Thalamo‐Lesional Connectivity Signatures of Bilateral Tonic–Clonic Seizures in Focal Cortical Dysplasia‐Related Epilepsy

ABSTRACT

Objectives

Focal cortical dysplasia (FCD) is the most common etiology of drug‐resistant epilepsy in children. Focal to bilateral tonic–clonic seizures (FBTCS) mark a high risk of drug‐resistant epilepsy and involve thalamocortical circuitry in their generation and propagation. Using FCD as a model of focal epilepsy, we examined the integrity of thalamocortical circuitry to characterize thalamo‐lesional connectivity signatures of FBTCS.

Methods

Forty‐three patients (14 with FBTCS) range 5–22 years, and 116 age‐ and sex‐matched healthy controls were included. A network‐driven thalamic functional parcellation was developed in healthy controls using Yeo 7‐network atlas. We calculated z‐scored thalamo‐lesional functional connectivity between thalamic functional parcels and dysplasia masks in reference to controls, which was used to classify the presence of FBTCS. Clinical associations of thalamo‐lesional connectivity were evaluated.

Results

Z‐scored thalamo‐lesional connectivity classified the presence of FBTCS with a median accuracy of 0.72. The thalamic parcel preferentially connected to the somatomotor‐network (SMN), comprising ventral‐posterolateral, pulvinar‐medial and intralaminar nuclei, contributed most to classification. Lower ipsilateral SMN‐thalamus connectivity was related to postsurgical seizure freedom and preserved ipsilateral thalamic volume.

Significance

Thalamocortical circuitry was examined between the seizure generator (lesion) and the amplifier (thalamus) in FCD patients. FBTCS are associated with thalamo‐lesional connectivity abnormalities that correlate to postsurgical outcomes, and demonstrate the SMN‐thalamus as a primary gatekeeper for secondary generalization. The successful implementation of our classifier suggests that these thalamo‐lesional signatures can serve as a potential neuromarker for seizure phenotyping, marking a step toward network‐informed, personalized treatment.

1. Introduction

Focal to bilateral tonic–clonic seizures (FBTCS) are a major risk factor for mortality; their presence indicates 27 times risk of Sudden Unexplained Death in Epilepsy [1]. Focal cortical dysplasia (FCD) is the most common etiology of surgically‐treatable, drug‐resistant epilepsy (DRE) in children [2, 3, 4]. FCD lesion‐network interactions may explain key clinical characteristics in FCD‐related epilepsy such as age of onset [5, 6], drug resistance [7, 8], neuropsychological deficits [9], and surgical outcomes [7, 10]. The presence of FBTCS in FCD is linked to DRE [8, 11] and worse surgical outcomes [12].

The thalamus is a deep, widely‐connected, multimodal integrative relay center [13] that has been proposed as a critical mediator for the FBTCS epileptogenic network [14]. The majority of basal ganglia output is inhibitory to the thalamus and striato‐thalamic connectivity disruptions may facilitate the propagation of ictal activity to widespread cortical networks for FBTCS [15]. Single‐photon emission computed topography shows thalamic hyperperfusion at the time of secondary generalization not present at seizure onset, demonstrating its key role in the genesis of FBTCS [16]. Similarly, stereo‐electroencephalography shows 100% of captured FBTCS had ipsilateral thalamic involvement [17]. Thalamic atrophy was found in patients with FBTCS, with the atrophy being more pronounced in the ipsilateral medial thalamus for patients with uncontrolled FBTCS than those with controlled FBTCS [18]. In addition to structural abnormalities, disrupted thalamocortical structural [19, 20] and functional connectivity [20, 21, 22, 23] are also associated with FBTCS. For example, disruptions in hippocampal to medial‐thalamic structural connectivity are observed in patients with FBTCS as compared to controls [20]. Moreover, ipsilateral medial‐dorsal thalamic nuclei are more uniformly functionally coupled with the rest of the brain in patients with FBTCS, suggesting greater thalamic‐mediated cortico‐cortical communication potentially facilitating secondary seizure generalization [22].

Despite advances in understanding FBTCS, gaps in knowledge remain. First, while epilepsy with FBTCS [24] is increasingly recognized as a brain network disorder, there is still a lack of understanding of the networks underlying thalamic facilitation of FBTCS. Second, while a few recent studies [22, 25] focused on the functional connectivity of thalamic subdivisions in FBTCS, no studies have determined how seizure onset zones interact with thalamic functional parcellations in FBTCS.

Using FCD as a natural model of focal epilepsy with well‐characterized seizure onset zones, we aimed to delineate the functional pathways between the known seizure onset zone (FCD lesion) and the seizure amplifier (thalamus), while capitalizing on a network‐driven thalamic functional parcellation scheme to gain nuanced network‐level insights. We adopted a novel technique of thalamic functional parcellation using the Yeo 7‐network atlas [26], enabling the investigation of thalamic functional subdivisions in context of widespread cortical functional networks. This allowed the identification of network neuromarkers of FBTCS in FCD‐related epilepsy. We utilized seed‐based resting‐state fMRI (rsfMRI) to characterize z‐scored thalamo‐lesional connectivity, which measures the functional connectivity deviation between FCD lesions and thalamic functional parcels from expected in healthy controls [7, 10]. Using this technique, we aimed to quantify how seizure onset zones interact with thalamic functional parcels in FBTCS. Then, we trained an artificial intelligence classifier to predict patients with FCD‐FBTCS from their fMRI‐derived thalamo‐lesional functional connectivity deviations and evaluated factors associated with those deviations.

2. Methods

2.1. Participants

This was an IRB‐approved retrospective cohort study performed at Children's National Hospital in Washington, DC. Surgical patients were identified from the epilepsy surgery database using the following inclusion criteria: 3T MRI‐confirmed FCD from January, 2011 to January, 2024; rsfMRI scans done at ages 5–22 years; at least 18 months of documented follow‐up after the MRI. Patients were excluded if there was dual pathology, hemimegalencephaly or tuberous sclerosis complex present in the imaging or patient history. All patients have documented FBTCS by clinical history as reported by epileptologists and/or confirmed by neurophysiology.

2.2. Imaging Acquisition

High‐resolution epilepsy‐protocol presurgical structural and functional MRI was obtained as part of routine clinical care on a 3T General Electric Discovery MR750 scanner with 8‐ and 32‐channel head coil from Children's National Hospital and reported by board‐certified pediatric neuroradiologists with experience in pediatric epilepsy surgical evaluations and reviewed independently by NTC. rsfMRI for healthy controls was obtained via existing data from our center, which was supplemented by 100 randomly‐chosen, age‐ and sex‐matched participants from Human Connectome Project‐Development (HCP‐D) project. To control for site differences, NeuroCombat [27] was used to harmonize 100 HCP‐D to the 16 controls from our site. A detailed description of imaging protocol can be found in our previous work [7].

2.3. Image Preprocessing

The data were preprocessed using fMRIprep v24.0.0 [28]. Briefly, the functional data were co‐registered to T1w structural data and then both were normalized to the Montreal Neurological Institute (MNI) space, after slice‐timing and motion correction. The functional data underwent de‐trending with removal of linear and quadratic trends, band‐pass filtering with a pass‐band of 0.01–0.08 Hz, and removal of nuisance regressors (six head‐motion parameters, mean of ventricular and white‐matter signal, and their temporal derivatives). The FCD lesions were manually segmented on preoperative 3TT1/T2/FLAIR images, then co‐registered to the T1w data and warped into the MNI space. No spatial smoothing was performed to increase spatial functional specificity in the subsequent thalamic parcellation analysis. We also employed FreeSurfer [29] to calculate thalamic and intracranial volume, and used normalized thalamic volume by dividing the raw volume with the intracranial volume.

2.4. Thalamic Functional Parcellation and Thalamo‐Lesional Connectivity Deviation

The thalamus was divided into seven nonoverlapping functional parcels using a winner‐take‐all approach based on each voxel's functional connectivity to the Yeo 7‐network cortical parcellation [26] using healthy controls' data [21]. Specifically, the Yeo 7‐network parcellation was applied to map the cortical regions into seven functional networks. While Yeo 7‐network parcellation was originally developed in adults, it captures the general network structure in children as young as 3 years old [30]. We averaged timeseries from each network and calculated the thalamo‐network connectivity between each of the thalamic voxels and Yeo networks, following Fisher r‐to‐z transformation. Harmonization was conducted on the r‐to‐z transformed connectivity using neuroCombat [27] to correct the site effect while preserving salient neurobiological information. We then divided the thalamus into seven nonoverlapping functional parcels, by assigning each voxel to the cortical network exhibiting the strongest group‐averaged connectivity. We also examined the relationship between those functional parcels and thalamic nuclei by overlapping our functional parcellations with the AAL3 atlas [31].

We used FCD lesion masks as seeds and calculate the functional connectivity between FCD lesion seeds and thalamic functional parcels, termed as thalamo‐lesional functional connectivity. We repeated this process across all FCD lesion locations. The healthy controls' thalamo‐lesional functional connectivity was Fisher's r‐to‐z transformed and harmonized, resulting in a matrix with the size of 116 controls × 43 FCD patients × 14 thalamic parcels (7 left and 7 right, calculated separately). To facilitate comparisons across FCD spatial locations, we followed our previous work [7] to calculate z‐scored thalamo‐lesional connectivity deviation for each FCD mask, using the controls' mean and standard deviation of thalamo‐lesional connectivity as references. An overall schematic to calculate z‐scored thalamo‐lesional connectivity is shown in Figure 1.

FIGURE 1.

Summary of thalamo‐lesional functional connectivity. The thalamus is divided into 7 nonoverlapping functional parcels using a winner‐take‐all approach based on each voxel's functional connectivity to the Yeo 7‐network cortical parcellation using healthy controls' data, resulting in 14 functional parcels for both hemispheres. The thalamo‐lesional functional connectivity is calculated between the thalamic functional parcels and the FCD lesion seed for patients and healthy controls. The mean and standard deviation of healthy controls' thalamo‐lesional functional connectivity are used to generate a z‐score for each FCD patient.

2.5. XGBoost Classifier Predicting FBTCS

An XGBoost classifier was trained to predict the presence of FBTCS using the z‐scored thalamo‐lesional connectivity deviation profile in R (https://cran.r‐project.org/web/packages/xgboost/index.html). A nested 5 × 5‐fold cross‐validation (5 outer folds, 5 inner folds) was used, with stratified k‐fold cross‐validation ensuring that the distribution of FBTCS cases in each fold was representative of the total study population. This process was repeated 50 times to ensure stability. The classification performance was evaluated using the unseen samples. 1000 permutation tests were conducted that randomly shuffled the labels to generate a null distribution for the classification to assess its statistical significance.

3. Results

3.1. Demographic Information and Clinical Characteristics

Forty‐nine patients with FCD and rsfMRI were identified. Six patients were excluded for the following reasons: three were removed due to prior resections; two removed due to excessive motion (mean frame displacement > 1 mm); one removed due to high signal dropout. The remaining 43 patients (22 females (51%), mean age = 14.2 ± 4.0 years old, range 5–22 years) were included in the subsequent analysis (see Figure 2A for lesion locations), among whom 14 (33%) had documented FBTCS by clinical history as reported by epileptologists and/or confirmed by neurophysiology. 116 healthy controls were used who were matched in sex, age, and head motion (16 from Children's National; 100 from the HCP‐D dataset). Demographic and clinical information are presented in Table 1.

FIGURE 2.

(A) Heat map of FCD lesion distribution for total cohort. Of note, this is anatomic convention (left is left, right is right). (B) Generation of thalamic functional parcellation. Thalamic functional parcellation map was generated from typically‐developing subjects (middle), color coded by based on their preferential connectivity to the Yeo 7‐network atlas. (C) Measuring structural‐functional overlap. A heat map showing the top three overlapping thalamic nuclei for each functional parcel with the percentage overlap. For example, the frontoparietal control network thalamic parcellation mostly overlaps with the ventral lateral nucleus (49.2%), considerably higher than the next two nuclei, that is, medial pulvinar (9.5%) and ventral anterior nucleus (7.9%). Network abbreviations: DAN, dorsal attention network; DMN, default mode network; FPN, frontoparietal control network; LIM, limbic network; SMN, somatomotor network; VAN, ventral attention network; VIS, visual network. Thalamic nuclear abbreviations: IL, intralaminar; MDm, mediodorsal medial; PuA, pulvinar anterior; PuI, pulvinar inferior; PuL, pulvinar lateral; PuM, pulvinar medial; VA, ventral anterior; VL, ventral lateral; VPL, ventral posterolateral.

TABLE 1.

Demographic, clinical, and scanner information.

	FCD	Controls	Statistics
Demographic information
Number	43	116
Age, mean ± SD, years	14.2 ± 4.0	13.7 ± 3.3	Two‐sample t‐test (p = 0.37)
Sex, M/F	21/22	56/60	Chi‐square test (p = 0.95)
Head motion, mean ± SD, mm	0.20 ± 0.17	0.19 ± 0.11	Two‐sample t‐test (p = 0.57)
Clinical characteristics
Age of seizure onset, mean ± SD, years	7.6 ± 4.8
Seizure duration, mean ± SD, years	6.6 ± 4.2
Seizure focus, left/right	24/19
FCD lobe, no. (%)
Frontal	11 (26)
Temporal	19 (44)
Parietal	5 (12)
Occipital	1 (2)
Other (multilobar, cingulate, and insula)	7 (16)
Pathology, no. (%)
Type I	7 (16)
Type II	15 (35)
Type III	8 (19)
Mild malformation of cortical development	2 (5)
Unavailable	11 (26)
Surgical outcome, no. (%)
Engel I	18 (42)
Engel II	10 (23)
Engel III	3 (7)
Engel IV	2 (5)
Unavailable	10 (26)
Scanner information
Software version, no. (%)
DV24.0_R01_1344.a	7 (16)
DV24.0_R02_1607.b	9 (21)
DV26.0_R01_1725.a	5 (12)
DV26.0_R03_1831.b	11 (26)
DV26.0_R05_2008.a	4 (9)
DV26.0_R06_2132.a	6 (14)
MR30.1_R01_2322.c	1 (2)
#Channels of head coil, no. (%)
8 Channels	35 (81)
32 Channels	8 (19)

3.2. Thalamic Functional Parcellation

The thalamus was parcellated into seven nonoverlapping functional parcels based on thalamic‐to‐Yeo functional connectivity and listed the top three overlapping nuclei for each functional parcel per the AAL3 atlas, as shown in Figure 2B. As expected, all functional parcels were associated with one or two dominant thalamic nuclei based on the percentage overlap (i.e., the volume of the overlap divided by the volume of the functional parcel), which suggests a strong correspondence between functional parcels and underlying anatomy.

3.3. Z‐Scored Thalamo‐Lesional Connectivity Predicts FBTCS

Thalamo‐lesional functional connectivity was calculated from patients and controls using FCD masks as seeds. Using controls' mean and standard deviation as the reference, we calculated the z‐scored thalamo‐lesional connectivity deviation for the patients with FCD. One‐way ANOVA also did not reveal any significant association between z‐scored thalamo‐lesional connectivity and FCD lobar location following FDR correction (all ps > 0.05), confirming the effectiveness of z‐scoring. One‐sample t‐tests on z‐scored connectivity did not reveal any significant difference from zero after multiple comparisons correction using FDR correction (all ps > 0.05).

We explored the multivariate z‐scored thalamo‐lesional functional connectivity pattern to investigate the neural underpinning of FTBCS. An XGBoost classifier trained using the z‐scored thalamo‐lesional connectivity deviation profile successfully predicted the presence of FBTCS, achieving a median accuracy of 0.72 (0.71 ± 0.04), significantly above the chance level (p _permutation = 0.025, Figure 3A), together with a median sensitivity of 0.85 (0.84 ± 0.05) and specificity of 0.47 (0.48 ± 0.06). We also quantified the relative contribution of seven thalamic functional parcels to the classification by examining feature importance. The SMN‐thalamus parcel, the functional parcel preferentially connected to the somatomotor network (SMN) and has structural overlaps with the ventral posterolateral, pulvinar medial and intralaminar nuclei, contributed more to FBTCS classification than any other thalamic parcel (Figure 3B, all ps < 0.001, Wilcoxon signed‐rank tests).

FIGURE 3.

Assessing the accuracy of XGBoost FBTCS classifier and the thalamic functional parcellation relative contributions. (A) Permutation tests demonstrating the significance of an XGBoost classifier prediction of FBTCS from rsfMRI data. (B) Each of the 7 thalamic parcels' contribution to the FBTCS classification. Of note, the somatomotor‐network (SMN) thalamus had a significantly higher contribution to this classification than any other functional parcels.

3.4. Post Hoc Statistical Analysis of Clinical Associations

We explored clinical associations of the z‐scored SMN‐thalamo‐lesional connectivity. Using a subset of patients with postsurgical outcomes (n = 33), post hoc logistic regression analysis examined the relationship between the seizure freedom (Engel I) and SMN‐thalamo‐lesional connectivity, while controlling for the presence of FBTCS, epilepsy duration, age of onset, and normalized thalamic volumes. The overall model was significant (χ ² = 15.5, df = 27, p = 0.008). We found that increased ipsilateral SMN‐thalamo‐lesional connectivity is associated with worse surgical outcomes (t = −2.25, 95% confidence interval [−16.18, −1.87], df = 27, p = 0.025, Figure 4A). We also found a significant interaction between ipsilateral thalamic volume and connectivity (t = 2.16, 95% confidence interval [−327.73, −3309.69], df = 27, p = 0.031). Specifically, decreased ipsilateral thalamic volume is associated with worse outcomes when SMN‐thalamo‐lesional connectivity is high (Figure 4B). No other factors were significant. Such a relationship was specific to ipsilateral connectivity and atrophy only, as the model using contralateral connectivity and atrophy did not reveal any significant results. To address potential anatomical and clinical confounds, we controlled for epilepsy duration and lesion location and our primary findings remained significant, suggesting that the observed connectivity signatures are not driven primarily by these factors although the sample size may limit full exclusion of anatomical confounding. There was also a significant correlation between ipsilateral SMN‐thalamo‐lesional connectivity and ipsilateral thalamic volume (Spearman rho = −0.36, p = 0.02) among all patients, and such correlation was significantly modulated by the presence of FBTCS (t = −2.45, 95% confidence interval [−458.2, −4808.7], df = 39, p = 0.019, Figure 4C). No such relationship existed for the contralateral SMN‐thalamo‐lesional connectivity or ipsilateral thalamo‐lesional connectivity for any other thalamic parcels (all ps > 0.05).

FIGURE 4.

Post hoc logistic regression to evaluate relationship between SMN‐thalamo‐lesional connectivity and postsurgical outcomes. (A) This logistic regression demonstrates that stronger ipsilateral SMN‐thalamo‐lesional connectivity is associated with worse surgical outcomes. Of note, this model controls for the presence of FBTCS and epilepsy duration. (B) The significant interaction between ipsilateral thalamic volume and SMN‐thalamo‐lesional connectivity such that ipsilateral thalamic atrophy (lower normalized thalamic volume) is associated with lower probability of postsurgical seizure freedom with a high ipsilateral SMN‐thalamo‐lesional connectivity (e.g., at 75th percentile). (C) The structural‐functional coupling between ipsilateral thalamic volume and SMN‐thalamo‐lesional connectivity is significantly modulated by the presence of FBTCS.

4. Discussion

This study used resting‐state functional connectivity analyses to identify thalamo‐lesional connectivity alterations as potential neuromarkers in patients with FCD‐related DRE, specifically identifying a novel functional pathway related to focal to bilateral tonic–clonic seizures. This is a novel method of quantifying thalamo‐lesional connectivity directly using a network‐driven thalamic functional parcellation based on the Yeo 7‐network atlas. Using z‐scored thalamo‐lesional connectivity, an XGBoost classifier was able to detect the presence of FBTCS with a cross‐validated accuracy of 72%, high sensitivity of 85% and modest specificity, with the SMN‐thalamus contributing more than other thalamic functional parcels. These findings were not directly related to FCD cortical lobar location, indicating the importance of this thalamic functional pathway in FCD‐DRE. Further post hoc analyses of SMN‐thalamo‐lesional connectivity revealed that stronger ipsilateral SMN‐thalamo‐lesional connectivity is associated with worse surgical outcomes. Structural‐functional coupling was demonstrated as ipsilateral SMN‐thalamo‐lesional connectivity was specifically associated with ipsilateral thalamic structural atrophy, and this correlation was driven by patients with FBTCS. The presence of this atrophy compounded the risk for worse surgical outcomes. These data harness the power of FCD as a natural model of focal epilepsy to define SMN‐thalamo‐lesional connectivity alterations as a potential network neuromarker with associated structural‐functional coupling, relating this functional pathway to the genesis of high‐risk FBTCS, thalamic atrophy, and more broadly to surgical outcomes in FCD‐DRE. The results demonstrate this method as a promising avenue toward precision epileptology.

These results build on accumulating and converging lines of evidence on the importance of the thalamus as a mediator of focal to bilateral (as well as generalized) tonic–clonic seizures. Most existing studies of fMRI as well as other structural and functional neuroimaging modalities are largely restricted to temporal lobe epilepsy or non‐lesional focal epilepsy patients.

Several studies have used global network analyses of rsfMRI to demonstrate network anomalies in patients with TLE and FBTCS. Patients with FBTCS exhibit more uniformly distributed thalamocortical connections, and those with uncontrolled FBTCS have abnormalities in basal ganglia‐thalamic subnetworks with described enhanced interactions between putamen and globus pallidus‐pars interna (GPi) and decreased thalamo–GPi interactions [22]. Those with FBTCS also have reduced thalamocortical functional connectivity in connections between dorsolateral pulvinar and bilateral parieto‐occipital lobes as compared to patients with TLE without FBTCS, where the abnormality remains unilateral [21]. Using a connectopic mapping technique, a recent study examining pediatric TLE patients showed that increased anterior thalamic‐basal ganglia connectivity was more pronounced in children with FBTCS [25]. Using dynamic functional connectivity analysis, left TLE patients with FBTCS show increased cortical/subcortical integration (left caudate and left frontoparietal control network, left cerebellum and left default mode network) and increased connectivity within the left anterior thalamic nucleus. Using task‐based fMRI, patients with TLE‐FBTCS had lower task‐related activation of bilateral thalami, and greater task‐related thalamotemporal and thalamomotor connectivity [32]. In non‐lesional focal epilepsy, FBTCS are associated with striato‐thalamo‐cerebellar and thalamocortical connectivity alterations. Here, we expand these analyses to evaluate temporal and extratemporal lesional focal epilepsy causing FBTCS with FCD‐DRE as the model. The seed‐based functional connectivity analysis allows us to measure connectivity from the known seizure generator (the FCD) to the FBTCS amplifier (thalamus). We show with a novel thalamic imaging technique that there are disruptions in the thalamo‐lesional circuitry, particularly between the SMN‐thalamus and FCD lesion, and that these connectivity alterations relate to thalamic atrophy, correlate to surgical outcomes, and contribute substantially to a machine learning algorithm that can classify FCD patients as having FBTCS or not based on rsfMRI.

There are several studies that describe volumetric changes in focal epilepsy patients with FBTCS with a variety of findings. The present results augment existing data concerning volumetric and white matter changes found in the thalami and cortices of patients with FBTCS. TLE with hippocampal sclerosis is associated with duration dependent medial thalamic atrophy that corresponds to cortical thinning of bilateral frontal and ipsilateral mesial temporal lobes. Structural MRI studies previously demonstrated varying patterns of thalamic atrophy, particularly medial nuclei (e.g., mediodorsal nucleus) in FBTCS patients [20, 23, 33]. Using structural covariance mapping, patients with focal epilepsy and FBTCS exhibit cortical volumetric loss in frontal, motor and somatosensory cortices that receive projections mainly from the ventral anterior nucleus [34]. These differences may arise from engagement of different pathways (e.g., TLE with FBTCS may cause different thalamic and extrathalamic atrophy patterns than extratemporal lobe generators). The present data demonstrate in a population of mostly extratemporal FCDs that there is ipsilateral thalamic volumetric atrophy. Diffusion weighted imaging studies find varying patterns of white matter anomalies in patients with FBTCS. TLE patients without FBTCS have more localized white matter anomalies in the temporal and frontal regions, whereas those with FBTCS have more widespread white matter changes. Additionally, TLE‐FBTCS patients have higher node abnormality when evaluating the white matter based connectome [24]. TLE patients with FBTCS have global white matter abnormalities in terms of diminished fractional anisotropy, and specifically altered white matter metrics in the hippocampal‐to‐thalamus tract that corresponds to medial thalamic anomalies [20]. Extratemporal and non‐lesional focal epilepsy patients with FBTCS show preservation in white matter integrity in thalamo‐occipital tracts, striato‐occipital tracts, and inferior fronto‐occipital fascicle compared to patients without FBTCS, implicating these segments in the genesis of FBTCS [35]. Future prospective multimodal imaging studies will more clearly delineate these networks.

Surgical outcomes for FCD‐DRE worsen with longer epilepsy duration [36]. As the lesions are grossly structurally static, this may be due to wider epileptogenic networks being established [10]. Duration‐dependent thalamic atrophy is documented in TLE. The present data harness the power of resting‐state functional connectivity analysis to correlate the known seizure focus to network‐level abnormalities and outcomes. We demonstrate that the strength of the SMN‐thalamo‐lesional functional connectivity is associated with thalamic atrophy ipsilateral to the FCD focus. This SMN‐thalamo‐lesional functional connectivity is associated with worse surgical outcomes that are compounded by the presence of ipsilateral thalamic atrophy and/or FBTCS, which cannot be explained by epilepsy duration alone. These findings fuse structural‐functional imaging findings and may contribute further to the idea of an expanding epileptogenic network across time, with thalamic atrophy as a potential neuromarker that should be further investigated.

These findings also demonstrate the expanded potential for machine learning‐derived classifiers in FCD‐related epilepsy. There are numerous classifiers in development to help with the identification of lesions on structural MRI [37, 38, 39] which have demonstrated ability to enhance detection and potentially improve surgical outcomes [40]. Here, we show a trained classifier is able to distinguish FCD‐DRE with FBTCS with high sensitivity using functional imaging data. Future multimodal classifiers that incorporate other modalities such as arterial‐spin‐labeling MRI [41, 42] or positron emission tomography [43] may further enhance precision.

This study is not without limitations. First, the modest sample size restricts the absolute predictive power of our machine learning pipeline. While the classifier's performance leaves room for optimization, its primary utility in this study was to pinpoint aberrant thalamic parcels associated with FBTCS. The classifier's lower specificity may be due to several reasons. FCDs are associated with widespread network changes; and the thalamus is a relay involved in focal seizures with and without secondary generalization. The classifier may also be affected by heterogeneity of the FCD spectrum: for example, the network abnormalities may differ within FCD subclasses and this study includes samples across the FCD spectrum. Second, single‐center data may not allow generalization of results. With a moderate sample size, our results should be confirmed in expanded and multicenter evaluations. Moreover, as abnormal structural connectivity has widely observed in patients with focal epilepsy [44], future multimodal analyses (e.g., diffusion weighted imaging) are needed to bridge functional alterations with underlying structural connectivity to delineate the anatomical propagation pathways of seizures. Third, while we employ the well‐established Yeo 7‐network atlas to derive a network‐driven thalamic functional parcellation, the functional specialization of the thalamus may not be discrete [45]. There may be developmental refinements of these networks across late childhood, but the fundamental architectures are stable and measurable in early childhood. A gradient approach [46] may offer a more holistic understanding of the thalamo‐lesion interaction. Fourth, the FBTCS classification is based on documentation, there is potential for underreporting. Fifth, not all patients are seizure‐free, raising the possibility that subtle FCD lesions may have been missed, limiting the anatomical specificity. Sixth, given the sample size and relatively small FBTCS subgroup, the study may be underpowered to fully exclude anatomical confounding effects on classification. Lastly, while we focus on the thalamus as a key node in seizure semiology, future research should evaluate whether similar connectivity signatures exist within basal ganglia–cortical circuits, which are also known to modulate seizure generalization and motor output in focal epilepsy.

5. Conclusions

The thalamus represents a critical subcortical region in the propagation of high‐risk focal to bilateral tonic–clonic seizures (FBTCS). By integrating novel functional thalamic parcellation with MRI analysis, this study examines the thalamocortical circuitry between the known seizure generator (lesion) and the amplifier (thalamus) in patients with FCD. We reveal that FBTCS generalization in FCD‐related epilepsy is underpinned by a distinct thalamo‐lesional signature. Specifically, our findings implicate the somatomotor‐linked thalamic parcel (SMN‐thalamus) as a key gatekeeper for secondary generalization, bridging the gap between focal cortical dysplasia and seizures with bilateral motor manifestations. We demonstrate a structural‐functional coupling of this functional parcel, as elevated SMN‐thalamo‐lesional connectivity is not only associated with FBTCS but also corresponds to ipsilateral thalamic atrophy, which suggests that chronic seizure propagation may drive progressive neuroanatomical degradation, and such degradation is more pronounced in patients with FBTCS. These connectivity anomalies are correlated to worse postsurgical outcomes that are compounded by the presence of ipsilateral thalamic atrophy. Finally, the successful implementation of a classifier suggests that the thalamo‐lesional signatures can serve as a potential neuromarker for seizure phenotyping. These results move beyond simple mapping to provide a framework for precision epileptology, where multimodal imaging can identify patients at risk for surgical failure and severe seizure characteristics.

Author Contributions

Hua Xie: conception and design of study, acquisition and analysis of data, drafting a significant portion of the manuscript or figures, statistical analysis; Xiaotong Li: acquisition and analysis of data, drafting a significant portion of the manuscript or figures; Venkata Sita Priyanka Illapani: acquisition and analysis of data, drafting a significant portion of the manuscript or figures; Leigh N. Sepeta: acquisition and analysis of data, drafting a significant portion of the manuscript or figures; Jason Schroeder: acquisition and analysis of data, drafting a significant portion of the manuscript or figures; Chima O. Oluigbo: acquisition and analysis of data, drafting a significant portion of the manuscript or figures; Madison M. Berl: acquisition and analysis of data, drafting a significant portion of the manuscript or figures; William D. Gaillard: conception and design of study, acquisition and analysis of data, drafting a significant portion of the manuscript or figures; Nathan T. Cohen: conception and design of study, acquisition and analysis of data, drafting a significant portion of the manuscript or figures.

Funding

H.X. is funded by the American Epilepsy Society Junior Investigator Award. Research reported in this publication was supported by the NIH/NINDS K23NS131522 (N.T.C.). This work was partially supported by P50 HD105328 (H.X. and W.D.G.). This work was also partially supported by the AAN Career Development Award and the Hess Foundation (N.T.C.). And this article was funded by the American Academy of Neurology.

Conflicts of Interest

The authors declare no conflicts of interest.

Funding Statement

This work was funded by National Center for Advancing Translational Sciences grant P50 HD105328; Hess Foundation ; National Institute of Neurological Disorders and Stroke grant K23NS131522; American Epilepsy Society ; American Academy of Neurology .

Data Availability Statement

Deidentified data are available upon request from any qualified investigator.