Thalamocortical hodology to personalize electrical stimulation for focal epilepsy

Abstract

Thalamic electrical stimulation offers a therapeutic approach for refractory epilepsy patients ineligibles for resective surgery. Yet, there is currently no consensus on the optimal stimulation site. We hypothesized that targeting thalamic nuclei with precise anatomical and functional connections (hodological matching) to cortical seizure onset zones (SOZs) could enhance neuromodulatory effects. Thus, we examined three subnuclei (pulvinar, anterior, and ventral intermediate/oral posterior, VIM/VOP) in 41 focal epilepsy patients. Combining neuroimaging and electrophysiology, we identified hodologically-matched targets: pulvinar for posterior-quadrant and temporal SOZs, anterior for frontotemporal, and VIM/VOP for rolandic. Analysis of 274 seizures revealed the critical role of matched nuclei in seizure termination. Stimulation of hodologically-matched (but not unmatched) nuclei immediately suppressed interictal epileptiform discharges. Finally, in 10 chronically implanted patients, matched stimulation substantially reduced seizure frequency (87.5%), compared to unmatched targets (8.3%) over time. Our results underscore the potential of hodological thalamic targeting to modulate epileptiform activity in focal epilepsy.

Introduction

Over 50 million people worldwide suffer from epilepsy, a neurological disorder characterized by recurrent seizures, stemming from synchronized and excessive paroxysmal electrical activity of neurons^1,2. In most of the cases (> 60%), said seizures originate and develop in restricted and focal cortical regions of the brain³, condition clinically referred to as focal epilepsy. Conversely, in generalized epilepsy, pathological brain activity stems from widespread areas of the brain, affecting both hemispheres^4,5. Independently of the focal or generalized onset, about 30% of epilepsy patients live with intractable, disabling seizures⁶, which cannot be controlled with medical therapy. The current standard of care for medically refractory epilepsy (MRE) relies on the resection of the seizure focus, i.e., the brain region associated with earliest electrophysiological changes during a seizure event⁷. However, if said cortical areas are not amenable for safe removal (or disconnection) because of overlap with eloquent cortices, extensive seizure foci, or a diagnosis of generalized epilepsy, neuromodulation approaches such as thalamic electrical stimulation might be an alternative^7,8,9,10,11. Recent studies have propelled thalamic stimulation for epilepsy to the forefront of neuroscience, underscoring its emerging role as a pivotal therapeutic strategy in the field¹².

The thalamus is a deep central structure crucial for its multifaceted functions and intricate organization. Indeed, through extensive connections with the cortex and other subcortical and peripheral structures, it is integral to a wide array of functions, including motor and sensory processing, consciousness, and sleep^13,14,15,16. This diverse functionality is attributed to the thalamus’s varied nuclear architecture and its intimate relationship with cortical areas¹⁷. Importantly, since the early 1950s, the thalamus was demonstrated to synchronize widespread cortical activity, hence contributing to the generation of pathological epileptic discharges¹⁸.

Indeed, altered thalamocortical synchrony was extensively observed during focal seizures and associated loss of consciousness¹⁹. Furthermore, structural connectivity variations were reported between the thalamus and the cortex in patients with generalized epilepsy²⁰.

Given these cortico-subcortical dynamics and epileptogenic interactions, thalamic electrical stimulation has been proposed since the early 80 s, with indications both in focal^9,21 and generalized epilepsy⁸. In this regard, thalamic sites characterized by diffuse connectivity as the centromedian nucleus (CM), have been hypothesized to benefit generalized seizures, which originate and develop through widespread neuronal networks. Indeed, in a recent clinical trial, Cukiert and colleagues reported a significant seizure reduction (> 50%) with CM electrical stimulation in 90% of tested patients diagnosed with generalized epilepsy²². CM stimulation, instead, has been less effective in patients with focal MRE²³. For this reason, thalamic nuclei with less diffuse connectivity, and mostly the anterior nucleus (ANT), have been explored as a possible target for neuromodulation in focal epilepsy. Already in the early 80 s, Cooper et al. showed a significant clinical control of seizures in 4 out of 6 patients with stimulation of the ANT²⁴. Since this seminal work, multiple studies have investigated the efficacy of ANT stimulation to treat focal epilepsy^{9,21,25,26,27,28}, including the pivotal SANTE trial that achieved 75% median seizure reduction at 7 years follow-up²⁹. Thanks to these works, ANT stimulation for focal epilepsy received FDA approval in 2018. However, clinical outcomes are highly variable depending on the anatomical location of the seizure onset zone (SOZ). Specifically, the most well-documented efficacy was observed in focal onset seizures occurring in the temporal and frontal lobe (78% and 86% seizure frequency reduction), whereas effects were limited for other cortical SOZs (39%)²⁹. This might be due to the ANT’s unique anatomical and functional connectivity to temporal and frontal cortical networks^30,31.

Exploration in other areas of the thalamus further suggests a possible link between thalamocortical anatomical connectivity and neuromodulation efficacy. Indeed, the pulvinar nucleus (PUL) of the thalamus, which has high connectivity towards the temporal and occipito-parietal lobes and presents different neural firing patterns during ictal progression and termination in patients with temporal lobe epilepsy^32,33,34,35, has been shown to be potentially effective for posterior quadrant epilepsy^36,37,38. Similarly, the ventrolateral portion of the thalamus (ventral intermediate/ventral oral posterior nuclei, VIM/VOP)³⁹ and the cerebellum⁴⁰, which present high structural connectivity to motor cortical areas, were tested, with positive outcome, for epilepsies organized in the agranular pre-central neocortical areas (rolandic epilepsy), which is particularly challenging for resective surgeries due to the high risks of motor deficits.

However, these studies remain isolated investigations lacking a comprehensive and systematic multimodal evaluation of precise thalamic targeting. Consequently, there is currently no consensus on which target is optimal for focal epilepsy and clinical indication on how to approach the heterogeneity of SOZs, with limited clinical outcome. To overcome this limitation, here we developed a multimodal framework integrating structural, functional, electrophysiological, and clinical data from focal epilepsy patients, enabling personalized, mechanistically informed thalamic targeting strategies. Based on prior evidence, here we argue that, for focal epilepsy, thalamic electrical stimulation should match the hodology of the thalamocortical pathway to enhance therapeutic outcome. In other words, targeting electrical stimulation to the thalamic subnucleus with preferential anatomical and functional connectivity to the SOZ (an approach that we refer to as hodological matching) will improve the efficacy of thalamic stimulation for the treatment of refractory focal epilepsy.

To test this hypothesis, we studied the anatomical and functional properties of neural activity of 41 patients with focal MRE. We focused on the study of three thalamic nuclei that exhibit more specific and less diffuse thalamocortical projections as compared to the CM: the ANT, the PUL, and the VIM/VOP. First, we defined a hodological map of the thalamocortical fibers from said thalamic nuclei towards the cortex using imaging and electrophysiology. In this way, we identified the cortical areas that match each nucleus, i.e., receiving major projections. We then characterized the functional involvement of these thalamic nuclei to ictal events occurring both in hodologically-matching and unmatching SOZs. Furthermore, we showed that electrical stimulation of only hodologically-matched thalamic nuclei immediately suppressed pathological discharges, i.e., interictal spikes, in the epileptogenic cortex, in acute settings. Finally, we verified that such effects led to significant clinical improvement by reducing seizure frequency in chronic stages. Specifically, we retrospectively evaluated 10 patients with a chronic thalamic stimulation implant (n = 7 matched, n = 3 unmatched). Importantly, our hodology-based approach achieved a mean reduction in seizure frequency of 87.5% which dramatically outperforms unmatched thalamic targeting (mean reduction in seizure frequency of 8.3%).

In summary, here we combined high-throughput imaging, electrophysiological, and clinical assessments to explore the feasibility of a novel connectivity-informed framework to guide thalamic neuromodulation. Our approach, i.e., targeting electrical stimulation to match the hodology of thalamocortical connections, provides a novel framework for tailoring neuromodulation therapy in focal refractory epilepsy. This strategy introduces innovative avenues for clinical targeting and therapeutic intervention, while challenging the prevailing paradigm that adopts a “one-nucleus-fits-all” SOZs.

Results

Study rationale and design

Here, we aimed at exploring whether targeting specific thalamic subnuclei might provide increased neuromodulatory effect in focal refractory epilepsy based on their hodological matching with the SOZ. To demonstrate our hypothesis, we deployed a large array of electrophysiological and imaging assessments in a unique cohort of 41 focal MRE patients (Fig. 1a, b and Supplementary Data 1 for demographic and clinical information for each patient). Specifically, we prospectively analyzed intracranial electrophysiological data from 37 patients (S01–S37) undergoing stereoelectroencephalography (SEEG) exploration for epilepsy monitoring and precise definition of the SOZs (Fig. 1c, d). We acquired preoperative structural magnetic resonance imaging (MRI) and postoperative computed tomography (CT) for precise SEEG contacts localization performed by a senior neurosurgeon (JGM) in all patients. All the participants had at least one SEEG electrode located in one of the nuclei of interest, namely the PUL, ANT, or VIM/VOP nuclei, and a diagnosis of focal epilepsy. Overall, 54 thalamic nuclei among those of interest were sampled in this cohort (Fig. 1c, d and Supplementary Data 2). Importantly, nearly half of our cohort presented simultaneous sampling of more than one thalamic nucleus of interest (Fig. 1c). We categorized each patient’s SOZ as frontal (meaning originating from frontal lobe regions such as the orbitofrontal cortex), temporal (including mesial and lateral temporal lobe structures), rolandic (including pre- and post-central peri-rolandic regions) or parietal/occipital cortex (namely originating from posterior quadrant areas) (Fig. 1d, Supplementary Fig. 1a, and Supplementary Data 3). In a subgroup of patients (n = 14), we acquired and analyzed high-resolution diffusion MRI data to corroborate the anatomical organization of thalamocortical fibers (Fig. 1g). Such connectivity pattern was also confirmed through electrophysiology with thalamocortical evoked potentials (n = 6). To study the interaction of the thalamus and the epileptic cortex, we then analyzed a total of 274 spontaneous seizures with concurrent sampling of the thalamic nuclei and cortical SOZ. Hence, each recorded seizure was categorized according to the SOZ anatomical location and the sampled thalamic nucleus during the event (Fig. 1d). We then performed electrophysiological testing in a subset of patients. Specifically, we used high-frequency thalamic stimulation (n = 17) to test immediate effects of hodologically-matched thalamic stimulation on interictal epileptiform activity (Fig. 1g). Finally, we retrospectively assessed long-term efficacy of our hodology based neuromodulation approach in 10 patients (S32–S41) with a chronic stimulation implant (Fig. 1e) and up to 3 years of follow-up (Fig. 1e, f).

Fig. 1: Study design and patients demographics.

a Left: Schematic of hodological matching of PUL, ANT, and VIM/VOP to specific cortical anatomical regions. Right: anatomical localization of PUL, ANT, and VIM/VOP nuclei on MNI space average T1-weighted MRI⁸⁷ with axial view and magnification of the thalamic nuclei. Thalamic parcellation is atlas-guided⁴² (A anterior, P posterior, M medial, L lateral). The figure was generated in Lead-DBS⁸⁸. b Study population composition. The cohort included n = 41 focal epilepsy patients, of whom 37 (S01–S37) underwent prospective SEEG assessment, and 10 (S32–S41) underwent retrospective chronic assessment. c Top left: axial view of reconstruction of SEEG electrodes targeting thalamic nuclei of interest, overlayed with atlas-guided thalamic parcellation for visualization purposes. Top right: magnification of the contacts located in the PUL (blue), ANT (yellow), and VIM/VOP (red). Bottom: grouped bar plot showing the proportion of patients with one thalamic nucleus versus more than one thalamic nucleus simultaneously implanted. d Stratification of SEEG patients (S01–S37). Top Left: pie chart of the lobar location of the SOZ. Top Right: pie chart of 54 thalamic SEEG contacts in PUL, ANT, and VIM/VOP. A higher number of thalamic contacts than patients implies that some patients had more than one nucleus of interest implanted. Bottom: Stratification of the 274 recorded spontaneous seizures according to the lobar location of different SOZs (left) and the respective thalamic coverage (right). A different distribution from the top panel implies a different number of seizures for each patient, as reported in Supplementary Data 2. e Top: Stratification of 10 chronically implanted patients (S32–S41) according to the lobar location of the SOZ (left) and implanted thalamic nucleus (right). Bottom: grouped bar plot showing the proportion of patients implanted with a matched versus unmatched neurostimulation system. f Histogram of follow-up years of chronically implanted patients. g Design and experiments of the study. 41 patients received SEEG implantation and were stratified according to their SOZ and thalamic coverage. Bottom, left to right: we investigated the anatomical (imaging and electrophysiology), and functional coupling during seizures involving the thalamus, hence defining hodologically matching and unmatching nuclei. We then tested the effect of acute thalamic stimulation and finally the clinical outcome following chronic thalamic stimulation was retrospectively evaluated in a subgroup of the study population.

Importantly, we also report the anatomical and functional characteristics of the CM nucleus both in focal (n = 4) and generalized (n = 1) epilepsy. Compared to the PUL, ANT, and VIM/VOP, CM showed widespread, yet sporadic, structural and functional connectivity (Supplementary Fig. 2), replicating well known evidence from the literature^8,11,22,41 and supporting its applicability as a target in generalized, rather than focal, epilepsy.

Overall, we designed and performed a comprehensive study to investigate the mechanisms of thalamocortical interactions and the effects of hodologically-matched thalamic stimulation in focal refractory epilepsy (Fig. 1g).

A specific hodological map of the thalamocortical pathway in the epileptic brain

While previous studies proposed a structural connectivity map of the thalamocortical projections^{17,42,43,44,45,46}, the majority of these studies were performed in healthy brains and were limited to imaging techniques⁴⁷, such as fiber tracking. Here, we confirmed the presence of a well-defined and highly specific anatomical map of the thalamocortical projections of PUL, ANT, and VIM/VOP in the epileptic brain with two complementary methodologies: neuroimaging and electrophysiology.

First, we performed high-definition fiber tracking (HDFT) of high-resolution diffusion MRI data to confirm previously reported anatomical organization of cortical connectivity patterns of the PUL, ANT, and VIM/VOP. We reconstructed all the likely axonal pathways between these nuclei and the cortical regions presenting seizures in our cohort (namely frontal, rolandic, parietal, occipital and temporal) (Fig. 2a). We quantified the relative strength of these connections by calculating the volume of thalamocortical projections from each nucleus to each cortical region normalized by the total volume of fibers (Fig. 2b). This analysis revealed a clear thalamocortical anatomical organization. Indeed, the PUL showed preferential connectivity towards the parietal and occipital regions, and the temporal lobe (mean volume value, frontal 0.07, rolandic: 0.05, parietal: 0.38, occipital: 0.32, temporal: 0.16); the ANT, instead, presented axonal projections going preferentially towards the frontal and temporal lobes (frontal: 0.55, rolandic: 0.03, parietal: 0.09, occipital: 0.08, temporal: 0.23); finally the VIM/VOP nuclei projected more remarkably to the rolandic cortex (frontal: 0.25, rolandic: 0.49, parietal: 0.12, occipital: 0.07, temporal: 0.05). This constricted structure of thalamocortical fibers from PUL, ANT, and VIM/VOP contrasts with the broad but sparse anatomical connectivity of the CM. Indeed, we found that the white matter connections between a cortical lobe and the anatomically-matched nucleus (ANT in frontal, VIM/VOP in rolandic, PUL in parietal and occipital, and PUL and ANT in temporal) presented significantly higher volume when compared to the CM (frontal: 0.52 vs 0.27, rolandic: 0.49 vs 0.19, parietal: 0.34 vs 0.16, occipital: 0.32 vs 0.11, temporal: 0.24&0.16 vs 0.11, for match vs CM nucleus respectively) (Supplementary Fig. 2a). Hence, our analysis confirmed that the fibers originating from the CM, while diffuse across all lobes, appear to be less specific and more sporadic, hence corroborating its indication as a more suitable target in generalized, rather than focal, epilepsy (Supplementary Fig. 2a, b).

Fig. 2: Anatomical connectivity of PUL, ANT, and VIM/VOP via neuroimaging and electrophysiology.

a Representative example of high-definition fiber tracking (HDFT) from PUL, ANT, and VIM/VOP nuclei (S01). b Volume of thalamocortical projections (mean ± standard error, n = 14 patients) from each nucleus to each cortical lobe normalized by the total volume of fibers projecting from each nucleus. c Group-level analysis of thalamocortical EPs. For each patient, the mean peak to peak amplitude is calculated across the contacts of a representative electrode for each cortical lobe. The normalized peak to peak amplitude of EP is color coded according to the stimulating nucleus evoking the response (blue for PUL, yellow for ANT, red for VIM/VOP). d, g, j Anatomical localization of SEEG implantation of representative patients (S01, S21, S13) that were tested for EPs with stimulation of PUL, ANT, and VIM/VOP, respectively. Each electrode is color-coded according to the brain lobe, and thalamic electrodes are indicated by an arrow. Figures were generated in Brainstorm⁸⁵. e, h, k Raw traces of stimulation triggered averages in different brain lobes from PUL, ANT, and VIM/VOP, respectively. Specific anatomical regions are: frontal operculum (pars opercularis), supplementary motor area, angular gyrus, inferior temporal sulcus (e, left to right); superior frontal sulcus, motor cingulate, posterior cingulate, posterior uncus (h, left to right); anterior cingulate, supplementary motor area, precuneus (k, left to right); f, i, l Violin plot of peak to peak amplitude of EPs in different brain lobes for PUL, ANT and VIM/VOP respectively. In all violin plot, the center line indicates the median, box bounds represent the first and third quartiles, whiskers extend to the most extreme data points within 1.5× the interquartile range, and points outside this range are plotted as outliers. For all panels, statistical significance and rejection of the null hypothesis of no difference of means with a 95% CI, was assessed with two-tail bootstrapping with Bonferroni correction. Triple asterisks indicate p << 0.0001.

We then confirmed the anatomical organization of PUL, ANT and VIM/VOP nuclei by analyzing thalamocortical evoked potentials (EPs) in the different cortical areas (Supplementary Fig. 3). For this, we applied low frequency (1 Hz) electrical stimulation to the different thalamic nuclei at 1–3 mA while recording cortical neural response evoked by each pulse in the different cortical lobes (Fig. 2d, g, j). We then analyzed stimulation triggered averages of EPs. Following thalamic stimulation, we observed robust cortical EPs in anatomically connected (i.e., matching), but not in unmatching, cortical regions (Fig. 2e, h, k). Indeed, analysis of EPs peak to peak amplitudes showed significantly larger responses in those areas that matched the previously identified projections, further confirming a precise anatomical organization of distinct thalamic nuclei (Fig. 2c, f, i, l).

In summary, we confirmed with multiple techniques a specific anatomical organization of the thalamocortical projections of PUL, ANT, and VIM/VOP in the studied population. We then define hodologically-matching (parietal, occipital, and temporal for PUL, frontal and temporal for ANT, and rolandic for VIM/VOP) and unmatching (frontal and rolandic for PUL, rolandic, parietal, and occipital for ANT, frontal, temporal, parietal, and occipital for VIM/VOP) regions.

Increased thalamocortical coupling during seizures in hodologically-matching SOZs

Given this specific structural organization of the thalamocortical fibers, we hypothesized a preferential functional involvement in the course of spontaneous seizures of the thalamic subnucleus that anatomically matches the SOZ (Fig. 3d). To test this hypothesis, we analyzed a total of 274 spontaneous (i.e., not induced by cortical electrical stimulation) ictal events. The anatomical location of the SOZ as well as the relevant time markers (pre-ictal, beginning, and end) of each seizure were determined by a certified clinical team. For each seizure, we computed the well validated non-linear h2 correlation coefficient^32,48,49,50 to identify time-resolved synchrony between each thalamic nucleus and each cortical SOZ. Specifically, we computed mean h2 coefficient during the ictal event (from seizure onset to termination) and the mean h2 coefficient during a 10 s baseline epoch spaced 2 min apart from the seizure’s onset³². We then compared mean h2 coefficient between the two phases separately for anatomically connected and non-connected thalamic nuclei to quantify the degree of specificity of thalamic involvement during spontaneous seizures.

Fig. 3: Functional coupling of the PUL, ANT, and VIM/VOP nuclei with SOZ during seizures.

a Representative examples of thalamocortical connectivity profiles of PUL (right, S03), ANT (middle, S23), and VIM/VOP (left, S25) with matched SOZ. For all nuclei, the top panel shows the cortical recording in the contacts of the SOZ, the middle panel shows the thalamic recording, and the bottom panel shows the h2 correlation coefficient. The gray shaded area includes the seizure duration (from onset to termination). b Boxplot of the mean h2 correlation for each seizure in matched and unmatched thalami, compared to mean h2 correlation during a baseline epoch preceding the seizure. Three representative patients are shown (right: S01, matched PUL, unmatched ANT; middle: S21, matched ANT, unmatched PUL; left: S12, matched PUL, unmatched VIM/VOP). c Raw traces of h2 correlation coefficient (smoothed for visualization) for a representative seizure for each patient of (b). The gray shaded area includes the seizure duration (from onset to termination). d Schematic illustrating the hodologically matching (bold colored lines) and unmatching (dashed gray lines) thalamocortical connections, created as detailed for Fig. 1a. e Violin plot showing the mean h2 coefficient across seizures for each matched (top) and unmatched (bottom) thalamus-SOZ pair. f Box plot of mean h2 coefficient across seizures during initiation, middle, and termination phase of each seizure. Each dot represents a patient. g Percentage difference of the change in h2 coefficient during seizures between matched and unmatched nuclei, across different frequency bands of interest. For all boxplots, the whiskers extend to the maximum spread not considering outliers, central, top, and bottom lines represent median, 25th, and 75th percentile, respectively. In all violin plot, the center line indicates the median, box bounds represent the first and third quartiles, whiskers extend to the most extreme data points within 1.5× the interquartile range, and points outside this range are plotted as outliers. For all panels, statistical significance and rejection of the null hypothesis of no difference of means with a 95% CI, was assessed with two-tail bootstrapping with Bonferroni correction. Triple asterisks indicate p << 0.0001.

As hypothesized, we found that while anatomically-matched nuclei showed a significant increase of synchrony between the thalamus and the SOZ, indicating functionally specific thalamic involvement during ictal events (Fig. 3a), minimal or no correlation was found in unmatched nuclei (Fig. 3b, c). This was consistent for all thalamic nuclei of interest and subjects (Fig. 3e, mean h2 was 0.13 vs 0.21 in matched with p < 0.001, and 0.12 vs 0.12 with p > 0.05 in unmatched, for baseline and seizure, respectively). Importantly, we also observed that thalamic nuclei had minimal or no correlation with non-SOZ regions (Supplementary Fig. 4a, b), demonstrating the specificity of the correlation with the anatomically-matched SOZ. Importantly, in 4 patients (S01, S22, S26, S33) that presented a thalamic contact also in the CM nucleus, we asked whether a thalamic site with more diffused but sparse connectivity (as the CM) would show a similar functional involvement as a thalamic nucleus with more defined and selective projections, as the hodological match. Interestingly, we found that the mean h2 coefficient during seizures of the SOZ with the CM, was either lower or unchanged compared to the h2 coefficient between the SOZ and the hodological matching nucleus (Supplementary Fig. 2e). Although preliminary, these results suggest that, while CM may remain a useful target in certain contexts, focal seizure networks may benefit more from stimulation of thalamic regions with stronger, selective connectivity to the SOZ. Additionally, for the single patient diagnosed with generalized epilepsy we found that PUL, ANT, and VIM/VOP nuclei had a lower correlation with the cortical regions exhibiting interictal and ictal epileptiform activity (0.16, 0.25, 0.21 mean h2 across 24 seizures and 13 channels for PUL, ANT, and VIM/VOP, respectively) as compared to the CM (0.3, Supplementary Fig. 2c, d). This confirms the more diffused profile of the CM connectivity (both anatomically and functionally), entailing its suitability for generalized, rather than focal epilepsy treatment. Overall, these functional outcomes robustly validated the thalamocortical anatomical organization observed with fiber tracking and thalamocortical EPs, further supporting the proposed definition of hodologically-matching and unmatching nuclei.

To further investigate possible mechanisms of thalamocortical interactions during focal seizures, we explored the coupling of the thalamus and SOZ in terms of seizure phases and frequency bands for matched thalamic nuclei. In particular, we aimed at verifying that the mechanism of thalamic involvement during seizures would not differ across the three nuclei of interest, since a different mechanism could imply a different design of electrical stimulation therapies. First, we asked whether maximal coupling of matched nuclei was specific to distinct seizure phases. To this aim, we computed the mean h2 correlation coefficient in three epochs (as previously reported³²): seizure initiation (including the 10 s preceding the first appearance of a tonic discharge in the SOZ contacts), middle part of the seizure (defined as the time interval separating seizure initiation and termination), and seizure termination (which includes the last 10 s of discharge). We found that the highest synchrony between the thalamus and SOZ occurred at seizure termination (Fig. 3f). Importantly, this phenomenon was present for all three thalamic nuclei when hodologically (i.e., anatomically and functionally) matched to the SOZ (Supplementary Fig. 4c). These findings support and validate the notion that synchronization loops between the thalamus and cortex might contribute to the termination of seizures, as suggested in previous studies in animals^32,51 and humans^52,53.

Additionally, we sought to determine if the observed synchrony presented a specific spectral profile. Hence, we compared the mean h2 coefficient computed in different frequency bands of the intracerebral recordings (delta, theta, alpha, beta, low gamma, high gamma, ripple, fast ripple). No spectral specificity was revealed: indeed, for all the frequency bands except the fast ripple band, the correlation between hodologically-matched thalamic nuclei and SOZ always increased significantly during the ictal event, while it showed no variation for unmatched thalamic nuclei (Supplementary Fig. 4d). Hence, we asked whether specific frequency bands would better distinguish hodologically-matched and unmatched nuclei potentially suggesting biomarkers for the refinement of the stimulation design. For this, we compared the increase in ictal correlation respect to baseline between hodologically-matched and unmatched nuclei in different frequency bands. Interestingly, we found that theta, alpha, and beta frequencies had a higher difference between the h2 correlation change during seizures between hodologically-matched and unmatched nuclei (Fig. 3g). These results show that the theta-beta frequencies might be more relevant to differentiate thalamic nuclei based on their connectivity to the SOZ.

All in all, these findings demonstrate that thalamic nuclei exhibit enhanced neural synchrony with hodologically-matched cortical SOZs, particularly at the period of seizure termination, and primarily driven by theta, alpha, and beta oscillations.

Thalamocortical synchrony reverses directionality during spontaneous seizures

While the h2 index revealed an increase in correlation between the hodologically-matched thalamic nuclei and the SOZ throughout the seizure course, here we sought to determine the directionality of this coupling. Importantly, previous studies reported a change in the directional coupling of thalamocortical interactions in temporal lobe epilepsy throughout the seizure course^32,54. Specifically, thalamic involvement was shown to occur early in a seizure, although it did not typically lead at the onset³². Additionally, animal studies have shown that the thalamus participates during seizures and plays a crucial role in their development, but it is less involved at seizure onset⁵¹. In order to explore whether a similar behavior was prevalent for all thalamic structures here investigated, we further expanded the analysis of hodologically-matched thalamocortical synchrony during spontaneous seizures by investigating the directionality of the coupling. To this purpose, we used Granger Causality (GC)⁵⁵ between the thalamic nuclei of PUL, ANT, and VOP/VIM and the hodologically-matching SOZs. This methodology could indeed provide insights in the understanding of propagation patterns of epileptic activity helping the future design of stimulation patterns aimed at decreasing epileptiform activity. Since our previous results suggest a crucial role in the theta-alpha-beta oscillations, for this analysis, we included only band-passed filtered (4–30 Hz) seizures that showed a clear increase in h2 coefficient in this frequency range and lasted at least 10 s. This resulted in a total of 123 seizures. For each seizure, we extracted two 10 s epochs of interest at seizure initiation and seizure termination and computed the GC matrix for the SOZ contacts and the hodologically-matching thalamic nucleus contacts (Fig. 4a). Importantly, GC of unmatched nuclei was not explored because of the lack of correlation with SOZ. To test whether the role of different thalamic nuclei was changing over the course of the seizure, we compared the GC coefficient from the thalamus towards the SOZ (th→SOZ) and from the SOZ towards the thalamus (SOZ→th).

Fig. 4: Analysis of thalamocortical versus corticothalamic interactions during seizures.

a Representative seizure recording (S01). The red bar indicates the full seizure duration, while the shaded gray regions show the initiation and termination epoch. On such windows, we calculated the GC coefficient matrix that illustrates connectivity from and to the thalamic nuclei to a matched SOZ. b Pie charts illustrating the proportion of seizures with higher corticothalamic (SOZ→th, light gray) or thalamocortical (th→SOZ, dark gray) GC coefficient for the initiation (left) and termination (right) epoch. c (Left) Plot of the evolution of corticothalamic and thalamocortical GC coefficients for 10 s epochs starting two minutes before a representative seizure. The arrows indicate the relative change of GC from initiation to termination. (Right) Quantification of the relative percentage change of GC coefficient from initiation to termination, for corticothalamic (light gray) and thalamocortical (dark gray) direction. d Boxplot of the relative percentage change of GC coefficient from initiation to termination for all subjects (n = 23 and n = 26). Outliers were removed with quartile method (see “Methods”). Each data point represents the mean value across all the seizures for each subject. e Pie charts illustrating the proportion of seizures with higher corticothalamic (SOZ→th, light color) or thalamocortical (th→SOZ, dark color) GC coefficient for the initiation (left) and termination (right) epoch, separated from the three nuclei of interest (PUL at left, ANT at middle, VIM/VOP at right). f Boxplot of the relative percentage change of GC coefficient from initiation to termination for all the seizures separated for the matched thalamic nucleus investigated (PUL at left, ANT at middle, VIM/VOP at right). For PUL n = 55 seizures, for ANT n = 19 and n = 20 seizures, for VIM/VOP n = 38 and n = 35 seizures. For all boxplots, the whiskers extend to the maximum spread not considering outliers, central, top, and bottom lines represent median, 25th, and 75th percentile, respectively. Outliers were removed with quartile method (see “Methods”). For all panels, statistical significance and rejection of the null hypothesis of no difference of means with a 95% CI, was assessed with two-tail bootstrapping with Bonferroni correction. Triple asterisks indicate p << 0.0001.

At seizure initiation, in the majority of the seizures (~ 70%), the GC coefficient of SOZ→th was higher than the GC coefficients th→SOZ (Fig. 4b for all subjects together, Supplementary Fig. 5 for each individual), suggesting a leading role of the cortex. In contrast, more than half of the events (~ 60%) presented a higher th→SOZ coefficient at the termination of ictal events, hence indicating a change in the leading anatomical structure. This could be due to (1) SOZ→th decreasing while th→SOZ remaining stable (or decreasing less) throughout the course of the seizure, (2) th→SOZ increasing and SOZ→th remaining stable (or increasing less). To test for this, we computed the relative change of the GC coefficient between termination and initiation and compared it between SOZ→th and th→SOZ. Interestingly, we noticed an increase in both the thalamocortical (th→SOZ) and corticothalamic (SOZ→th) coupling throughout the duration of the event (Fig. 4c). However, the relative change was higher for the th→SOZ consistently across all the subjects, further demonstrating a leading role of the thalamus at seizure termination (113.13 vs 321.01%, for SOZ→th and th→SOZ, respectively) (Fig. 4c, d).

Importantly, we verified that the same characteristics of thalamocortical interactions were not specific to a particular nucleus. As a matter of fact, in all studied nuclei, the absolute GC value in the initiation epoch and termination epoch showed a reverse trend, with a primary leading role of the cortex at first (75% for PUL, 59% for ANT, 60% for VIM/VOP), followed by a predominantly leading role of the thalamus at termination (54% for PUL, 68% for ANT, 52% for VIM/VOP) (Fig. 4e). Additionally, PUL, ANT and VIM/VOP showed a significantly higher relative change for th→SOZ throughout the seizure (Fig. 4f).

In summary, here we investigated the underlying mechanisms of thalamocortical and corticothalamic coupling involving the ANT, PUL, and VIM/VOP with hodologically-matched SOZ during spontaneous seizures. Our findings suggest a potential trend where cortical regions exert a dominant influence at seizure initiation, while the thalamus may assume a more prominent role as the ictal event progresses. These findings may have clinical relevance in neuromodulation strategies for stimulation: given that all three nuclei exhibited comparable patterns of activation during ictal events, our data suggest that the temporal dynamics of thalamic influence on cortical networks may be a common phenomenon across thalamic structures. As such, tailoring the modality of stimulation (such as the timing of stimulation onset in relation to the seizure onset) to each specific nucleus may not be essential or necessary.

Matched thalamic stimulation acutely suppresses epileptiform discharges

Given the peculiar hodology that we previously found both in the structural connectivity of the PUL, ANT, and VIM/VOP, as well as in the functional coupling observed during ictal events, we hypothesized that targeting electrical stimulation to the thalamic nucleus that hodologically-matches the cortical SOZ might be more effective in reducing epileptiform activity than unmatched nuclei. To test this hypothesis, we delivered bipolar thalamic stimulation through the SEEG electrodes in 17 patients in acute settings (in the epilepsy monitoring unit, EMU). For clinical reasons, the stimulation testing was always performed at the end of the clinical EMU evaluation, after the patient had shown spontaneous seizures. We evaluated the effect of hodologically-matched and unmatched stimulation on interictal epileptiform discharges (IEDs), namely interictal spikes. We used an automatic validated algorithm^56,57 to extract IEDs from cortical recordings in the SOZ (Fig. 5a), previously subdivided in 5 s epochs and denoised (see “Methods”). We aimed at quantifying the difference between baseline (including the epochs in the 2 min preceding the first stimulation applied) and stimulation epochs (see “Methods”). For this, we performed two distinct analysis: (a) in n = 8 patients with continuous evident pathological spiking (Fig. 5a), we computed the rate of IEDs per minute in baseline (Stim OFF) and stimulation (Stim ON) epochs; (b) in n = 9 patients with more than half of the baseline epochs without pathological spikes, we computed the probability of suppressed IEDs in stimulation epochs and compared it to the probability of suppressed IEDs in baseline epochs, hence obtaining a percentage change in IED suppression probability (Supplementary Data 2). In the first analysis, a positive effect of the stimulation (meaning reducing pathological IEDs) should result in a smaller IEDs rate during stimulation epochs than during baseline (stimulation OFF) epochs; whereas in the latter, a positive number indicated that the stimulation epochs are more likely to present no IEDs than the baseline epochs (without stimulation).

Fig. 5: Hodologically-matched electrical stimulation of the thalamus better reduces pathological IEDs.

a Representative example of an interictal SEEG recording of a SOZ contact, with clearly visible IED (spikes) and their time-frequency representation exploited for automatic detection. b (Left) boxplot for IED rate/minute in stimulation OFF (n = 18 epochs) and stimulation ON (n = 8 epochs) of the PUL (matched nucleus). (Right) representative stim OFF (top) and stim ON epoch (bottom) with matched PUL stimulation shows immediate suppression of IED. c Representative example, as in (b), for matched ANT stimulation (n = 7 and n = 24 epochs for stimulation OFF and ON, respectively). d Representative example, as in (b and c), for matched VIM/VOP stimulation (n = 8 and n = 21 epochs for stimulation OFF and ON, respectively). e Boxplot of representative examples of unmatched thalamic stimulation on IED rate for ANT (left, n = 24 vs n = 11 for stimulation OFF and ON, respectively), and VIM/VOP (right, n = 19 vs n = 41 for stimulation OFF and ON, respectively). f Percentage change relative to baseline of the probability of IEDs suppression in matched and unmatched thalamic stimulation in two representative patients. g All-subject analysis of the IED rate percentage variation to baseline with matched (left) and unmatched (right) stimulation. h All-subject analysis of the percentage change relative to baseline of the probability of IEDs suppression in matched (left) and unmatched (right) stimulation. For all boxplots, the whiskers extend to the maximum spread not considering outliers, central, top, and bottom lines represent median, 25th, and 75th percentile, respectively. For all panels, statistical significance and rejection of the null hypothesis of no difference of means with a 95% CI, was assessed with two-tail bootstrapping with Bonferroni correction. Triple asterisks indicate p << 0.0001.

As hypothesized, we found that electrical stimulation of the PUL, ANT and VIM/VOP significantly suppressed cortical IEDs when the SOZ was in a cortical region that hodologically-matched the thalamic nucleus stimulated (Fig. 5b–d, g and Supplementary Fig. 6). Overall, our analysis revealed that hodologically-matched but not unmatched thalamic stimulation significantly reduced the IED rate in all tested patients (for matched stimulation, IED variation to baseline ranged from −29.12% to −73% with p < 0.01 and p < 0.001, while for unmatched stimulation, it ranged from −4.2% to −25%) (Fig. 5g). Here, we highlight the IED rate values for a representative patient for each nucleus of interest: we observed a reduction of the IED rate of 132 vs 78 spikes/minute for the PUL (median rate at stimulation OFF vs stimulation ON in S01, p < 0.001) (Fig. 5b), 144 vs 60 spikes/minute for the ANT (in S23, p < 0.001) (Fig. 5c), and 132 vs 48 spikes/minute in the VIM/VOP (in S13, p < 0.001) (Fig. 5d). Similarly, when evaluating the percentage change in the probability of IED suppression, we found that hodologically-matched thalamic stimulation epochs were more likely to present no IEDs when compared to unmatched thalamic stimulation epochs. Indeed, In 5 out of 6 patients that received stimulation in the nucleus matching their SOZ, we found at least a 10% increase (up to over 200%) in the probability of epochs with no IEDs with stimulation on with respect to stimulation off (Fig. 5h); whereas the stimulation effect was more inconsistent (−22.6 to 10.66%) when the targeted nucleus was not matching the hodology of the thalamocortical fibers towards the SOZ (Fig. 5h). This finding further suggests that the more effective thalamic target to suppress IEDs is the one matching the hodology of the SOZ.

Overall, we showed that only targeted electrical stimulation of the nuclei hodologically-matching the SOZ is significantly reducing the rate of IEDs and increasing the probability that stimulation epochs present no IEDs (complete suppression). These findings are crucial in guiding approaches to address the heterogeneity of SOZs and in clinically selecting optimal stimulation targets for the electrical neuromodulation of focal epilepsies.

Matched thalamic electrical stimulation in chronic epilepsy treatment

Finally, to verify whether the observed immediate effects on IEDs could be translated in a long-term clinical improvement in epileptic patients, we retrospectively evaluated 10 patients (six of which being part of the SEEG cohort) that were chronically implanted for thalamic stimulation (in the PUL, ANT, or VIM/VOP) as part of their clinical care. Specifically, n = 7 patients received a chronic implant in the matched thalamic nucleus (S33–S39) and n = 3 patients received the implant in an unmatched thalamic nucleus (Fig. 1e and Supplementary Fig. 1b). Across all matched patients, 6 (S3, S35–S39) presented focal MRE originating from the posterior quadrant regions and hence received a neurostimulation system in the PUL, while one (S34) was implanted in the ANT to treat a frontal onset MRE (Supplementary Table 1). On the other hand, the unmatched chronic cohort received a chronic implant in the ANT with SOZs located in the rolandic (S32) and posterior quadrant regions (S40–41). Overall, in the matched group we observed a mean reduction in seizure frequency of 87.5 %, which dramatically outperforms the mean reduction of 8.3% observed in the unmatched group (Fig. 6a). As a representative example, for S33 seizures reduced from 5 per week to 1 per month, and the patient also self-reported a significant reduction in seizure severity, noting no missed workdays since the implant of the stimulation device, compared to missing 1–2 days per month prior to surgery. Additionally, the anti-seizure medication (ASM) intake for each patient, that we report here as a supportive observation, decreased in number or dose for 5/7 patients since the matched neurostimulation system was implanted, and did not increase for any of the participants (Supplementary Table 2). No changes of ASM were reported in the unmatched group. Importantly, albeit preliminary, our results align with earlier research, limited to the study of ANT stimulation, as reported from the SANTE trial, at 7 years follow-up. Indeed, in this previous study, unmatched stimulation (ANT stimulation for seizure onsets other than frontal or temporal) only achieved 39% reduction in seizure frequency, whereas outcome for matched SOZ were comparable to the ones we observed (78% and 86% for temporal and frontal onset, respectively)²⁹ (Fig. 6b). While direct comparisons are limited by methodological and anatomical differences, the observed consistency strengthens the hypothesis that hodology-informed targeting may improve the efficacy of chronic neuromodulation.

Fig. 6: Hodologically-matched thalamic stimulation as a neuromodulation therapy in chronic epilepsy treatment.

a Bar plot of the percentage of seizure frequency reduction for chronically implanted patients S32–S41. Purple bars indicate hodologically-matched thalamic targeting, while grey bars indicate hodologically-unmatched thalamic targeting. b Comparison of results with available literature: median seizure frequency reduction reported by the SANTE trial at 7 years follow-up for frontal, temporal (matched, in purple), and other lobes (unmatched, in grey).

In conclusion, our results suggest that targeting thalamic stimulation to match the hodology of the thalamocortical projections could translate in significant therapeutic benefits, including notable reductions in the frequency and severity of disabling seizures. This approach, albeit necessitating future larger studies to infer clinical efficacy, lays the groundwork for more personalized and precise target selection in neuromodulation for medically refractory focal epilepsy, particularly in patients who are not candidates for curative resective or ablative interventions.

Discussion

Electrical stimulation of the thalamus is undergoing a rapid advance as a treatment opportunity for patients with MRE^12,58,59, with indications in both generalized^22,58,60 and focal epilepsy^{9,61,62,63,64,65}. While in generalized epilepsy the CM nucleus effectively reduces seizure events, in focal epilepsy, this expansion requires a profound understanding of the intricate anatomical and functional thalamocortical interactions with highly heterogeneous and individualized epilepsy phenotypes to improve clinical outcome. While in recent years there has been an increasing interest in expanding beyond the ANT^38,66,67, recent isolated studies have often been limited by small sample sizes, single nucleus targeting, and lack of systematic hypothesis-driven testing^37,39,68,69. Here, we proposed a novel approach to select the optimal targets for thalamic electrical stimulation in focal epilepsy, considering the specificity of SOZs in relation to cortical and subcortical anatomy. This framework leverages a novel integration of neuroimaging-derived structural connectivity patterns, then cross-validated with intracranial electrophysiological functional data and clinical assessments tested on an unprecedent large dataset. In contrast to traditional investigational methods (such as fMRI), which often rely on single modalities and lack patient-specific electrophysiological validation, our hodologically-based targeting features improved spatial and temporal precision, embracing both the evolving dynamics of thalamocortical interactions and the subject-specific nuances of anatomy and pathology. In this work, we demonstrated preliminary evidence that leveraging the hodology-based interactions between certain subnuclei of the thalamus and the epileptogenic cortex results in stronger neuromodulator effect. This approach highlights the importance of thalamocortical hodology in individualizing stimulation targets, potentially advancing therapeutic outcomes, and paving the way for more effective neuromodulation strategies in patients with focal MRE who are not candidates for curative interventions.

Since the early stages of epilepsy surgery and the first mechanistic studies about epileptogenicity, focus has increasingly turned towards deeper brain structures, particularly the thalamus^70,71. Early animal studies by Penfield and Jasper in 1947 identified the thalamus as a potential generator of epileptic cortical activity, proposing the theory of “centrencephalic” epilepsy^18,72. Indeed, this structure mediates reciprocal cortical-subcortical connections, functioning as an ‘integrative hub’ for functional brain networks⁷³. Recently (challenging previous theories), the thalamus has been implicated not in the generation, but in the propagation of generalized and focal onset seizures, due to its role in the bi-hemispheric cortical spread of epileptiform activity^35,52,74. In addition, more recent evidence indicates that the thalamus plays a critical role also in the network dynamics of focal-onset epilepsies^19,32,34, but lacked a consistent and systematic analysis of the hodological patterns of propagation from different cortical and thalamic regions during ictal events. Our results contribute to this discussion. Indeed, we detected significant correlation with ictal activity in all recorded thalamic nuclei, further suggesting the thalamus as a potential neuromodulation target to treat focal MRE. However, correlation between thalamic and cortical SOZ activation was specific to thalamocortical anatomical connection, as identified by high-density tractography as well as electrophysiological assessment, supporting the hypothesis that epileptic propagation patterns during seizures obey predictable anatomo-functional patterns and highlighting the need to carefully select thalamic nuclei for neuromodulation according to individual clinical and electrophysiological data.

We applied this mechanistic understanding of the thalamocortical interactions to inform hodology-based electrical stimulation aimed at suppressing epileptiform activity. Our results demonstrated that hodologically-matched thalamic stimulation immediately suppresses IEDs. Additionally, this targeted stimulation markedly enhances the likelihood of achieving complete suppression of IEDs during stimulation epochs, suggesting robust potential for targeted thalamic stimulation as a therapeutic strategy in epilepsy neuromodulatory management. Immediate effects in n = 17 subjects, paralleled with the drastic clinical improvement in seizure frequency with chronic matched thalamic stimulation (up to 95%) in seven patients, provide promising evidence that a hodology-based stimulation would improve clinical outcome for patients with focal epilepsy. Importantly, our comparative analysis revealed a markedly greater seizure reduction with hodology-matched thalamic stimulation compared to unmatched targets (mean: 87.5% vs. 8.3%). These findings underscore the limitations of commonly applied clinical practices, which often apply a uniform stimulation target across diverse SOZs, without accounting for individual variations in the anatomical location and propagation patterns of epileptic networks.

Mechanistically, our results identified a trend suggesting that the epileptic cortex is the dominant structure at the initiation of seizures, while the thalamus assumes a leading role throughout the course of ictal events, showing higher synchronization towards the middle and termination of the ictal events. Importantly, this pattern was consistent across all thalamic regions investigated. In this regard, in an animal model of limbic epilepsy induced by GABA antagonist (i.e., bicuculline), Aracri and colleagues demonstrated hippocampal activation first, progressing to hippocampus-thalamus synchronization, and then thalamic bursts leading to the termination of seizures⁷⁵. Destexhe and colleagues also highlighted the role of cortically induced coherence of thalamic-generated oscillations⁷⁶. They implicated temporal and spatial thalamocortical interactions, suggesting that, at the onset of seizures, thalamocortical cells are hyperpolarized due to the greater power of excitatory cortical projections to reticular neurons compared to thalamic relay neurons. Once the initial cortical fast phase is abolished, thalamic relay neurons projecting to the cortex depolarize synchronically, reinforcing the coherence of seizure activity arising from the cortex, and promoting the highly synchronous cortical spike phase observed during the middle and end of seizures. Our findings related to the directionality of epileptic activity through thalamocortical circuits align with this mechanistic theory, where information flows to the thalamus from the epileptic cortex at seizure initiation and back to cortical areas during the late seizure phase. This supports the hypothesis that the thalamus plays distinct roles in this dynamic process⁷⁷. Furthermore, it suggests the possibility to design closed-loop stimulation paradigm to restrict the stimulation to precise moment of the seizure development with potentially stronger neuromodulatory effects. This highlights the need for future studies to explore finer temporal parcellations of the thalamocortical interactions dynamics throughout ictogenesis and seizures development that could inform phase-specific neuromodulation strategies. In this regard, our findings indicate that the theta to beta frequency range represents a key hallmark of thalamocortical connectivity⁷⁸, which could serve as a biomarker for seizure detection and closed-loop monitoring.

The primary limitation of this study lies in the inherent sampling bias and protocol variability characteristic of SEEG-based clinical research. Nevertheless, our cohort is distinctive in its inclusion of a broad range of thalamic nuclei and SOZ combinations—both hodologically matched and unmatched (Supplementary Fig. 1a). Notably, the study design incorporates extensive and well-characterized intra- and inter-subject comparative controls, providing a level of analytical rigor and granularity that distinguishes it from prior investigations. This, combined with the integration of anatomical, electrophysiological, and stimulation data, provides an unprecedented level of detail and opens new avenues for the personalization of neuromodulation therapies. Importantly, another limitation of this study is that the clinical assessment of electrical stimulation as a chronic therapy was evaluated in just 10 patients as an observational retrospective report, and although compelling must not be interpreted as a definitive efficacy proof. Hence, inhomogeneous evaluation of all thalamic nuclei and SOZs occurred. However, these results are paralleled to a thorough investigation of the mechanisms of thalamocortical coupling in focal epilepsy, which support our findings. Additionally, even if limited in number, our evaluation of the effect of matched stimulation in a chronic stage of the treatment results in a drastic clinical improvement well above other literature reports. While we acknowledge that the working principles of DBS (open-loop) and RNS (closed-loop) are fundamentally different, a recent meta-analysis⁷⁹ on over 50 studies reported no significant different in seizure reduction or responder rate across the two systems. Yet, larger controlled clinical studies aimed at assessing clinical efficacy are now necessary to confirm the efficacy of hodologically-matched stimulation in the reduction of seizure frequency and improvements in patients’ quality of life in chronic settings. Another study limitation includes sampling bias associated with explorations of the human brain using intracranial electrodes, where conclusions can only be drawn from areas where electrodes were implanted. Nevertheless, strict inclusion and exclusion criteria minimized this limitation by focusing on subjects with highly localized seizures, many of whom became seizure-free after restricted focal resections (Supplementary Data 1). More granular parcellation of both cortical and thalamic sub-regions may serve in the future to optimize hodological resolution and enhance the anatomical precision of connectivity-based targeting. Additionally, the applied parameters of acute stimulation were restricted to few parameters (100 Hz only, 1–3 mA, continuous stimulation), mainly because the focus of the current study was related to the thalamic targeting and not to the different paradigms of electrical stimulation. However, several studies previously identified the optimal stimulation parameters in the high-frequency range (> 100 Hz)^80,81,82, as in our chronic participants. Finally, while all acute stimulation was unilateral and not designed to compare laterality effects, our findings support the feasibility of hodologically guided unilateral DBS in patients with well-lateralized seizure networks, warranting further investigation. Future studies will need to better refine stimulation parameters and identify optimal stimulation frequencies.

In summary, understanding the thalamocortical hodological principles governing the organization of focal seizures and their potential clinical applications is invaluable. Personalizing therapies based on the connectivity of the SOZ to thalamic sub-nuclei and individual seizure characteristics may lead to improved efficacy and more favorable outcomes in the neuromodulation field for epilepsy. Our work addresses the critical unmet clinical need to identify optimal neuromodulation target by demonstrating predictable patterns of connectivity in thalamocortical interactions and guiding optimal thalamic targeting based on location of ictogenesis. This hodology-based approach would impact the current clinical standard in which one target fits all the SOZ (ANT in focal epilepsy, CM in generalized epilepsy), and lay the foundations for highly individualized neuromodulation treatments.

Methods

Participants information

All participants included in this study were diagnosed with drug-refractory epilepsy and underwent SEEG implantation at the University of Pittsburgh Medical Center as a part of their standard clinical care. They were hospitalized in the EMU from January 2020 to May 2025. All patients received a comprehensive neurological assessment, neuropsychological testing, routine MRI and CT and SEEG implantation. Informed consent was obtained from all patients or legal representatives. In order to prospectively study the anatomical and functional properties of the thalamus and the epileptic cortex via SEEG monitoring, patients were included in the study (S01–S37) in the present study if: (1) at least one SEEG electrode contact was in the PUL, ANT or VIM/VOP, (2) at least one of the electrophysiological recordings analyzed in the study was performed (minimum 2 spontaneous seizures, thalamocortical evoked potentials, HDFT, thalamic stimulation), (3) the patient had been diagnosed with focal epilepsy characterized by well-localized focal seizures as determined by SEEG, with seizure onset confined to a limited number of contiguous electrode contacts and without rapid early propagation. Exclusion criteria encompassed refusal in participating in the study and SEEG implantation complication (severe intracranial hemorrhage). We also retrospectively evaluated 10 participants (S32–S41, six of whom participated also in the prospective SEEG study component) that received a chronic thalamic stimulation system as part of their clinical care to evaluate long-term efficacy of thalamic stimulation. To be included in this retrospective portion of the study, the patients needed to fulfill the following inclusion criteria: (a) patients received a diagnosis of MRE focal epilepsy; (b) patients were not candidates for resective surgeries; (c) patients received a chronic implant in either the PUL, ANT, or VIM/VOP between January 2020 and March 2025.

A comprehensive overview of the patients’ demographics and epilepsy categorization, summary of the experiments, and a summary of the data collected for each patient are reported in Supplementary Data 1 and 2. Overall, we included 37 SEEG participants (S01–S37, 27 males, 10 females) and 10 chronically implanted patients (S32–S41), of age 36 ± 11.3 (mean ± SD). The age at seizure onset was 16.89 ± 14.2 years-old, while the duration of the epilepsy disease was 19.13 ± 13.5 years (mean ± SD). Additional information can be found in the Supplementary Data 1 and 2 and Supplementary Table 1.

SEEG implantation and localization

The number and location of SEEG electrodes implanted was pre-operatively planned individually for each patient as part of their clinical care and was not influenced by this study. The implant procedure was performed with robotic stereotactic guidance (ROSA, Zimmer-Biomet, Warsaw, IN, USA), applying bone fiducial registration with accuracy < 0.5 mm). Overall, we implanted a total of 7547 contacts (Microdeep® SEEG Electrodes, 12 to 18 channels, DIXI Medical, Marchaux-Chaudefontaine, France), with an average per patient of 203.4 ± 29.19. To target the thalamic nuclei of interest, we used the following stereotactic coordinates: (1) for PUL, in relation to the AC/PC reformatted planes, and having the PC (posterior commissure) as the reference point, coordinates are X: 2–10 mm lateral to the midline plane; Y: 0 to −6 mm posterior to PC; Z: 0 to −6 mm below the AC/PC defined horizontal plane; (2) for ANT, in relation to the AC/PC reformatted planes, and having the MC (mid commissural point) as the reference point, coordinates are X: 6–8 mm lateral to the midline plane; Y: 6 to 7 mm anterior to the MC; Z: 2 to 4 mm above the AC/PC defined horizontal plane; 3) for VIM/VOP, lateral (from about 5 to about 15 mm lateral to the AC/PC line); anterior/posterior (from about 6 to about 7 mm anterior to PC); and dorsal/ventral (from about +1 to about −2 mm from the AC/PC plane.

To reconstruct the position of the SEEG electrode, we used CURRY software (Compumedics NeuroScan, Hamburg, Germany). First, we marked key reference points on each pre-implantation MRI T1 scan (namely the AC, PC, and the nasion). These images were then co-registered with post-implantation CT, and a clinical team member manually identified and labeled all the SEEG contacts from the CT. All the contact locations of the subject MRI were visually inspected and confirmed by a certified neurosurgeon (JGM).

For all subject analysis (Fig. 1b), the electrode coordinates in the subject space MRI were then translated to coordinates in MNI space for each patient. For this, we first co-registered the post-implantation CT scans with the T1-weighted MRI using linear (affine) registration. We then registered the T1 MRI to MNI152 standard using affine registration by SPM mutual information algorithm. SEEG contact coordinates identified in the patient’s native space (CT) were thus transformed into MNI space by applying the concatenated transformations (CT-to-MRI affine + MRI-to-MNI affine). This procedure ensures precise anatomical localization of contacts across patients in a common stereotactic space, allowing for group-level visualization and analysis. Due to poor image quality, the electrode reconstruction was not successful for S36 and S37. The contact location within the thalamic nuclei of interest was visually confirmed for each patient by an experienced neurosurgeon (JGM), using the individual patients’ images to account for intersubject anatomical variability. The use of atlas-guided thalamic parcellations served as an auxiliary visualization tool. The assessment of SOZ contacts (as well as non-SOZ contacts used in the following analysis) was confirmed by an experienced neurosurgeon (JGM) (Supplementary Data 3). Specifically, the SOZ selection was based on clinical seizure annotations and electrophysiological review. If multiple SOZ contacts were present, we selected the contact with the earliest seizure onset pattern (e.g., low-amplitude fast activity or sharply demarcated onset) and minimal noise or artifact. As defined in the selection criteria, the SOZ was relatively focal in all patients, making it feasible to select a single representative contact pair for the analysis.

SEEG recording and stimulation

SEEG data were recorded with Natus Quantum System EEG diagnostic and monitoring system (Natus, Pleasanton, A, USA), with a sampling rate of 2048 Hz (for S25, S26, S37, the sampling rate was 1024 Hz) during the extra-operative monitoring at the EMU of the University of Pittsburgh Medical Center (Presbyterian Hospital). For all analysis, we applied SEEG bipolar montage to increase spatial selectivity and reduce noise levels. To deliver electrical stimulation in the thalamic contacts, we used the Nicolet Cortical Stimulator (Natus, Pleasanton, A, USA), which delivers biphasic pulses up to 100 Hz to ensure comparability of neuromodulatory effects (amplitude was set to 1–3 mA and pulse duration to 60–300 μs for all patients, individually adjusted to accommodate patient-specific thresholds and comfort levels) with bipolar electrode configuration.

High-definition fiber tracking

To estimate anatomical projections from the thalamus to various cortical areas, we first performed HDFT of diffusion MRI data. The diffusion images were acquired on a SIEMENS Prisma Fit scanner using a diffusion sequence (2 mm isotropic resolution, TE/TR = 99.2 ms/2490 ms, 257 diffusion sampling with maximum b-value 4010 s/mm²). Diffusion tensor estimation and tractography were performed using DSI Studio (https://dsi-studio.labsolver.org). The accuracy of b-table orientation was examined by comparing fiber orientations with those of a population-averaged template. The tensor metrics were calculated using DWI with b-value lower than 1750 s/mm². For fiber tracking, random values were used for tracking threshold (from 0.5 to 0.7 of Otsu’s threshold), angular threshold (from 15 to 90°), and step size (from 0.5 to 1.5 voxel distance). We utilized seed regions in thalamus to create white matter tracts to regions of interest in the cortex (seed-to-ROI fiber tracking). Seed regions were selected in ANT, VIM/VOP, PUL, and CM based on the extended Human Connectome Project multimodal parcellation atlas. Regions of interest were selected in the frontal lobe, rolandic area, parietal lobe, occipital lobe, and temporal lobe. Tracks with lengths shorter than 30 mm or longer than 1000 mm were discarded. A total of 10,000 tracts were placed. Topology informed pruning was applied to the tractography with 2 interactions to remove false connections. We then quantified the volume of white matter tracts projecting from each thalamic nucleus to each cortical area. For this, we calculated tract volume using the method described by Yeh⁸³, which converts the tractography streamlines into a set of unique voxels they occupy; the volume is then estimated by multiplying the number of these voxels by the voxel size. We normalized the volume of each white matter projection by the total volume projections from each thalamic nucleus (Fig. 2a, b and Supplementary Fig. 2a, b).

Thalamocortical evoked potentials

In order to confirm the anatomical organization of thalamocortical fibers with electrophysiological techniques, thalamocortical evoked potentials were elicited by thalamic stimulation of PUL, ANT, and VIM/VOP (frequency = 1 Hz, amplitude = 1–3 mA, pulse width = 300 μs) and recorded from all available SEEG channels with Natus recording system. SEEG intracerebral data from all cortical channels were filtered with a band-pass Butterworth 2nd order filter (1–1000 Hz) and DC offset was removed from all recordings. From each stimulation pulse, we extracted 520 ms epochs (20 ms prior and 500 ms after the stimulus) and computed stimulation triggered averages. Each epoch was baseline corrected with the pre-stimulus interval. We calculated the peak-to-peak amplitude of evoked potentials in each epoch as the difference between the maximum and minimum voltage value observed within the first 400 ms from the stimulus. We performed this analysis for all available channels that presented sufficient signal to noise ratio, that were located in the regions of interest for this study (not in thalamus) and that were not located in the white matter (Fig. 2d–l and Supplementary Fig. 3).

Acquisition and analysis of spontaneous seizures

Spontaneous seizures acquisition

During the extra-operative monitoring, we recorded 274 spontaneous seizures from 34 patients (duration mean ± std = 61.68 ± 67.96 s). The remaining 3 (S14, S18, S31) patients either presented technical problems that invalidated the inclusion of the seizures or did not present any spontaneous events during the hospitalization. For our analysis, we collected recordings from 2 min before the seizure onset up to 30 s after the seizure termination. These time events (onset and termination) were manually marked for each event by certified medical personnel. We used such markers to define 5 epochs of interested, used in subsequent analysis: (1) baseline (corresponding to 20 s occurring two minutes before the seizure onset); (2) seizure initiation (including the 10 s preceding the first appearance of a tonic discharge in the SOZ contacts); (3) seizure termination (including the last 10 s of discharge); (4) middle of seizure (as the epoch between the previous two intervals (5) total seizure duration (defined as the interval between onset and termination).

Non-linear correlation analysis

For each recorded seizure, we computed the h2 non-linear correlation coefficient, as implemented in Anywave⁸⁴ for each matched and unmatched pair of thalamus-SOZ bipolar contacts. The h2 coefficient between two signals is a time-resolved measure of their non-linear dependence. The coefficient was computed over a 2 s time window, sliding by steps of 1 s^19,48. We computed the h2 across the full SEEG spectrum (1–500 Hz), as well as separate frequency bands: delta (1–4 Hz), theta (4–8 Hz), alpha (8–15 Hz), beta (15–30 Hz), gamma (30–45 Hz), gamma2 (55–90 Hz), ripple (80–250 Hz), fast ripple (250–500 Hz). For each time interval of interest (seizure duration, seizure initiation, middle of seizure, seizure termination), we computed the mean h2 (Fig. 3b, e, f).

To compute the change in h2 between matched and unmatched nuclei (Fig. 3g), we calculated, for each matched and unmatched pair, the percentage increase of h2 in the seizure duration with respect to the baseline epoch, as:

$$h2\,{change}(\%)=100 * ({mean}(h{2}_{{seizure}})-{mean}(h{2}_{{baseline}}))/{mean}(h{2}_{{baseline}})\%$$

(1)

We then performed the subtraction between the percentage h2 changes calculated in matched thalamus-SOZ and unmatched thalamus-SOZ pairs.

For the correlation between the thalamus and non-SOZ regions, we randomly selected two seizures for each patient. We computed the mean h2 of the thalamus with 3 distinct non-SOZ pair of contacts (Supplementary Data 3).

Granger Causality analysis

To analyze the directionality of the interactions between the thalamus and the cortex, we used the Granger Causality method⁵⁵, as implemented in Brainstorm⁸⁵. First, we selected all the seizures that lasted at least 10 s and showed an increase in the h2 correlation coefficient in the theta to beta band. We filtered the raw SEEG signals in the 4–30 Hz range and divided each seizure recording in 10 s non-overlapping windows. For each window, pairwise Granger Causality was estimated using the default model order selection and parameters provided by Brainstorm, for each matched thalamus-SOZ pair. This computation results in a 4 × 4 matrix for each epoch, where the second diagonal contains the thalamocortical (th→SOZ) and corticothalamic (SOZ→th) GC value. In addition to the absolute GC value, we quantified per percentage change in GC from initiation to termination (Fig. 4c, d, f) as:

$$\Delta {GC}(\%)=100 * (G{C}_{{termination}}-G{C}_{{initiation}})/G{C}_{{initiation}}\left.\right)\%$$

(2)

where termination and initiation are the 10 s epochs preceding the seizure onset and the seizure termination time stamps, respectively. To avoid misinterpretations given by the variability in the number of seizures recorded from each patient, we highlight the single-subject results in Supplementary Fig. 5. In a total of 29 patients, 22 showed an increase in the proportion of seizures with the thalamus leading at termination (pie charts in Supplementary Fig. 5) and 22 showed higher $\Delta {GC}$ in the thalamocortical (and not corticothalamic) coefficient (bar plot in Supplementary Fig. 5). Such finding robustly confirms population-observed trends.

Case report: CM connectivity in generalized epilepsy

We had the unique opportunity to justify the exclusion of the CM in our study (Supplementary Fig. 2c, d), focused on focal epilepsy, by performing structural and functional connectivity analysis of the CM, as well as the PUL, ANT, and VIM/VOP on a patient affected by generalized epilepsy. In particular, the patient was a female in her late 20 s, with a severe MRE, with more than 10 generalized seizures per day. The patient underwent SEEG exploration to investigate the appropriate thalamic target for neuromodulation as a part of her clinical care.

We performed h2 correlation analysis of 24 seizures recorded with simultaneous sampling of CM, PUL, ANT, VIM/VOP as well as 13 bipolar electrodes sampling the following brain structures: insula, frontal operculum, superior temporal sulcus, superior temporal gyrus, middle temporal gyrus. We computed the mean h2 coefficient for each thalamic nucleus with the cortical electrodes. Unlike the focal epilepsy patients, where the SOZ only comprises a pair of cortical electrodes, here we averaged the h2 values of each thalamus-cortex pair.

High-frequency acute thalamic stimulation

Stimulation testing and protocol

During post-operative monitoring in the EMU, we had the unique opportunity to perform acute thalamic stimulation testing on 17 patients of our cohort. Importantly, we could stimulate in both matched and unmatched thalamic nuclei. We applied continuous electrical stimulation with bipolar configuration on the thalamic site of interest, at a frequency of 100 Hz. This value was chosen because it was the highest allowed by the stimulator used, and it is in line with commonly used parameters for thalamic electrical stimulation (≥ 100 Hz). The amplitude and pulse width were adjusted individually for each patient according to their comfort levels, but never exceeded the ranges of 1–3 mA and 60–300 us respectively.

For clinical safety reasons, the stimulation testing occurred only after the patients exhibited spontaneous seizures and in seizure-free days. All patients were awake during stimulation testing to minimize potential changes in arousal state that might have influenced electrophysiological assessments such as IEDs rate. The duration of the stimulation train varied from a minimum of 5 s to a maximum of 30 s, repeated over 1–10 times depending on the patient comfort and clinical state.

Preprocessing and IED detection

To analyze the immediate effect of thalamic stimulation we extracted the following data: (i) for baseline (Stim OFF), we collected 2 min of interictal activity in the SOZ channels before any kind of stimulation was delivered to the patient; (ii) for active stimulation period (Stim ON), we included the SOZ recording for the whole duration of stimulation up to 10 s after stimulation was stopped³⁶. We excluded the first two seconds after the stimulation was delivered to reduce false positive detection due to transient stimulation-induced burst⁸⁶. Unfortunately, for technical reasons, 5 patients had a shorter baseline period (average of 45 s). Prior epoching, we preprocessed the data in Anywave⁸⁴ by applying a band-pass filter (1–1000 Hz) and a notch filter (60 Hz) and manually removing channels affected by stimulation artifact or other noise sources. On the remaining channels, we applied independent component analysis (ICA) and extracted n components (where n = number of channels-1). This method is a common technique used to separate independent sources linearly mixed in several sensors. We visually inspected all the components and rejected the ones associated with stimulation artifacts. In this way, we could reliably apply an automatic spikes detection algorithm on the stim ON and stim OFF recordings.

Each intracranial recording of interest was then divided in 5 s non-overlapping epochs, on which we computed the detection of epileptiform discharges. To avoid any confound due to the stimulation artifact, we considered as IED only interictal spikes, and not high-frequency oscillations. For this, we used Delphos software^56,57, which detects interictal spikes from SEEG recording from their time-frequency representation. This detector was previously demonstrated to allow for almost 100% specificity and more than 80% sensitivity. To ensure correct spikes detection, we (1) computed the analysis also on non-SOZ channels, to confirm that no spikes would be detected and (2) visually inspected the data. We saved the IED number for each Stim ON and Stim OFF epoch.

IED analysis

To quantify the impact of acute thalamic stimulation on the IED rate, we computed two complementary analyses, based on the data quality: (a) if the IED were clearly detectable and frequent (occurring in at least 50% of the baseline epochs), we calculated the IED rate as the number of IED per minute, in each Stim ON and Stim OFF epoch; (b) if more than half of the baseline epochs presented no IED, we calculated the probability of finding no spikes across all the Stim ON and Stim OFF epochs (Supplementary Data 2). Even if consistent spiking was observed in baseline epochs, we included in analysis b also a subject (S25) that presented complete suppression of IEDs in stimulation epochs, hence precluding statistical analysis. For both analyses, we computed the percentage difference of these quantities for Stim ON vs Stim OFF to characterize the stimulation effect. A beneficial effect of stimulation on epileptiform discharges would result in a smaller IED rate with stimulation ON for analysis a, and a positive change in analysis b, prompting to show that stimulation epochs would more likely show complete suppression of IED.

Clinical outcome of chronic thalamic stimulation

We retrospectively evaluated clinical outcome of hodologically matched (n = 7) and unmatched (n = 3) stimulation in 10 patients (S32–S41) who were not candidates for resective surgery (Supplementary Data 1) and hence received a chronic neuromodulation system as part of their clinical care. Six of these patients (S32–S37) previously underwent SEEG monitoring and are also included in the first prospective part of this study.

Demographics and epilepsy information for these patients is reported in Supplementary Data 1. All patients received a bilateral implant of the PUL or ANT, according to their SOZ location and as indicated by their clinical team following SEEG monitoring. Nine patients (S32–S37, S39–S41) received a Responsive Neurostimulator (RNS), while one patient (S38) received a DBS system (Medtronic Percept). Stimulation parameters were individually tailored for each patient according to their clinical need (Supplementary Table 1). No clinical decisions were based upon this research.

We retrospectively evaluated the percentage reduction in disabling seizures’ frequency and the change in antiseizure medication. The follow-up window ranged from 3 to 36 months, during which patients were maintained on stable medication regimens, and no substantial changes in dose or drug class were made. Seizure frequency was assessed using RNS device recordings for the nine patients treated with RNS, providing objective, device-captured data, and integrated with clinical notes. For the single patient who received DBS, seizure frequency was derived from clinical notes and patient-reported history, reflecting standard documentation practices in real-world settings.

Data analysis and statistical procedures

All figures and analysis were performed using Matlab 2023b (Mathworks, California, US). For all box plots reported in this manuscript, the whiskers extend to the maximum spread not considering outliers, central, top, and bottom lines represent median, 25th, and 75th percentile, respectively.

For all the analysis, we used the bootstrap method, which does not rely on distributional assumptions of the data, but rather resamples the quantities of interest to achieve empirical confidence intervals. For each comparison, we created a bootstrap sample (for n = 10000 repetitions) by drawing a sample with replacement from the actual data points (n = 10000 repetitions), and calculated the difference in means of the resampled data. We then applied two-tailed bootstrapping with significance levels of 0.05 (95% confidence interval). If multiple comparisons were performed at once, we used a Bonferroni correction by dividing the alpha value by the number of pairwise comparisons being performed. To extract p-values, we approximated the sampling distribution of the mean difference from the bootstrap replicates and computed the probability of observing a difference at least as extreme as zero under this Gaussian distribution (two-tailed test). For visualization purposes, we indicated *** when p < 10^−4.

Ethics

All experimental protocols were approved by the University of Pittsburgh Institutional Review Boards (IRB) (protocol STUDY20070113 for SEEG cohort and STUDY21020058 for chronic cohort). Each participant gave informed written consent.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.