Altered Interictal Bed Nucleus of Stria Terminalis Connectivity in Patients With Temporal Lobe Epilepsy

Abstract

Background and Objectives:

Temporal lobe epilepsy (TLE) is a highly prevalent neurological disorder, with 30–50% of patients developing drug-resistant epilepsy (DRE). Pharmacoresistant seizures remodel critical arousal and respiratory networks, impairing autonomic function and chemoreception, and putting patients at increased risk for adverse respiratory events and sudden unexpected death (SUDEP). Given that the bed nucleus of stria terminalis (BNST) serves as a key relay between brainstem respiratory nuclei and cortical arousal networks, we characterized interictal BNST connectivity alterations in patients with TLE.

Methods:

We conducted a case-control study of patients with drug-resistant TLE evaluated for epilepsy surgery at Vanderbilt University Medical Center, compared to healthy controls with no history of neurological disease. Inclusion criteria for patients included clinical TLE diagnosis and age 18–65 years. Using resting-state fMRI (multiband factor=3, TR=1.3s), we measured functional connectivity (FC) and effective connectivity via Granger causality (GC) between BNST and whole-brain, cortical networks, and brainstem nuclei. Graph theoretical network metrics assessed BNST hub properties. Statistical analyses employed multiple comparison corrections and age-corrected z-scores.

Results:

Thirty-seven patients with TLE (mean age 42.5 ± 12.1 years, 43.2% female) and 33 healthy controls (mean age 36.2 ± 12.0 years, 54.5% female) were studied. Patients demonstrated bilateral reductions in BNST connectivity and causal influence with the whole brain (FC: −2.31 ± 2.87, P = 0.0032; GC: −0.18 ± 0.08, P = 0.0025). While FC showed preserved BNST-brainstem connectivity, GC revealed ipsilateral disruptions in BNST influence over ventral tegmental area (0.023 ± 0.026, P = 0.0067), median raphe (−0.009 ± 0.029, P = 0.0038), and cuneiform nuclei (0.012 ± 0.062, P = 0.0153). Critical respiratory circuits showed divergent reorganization: dorsal raphe-parabrachial complex pathways exhibited 57.2% efferent reduction (P = 0.0028) with 204.6% compensatory afferent increase (P = 0.0020), while dorsal raphe-locus coeruleus circuits showed bilateral deterioration (66.2% reduction DR→LC, P = 0.0015; 56.4% reduction LC→DR, P = 0.0189). Graph analyses confirmed compromised BNST network integration bilaterally (P < 0.05).

Discussion:

Our findings reveal network reorganization in TLE that compromises autonomic and arousal circuit integrity, leading to failed respiratory-autonomic integration that may underlie respiratory vulnerability and increased SUDEP risk; however, we did not directly study SUDEP cases.

Introduction

Epilepsy is one of the most prevalent neurological disorders worldwide, affecting approximately 50 million people, with temporal lobe epilepsy (TLE) being the most common form¹. A particularly debilitating facet of TLE is pharmacoresistance, which manifests in 30–50% of those afflicted by the disease^2,3; indeed, drug-resistant epilepsy (DRE) affects deep subcortical territories where recurrent ictal discharges remodel critical arousal and autonomic circuitry⁴. The mechanisms driving these pathological changes remain elusive, but advances in neuroimaging and electrophysiology have now enabled examination of previously inaccessible subcortical regions. Prior work has outlined the Extended Network Inhibition Hypothesis (ENIH)⁴, which posits that repeated seizure activity at the seizure onset zone (SOZ) spreads to adjacent subcortical structures responsible for neocortical activation and respiratory control, disrupting their vital regulatory functions⁴. Interestingly, these same regions demonstrate abnormal interictal functional connectivity (FC) to their neocortical counterparts^5–7, highlighting their aberrancy even in the absence of an ongoing seizure. In patients with TLE, and especially those with DRE, repeated ictal insults may damage respiratory brainstem nuclei, leading to defective central carbon dioxide (CO₂) chemoreception^8,9, thereby increasing their risk for sudden unexpected death in epilepsy (SUDEP)⁹. SUDEP typically follows a focal to bilateral tonic-clonic seizure (FBTCS) and causes postictal brainstem dysfunction, affecting the central respiratory command localized in the pons and medulla^10,11. When FBTCS does not immediately precede SUDEP, brainstem disturbances may result from cumulative damage to critical respiratory nuclei, as postulated by the ENIH.

SUDEP is the second-leading cause of potential years of life lost among neurological disorders, second only to stroke, and it is the leading cause of death in young patients with DRE who have generalized seizures¹². While SUDEP accounts for 7–17% of deaths among all epilepsy patients, this figure soars to 50% of deaths in DRE patients^13,14. Although our investigation does not directly study SUDEP cases, we focus on characterising the networks that regulate respiratory function and arousal—systems whose dysfunction likely represents the pathophysiological substrate underlying the respiratory vulnerability (e.g., ictal and postictal apnea) seen in patients with TLE^15,16. By examining how these circuits become compromised in TLE, we aim to identify potential mechanisms that may contribute to autonomic instability and SUDEP risk.

Recent evidence has revealed the bed nucleus of the stria terminalis (BNST), a forebrain structure within the extended amygdala, as a critical integration hub regulating hypercapnic arousal responses. The BNST’s strategic anatomical position enables it to receive convergent inputs from multiple brainstem regions containing intrinsically chemosensitive neurons^17,18: parabrachial complex (PBC)¹⁹, locus coeruleus (LC)²⁰, and nucleus tractus solitarius (NTS)²¹, allowing it to process and coordinate chemosensory information to modulate arousal responses. Moreover, stimulation of GABAergic BNST neurons projecting to the VTA rapidly induces sleep-to-wake transitions²². Given that 90% of SUDEP cases occur while patients are sleeping²³, this BNST-VTA circuit may directly regulate hypercapnia-induced arousal, with its degree of dysfunction potentially representing a biomarker for SUDEP risk. Lastly, BNST neurons projecting to the PBC demonstrate hypoexcitability in Dravet Syndrome, an epileptic encephalopathy with increased risk of SUDEP²⁴.

We thus sought to characterize how TLE affects this critical arousal integration hub by examining interictal BNST connectivity alterations using high temporal resolution functional MRI (fMRI). Our approach utilized Granger causality (GC) to assess directed connectivity, which proved essential in revealing cases where functional connectivity (FC) remained intact while directional influence became critically disrupted, a nuanced pathology undetectable through conventional methods. We investigated global BNST connectivity disruptions, characterized bidirectional BNST-cortical rewiring, examined directional pathology between BNST and subcortical arousal nuclei, investigated brainstem respiratory connectivity alterations, and employed graph theoretical metrics to quantify the BNST’s compromised role as a network integration hub in TLE.

Methods

Participants

Thirty-seven adult patients were referred to and evaluated for epilepsy surgery at Vanderbilt University Medical Center between 2018–2024, compared to 33 healthy controls. All patients underwent comprehensive presurgical evaluation including video-EEG monitoring, high-resolution MRI, neuropsychological testing, and when indicated, additional studies such as FDG-PET, ictal SPECT, or invasive monitoring. Clinical TLE diagnosis was established through standard epileptological evaluation including seizure semiology, interictal and ictal EEG findings, and neuroimaging, with epileptogenic zone lateralization determined by consensus of the multidisciplinary epilepsy surgery team incorporating all available clinical, electrophysiological, and neuroimaging data. Patients were classified as having left TLE (n =13, 35.1%), right TLE (n =17, 45.9%), or bilateral/unknown TLE (n =7, 18.9%). Inclusion criteria for patients included age 18–65 years and presurgical evaluation candidacy, while controls had no history of neurological disease. The patient group was older (mean age: 42.5 ± 12.1 years, range: 19–64 years) than controls (mean age: 36.2 ± 12.0 years, range: 22–61 years), with sex distribution including 21 males (56.8%) and 16 females (43.2%) in patients, compared to 15 males (45.5%) and 18 females (54.5%) in controls. To account for age differences, all analyses used age-corrected z-scores.

Standard Protocol Approvals, Registrations, and Patient Consents

All procedures were approved by the Vanderbilt University Institutional Review Board. All participants gave written informed consent for the study.

Imaging

We performed imaging with a Philips Achieva 3T MRI (Philips Healthcare, Best, the Netherlands) and 32-channel head coil. MRI sequences included (1) 3D T1-weighted whole-brain images for tissue segmentation and interparticipant normalization (1.0 × 1.0 × 1.0 mm³), (2) 2D T1-weighted axial images for functional-to-structural image coregistration (1.0 × 1.0 × 2.5 mm³), and (3) 2D T2*-weighted functional MRI, multiband factor = 3 (2.5 × 2.5 × 2.5 mm³), TR = 1300 ms. All resting-state fMRI sessions included instructions to lie at rest with eyes closed for the scan. We performed physiologic monitoring of respiratory and cardiac rates at 496 Hz.

Regions of Interest

Regions of interest (ROIs) included 127 cortical and subcortical regions defined in MNI template space using the CONN toolbox atlas and Harvard-Oxford Cortical Atlas (HOA). Cortical nodes were grouped into networks according to the HOA. For subcortical nuclei, we utilized the Harvard Ascending Arousal Network (AAN) Atlas, which includes detailed segmentation of the following nuclei: Cuneiform/Subcuneiform Nucleus (CSC), Dorsal Raphe (DR), Locus Coeruleus (LC), Median Raphe (MR), Parabrachial Complex (PBC), and Ventral Tegmental Area (VTA). Additionally, we incorporated a specific mask for the bed nucleus of stria terminalis (BNST) defined based on recent high-resolution structural imaging protocols²⁵.

Non-Directed Functional Connectivity Analysis

The fMRI data were preprocessed and analyzed using SPM8 and MATLAB R2021b. Preprocessing steps included correction for respiratory and cardiac noise using RETROICOR, slice timing correction, motion correction, segmentation into white and gray matter and cerebrospinal fluid, and spatial normalization to a standard MNI template. Regional time series were extracted using 127 cortical and subcortical ROIs from multiple atlases defined in MNI space and underwent bandpass filtering between 0.0067 and 0.1 Hz to retain the relevant frequency range characteristic of intrinsic brain network oscillations. Images were then spatially smoothed using a 6mm full-width half-maximum Gaussian kernel.

Functional connectivity was computed using partial Pearson correlation between ROI time series, with 6 motion parameters and mean white matter signal as confounds. Correlations were Fisher z-transformed. Age correction was implemented by building a linear regression model from healthy controls that used age as an independent variable and regionwise connectivity as a dependent variable. After correction, data matrices were arranged according to ipsilateral and contralateral epileptogenic sides rather than left or right hemisphere designations.

Effective Connectivity Analysis

To assess the directional influence between BNST and other brain regions, we implemented Granger causality analysis following the methodology described by Rogers et al. (2010)²⁶. Time series for GC analysis underwent identical preprocessing to the FC analysis. Prior to GC computation, regional time series were further preprocessed with linear detrending and z-score normalization to ensure stationarity. Age correction was applied using the same linear regression model developed from healthy controls. For each pair of time series (BNST and target region), we computed bidirectional Granger causality using a first-order bivariate autoregressive model.

First, for calculating F(X→Y) (BNST influencing target region):

1. A restricted univariate model was fitted: $Y(t)=aY(t-1)+u$ (1)

2. An unrestricted bivariate model was fitted: $Y(t)=aY(t-1)+bX(t-1)+w$ (2)

3. Granger causality was calculated as $F(X\to Y)=log(var(u)/var(w))$ (3)

F(Y→X) (target region influencing BNST) was calculated with variables reversed.

To quantify the net directional influence between BNST and each brain region, we calculated the Granger causality difference: $[\text{F}(\text{X}\to \text{Y})-\text{F}(\text{Y}\to \text{X})]$. A positive value indicated a predominant influence from BNST to the target region, while a negative value indicated a predominant influence in the opposite direction.

Seed-to-Voxel Analysis

We performed seed-based voxelwise analysis using the CONN toolbox. A voxelwise analysis was used to determine the anatomical distribution of FC differences between the BNST and whole-brain voxels in patients with TLE compared to controls. BNST time series were correlated with every brain voxel using the same preprocessing pipeline and covariates as the ROI-to-ROI analysis, with correlations Fisher z-transformed and age-corrected. Group comparisons used two-sample t-tests at each voxel, thresholded at p < 0.05 with cluster-level correction for multiple comparisons.

Percent Change in Information Flow Analysis

To quantify bidirectional connectivity alterations between BNST and cortical networks, we separately analyzed age-corrected, z-scored, and detrended directional GC values. Unlike our net directional influence approach using GC differences, we examined GC values in each direction (ascending and descending) independently. For each direction, statistical significance was assessed using two-sample t-tests on the age-corrected GC values themselves. We then calculated percentage changes from the group means of the age-corrected values for reporting effect sizes. Results were reported as percentage changes relative to control group means with corresponding p-values.

Clinical and Medication Data

Patient demographics, seizure semiology, and epilepsy details—including seizure type and frequency, epilepsy duration, and history of generalized tonic-clonic seizures—were obtained from epileptologist clinical assessments. Antiseizure medication (ASM) load and seizure frequency data were collected, and correlation testing revealed no significant correlations with BNST connectivity measures.

Statistical Analyses

Parametric tests compared FC and GC between groups. Outliers were excluded using the interquartile range method (1.5×IQR). Two-sample t-tests assessed group differences, with separate analyses for ipsilateral and contralateral BNST connections. Multiple comparison correction used Bonferroni-Holm correction for hypothesis-driven percent change analyses targeting specific directional connectivity pathways, where stringent Type I error control was essential. False discovery rate (FDR) correction was applied to exploratory analyses involving multiple brain regions or network connections. Statistical analyses were performed with MATLAB R2021b and significance was defined as P < 0.05 for all tests.

Graph Theory Analyses

Graph theoretical analyses quantified BNST network integration using binary adjacency matrices created from age-corrected FC. For each subject, the top 15% of connections were retained and others were set to 0, maintaining only strongest connections. We measured Flow Coefficient (information transfer efficiency), Betweenness Centrality (network bridge importance), and Clustering Coefficient (local neighbor connectivity that bypasses the BNST).

Results

Participants

Thirty-seven patients with TLE and 33 healthy controls were included in the study. Demographic and clinical data are summarized in Table 1. Patients were classified as having left TLE (n = 13, 35.1%), right TLE (n = 17, 45.9%), or bilateral/unknown TLE (n = 7, 18.9%). The patient group was older (mean age: 42.5 ± 12.1 years, range: 19–64 years) than controls (mean age: 36.2 ± 12.0 years, range: 22–61 years). Sex distribution included 21 males (56.8%) and 16 females (43.2%) in patients, compared to 15 males (45.5%) and 18 females (54.5%) in controls. To account for age differences, all analyses used age-corrected z-scores. Comprehensive correlation analyses revealed that antiseizure medication load, disease duration, and seizure frequency showed weak correlations with BNST connectivity measures (r = −0.230 to 0.365, all P > 0.05).

Table 1.

Subject Demographics and TLE Side

Demographics	Patients (n = 37)	Controls (n = 33)	P value
Gender
Male	21 (56.8%)	15 (45.5%)	0.345
Female	16 (43.2%)	18 (54.5%)
Age (years)
Mean ± SD	42.5 ± 12.1	36.2 ± 12.0	0.0321*
Range	19–64 (45)	22–61 (39)
TLE Side
Left	13 (35.1%)	N/A
Right	17 (45.9%)	N/A
Other/Unknown	7 (18.9%)	N/A

Demographic and clinical data for temporal lobe epilepsy patients and healthy controls included in the study. Data are presented as number (percentage) for categorical variables and mean ± standard deviation for continuous variables. Statistical comparisons were performed using chi-squared test of the proportions between groups for categorical variables and independent t-tests for continuous variables. N/A = not applicable; SD = standard deviation; TLE = temporal lobe epilepsy.

^*P < 0.05.

BNST Global Connectivity in Patients with TLE

Given the BNST’s purported role as a critical forebrain relay node coordinating protective arousal responses between chemosensitive brainstem nuclei and cortical attention networks, we first interrogated its global connectivity architecture to quantify overall integrative capacity before dissecting circuit-specific pathology. Patients with TLE demonstrated significant bilateral reductions in BNST connectivity, with ipsilateral BNST showing diminished FC with the whole brain (0.00 ± 2.98 controls mean ± SD, −2.31 ± 2.87 patients, P = 0.0032, two-sample t test; Figure 1A) and similarly compromised contralateral patterns (eFigure 1). Directional analyses using GC revealed more disruptions, with ipsilateral BNST exhibiting reduced net causal influence over global brain networks (−0.12 ± 0.07 controls, −0.18 ± 0.08 patients, P = 0.0025, two-sample t test; Figure 1B). Complementing these global findings, seed-to-voxel analyses revealed widespread connectivity reductions between ipsilateral BNST and temporal, thalamic, and brainstem territories (all P < 0.05, cluster-corrected; Figure 1C), with contralateral BNST demonstrating parallel deterioration (eFigure 1).

Figure 1. Patients with TLE demonstrate bilateral reduction in BNST FC and GC to whole brain.

(A) Ipsilateral BNST to whole brain connectivity shows significantly reduced FC in patients after outlier removal (P = 0.0032). (B) Ipsilateral BNST to whole brain Granger causality shows significantly more negative net directionality in patients (P = 0.0025). (C)Voxel-wise functional connectivity analysis reveals significant alterations in BNST connectivity patterns. Ipsilateral BNST (left panel is left TLE; right panel is right TLE) shows marked reduction in FC with temporal regions, thalamus, and brainstem structures (all P < 0.05, cluster corrected), color bar indicates average connectivity differences between controls and patients with cluster correction across subjects, with blue indicating decreased connectivity and red indicating increased connectivity in patients relative to controls. * P < 0.05, ** P <0.01.

Bidirectional BNST-Cortical Communication

Having observed that the BNST is poorly functionally connected to the whole brain, and that it appears to lose causal influence on other brain regions in TLE, we next sought to determine the percentage of information lost between the BNST and specific cortical networks. We implemented a percent change in information flow analysis using GC measures to quantify these directional disruptions. This approach provides a standardized method to represent TLE connectivity disruptions as percentage reductions relative to healthy controls, allowing for direct comparisons across different network connections.

After applying multiple comparisons correction, our analysis revealed reductions in ascending regulatory control from the BNST to multiple cortical networks in patients with TLE compared to healthy controls (Figure 2A). Ipsilateral BNST exhibited significant decreases in information outflow to Frontoparietal (63.4%, P = 0.0242, two-sample t test), Somatomotor (65.4%, P = 0.0116, two-sample t test), and Ventral Attention (VAN) networks (66.6%, P = 0.0044, two-sample t test). Contralateral BNST showed similar reductions to Frontoparietal (73.0%, P = 0.0061, two-sample t test), Somatomotor (70.1%, P = 0.0086, two-sample t test), and Ventral Attention (72.9%, P = 0.0025, two-sample t test) networks, plus reduced connectivity to the Visual network (Figure 2A). Examining the opposite direction revealed increased top-down control (Figure 2B), with the Dorsal Attention Network (DAN) showing increased information flow to both ipsilateral (46.1% increase, P = 0.0492, two-sample t test) and contralateral BNST (50.5% increase, P = 0.0255, two-sample t test) (Figure 2B).

Figure 2. Bidirectional Disruption in BNST-Cortex Communication Pathways in TLE.

(A) Patients with TLE exhibit disruption in ascending control from BNST to cortical networks. Both ipsilateral and contralateral BNST show substantial reductions in regulatory output to critical attention-related networks: Frontoparietal (63.4%, P = 0.0242), Somatomotor (65.4%, P = 0.0116), and Ventral Attention (66.6%, P = 0.0044) networks. Contralateral BNST also showed pronounced reductions in ascending control to Frontoparietal (73.0%, P = 0.0061), Somatomotor (70.1%, P = 0.0086), and Ventral Attention (72.9%, P = 0.0025) networks. (B) Patients with TLE demonstrate top-down cortical control over the BNST. Ipsilateral BNST shows increased input from the Dorsal Attention (46.1% increase, P = 0.0492) network. Contralateral BNST demonstrates similar pattern with significant increase from Dorsal Attention network (50.5% increase, P = 0.0255). Values represent percent decrease or increase in Granger causal influence from cortical regions to BNST in patients compared to controls. P-values reflect Bonferroni-Holm correction for multiple comparisons, * P < 0.05, ** P <0.01.

BNST Influence on Ascending Arousal Network Nuclei

After quantifying BNST-cortical disruptions, we investigated BNST communication with ascending arousal network (AAN) nuclei that regulate consciousness and respiration. Notably, FC analyses revealed no significant disruptions between ipsilateral BNST and brainstem nuclei after FDR correction (Figure 3A). However, when we applied directional GC to the same circuits, our analyses demonstrated disrupted information flow from ipsilateral BNST to three key arousal centers after FDR correction: VTA (0.049 ± 0.042 controls, 0.023 ± 0.026 patients; P = 0.0067, two-sample t test), MR (0.026 ± 0.054 controls, −0.009 ± 0.029 patients; P = 0.0038, two-sample t test), and CSC (0.050 ± 0.056 controls, 0.012 ± 0.062 patients; P = 0.0153, two-sample t test). Quantification of these disruptions revealed reductions in information flow from ipsilateral BNST to these critical brainstem nuclei (Figure 3C). Unlike the cortical-BNST pathways where we observed significant increases in descending control, the brainstem nuclei showed no evidence of upregulation to the BNST (Figure 3D). In contrast to the ipsilateral BNST findings, contralateral BNST maintained normal FC and GC with these same brainstem nuclei (eFigure 2).

Figure 3. Standard Functional Connectivity Masks Directional Pathology Between BNST and Brainstem Arousal Nuclei in TLE.

Patients with TLE show preserved FC but significantly disrupted causal influence between BNST and brainstem nuclei. (A) Standard FC analysis shows no significant differences after FDR correction. (B) Granger causality reveals significant reductions in ipsilateral BNST’s influence on key arousal nuclei: VTA (P = 0.0067), MR (P = 0.0038), and CSC (P = 0.0153). (C) Ipsilateral pathways show substantial deficits: VTA (52.4% reduction, P = 0.0067), MR (135.1% reduction, P = 0.0038), and CSC (76.3% reduction, P = 0.0153). (D) Brainstem nuclei show no compensatory signaling to BNST: VTA (4.8% decrease, P = 0.8871), MR (39.3% increase, P = 0.3768), and CSC (15.3% increase, P = 0.6884), * P < 0.05, ** P <0.01.

Brainstem Respiratory Network Connectivity Alterations

After examining BNST connectivity with AAN nuclei, we next sought to examine the FC and GC of respiratory nuclei critical for maintaining vigilance during periods of physiological challenge. We specifically investigated connectivity between the DR, a major serotonergic nucleus essential for respiratory chemosensitivity^10,27–29, and two key respiratory regulatory centers: the PBC^19,24,30,31 and LC^32,33. Similar to our BNST findings, standard FC analysis revealed no differences between patients with TLE and healthy controls in DR-PBC (P = 0.7846) and DR-LC (P = 0.9106) connectivity after FDR correction (Figure 4A). However, when we applied GC analysis to these same circuits, we uncovered pathology that remained undetected by conventional FC measures (Figures 4B and 4C). We observed reduction in outflow from DR to PBC (0.024 ± 0.023 controls, 0.010 ± 0.009 patients; 57.2% reduction, P = 0.0028, two-sample t test, FDR correction) coupled with an increase in signaling from PBC to DR (0.011 ± 0.013 controls, 0.034 ± 0.035 patients; 204.6% increase, P = 0.0020, two-sample t test, FDR correction; Figure 4B). In contrast, the DR to LC effective connectivity not only showed impairment (0.062 ± 0.068 controls, 0.021 ± 0.021 patients; 66.2% reduction, P = 0.0015, two-sample t test, FDR correction), but this deficit was accompanied by a significant reduction in the reverse direction from LC to DR (0.032 ± 0.039 controls, 0.014 ± 0.017 patients; 56.4% reduction, P = 0.0189, two-sample t test, FDR correction; Figure 4C).

Figure 4. Critical Respiratory Control Circuits Show Preserved Functional Connectivity Despite Profound Directional Signaling Deficits in TLE.

(A) Standard functional connectivity analysis shows no significant differences between patients and controls for both DR-PBC connectivity (P = 0.7846) and DR-LC connectivity (P = 0.9106). (B) Granger causality analysis reveals significant bidirectional disruption between DR and PBC in patients with TLE. DR→PBC causal influence is markedly reduced (57.2% reduction, P = 0.0028), while PBC→DR signaling shows substantial increase (204.6% increase, P = 0.0020). (C) Granger causality analysis between DR and LC demonstrates significant bidirectional deficits in patients with TLE. Both DR→LC (66.2% reduction, P = 0.0015) and LC→DR (56.4% reduction, P = 0.0189) causal influences are significantly diminished, * P < 0.05, ** P < 0.01.

Graph Theoretical Analysis of BNST Network Properties

We next characterized the BNST’s altered role within the broader brain network architecture using graph theoretical analyses. Patients with TLE demonstrated significantly reduced BNST flow coefficient bilaterally (ipsilateral 0.0304 ± 0.0084 controls, 0.0250 ± 0.0100 patients; P = 0.0268; contralateral 0.0304 ± 0.0087 controls, 0.0252 ± 0.0101 patients; P = 0.0343, two-sample t test), indicating impaired information transmission efficiency (Figure 5A). Similarly, BNST betweenness centrality was significantly reduced bilaterally (ipsilateral 3.8307 ± 1.0202 controls, 3.2069 ± 1.2370 patients; P = 0.0338, two-sample t test; contralateral 3.8118 ± 0.9969 controls, 3.2181 ± 1.2506 patients; P = 0.0424, two-sample t test), suggesting diminished importance as a network bridge (Figure 5B). Lastly, we examined how this loss of global integration affects local network organization around the BNST. Clustering coefficient was significantly increased bilaterally (ipsilateral 0.9696 ± 0.0084 controls, 0.9750 ± 0.0100 patients; P = 0.0268, two-sample t test; contralateral 0.9696 ± 0.0087 controls, 0.9748 ± 0.0101 patients; P = 0.0343, two-sample t test), indicating that in TLE, the BNST’s neighbors establish direct connections that bypass the BNST as an intermediary (Figure 5C).

Figure 5. Graph Theory Analyses Reveal Compromised BNST Network Integration in TLE.

Patients with TLE demonstrate bilateral BNST integration deficits with compensatory contralateral hyperconnectivity to cortical networks. (A) BNST flow coefficient is significantly reduced bilaterally in patients with TLE: ipsilateral BNST (P = 0.0268) and contralateral BNST (P = 0.0343). (B) BNST betweenness centrality shows bilateral reductions in patients with TLE: ipsilateral (P = 0.0338) and contralateral (P = 0.0424), indicating diminished importance as a network integration hub. (C) BNST clustering coefficient is significantly increased bilaterally in patients with TLE: ipsilateral BNST (P = 0.0268) and contralateral BNST (P = 0.0343), indicating increased local clustering among neighbouring nodes, * P < 0.05.

Discussion

Our findings point to a pattern of network reorganization in TLE that compromises autonomic and arousal circuit integrity. Figure 6 integrates our connectivity results to illustrate how disrupted BNST and subcortical brainstem connectivity compromises their roles in arousal maintenance and chemosensation, thereby leading to failed respiratory-autonomic integration that may underlie respiratory vulnerability and increased SUDEP risk. Specifically, bilateral reductions in BNST FC and GC with the whole brain suggest a fundamental loss in its regulatory capacity, with severe disruptions in attentional networks supporting the BNST’s crucial role in mediating arousal responses^22,24,34. This dysregulation manifests through bidirectional rewiring with compromised ascending control over the VAN counterbalanced by enhanced top-down signaling from the DAN. The distinct disruption patterns between ventral and dorsal attention networks align with each network’s established role³⁵. The VAN primarily mediates stimulus-driven attention and serves as an alerting system for detecting respiratory distress signals³⁵. In TLE, this disruption occurs specifically in its natural bottom-up direction (BNST→VAN), potentially compromising respiratory distress detection. Conversely, the DAN demonstrates increased connectivity to the BNST in its natural top-down direction³⁵, suggesting compensatory strengthening of conscious attentional control over this critical forebrain relay node.

Figure 6. Extended Network Inhibition Hypothesis: A Working Model of Compromised Arousal Circuitry in TLE.

(A) Under normal conditions (left), the BNST maintains robust regulatory control over cortical networks (solid orange arrows) and key brainstem arousal nuclei (solid purple arrows), integrating important chemosensory information to modulate arousal. (B) In the interictal period, accumulated seizure-induced damage from the epileptogenic zone (SOZ, red) disrupts BNST connectivity, resulting in attenuated ascending control to cortical attention networks (dashed orange arrows) and compromised ipsilateral communication with brainstem respiratory nuclei (dashed purple arrows). Our findings reveal compensatory mechanisms including enhanced descending signaling from dorsal attention networks to BNST (thick green arrow) and differential reorganization of subcortical pathways, with the DR-PBC circuit exhibiting enhanced compensatory connectivity (thick purple arrows) while DR-LC circuits demonstrate complete bilateral deterioration with no adaptive response. This pattern of failed integration may underlie the respiratory vulnerability observed in patients with TLE at risk for SUDEP. VTA = ventral tegmental area; DR = dorsal raphe; LC = locus coeruleus; PBC = parabrachial complex.

While FC showed preserved connectivity between BNST and brainstem nuclei, directional analyses revealed disruptions in ipsilateral BNST’s causal influence over key arousal-regulating centers including the VTA, MR, and CSC. The disruption of BNST-VTA connections is particularly alarming given recent evidence demonstrating that optogenetic stimulation of GABAergic BNST neurons that project directly to the VTA induces wakefulness from sleep²². The observed deterioration of the BNST’s regulatory control over the VTA in our TLE cohort may therefore directly compromise patients’ ability to mount appropriate arousal responses to hypercapnic conditions during sleep, potentially explaining the fatal respiratory suppression that characterizes SUDEP cases. Similarly, disrupted BNST influence over the MR—which regulates REM sleep and integrates hypothalamic signals affecting arousal³⁶—and the CSC—which coordinates autonomic responses to respiratory stressors as part of the mesencephalic locomotor region³⁷—likely further impairs protective arousal mechanisms during nocturnal respiratory challenges in patients with TLE. Notably, contralateral BNST demonstrated no significant disruptions in either FC or GC to these same brainstem nuclei, highlighting the lateralized nature of subcortical network disruptions in TLE.

We next examined connectivity patterns within these brainstem respiratory control circuits; similarly, critical respiratory control nodes demonstrated preserved FC despite directional signaling deficits. The DR-PBC circuit, where serotonergic DR neurons normally maintain respiratory output and regulate arousal responses to hypercapnia³⁸,showed a compensatory increase in reverse signaling (PBC→DR) despite efferent disruption (DR→PBC). In contrast, the DR-LC pathway, where noradrenergic LC neurons activate brainstem respiratory centers that modulate arousal and respiratory function³³, exhibited complete bilateral deterioration with no adaptive response. These differential patterns of network reorganization suggest that some respiratory circuits maintain functional capacity through compensatory mechanisms while others show more extensive disruption, potentially contributing to the complex pathophysiology underlying respiratory dysfunction in TLE.

Our findings regarding the circuit vulnerabilities in respiratory control pathways are corroborated by multiple lines of evidence from previous investigations in epilepsy patients. Recent work has demonstrated markedly suppressed firing of serotonergic neurons in the medullary raphe during and after seizures, coinciding with significant decreases in respiratory rate, tidal volume, and minute ventilation³⁹—a dysfunction that directly parallels our observations of compromised DR→LC and DR→PBC pathways in patients with TLE. Moreover, our proposed ENIH model (Figure 6) gains compelling validation from research demonstrating that ictal apnea begins quickly after seizure spread to the amygdala⁴⁰; specifically, our findings suggest that repeated ictal insults damage these critical arousal and respiratory structures, putting patients with TLE at increased risk of arousal and autonomic failure. Our findings are further supported by work demonstrating that focal epilepsy produces significant brainstem atrophy that destabilizes brainstem-brain interactions, leading to impaired cortical activation and autonomic control⁴¹. Together, these convergent findings support our observation that TLE disrupts the integrative role of the BNST as key relay node between brainstem respiratory nuclei and cortical arousal networks.

Graph theoretical analyses revealed that TLE alters the BNST’s integration within the global brain network. These analyses demonstrated reduced flow coefficient and betweenness centrality in both ipsilateral and contralateral BNST, indicating impaired capacity to efficiently process and relay information. Complementing these findings, significantly increased clustering coefficient in patients with TLE demonstrated that the BNST’s immediate neighbors establish direct connections that bypass the BNST as an intermediary. This elevated local clustering is consistent with previous findings showing increased clustering coefficient in limbic structures, the insula, superior temporal regions, and thalamus in TLE⁴², and reflects a fundamental shift from hub-dependent global integration to direct neighbor-to-neighbor communication^43,44. Put together, these results suggest that the BNST transitions from a critical hub facilitating information flow between distant brain regions to a bypassed node whose compensatory local clustering cannot substitute for the loss of global coordination capacity essential for arousal responses during respiratory challenges.

The identification of these critical circuit vulnerabilities was enabled by our application of GC, a methodological approach that warrants attention given ongoing discussions about its suitability for fMRI data. While FC quantifies temporal correlations between regions, GC provides crucial insights into directional information flow by determining whether activity in one region predicts future activity in another beyond what the second region’s own past would predict. This distinction proved essential in our study, revealing networks where correlation strength remained intact while directional influence became critically disrupted—a nuanced pathology undetectable through conventional methods. We addressed methodological concerns raised by Wen and colleagues⁴⁵ by implementing comparative analyses rather than focusing on absolute causality magnitudes, examining patient-control differences, and comparing bidirectional influences, with careful preprocessing ensuring time series stationarity through detrending and z-scoring. GC differences [F(X→Y) - F(Y→X)] further mitigated concerns by focusing on directional asymmetry rather than absolute values. While hemodynamic response function differences between brainstem and cortex represent a legitimate methodological challenge, our findings within brainstem nuclei likely reflect true neuronal signaling alterations rather than vascular confounds for two reasons: these nuclei occupy limited spatial volume within the brainstem, minimizing regional hemodynamic variations, and share vascular supply predominantly from the basilar artery and its branches, creating a relatively homogeneous perfusion environment⁴⁶.

Given its potential role in dysfunctional hypercapnia-induced arousal, the BNST represents a particularly promising target for neuromodulatory intervention in patients at elevated risk for SUDEP. Clearly, human electrophysiological studies and post-mortem histological examination of BNST tissue from SUDEP patients are needed to further confirm the BNST’s critical role in these respiratory and arousal processes. If the BNST’s function as a critical arousal integration center is compromised in humans with TLE, as our study demonstrates and as established murine models have shown through disrupted BNST signaling that severely impairs hypercapnic arousal responses⁴⁷, then targeted intervention could potentially restore this vital protective mechanism. BNST deep brain stimulation has already demonstrated both safety and efficacy for treatment-resistant obsessive-compulsive disorder⁴⁸, providing a translational pathway for clinical implementation in appropriately selected epilepsy patients. Alternatively, rather than targeting the BNST directly, precision neuromodulatory interventions could selectively enhance signaling within circuits demonstrating failed compensatory adaptation; this circuit-specific therapeutic strategy would represent a significant advancement in personalized epilepsy surgery, potentially restoring critical arousal and respiratory protective mechanisms in patients at heightened risk for autonomic collapse and SUDEP.

Our investigation of BNST connectivity in TLE provides compelling evidence for fundamental reorganization of arousal and respiratory circuitry that may directly contribute to respiratory vulnerability and SUDEP pathophysiology. Our findings align with emerging evidence on the role of the BNST and monoaminergic neurons in respiratory regulation, where dysfunction of serotonergic, noradrenergic, and dopaminergic systems can critically impair both arousal responses to hypercapnia and protective breathing reflexes. Our observed disruptions in BNST-VTA, DR-PBC, and DR-LC signaling parallel the monoaminergic circuit vulnerabilities implicated in sleep-disordered breathing and attenuated chemoreceptive reflexes. These findings advance our understanding of neurobiological mechanisms underlying respiratory vulnerability in TLE and suggest that tailored neuromodulatory approaches targeting the BNST and compromised circuits may represent promising patient-specific therapeutic strategies to mitigate SUDEP risk in carefully selected patients with refractory epilepsy.

Limitations

We did not study SUDEP cases, limiting causal inferences regarding SUDEP pathophysiology. Additionally, language dominance and severity of hippocampal sclerosis were not accounted for. Our resting-state approach cannot capture dynamic changes during sleep or seizure states when SUDEP risk is highest, and we lack measures of BNST neurochemistry, neurotransmitter receptor expression, or connectivity with specific chemosensitive brainstem populations. Human electrophysiological studies and post-mortem histological examination of BNST tissue from SUDEP patients are needed to confirm the BNST’s role in hypercapnic arousal. Technical limitations in brainstem imaging, including susceptibility artifacts, physiological noise, and spatial resolution constraints, may affect signal quality and nuclear delineation despite preprocessing. While our multiband acquisition achieved TR=1.3s, higher temporal resolution would provide better estimates of directional connectivity. Future investigations employing stereoelectroencephalography during sleep, objective respiratory monitoring, and prospective validation in cohorts with documented respiratory events will be essential before clinical implementation of biomarkers for SUDEP risk stratification.

Conclusions

We documented alterations in BNST connectivity in TLE that may contribute to arousal dysfunction, respiratory vulnerability, and SUDEP. This dysregulation manifests through bidirectional rewiring with compromised ascending control over the VAN counterbalanced by enhanced top-down signaling from the DAN. While FC showed preserved connectivity between BNST and brainstem nuclei, GC analyses revealed disruptions in ipsilateral BNST’s influence over brainstem respiratory and arousal centers. Our findings align with models demonstrating hypoexcitable BNST neuronal projections in epileptic encephalopathies for SUDEP. Graph theoretical analyses confirmed the BNST’s compromised network integration in TLE, with the BNST demonstrating diminished capacity to relay information between respiratory and arousal systems. Seizure-induced damage to arousal circuits likely contributes directly to respiratory vulnerability underlying sudden death in epilepsy and provides evidentiary support for BNST neuromodulation as preventive intervention in high-risk SUDEP patients.

Supplementary Material

Supplementary Figure 2. eFigure 2. Contralateral BNST shows preserved functional connectivity and Granger causality with brainstem arousal nuclei in TLE.

(A) Functional connectivity analysis (measured as z-scores) shows no significant differences between patients with TLE and healthy controls in connectivity between contralateral BNST and brainstem nuclei, including cuneiform/subcuneiform nucleus (CSC), median raphe (MR), and ventral tegmental area (VTA). (B) Granger causality analysis demonstrates no significant disruptions in directional information flow from contralateral BNST to the same key brainstem nuclei. Unlike the ipsilateral BNST, which showed significant reductions in causal influence over these arousal centers, contralateral BNST maintains relatively normal communication patterns with these critical respiratory and arousal-regulating nuclei. This pattern suggests that pathological changes in TLE primarily affect the hemisphere ipsilateral to seizure onset, with relative preservation of contralateral regulatory circuits.

Supplementary Figure 1. eFigure 1. Patients with TLE demonstrate bilateral reduction in contralateral BNST FC and GC to whole brain.

(A) Contralateral BNST to whole brain connectivity shows significantly reduced FC in patients (t-test, *** P < 0.001). (B) Contralateral BNST to whole brain Granger causality demonstrates significantly more negative net directionality in patients (** P < 0.01), indicating reduced outward information flow. (C) Voxel-wise functional connectivity analysis reveals significant alterations in contralateral BNST connectivity patterns. Left panel shows left BNST in right TLE; right panel shows right BNST in left TLE. Both demonstrate marked reduction in FC with thalamic regions bilaterally, temporal regions (including inferior and middle temporal gyri), precuneus, superior frontal regions, and insular cortex (all P < 0.05, cluster corrected). Color bar indicates average connectivity differences between controls and patients with cluster correction across subjects, with blue indicating decreased connectivity and red indicating increased connectivity in patients relative to controls. ** P < 0.01, *** P < 0.001.

Data Availability Statement

Data may be made available upon reasonable request.