onset alters the distribution of energy resources across the hemispheres. Also, higher frequency of seizures was related to both higher controllability and higher hubness (EC) in the ipsilateral parietal lobule (areas 1/2/3), but with lower controllability in the ipsilateral insula (dorsal, granular). Note, the relationship between seizure frequency and energy consumption across individual parcels was predominantly positive, and mixed in terms of hemispheric laterality (i.e., ipsilateral and contralateral). Interestingly, longer illness duration displayed the most associations, involving both higher controllability and higher hubness. Five out of the six associations involved contralateral nodes, mostly midline cortical (AC and EC of precuneus, EC of orbital) or midline/contralateral subcortical in nature (EC of the thalamus and parahippocampus). These correlations suggested that as the TLE disease progresses there is an increased risk for connectome disruptions of contralateral midline, white matter structures. Our findings also showed a positive correlation between illness duration and energy consumption, with longer illness duration linked to higher energy consumption, findings that argued against the possibility that network entrainment from seizure chronicity would reduce transition energy costs. Interestingly, out of the ten correlations involving the clinical variables only one served as a source or target in the three transitions highlighted in our findings.

While the abnormal transitions we report for TLE can be said to bear relevance to the expression of seizures, our data remains agnostic as to whether these abnormalities are a direct depiction of epileptogenic networks or represent some secondary network effects related to the core neurobiologic impact of TLE. It is also important to note that all our transition energy measures were, by definition, linked to underlying structural and functional abnormalities (dALff). These regions were both functionally more active (abnormal during the intrinsic state) and structurally more inefficient and energy consuming during the brain state transitions we identified. Thus, the regions we reported to be sources and/or targets in the state transitions can be said to have coupled structural and functional abnormalities. Our data also made clear that in TLE such function/structure coupling was strongly ipsilateral to the seizure onset zone (see Table 3; orbital gyrus, occipital cortex, precuneus), yet largely outside the ictal temporal lobe. Accordingly, our findings are consistent with multiple other studies showing that many of the key deleterious impacts of seizures have proximity to the seizure onset zone (i.e., ipsilateral), but are largely extra-temporal in nature (Javidi et al., 2024; Vytvarova et al., 2017; Bernhardt et al., 2016; Bernhardt et al., 2010).

Our transition energy findings have implications that go beyond the nodal, pairwise connectivity implied by our hubness and controllability measures. That is, state transition energy captured higher order interactions among regions as it required regions to be jointly active and reconfiguring to form a new target network. Accordingly, our transition energy data indicated that large-scale synergistic, regional group interactions are characteristic of TLE seizures and their effects (see related work (Santoro et al., 2024)). Another difference was that our hubness and controllability were based upon time invariant structural properties, while transition energy was based upon abnormalities in the time variant activations present in the intrinsic, resting state signal.

In terms of clinical implications, our findings can affect clinical management in five ways. First, detecting the intrinsic brain state transitions that are abnormal provides a new perspective on the etiology of the cognitive, mood, behavioral, sleep and other comorbidities associated with epilepsy. Second, the method we describe (specifically our controllability data) points to the specific brain regions that disproportionately contribute to these abnormal network transitions, providing a potential target for therapeutic interventions. Third, there is increased understanding that even focal epilepsies are a network disorder. As a result, stopping epilepsy will require identifying and targeting not just an epileptogenic region, but also a broader epileptogenic network(s). With this in mind, our findings can be seen as providing a hypothesis, a narrowing down of the brain networks (i.e., brain states) closely aligned epilepsy. Thus, transition energy provides another tool in the armamentarium we utilize in our quest to both identify the epileptogenic network and achieve early detection of its accruing effects. Fourth, there is evidence indicating that in a state of health the brain aims toward energy efficiency in order to function and do its work. Thus, a method that identifies networks operating inefficiently provides a unique window into those networks experiencing functional compromise in need of therapeutic intervention. Fifth, we do not know why individuals with similar brain injuries can display different functional deficits and have different treatment outcomes. Having access to measures that tap alternative or previously unmeasured brain mechanisms may help explain these individual differences. Clearly, however, determining the function and nature of networks displaying abnormal transition energy will require future study.

There are several notable limitations to our findings. State transitions in the setting of epilepsy are complex as there may be changes associated with sleep, medication, and inter-ictal epileptiform activity that are present in the intrinsic resting state. Such transition, if present during scanning, would likely differ between individuals and, therefore, cancel out in our group-level analyses. Nonetheless, a key limitation of our study is the absence of simultaneous EEG recording during the fMRI acquisition, restricting our ability to definitively determine the neurophysiological state of the brain at the time of scanning. In our analysis, we extracted the brain states from rest-state fMRI data, ensuring that the identified states were definitely present in both our TLE and HP samples. However, the limited duration (16 min) and resolution (0.8 s) of the fMRI data certainly hindered our ability to capture interictal seizure network activities, which operate on a much shorter time scale (Khambhati et al., 2024; Chen et al., 2022; Khambhati et al., 2015). It will be important to use methods with higher temporal resolution to map state transitions that are closer to the underlying temporal dynamics of seizures. For instance, it remains to be seen whether brain state transitions more definitively linked to the hypersynchronous activity of seizures are, in contrast to our data, associated with low energy requirements (Barron et al., 2014; Gu et al., 2015), and governed by the well-known initiators of brain state transitions in healthy normals (e.g., precuneus (Hwang et al., 2017)) or the high connection density regions of the brain (e.g., thalamus). It is also important to note that the regional topology of abnormal transitions in TLE, regardless of whether they are metabolically costly or more efficient than those in healthy individuals, will likely vary in different types of epilepsy and the subtypes of TLE (e.g., focal seizures only versus those with additional focal-to-bilateral, tonic-clonic events). Such distinctions warrant further study. Lastly, network control theory metrics such as transition energy assume linear dynamics in the brain, but nonlinear models need to be tested to determine if they capture better the complex dynamics between the individual neurons that are seizing and the state transitions they may initiate or perturb.

Our findings report the abnormal energy state transitions found in TLE after interrogation of a large pool of brain states. We demonstrated that TLE patients relied on a more restricted and energy-intensive network to execute these state transitions, in contrast to the efficient and more externally mediated features of these transitions in healthy individuals. Our findings highlight the potential abnormal network mechanisms that accrue from TLE seizure activity, which may help explain the well-known cognitive inefficiencies and other deficits of the TLE disorder. While nodal connection density and nodal control of transitions usually correlate and overlap with the networks involved in transitions (e.g., either as sources or targets), we showed that this is not the case in TLE. The sources in the abnormal state transitions we identified leveraged brain graph topology to diffuse patterns of activity toward a target state without utilizing connection density or inherent high controllability to minimize the size of their output (energy) to achieve the desired state transition, noting the minimizing such output is the normal state of affairs in the healthy brain (Khambhati et al., 2024). We conclude that the TLE brain does not perturb brain wiring just to generate seizures. Other perturbations occur to support an inefficient switching between brain states, a feature which likely contributes to TLE cognitive and behavioral deficits. Thus, the signs of seizures we see in our data, whether they represent seizure generation, spread, or indirect seizure effects and sequelae, point to a coupling of functional and structural abnormalities that produce inefficient, costly transitions through networks that depend largely on structures outside the ictal temporal lobe and occur without the use of the mediation observed in healthy participants.

We demonstrated that the topology of the TLE connectome is not constrained to optimize efficient signaling and minimize wiring costs in support of the variety of brain state transitions it undertakes. By interrogating for unique energy-based transition dynamics, we provide a framework for identifying the regions/networks that drive the brain to seize, and for exploring the impact of seizures on cognitive and other brain networks, offering a foundation for future research into epilepsy-related dysfunctions.

Supplementary Material

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.nbd.2025.107089.

Data availability

The data that support the findings of this study are available upon request. The data are not publicly available due to privacy or ethical restrictions.