Skip to main content
Brain logoLink to Brain
. 2026 Apr 3;149(11):3806–3820. doi: 10.1093/brain/awag120

Intracranial biomarkers for anterior thalamic deep brain stimulation in epilepsy: a long-term observational study

Giovanna Aiello 1,2, Adham Elshahabi 3, Lennart Stieglitz 4, Anina Belvedere 5, Tena Dubcek 6,7, Rafael Polania 8,9,#, Lukas Imbach 10,11,✉,#
PMCID: PMC13730554  PMID: 41934257

Abstract

Deep brain stimulation (DBS) of the anterior nucleus of the thalamus (ANT) is an established treatment for patients with drug-refractory epilepsy, yet long-term therapeutic outcomes are highly variable and challenging to predict. This variability is compounded by the delayed and gradual effects of DBS, the difficulty of consistent seizure monitoring, and the absence of physiological biomarkers to inform treatment. In this study, we analysed longitudinal intracranial recordings over an observational period of 4 years from a cohort of 22 patients with ANT DBS. Our primary goal is the identification of neurophysiological signatures that could predict and track clinical DBS response.

Our results show that DBS responders and non-responders exhibit distinct ANT spectral trajectories over time, with responders showing a progressive increase in higher frequencies (β1 and γ) and decreased lower-frequency (δ and θ) activity in comparison to non-responders. Notably, these dynamic biomarkers, particularly high- to low-frequency ratios (e.g. β1/θ), enabled the early discrimination of clinical outcomes. Additionally, we provide evidence of robust circadian and multidien rhythms in ANT local field potentials, with further analyses supporting the feasibility of adaptive stimulation protocols to improve therapeutic outcomes.

Together, these findings propose ANT spectral dynamics in outpatient settings as a promising tool for early prediction of therapeutic efficacy and pave the way for biomarker-guided optimization of DBS therapy in epilepsy.

Keywords: epilepsy, deep brain stimulation, biomarker, thalamus, neuromodulation


Aiello et al. show that thalamic activity recorded from epilepsy patients receiving deep brain stimulation can predict treatment success early on. These biomarkers could guide personalized therapy, offering new hope for improving outcomes in drug-resistant epilepsy.

Introduction

Drug-refractory epilepsy affects approximately one-third of epilepsy patients.1,2 For those not eligible for resective epilepsy surgery, owing to multifocal or generalized seizure onset, deep brain stimulation (DBS) is an alternative treatment option,3-5 whose safety and efficacy were validated in a double-blind, randomized, multi-centre controlled clinical trial (SANTE trial).6

However, we still observe a large variability in responsiveness to anterior thalamic (ANT) DBS.7 For instance, median seizure reduction at 5 years post-implantation is reported at 65% in the SANTE trial8 and 56% in the recently published MORE registry,9 and seizure freedom is achieved in only 16%10 of patients.11,12 Given that the accepted concept of ‘DBS responsiveness’ applies only to patients who experience a reduction in seizure frequency of ≥50%, optimal treatment strategies should aim to maximize seizure reduction rates in order to improve quality of life significantly.

Growing evidence suggests that DBS chronically modulates epileptic networks rather than acting as an acute seizure suppressor.13-16 This view is supported by the gradual seizure reduction observed over years in treated patients.6,8,11,17,18 In the SANTE trial, the median seizure reduction increased up to 75% at 7 years follow-up. Such delayed benefits suggest that thalamic neuromodulation engages neuroplastic mechanisms, including neuronal reorganization, long-term potentiation or altered gene expression.19,20 These observations have motivated efforts to identify neurophysiological signatures predictive of long-term treatment response by capturing favourable network adaptations. However, the slow evolution of these changes complicates early assessment of therapeutic efficacy and decisions regarding treatment adjustment or re-evaluation.

In addition to neuroplastic changes driven by chronic neuroplasticity, implanted patients undergo repeated device-tuning sessions to identify stimulation parameters that maximize seizure reduction. However, parameter optmization typically relies on trial-and-error approaches owing to the absence of standardized guidelines. This process is challenged further by unreliable seizure reporting, because approximately one-third of patients fail to report events consistently, and nearly 50% of seizures go undocumented.21-23 Moreover, reliable evaluation of seizure reduction requires weeks to months of monitoring, potentially delaying meaningful improvements,24 given the gradual emergence of therapeutic effects.

For these reasons, there is an urgent need for a framework to define biomarkers that are correlated with the variability in patient responsiveness to optimize therapeutic outcomes on an individual basis. This challenge involves multiple dimensions, including: (i) determining whether the neurophysiological profile of a patient influences their likelihood of responding to therapy25 (predictive biomarkers); and (ii) identifying quantitative indicators that track treatment efficacy online.

Recent technological and conceptual advancements in DBS have opened new avenues towards personalization of therapy. In particular, the emergence of stimulation devices able to both record and stimulate (i.e. bidirectional) offers the possibility to monitor neural activity chronically and adapt stimulation based on real-time or long-term brain dynamics. These tools have fuelled interest in defining neurophysiological biomarkers that can optimize therapy in a data-driven, patient-specific manner.26-31 Furthermore, the ability to record from the implanted device enables continuous data collection outside the clinical setting.

Although such biomarkers might aid patient selection and capture the slow, long-term network modulation, DBS personalization must also account for faster, cyclical fluctuations in brain excitability that shape seizure risk. Seizure likelihood varies with intrinsic biological rhythms, including circadian (24 h),32-34 multidien, seasonal and menstrual cycles, which influence seizure timing, severity35-41 and clustering.32,36,42-44 Beyond these predictable rhythms, growing evidence points to pro-ictal states (transient brain states that increase seizure propensity45,46) and their potential impact on stimulation efficacy.

Importantly, aligning neuromodulation protocols with temporal dynamics could enhance therapeutic outcomes through state-dependent stimulation strategies.15,47-49 However, it remains to be elucidated whether ANT activity can reliably detect such states and provide meaningful information for adaptive strategies.

Here, we investigate how ANT activity-based biomarkers can serve three distinct purposes. First, anterior thalamic local field potentials (ANT-LFP) evolve alongside treatment progression, offering a reliable quantitative framework of chronic neuromodulation. Second, their long-term correlation with therapeutic efficacy might enable early prediction of clinical DBS outcomes. Third, the presence of specific neurophysiological oscillations can be used to inform state-dependent neuromodulation therapies.

Materials and methods

Patient population and surgical implantation

A cohort of 22 patients (10 females) diagnosed with drug-resistant focal epilepsy, with an average age of onset of 16.27 ± 14.47 years, underwent bilateral implantation of DBS leads in the anterior nucleus of the thalamus (ANT; Table 1). The average age at implantation was 37.59±12.12 years. All surgeries were conducted between 2018 and 2024 at the Department of Neurosurgery, University Hospital Zurich by stereotactic implantation with the neuromate® robotic system (Renishaw Mayfield SA). Detailed information about the presurgical evaluation and surgical implantation procedures is available in our previously published work.50 Lead localization was confirmed postoperatively by analysing target coordinates in the AC-PC space for each patient (superimposing the individual MRI and SureTune™ for each subject; Supplementary Table 1). Informed consent was obtained from all patients or their legal representatives in compliance with the Declaration of Helsinki, and the study was approved by the local ethics committee (Kantonale Ethikkommission, KEK Zurich). Treatment response was evaluated based on self-reported seizure frequency at last follow-up.

Table 1.

Patient demographics

ID/sex Age/age of onset Age at t=I ASMt  =  PreI/ASMt  =  end Lobes in SOZ Engel/ILAE (t = end) Response (t = end) Type of recordings (LS/CS) First recording PI (days) Last recording PI (days)
S1/male 31–35/1–5 21–25 2/2 T1, T2, F2 4/5 0 LS 70 533
S2/male 46–50/21–25 36–40 3/3 T1 T2 3A/4 2 LS/CS 97 944
S3/male 31–35/1–5 26–30 4/6 T1 T2 F1 F2 4B/5 1 LS/CS 270 1452
S4/female 71–75/61–65 66–70 3/2 T1 T2 P1 P2 1A/1 3 LS 260 260
S5/female 51–55/16–20 46–50 4/4 Other 3A/4 3 LS/CS 43 844
S6/female 56–60/16–20 51–55 2/2 T1, T2 2B/3 3 LS 4 901
S7/male 46–50/6–10 41–45 4/4 F1 F2 3A/4 2 LS/CS 112 1092
S8/female 26–30/1–5 26–30 5/2 F1 4B/5 1 LS 63 594
S9/female 56–60/1–5 51–55 6/6 T1 T2 F1 F2 3A/4 1 LS 1 742
S10/female 51–55/16–20 46–50 2/2 T1 T2 P2 F2 2B/3 3 LS 64 147
S11/male 46–50/11–15 41–45 5/5 F1 F2 3A/4 2 LS/CS 641 879
S12/male 36–40/16–20 31–35 4/4 T1 T2 3A/4 2 CS 235 235
S13/male 26–30/11–15 26–30 2/2 T1 T2 3A/4 3 LS/CS 214 1286
S14/male 51–55/41–45 51–55 3/2 T1 F1 F2 3B/4 3 LS/CS 110 831
S15/female 31–35/16–20 26–30 6/4 T2 P2 O2 3A/4 2 LS 1 418
S16/female 26–30/6–10 26–30 4/3 F1 F2 4B/5 0 LS 76 600
S17/male 36–40/1–5 36–40 3/2 T1 T2 3A/4 2 LS/CS 7 706
S18/male 26–30/26–30 26–30 3/2 T2, AHR 4B/5 1 LS/CS 6 484
S19/female 36–40/11–15 36–40 2/1 F1 F2 3A/4 2 LS 2 376
S20/male 26–30/6–10 26–30 3/3 T1 F1 3A/4 2 LS 8 231
S21/female 21–25/16–20 16–20 4/3 P1 P2 O1 T1 T2 3A/4 2 CS 14 14
S22/male 36–40/16–20 36–40 4/4 T1 T2 4/5 0 LS/CS 21 111

Patient demographics (including sex, age of onset, age at implantation, seizure onset lobes, Engel/ILAE at last follow-up, responsiveness and types of recordings) are reported. Ages are reported as non-overlapping age ranges. Lobes included in the seizure onset: F = frontal; O = occipital; P = parietal; T = temporal; 1 = left; 2 = right. Types of recording: CS = chronic sensing; LS = local field potential streaming. PI = post-implantation; ‘t = I’ = surgery date (implantation); ‘t = PreI’ = pre-implantation; ‘t = end’ = last clinical follow-up. Does not necessarily include a recording. Response is indicated as one of four levels: 0 = 0%–25%; 1 = 25%–50%; 2 = 50%–75%; 3 > 75%. Percentages indicate seizure frequency reduction in comparison to pre-implantation baseline. ASM = anti-seizure medication; ILAE = International League Against Epilepsy; SOZ = seizure onset zone.

Deep brain stimulation

All participants were implanted with quadripolar DBS leads (3389, Medtronic) connected to a Medtronic PERCEPT™ system. Chronic continuous (non-cycling) stimulation was administered, with a default stimulation frequency of 145 Hz. Stimulation parameters, including active contact, stimulation modality (monopolar or bipolar) and amplitude, were tailored individually based on periodic follow-ups and treatment response (Supplementary Fig. 2).

Data acquisition and selection

During regular outpatient follow-ups, patients underwent the following two types of recording to monitor treatment response.

Local field potential streaming

During outpatients visits, we recorded local field potential (LFP) data from three non-adjacent contact pairs in each side (0–3, 1–3, 0–2), without ongoing stimulation (i.e. passive streaming mode), in standardized recording settings: patients were awake, relaxed, with eyes closed and in a sitting position. The average recording time was 75.73±32.43 s. The raw signal (LFP; in microvolts) has a sampling frequency of 250 Hz.

Chronic sensing

In this modality, patients were monitored for a maximum consecutive duration of 60 days, gathering one data sample every 10 min (sampling frequency of 144 samples/day) in a predefined frequency band tailored on an individual basis based on the most prominent peak in the ANT-LFP periodogram [band-width: 5 Hz and from one pre-selected montage (bipolar), adjacent to the active (stimulation) contact51].

Patients were monitored in both modalities multiple times at different periods post-implantation (≤4 years; Fig. 1). LFP streaming data were gathered during all outpatient visits.

Figure 1.

For image description, please refer to the figure legend and surrounding text.

Study design, recording modalities and clinical responsiveness. (A) Schematic diagram of the experimental design. Each of the n = 22 patients implanted with deep brain stimulation (DBS) of the anterior nucleus of the thalamus (ANT) was recorded throughout multiple years. ‘START’ indicates the implantation date. (B and C) Both recordings in clinic (local field potential streaming) and chronic sensing were performed. (D and E) Distribution of the available recordings and sensing configurations in the two modalities. (F) For each bipolar montage (indicated with curved lines connecting the cathode and anode), the mean periodogram across subjects is represented (together with the standard error of the mean, as a shaded area). The relevant frequencies in the periodogram are indicated with bold dashes and zoomed in with the peak detections shown for θ, α, β1 and β2. (G) The number of subjects with a peak detected for each bipolar montage and each frequency is reported in the heat map. (H) Distribution of the four classes of responsiveness, together with their representation over time. LFP = local field potential; SF = seizure frequency.

Data processing and extracted features

LFP streaming

In the LFP streaming modality, three voltage differences were measured at the contact pairs 0–3, 1–3 and 0–2 (hereafter, V0|3,V1|3andV0|2). We re-referenced to neighbouring bipolar configurations by calculating the linear combinations of V0|3,V1|3andV0|2; for example: V0|1=V0−V1=V0−V3−(V1−V3)=V0|3−V1|3.51,52 Following the same procedure, we derived V0|1,V1|2 and V2|3. The re-referencing was performed to ensure a constant distance between bipolar contacts to reduce the bias in amplitude baseline differences and to remove the influence of distance-based filtering of high- or low-frequency components of the signal.53 Extracting these three combinations allowed us to compare spectral properties of different locations within the ANT.

Although we label contacts 0, 1, 2 and 3 according to the default electrode configuration, the coordinates in space of each contact differ across patients owing to substantial inter-thalamic anatomical variability and differences in surgical implantation trajectories. To address this, post-operative CT scans were co-registered with pre-operative T1-weighted MRI, and electrode positions were projected into Montreal Neurological Institute (MNI) space, allowing a visualization of all electrode locations and the LFP biomarkers across subjects (Supplementary Fig. 1). Because no clear inferior–superior gradient was observed in the LFP biomarkers, we selected contact pair 1–2 as the representative site for analysis. According to the SureTune reconstructions, this pair lay within the ANT for every patient (Supplementary Table 1).

The first measurement for each patient post-implantation (t=t0) was considered as the ‘post-implantation’ baseline. We did not analyse pre-implantation ANT measurements via, for instance, stereoelectroencephalography in the presurgical evaluation phase54,55 or intraoperatively during the implantation.

Power spectral analysis

We preprocessed the raw data to extract the power spectral density by initially applying a notch filter at 50 Hz and its harmonics, followed by a fourth order Butterworth bandpass filter in the frequencies of 1–70 Hz. We then epoched the data in 4 s windows with no overlap.50 A band-wise power spectrum was then calculated by averaging the power in six bands of interest: δ (1–4 Hz), θ (4–8 Hz), α (8–12 Hz), β1 (12–20 Hz), β2 (20–30 Hz) and γ (30–70 Hz). To ensure across-subjects comparison, we normalized the power spectra (PS) to their maximum value.

PS(f)=PS(f)/maxfPS(f) (1)

Initially, we assessed the change over time of the analysed bands by grouping the time post-implantation into the following six ‘intervals’: (i) post-implantation to 6 months; (ii) from 6 months to 1 year; (iii) from 1 to 1.5 years; (iv) from 1.5 to 2 years; (v) from 2 to 3 years; and (vi) from 3 to 4 years. We then normalized the power spectrum of each frequency band (fb) by dividing it by the power spectrum on the first observation post-implantation, to assess its evolution for each thalamus j(Tj) separately:

PSTj(fb)=PSTj(fb)/PSTj(fb)(t=t0) (2)

Chronic sensing data

Data recorded in chronic sensing was under continuous stimulation and gathered from only one non-adjacent bipolar montage [surrounding the stimulating contact(s)].51 As an example, if the stimulating contact is contact 2, the signal will be recorded in the contact pair 1–3 (V1|3). Based on the individual power spectrum of each subject, the most prominent peak in the ANT was detected, and its specific frequency (with a band width of 5 Hz) was monitored continuously, sampling every 10 min during a maximum of 60 days.51,52 Separately, we analysed thalamic recordings from both sides. Moreover, the same patient might have more than one 60 day recording over time (although not necessarily collected consecutively).

To assess the presence of ultradian, circadian and multidien rhythms, we initially preprocessed the signal by removing any outliers. Given the average (μ) of the signal, the standard deviation (σ) of the signal, and each observation x, an outlier was marked if abs(x)≥μ±6σ, and replaced by linear interpolation with its direct neighbours. The data were then z-scored to ensure comparability between subjects. The data were detrended by subtracting a first-degree polynomial fitted using least squares. A Welch periodogram was then deployed with a window size of 512 samples [with an overlap of 128 samples and a non-uniform Fast-Fourier Transform (NFFT) of 4096], with samples collected every 10 min. The resulting power spectrum was normalized, dividing by its total sum [i.e. ∑f=1endPS(f)]. We then assessed how many signals showed a peak around circadian, ultradian and multidien frequencies. Specifically, we focused on 24 h periodicity (i.e. circadian), periodicity over multiple days (i.e. multidien) and periodicity within 24 h (i.e. ultradian). The heights of these peaks were computed to assess the strength of the periodic rhythm with that frequency.

Assessment of responsiveness

We divided patients into four classes based on their percentage of self-reported seizure reduction at their last follow-up in comparison to baseline (i.e. pre-implantation): (i) super (or level 3) responders (75%–100%); (ii) level 2 responders (50%–75%); (iii) level 1 responders (25%–50%); and (iv) non-responders (<25%) responsiveness levels.

In analogy to previous studies, classes (i) and (ii) were merged as the ‘responders’ group, and classes (iii) and (iv) were classified as ‘non-responders’.

Furthermore, we assessed the total number of status epilepticus (SE) before and after surgery to correlate the seizure reduction effect of ANT DBS with clinical risk for SE, as described previously.56 We also calculated the annual risk of emergency department (ED) visits attributable to serial seizures or SE and the reduction of anti-seizure medications (ASMs) in all patients prior to and post DBS implantation.

Statistical and data analysis

Group comparisons

Initially, we analysed differences in responsiveness independent of time evolution, focusing on: (i) post-implantation measurements to identify early predictive biomarkers; and (ii) all data points over time to test for baseline shifts in power spectra. Two-level comparisons (responders versus non-responders, super-responders versus others), used Mann–Whitney U-tests. Four-level responsiveness was assessed with Spearman’s correlation and N-way ANOVA with Tukey’s HSD post hoc correction.

Time evolution

Temporal correlations between spectral power and time were modelled using generalized additive mixed models (GAMM), accounting for non-linear effects with splines and inter-thalamic variability as a random effect. Splines are smooth, flexible functions made by connecting polynomial segments at certain points called knots. They allow the modelling of non-linear relationships without predefining a default non-linear function (for instance, a quadratic or exponential function).56 The mgcv package was used in R Statistical Software (v.4.3.2, R Core Team).57

Specifically, we modelled a baseline group difference between responders and non-responders, separate time trends for each group, and thalamic-level random intercepts to account for individual variability.

PS∼Responsiveness+s(Time,by=Responsiveness)+s(thalamiID,bs=′re′) (3)

A model was fitted for each frequency band of interest. Here, PS is the spectral power, responsiveness is a fixed effect (with either two or four levels of responsiveness), ‘s’ indicates a spline (smooth curve) over time, which is fitted to the data for each of the two levels of responsiveness. The thalamus is modelled using a basis (bs) as a random effect (′re′) that models a thalamus-specific intercept.

Only data ≤2 years post-implantation were considered. Pairwise differences in spectral trajectories with non-overlapping confidence intervals indicated significant divergence over time. In the Results section, we present the earliest time at which discrimination could be identified for each band and contact.

Circadian data

Peak heights of circadian and ultradian rhythms were compared using Mann–Whitney U-tests. Analyses were performed in MATLAB (using the fieldtrip58), Python (using MNE59) and R (using mgcv57).

Results

A cohort of n = 22 epilepsy patients implanted bilaterally in the ANT (Fig. 1A) underwent longitudinal LFP recordings either during outpatient visits (‘LFP streaming’) (nrec=104, in nsubj = 20) or via chronic sensing mode (nrec=27, in nsubj = 12) (Fig. 1B and C). Patients were followed up for ≤4 years post-implantation (Fig. 1D and E). The first post-implantation recording (baseline) occurred at different points in time for each patient (μ=138.15,median=63.5, minimum = 1 and maximum = 641 days post-implantation), and the final recording occurred at 621.81±386.26 days post-implantation.

Across subjects with LFP streaming data nsubj = 20), ANT activity exhibited a characteristic aperiodic 1/f component, with intermittent peaks in δ, θ, α and β bands, variably expressed across patients and bipolar montages (Fig. 1F and G).

Patients were stratified into four responsiveness levels based on seizure-frequency reduction (−ΔSF%) at last follow-up: super-responders (i.e. level 3 responders; −ΔSF% = 75%–100%, nsubj/nall = 27%), level 2 responders (−ΔSF% = 50%–75%, nsubj/nall = 41%), level 1 responders (−ΔSF% = 25%–50%, nsubj/nall = 18%) and non-responders (−ΔSF% = 0%–25%, nsubj/nall = 14%) (Fig. 1H). Correlation of occurrences of SE, ASM change and ED visits with DBS response revealed a meaningful trend (i.e. a decrease in any of these measures was related to a better responsiveness), albeit not statistically valid (SE: r = −0.09, P = 0.69; ED: r = −0.25, P = 0.26; and ASM: r = −0.19, P = 0.40).

Demographic and clinical details are reported in Table 1.

ANT activity changes over time, concurrently with patient-specific therapeutic efficacy

To assess longitudinal changes in ANT activity, we averaged normalized LFP power within predefined intervals post-implantation (Fig. 2A). At least 10 patients contributed with data ≤3 years post-implantation, whereas fewer patients were followed up beyond this time window (Fig. 1H). Power spectra were normalized to the first post-implantation recording of each thalamus PS(t=t0), and analyses were performed separately across bipolar montages, with contact V1|2 reported in the main figure owing to its consistent localization within the ANT. Results from two other bipolar derivations are reported in the Supplementary material (unless specified).

Figure 2.

For image description, please refer to the figure legend and surrounding text.

Anterior thalamic activity changes over time and is related to patient-specific therapeutic efficacy. (A) Evolution of the normalized power spectra (PS) for each band (mean ± standard error of the mean) in contact pair 1–2. (B) Post-implantation PS for each of the two classes, only for the montage 1–2. n represents the number of subjects in each group. (C) Raw evolution of the normalized PS for each band and each responsiveness group. Data were averaged for each time interval (6 months long). NTHAL represents the number of thalami. (D) Population-level predictions from the generalized additive mixed model (GAMM) for responders and non-responders. A grey box indicates significance (i.e. non-overlapping CIs). The earliest day of non-overlapping CIs is reported. (E) For each fixed effect of the model (evolution for responders, evolution for non-responders and baseline difference) and each band, significance as a predictor is reported (P < 0.05) as a black box. A white/empty box represents no statistical significance for that fixed effect (P > 0.05). (F) GAMM predictions for the ratio β1/θ. A grey box indicates significance, and the earliest day of non-overlapping CIs is reported. (G) For each band and each contact pair, the earliest time for discrimination is reported in months post-implantation. (H) Merged observations over time for the two groups with reported Mann–Whitney U-tests and statistical significance. **P < 0.01, ****P < 0.0001. CI = confidence interval; s.e.m. = standard error of the mean; SF = seizure frequency.

Lower frequency bands (δ and θ) exhibited a decreasing trend over time, whereas higher frequencies (α, β1, β2 and γ) increased (Fig. 2A), a pattern that was consistent across the dorsoventral ANT axis (Supplementary Fig. 3). However, substantial inter-thalamic variability precluded detection of a significant time effect at the group level.n

We therefore hypothesized that this variability reflected differences in therapeutic response and divided patients using a 50% seizure-reduction threshold [responders (−ΔSF%≥50%) = 13; non-responders (−ΔSF%≤50%) = 7].6 Patients with a seizure frequency reduction of <50% will hereafter be referred to as ‘non-responders’.

No spectral differences were observed between groups at the first post-implantation measurement (Fig. 2B; P < 0.05, Mann–Whitney U-test), suggesting that early recordings could not be predict outcome. In contrast, longitudinal trajectories diverged between responders and non-responders, suggesting outcome-specific modulation of ANT dynamics by chronic DBS (Fig. 2C).

To test this formally, we applied a GAMM incorporating responsiveness, non-linear time effects and thalamic-level random intercepts. This model assessed the significance of time in predicting changes in the power spectrum when accounting for the responsiveness of the patients. We modelled the power spectral evolution as a function of responsiveness (with two possible levels), adding an offset in the interpolating spline, and the non-linear effect of time on power spectrum (by considering one spline for each responsiveness level, assuming that responders and non-responders presented potentially independent dynamics). Inter-thalamic variability was accounted for by adding it as a random effect. Owing to limited follow-up in non-responders beyond 2 years, analyses were restricted to this period. Using the GAMM, we found that PS changes significantly over time but, most importantly, does so in a responsiveness-dependent manner. Given the model:

PS∼Responsiveness+s(Time,by=Responsiveness)+s(thalamiID,bs=′re′) (4)

we hereby report the significance of the predictors for the contact pair 1–2.

We observed that responders and non-responders presented diverging behaviours over time, as represented by the population-level predictions from the GAMMs in different frequency bands (Table 2 and Fig. 2D). Statistically significant differences across the two groups were arising over time, with a minimum time lag of discrimination of the two groups at 459 days post-implantation (δ) and a maximum of 575 days post-implantation (β2).

Table 2.

Generalized additive mixed model (50% cut-off responsiveness)

V 1|2 Responsiveness (fixed effect): responsiveness_1-responsiveness_0 s(Time):responsiveness_0 s(Time):responsiveness_1 Model fit
δ Estimate: −2.181
Pr(>|t|): 0.0309*
P-value: 0.035 *
edf: 1.715
P-value: 0.212
edf:1.000
Radj2=0.492
Deviance = 58.7%
θ Estimate: −0.017
Pr(>|t|) = 0.218
P-value: 0.00875 **
edf: 1.868
P-value: 0.01613 *
edf: 1.000
Radj2=0.47
Deviance = 56.4%
α Estimate: 0.0186
Pr(>|t|) = 0.093
P-value: 0.973
edf: 1.000
P-value: 0.590
edf: 1.615
Radj2=0.523
Deviance = 61%
β1 Estimate: 0.0262
Pr(>|t|) = 0.0259*
P-value: 0.123
edf: 1.000
P-value: 0.314
edf: 5.584
Radj2=0.47
Deviance = 57.9%
β2 Estimate: 0.03506
Pr(>|t|) = 0.154
P-value: 0.0142 *
edf: 1.622
P-value: 0.1372
edf: 1.000
Radj2=0.838
Deviance = 87.4%
γ Estimate: 0.0235
Pr(>|t|) = 0.419
P-value: 7.5 × 10−5 ***
edf: 2.894
P-value: 0.0231 *
edf: 3.874
Radj2=0.658
Deviance = 73.8%

Results of the generalized additive mixed model for each frequency band and contact 1–2. For the splines, effective degrees of freedom (representing the complexity of the spline, edf) and P-values are reported. The model fit is reported using adjusted R2 and deviance. In the responsiveness fixed effect, ‘Pr(>|t|)’ indicates the P-value from the regression-based t-statistic, and ‘Estimate’ indicates the estimate of the linear fit (if positive, a higher power spectrum is correlated with a better responsiveness). If the responsiveness fixed effect is significant, the two levels of responsiveness show a significantly different power spectrum (independent of time). If the splines are reported as significant for a specific responsiveness group, that group has a significant evolution over time (i.e. anterior nucleus thalamus dynamics changes in a significant way). Significant factors are reported in bold.

* P < 0.05.

** P < 0.01.

*** P < 0.001.

Analysing the effect of time on the power spectral dynamics separately for each group, we found that non-responders showed a significant change in spectral power in δ, θ, β2 and γ. These effects were strongly non-linear, as shown by the estimated degrees of freedom (edf; Table 2). Responders, in contrast, presented significant changes in PS only for θ (linear decrease, edf = 1.000) and γ (non-linear). Responsiveness alone had a (significant) fixed effect in δ and β1, meaning that there was a baseline difference between the two groups in these two frequency bands. For all frequency bands, the random effect of the thalamus (represented by the thalamic ID), was significant (P < 2 × 10−16), meaning that each thalamus has a non-negligible different baseline level in modelling the predictors–outcome relationship. Significant fixed effects of the model (P < 0.05) are summarized graphically in Fig. 2E. Interestingly, LFP power spectrum in θ was the only frequency band showing a significant change both in responders and in non-responders.

Given that lower frequencies (δ and θ) and high frequencies (≥α) expressed opposite relationships to therapeutic efficacy, we formulated a high-to-low ratio between high and low band power to maximize the difference between responders and non-responders. The ratio in power spectrum between β1 and θ (β1/θ) showed a significant fixed effect of the responsiveness level [Estimate = 0.29, Pr(>|t|) = 0.0131] and a significant effect of time for both responders and non-responders [s(Time):Responsiveness_0, edf = 1.626, P = 0.00899; s(Time):Responsiveness_1, edf = 1.000, P = 0.04628] (Fig. 2E). Interestingly, when observing the population-level predictions of the GAMM of β1/θ over time, differences between responders and non-responders arose as early as 375 days post-implantation (Fig. 2F), almost 3 months earlier than when considering δ power only.

We then analysed the earliest time point post-implantation at which discrimination between the two is present (defined by non-overlapping confidence intervals of the population-level GAMM predictions). Among the contact pairs, the uppermost bipolar montage (3–2) showed the weakest discriminatory power, whereas contacts 0–1 and 1–2 demonstrated the ability to discriminate between groups, with a minimum discernible difference observed at 18 months post-implantation. However, when examining high-to-low frequency ratios, the minimum discrimination time was reduced by half, with contact 0–1 allowing for group differentiation as early as 9 months post-implantation (Fig. 2G).

When merging all the collected samples over-time, and comparing the two groups (Mann–Whitney U-test, P < 0.0083 for significance with Bonferroni multiple comparison correction), we found significant differences among the two groups in δ, θ and β1 (P-values: δ,0.002;θ,0.006andβ1,2.6×10−4; Fig. 2H).

ANT dynamics predicts super-response to therapy by month 1 post-implantation

Building on the binary classification (≤50% versus >50% response), we next investigated the temporal dynamics of ANT-LFP in a more granular analysis using the four-level responsiveness scale as defined above.

In post-implantation measurements, we found that super-responders (level 3) had the lowest δ and θ LFP power, and the highest power in α, β1, β2 and γ frequency bands. By deploying Spearman’s correlation, we assessed the presence of a positive relationship between the PS (calculated from V1|2) at post-implantation and the efficacy of stimulation. We noticed that δ power was significantly lower in super-responders in comparison to the three remaining levels of responsiveness (Fig. 3A), and γ power presented the opposite relationship (i.e. higher in super-responders), when comparing across the four levels of responsiveness with an N-way ANOVA (and Tukey’s post hoc HSD test). This implies that super-responders show specific neurophysiological properties already at baseline measurements.

Figure 3.

For image description, please refer to the figure legend and surrounding text.

Anterior thalamic dynamics predicts super-response to therapy. (A) Post-implantation power spectra (PS) for each of the four classes, only for the montage 1–2. n represents the number of subjects in each group. A dashed line in β1 indicates a significant Spearman’s correlation (ρ and P-value are reported). Tukey’s post hoc HSD test results are reported. (B) Raw evolution of the normalized PS for each band and each responsiveness group. Data were averaged for each time interval (6 months long). (C) The minimum time of discrimination across the four classes is reported as days post-implantation. A black empty box indicates the class with the lowest time for discrimination. (D) Generalized additive mixed model (GAMM) predictions for the ratio β1/θ for the four levels. (E) For contact pair 1–2, the earliest time for discrimination is reported in months post-implantation for super-responders versus each of the other groups. (F) Population-level predictions from the GAMM for super-responders versus suboptimal responders. A grey box indicates significance [i.e. non-overlapping confidence intervals (CIs)]. The earliest day of non-overlapping CIs is reported. (G) For each fixed effect of the model (evolution for super-responders, evolution for suboptimal-responders and baseline difference) and each band, significance as a predictor is reported (P < 0.05) as a black/grey box. A white/empty box represents no statistical significance for that fixed effect (P > 0.05). (H) GAMM predictions for the ratio β1/δ. A grey box indicates significance, and the earliest day of non-overlapping CIs is reported. (I) For each band and each contact pair, the earliest time for discrimination is reported in months post-implantation. (J) Merged observations over time for the four groups with reported ANOVA-N with Tukey’s post hoc HSD test and statistical significance. *P < 0.05, ****P < 0.0001. SF = seizure frequency.

Following the same analytical approach as above in the dichotomous analysis, we examined the evolution of the LFP power spectrum over time, grouping observations into predefined 6 month intervals. Level 2 responders and super-responders (light and dark green, respectively) showed progressively diverging patterns over time when compared with level 1 and non-responders (orange and red), as emphasized by the black box in the panel (Fig. 3B). This reinforces the idea that DBS shapes ANT-LFP dynamics in diverging ways for responders and non-responders.

To determine the earliest time point at which the group of responders could be differentiated, we examined the population-level predictions generated by the GAMM, focusing on the first time interval where confidence intervals no longer overlapped (Fig. 3C). We found that super-responders could be discriminated earlier than other groups by identifying the minimum time of discrimination across all contacts and frequency bands. Additionally, when analysing high-to-low frequency ratios (e.g. β1/θ), super-responders again displayed distinct early dynamics in comparison to the other three response levels (Fig. 3D). Focusing on LFP data from the contact pair 1–2, we found that super-responders could be differentiated from level 1 and level 2 responders as early as 6 months post-implantation, based on the β1/θ ratio. Discrimination from non-responders was achieved slightly later, at 9 months post-implantation (Fig. 3E).

To contrast optimal versus suboptimal clinical trajectories further, we grouped level 1 and 2 responders and non-responders into a single ‘suboptimal outcome’ group (n = 14) and compared them against the super-responders (n = 6) (Table 3). Strikingly, population-level GAMM predictions revealed early divergence in trajectories, with δ power showing significant separation as early as 154 days post-implantation (Fig. 3F). Here, all bands except α and β2 showed progressive divergence in time.

Table 3.

Generalized additive mixed model (super-responders versus all)

V 1|2 Responsiveness (fixed effect): responsiveness_Super-responsiveness_0 s(Time):responsiveness_0 s(Time):responsiveness_super Model fit
δ Estimate: −0.133
Pr(>|t|): 8.25 × 10−4***
P-value: 0.2599
edf: 1.000
P-value: 0.20
edf:1.000
Radj2=0.47
Deviance = 56.4%
θ Estimate: −0.022
Pr(>|t|) = 0.13
P-value: 0.1544
edf: 1.000
P-value: 0.0035 **
edf: 1.000
Radj2=0.45
Deviance = 53.9%
α Estimate: 0.011
Pr(>|t|) = 0.298
P-value: 0.56
edf: 1.000
P-value: 0.74
edf: 1.000
Radj2=0.52
Deviance = 60.0%
β1 Estimate: 0.039
Pr(>|t|) = 0.0015**
P-value: 0.42
edf: 1.35
P-value: 0.28
edf: 1.000
Radj2=0.46
Deviance = 55.5%
β2 Estimate: −0.001
Pr(>|t|) = 0.96
P-value: 0.13
edf: 2.02
P-value: 0.088
edf: 1.000
Radj2=0.83
Deviance = 86.8%
γ Estimate: 0.10
Pr(>|t|) = 4.10 × 10−4***
P-value: 0.11
edf: 1.000
P-value: 0.004 **
edf: 1.000
Radj2=0.598
Deviance = 67.8%

Results of the generalized additive mixed model for each frequency band and contact 1–2. For the splines, effective degrees of freedom (representing the complexity of the spline, edf) and P-values are reported. The model fit is reported using adjusted R2 and deviance. In the responsiveness fixed effect, ‘Pr(>|t|)’ indicates the P-value from the regression-based t-statistic, and ‘Estimate’ indicates the estimate of the linear fit (if positive, a higher power spectrum is correlated with a better responsiveness). If the responsiveness fixed effect is significant, the two levels of responsiveness show a significantly different power spectrum (independent of time). If the splines are reported as significant for a specific responsiveness group, that group has a significant evolution over time (i.e. anterior nucleus thalamus dynamics changes in a significant way). Significant factors are reported in bold.

** P < 0.01.

*** P < 0.001.

Super-responders exhibited a significant linear (as shown by edf = 1.000) decrease in θ power and a linear increase in γ power over time, reinforcing prior findings (Table 1). This suggests that a reduction in θ power over time and an increase in γ power (compared with baseline) is a prognostically good indicator for treatment response. Baseline differences between groups were evident in δ, β1 and γ power. Incorporating high-to-low frequency ratios further enhanced discriminatory power, revealing both distinct baseline values and temporal dynamics in super-responders across all proposed ratios. Among patients with a suboptimal outcome, non-responders showed a significant evolution only in the γ/δ ratio (Fig. 3G).

We report as an example the evolution of β1/δ across the three contact pairs. Here, discrimination across the two groups was possible as early as 26 days post-implantation in contact 2–3 (Fig. 3H).

Among all contact pairs, the most superior contact pair exhibited the earliest discrimination between groups (occurring <3 months post-implantation), and the inferior contacts revealed later discrimination, with separation in the β1/δ ratio emerging only after 18 months (Fig. 3I).

When aggregating data across the full follow-up period, the trends observed post-implantation became more pronounced. Super-responders demonstrated significant and robust differences (P < 0.0001) from all other groups in δ, β1 and γ power (Fig. 3J), confirming the stability and strength of these spectral biomarkers over time.

ANT activity shows relevant circadian and multidien rhythms, which might be relevant for therapeutic outcomes

We analysed long-term chronic sensing recordings from 12 subjects selected individually and unevely distributed (Fig. 4A): θ (nsamples = 33), γ ( nsamples = 10), α (nsamples = 5) and β1 (nsamples = 2).

Figure 4.

For image description, please refer to the figure legend and surrounding text.

Anterior thalamic activity shows relevant circadian and multidien rhythms. (A) Sensing frequencies distribution. (B) raw local field potential (LFP) power for two patients over 1 month (left), day–time of day distribution of power (middle) and polar histogram over time of the day (right). (C) For two patients, raw LFP is reported to show multidien rhythms as in B. The day–time of day distribution is also reported, together with the average power spectra (PS) over days (left inset). (D) Left: Polar distribution of the band power over time of the day. The thick, blue line indicates the mean value, the dashed blue lines indicate the standard error of the mean, and the grey lines indicate each recording separately. Right: Welch’s periodogram of the relevant rhythms. The thick blue line indicates the mean, and the grey lines indicate each periodogram separately. Detected peak times are reported as black dots. (E) For each band, the percentage of recordings showing ultradian, circadian and multidien rhythms is reported for each relevant frequency. The peak height at each of the three groups is also shown. (F) For θ and γ bands, the same results are reported separately for responders [i.e. class (i) and (ii)] and non-responders [i.e. class (iii) and (iv)], together with the population peak height at 24 and 12 h and their statistical significance (Mann–Whitney U-test, P < 0.05). *P < 0.05, **P < 0.01, ***P < 0.001. NR = non-responders.

ANT-LFP activity was dominated by circadian rhythms, with additional multidien patterns. Examples from two patients (Fig. 4B, left side), illustrate θ and γ evolution over 1 month, showing day- or night-specific prominence (Fig. 4B), and multidien modulation as the baseline shifts over several days (Fig. 4C). Across patients, α, β1 and γ power were concentrated during daytime, whereas θ was more evenly distributed (Fig. 4D).

Periodogram analysis revealed circadian (24 h), ultradian (<24 h) and multidien (>24 h) peaks. The θ band recordings predominantly displayed circadian rhythms, with 20% of the recordings showing 4 day multidien cycles. In the α band, peaks at 12, 24, 2 and 3 days were visible, with the highest power in circadian oscillations. In the β1 band, both observations had peak at 24 and 6 h. Finally, when analysing γ band activity, no multidien rhythm was detected but all the observations presented a circadian rhythm (Fig. 4E).

Intriguingly, in θ band data from two non-responders and five responders, non-responders exhibited a stronger circadian modulation, suggesting that individual rhythmicity might relate to DBS response.

Discussion

ANT DBS is a promising approach for patients with drug-refractory epilepsy, yet it yields a moderate median seizure reduction of 50%–65% at 5 years post-implantation, underscoring substantial interindividual variability in clinical outcomes. Two key challenges contribute to this variability: the absence of validated neurophysiological biomarkers for predicting and monitoring therapeutic efficacy, and the current inability to implement adaptive DBS strategies in epilepsy owing to limited understanding of ANT neurophysiological dynamics. Identifying robust, patient-specific biomarkers within ANT activity, while accounting for intrinsic physiological rhythms such as circadian fluctuations, is therefore essential to optimize stimulation strategies and long-term therapeutic success.

In this study, we leveraged a large monocentric longitudinal dataset of chronic and intermittent ANT-LFP recordings collected over ≤4 years in an outpatient setting. As a core finding, we identified outcome-specific spectral signatures of ANT activity that distinguish responders from non-responders both at baseline post-implantation and, more prominently, through their longitudinal evolution. Importantly, ANT spectral dynamics in super-responders followed distinct trajectories that emerged earlier than seizure-diary-based clinical assessments, suggesting that neurophysiological markers might precede observable clinical improvement. These findings support the concept that ANT DBS exerts its therapeutic effects through chronic neuromodulation of thalamocortical circuits, accompanied by progressive, outcome-dependent spectral reorganization. Additionally, we provide new insights into rhythmic fluctuations in ANT activity that might be leveraged to refine stimulation strategies.

Together, these results highlight the potential for earlier, physiologically guided optimization of DBS parameters in a therapy that is otherwise dominated by slow and empirical trial-and-error adjustments.

ANT activity holds predictive biomarkers of therapeutic efficacy

A key result of this work is that post-implantation ANT oscillatory activity evolves along distinct temporal trajectories depending on the clinical response. Responders exhibited a progressive increase in higher-frequency activity (especially β1 and γ bands), alongside a reduction in lower-frequency activity (δ and θ), whereas non-responders showed opposite maladaptive trends.

Using the clinically established 50% seizure-reduction threshold, we observed no significant spectral differences at baseline between responders and non-responders. However, their longitudinal spectral trajectories diverged robustly, suggesting that the direction and dynamics of spectral change, rather than the initial post-implantation state, carry predictive value for therapeutic outcome. This notion is especially relevant given that ANT DBS effects emerge gradually and are thought to rely on long-term network reorganization and plasticity. The progressive differentiation of spectral profiles over time is consistent with this framework, suggesting that effective neuromodulation induces specific, measurable adaptations within thalamocortical networks.13 Notably, responders showed a significant linear decrease in θ power and a non-linear modulation of γ power.

The importance of θ oscillations in describing the diverging evolution of ANT dynamics in responders and non-responders resonates with our previous findings, where high ANT θ band activity during intraoperative DBS electrode implantation was found to be predictive of a good therapeutic outcome.50 Recently, we also established that ANT θ band activity is correlated with seizure likelihood in a stereoelectroencephalography study during presurgical evaluation.55 This suggests that higher θ band activity in the ANT might reflect both a pathological biomarker of epileptic networks and be correlated with high seizure burden. Accordingly, θ band activity could serve as a neurophysiological proxy for the identification of the optimal stimulation site and the seizure-suppressing network effect.

In the present study, our ‘baseline’ measurements correspond to the first post-implantation recording, after stimulation had already commenced. Here, we observed a marked reduction in θ oscillations, with a continuous decrease over the long-term evolution in clinical responders. This supports the hypothesis that chronic ANT DBS can modulate and reduce pathological θ activity, particularly in patients with initially elevated θ power, reinforcing its potential as a targeted biomarker of therapeutic efficacy.

Group differences between cohorts were also observed in δ and β1 frequency bands, consistent with our previous findings showing that these bands are key in discriminating epileptogenic from non-epileptogenic ANT-based networks.55 The importance of β1 oscillations has been further supported by pre-operative stereoelectroencephalography studies, which demonstrated that ANT networks encompassing the seizure onset zone were orchestrated by β1 oscillations.55 Moreover, during DBS implantation (before stimulation), β1 activity was lower in responders compared with non-responders.50 In the present study, the first post-implantation measurement showed high β1 activity in responders, increasing over time. In contrast to θ oscillations, β1 activity might therefore reflect a physiological network progressively reinforced by ANT DBS. The diverging dynamics between lower (δ and θ) and higher frequencies (≥α) in both responders and non-responders validated our previous results linking the ANT power spectrum to the time to next seizure.55 This dichotomy suggested that therapeutic discrimination is enhanced by combining frequency bands. Accordingly, high-to-low frequency ratios (e.g. β1/θ), improved the predictive performance of our models.

These biomarkers have important translational potential as complements to seizure diaries, which are often incomplete. If validated against longitudinal seizure evolution, ANT-LFP biomarkers could provide earlier and more objective estimates of treatment efficacy, enabling timely adjustments to therapy.

Although a 50% seizure reduction remains the standard criterion for responsiveness,6 it does not capture outcome heterogeneity. Patients exceeding this threshold can differ substantially in long-term benefit. Moreover, achieving a ‘super-response’, i.e. a high seizure frequency reduction, is the clinical outcome that truly impacts quality of life. Identifying patients who show signs of ‘super-responsiveness’ or deviate from this trajectory might provide the means to enable therapy optimization before the standard clinical re-assessment window. These patients exhibited a distinctive neurophysiological signature already at post-implantation measurements (δ and γ). Their longitudinal ANT dynamics diverged earlier from other groups, especially in δ, β1 and γ, with high-to-low ratios (e.g. β1/δ) emerging as especially robust markers, detecting group-level effects as early as 26 days post-implantation.

Although the β1/δ ratio was presented as an illustrative example, the broader concept of a high-to-low frequency ratio provides a succinct index that captures both the enhancement of faster oscillations (potentially linked to cortico-thalamic engagement and plasticity) and the attenuation of slower rhythms (often associated with pathological synchrony). By inherently normalizing within each subject, this approach also considers inter-individual differences in absolute power values. Although chosen as a representative measure, its utility is also supported by its ability to reflect baseline differences between responders and non-responders and to track outcome-related dynamics over time, highlighting its potential role as both an early post-implantation predictor and a longitudinal marker of therapeutic progression.

Although we systematically reconstructed electrode locations in MNI space and examined biomarker distributions along the dorsoventral ANT axis, our analyses did not reveal a consistent spatial gradient, suggesting that anatomical position within the nucleus alone is insufficient to explain the variability in ANT dynamics. This absence of a clear location-dependent pattern might reflect the substantial inter-individual anatomical variability of the thalamus, differences in how each hemisphere participates in the epileptic network of the patient, and additional physiological or clinical factors captured by the significant random effects in the GAMM. Moreover, although MNI-based reconstructions provide a valuable cross-subject reference frame, the process inherently introduces approximations that can blur fine-grained spatial distinctions. Nonetheless, given the known heterogeneity of ANT microcircuitry and the possibility that more nuanced spatial–functional relationships exist, further investigation using higher-resolution anatomical models, individualized thalamic atlases and refined lead-localization approaches will be important to determine whether specific subregions of the ANT exhibit distinct neuromodulatory or biomarker profiles. Such work could ultimately help to clarify whether location-specific dynamics contribute meaningfully to clinical responsiveness and inform future strategies for patient-specific stimulation targeting.

Another key observation is the heterogeneity of thalamic signal baselines across patients in both models (50% or 75% cut-off), reflected in the significant random effect of thalamic ID. This underscores the need for patient-specific benchmarks rather than fixed thresholds for interpreting biomarkers and supports the potential utility of adaptive or closed-loop DBS, which can tailor stimulation to individual neurophysiology over time.

ANT rhythmicity: circadian and multidien oscillations

Extended recordings revealed that ANT activity follows ultradian, circadian and multidien rhythms, with circadian oscillations (24 h periodicity) predominant across all frequency bands. The α, β1 and γ activity, associated with arousal and cognitition, peaked during the daytime, whereas θ rhythms were more variable across individuals. Multidien rhythms might reflect homeostatic or adaptive processes connecting ANT signals to seizure activity, suggesting a potential role as biomarkers for pro-ictal states.

Importantly, stronger circadian rhythms in θ power were more prevalent in non-responders, whereas one responder showed pronounced ultradian (12 h) rhythms. Although limited by sample size and seizure data, these findings suggest that individual rhythmicity might influence therapeutic outcomes and could guide state-informed, personalized stimulation protocols. For instance, if θ band circadian fluctuations were correlated to seizure occurrence and seizures also were subject to a circadian rhythmicity, a loss of such 24 h rhythms could be beneficial.

Limitations and future directions

The present study is subject to several limitations. The most notable limitation is the relatively small sample size for long-term outpatient recordings (yet, the largest of its kind), particularly among non-responders, which limits the generalizability of the findings. The study was conducted with an observational design in a clinical cohort; therefore, follow-up timing and modalities were inherently inconsistent and heterogeneous across patients. This resulted in a progressive decline in the number of patients over time, with no non-responders represented >3 years post-implantation. To minimize potential biases, analyses of longitudinal trends were restricted to data within the first 2 years, when sample sizes were sufficient for meaningful statistical comparisons. Furthermore, inconsistent follow-up timing necessitated the implementation of model-based interpolation, thereby introducing complexity to the assessment of continuous trends. Although the therapeutic effects are anticipated to manifest over extended periods, the identification of short-term biomarker changes has the potential to facilitate more expeditious treatment decisions. Future studies with larger cohorts, standardized follow-up intervals and more balanced distributions across responsiveness levels will be essential to confirm the robustness and reproducibility of these results and to characterize the inter-individual variability better. Until such data are available, the present findings should be interpreted with caution as promising findings supporting biomarker-guided personalized approaches.

Another methodological consideration concerns the unit of analysis. Because hemispheric epileptic involvement can differ substantially within the same patient, each thalamus (left and right) was analysed independently. This design choice acknowledges that a thalamus ipsilateral to a seizure focus might not exhibit the same neurophysiological dynamics as its contralateral counterpart. To ensure statistical validity under this framework, thalamus (rather than patient) was modelled as the random effect, thereby accounting for intra-subject variability and avoiding underestimation of variance that could arise from pooling hemispheric data.

Patients were grouped based on their final responsiveness level, rather than correlating biomarkers with seizure frequency at each time point; an approach that could strengthen short-term predictive value. To address the known limitation of under-reported seizures, we also evaluated clinical parameters including ASM reduction, ED admission risk and SE occurrence; these measures showed trends consistent with higher responsiveness but were not included in the primary analysis owing to limited statistical power, representing a potential direction for future studies linking neurophysiological biomarkers to broader clinical outcomes. In addition, integrating standardized quality-of-life assessments would be a valuable future direction to capture patient-centred benefits that might not be reflected fully in clinical or electrophysiological metrics. An additional level of complexity comes from the fact that the variability in stimulation parameters (particularly amplitude and active contact location) increases the difficulty in interpretating across-subject results and necessitates further investigation into their influence on ANT dynamics.

Chronic sensing was limited to 60 day windows, restricting the ability to analyse longer-period cycles, such as multidien rhythms relevant for seizure clustering. Future studies should incorporate longer recordings, standardized follow-up protocols and larger cohorts.

Finally, although strong correlations were observed between ANT activity and clinical response, the question of causality remains to be resolved. The integration of seizure and rhythm data, in conjunction with acute stimulation studies, could enhance our understanding of pro-ictal states and improve stimulation strategies.

Clinical implication and conclusion

Our findings demonstrate that ANT activity carries robust, time-evolving spectral biomarkers with significant clinical implications. By moving beyond cross-sectional analyses and embracing the longitudinal complexity of neural dynamics, we have demonstrated that spectral biomarkers, especially when framed as evolving features, can help to distinguish therapeutic outcomes earlier in the course of treatment. Importantly, the translational relevance of these results spans two complementary domains: (i) the use of frequency-based biomarkers to guide stimulation programming and titration; and (ii) the exploitation of intrinsic temporal dynamics for adaptive, state-dependent DBS strategies.

In the first domain, biomarkers such as the β/θ ratio could be explored in relationship to treatment strategy decisions, particularly for the programming and refinement of stimulation parameters. Our findings suggest that future approaches could leverage these physiological biomarkers to guide parameter selection and titration in a more systematic and personalized way. Rather than relying solely on empirical adjustments, clinicians could use ANT spectral dynamics to steer stimulation settings towards patient-specific optima. Their predictive value within the first year post-implantation supports a shift towards earlier, neurophysiologically informed calibration of DBS therapy, replacing the current trial-and-error approach with more targeted and timely interventions. In the second domain, the presence of circadian and multidien rhythmicity in ANT activity introduces a dynamic dimension that extends beyond static biomarkers. These temporal fluctuations in thalamic excitability could inform adaptive, state-dependent DBS strategies,24 in which stimulation is modulated according to the state of the patient or phase of seizure susceptibility (i.e. individual dynamics of high and low seizure propensity). Accounting for the intrinsic temporal structure of ANT activity would allow stimulation to adapt dynamically to fluctuations in seizure propensity, aligning therapeutic delivery with endogenous neural timing.

Together, these two aspects underscore the potential of integrating both static and dynamic neurophysiological information into DBS management. Future work should validate these findings in larger and independent cohorts, explore the mechanistic underpinnings of these spectral changes and link these biomarkers to seizure rhythms. In conclusion, these findings offer a framework for integrating objective neural signatures into DBS programming and highlight the potential for adaptive, patient-specific stimulation strategies. By embracing the dynamic nature of thalamic neuromodulation, this work opens new possibilities for precision neurostimulation in epilepsy.

Supplementary Material

awag120_Supplementary_Data

Acknowledgements

We are deeply grateful to the patients who participated in this study and consented to collect their data for the advancement of knowledge and for a better future for epilepsy patients. The authors would like to thank Dr Kai Lutz and Dr Gaetano Leogrande (Medtronic) for sharing technical information on the system. Medtronic had no impact on the study design, data acquisition or writing of the manuscript. The funder had no role in the experimental design, analysis, manuscript preparation or submission. All authors had complete access to data. All authors authorized the submission of the manuscript.

Contributor Information

Giovanna Aiello, Department of Health Science and Technology, ETH Zurich, Zurich 8092, Switzerland; Department of Epileptology, Swiss Epilepsy Center Klinik Lengg, Zurich 8008, Switzerland.

Adham Elshahabi, Department of Epileptology, Swiss Epilepsy Center Klinik Lengg, Zurich 8008, Switzerland.

Lennart Stieglitz, Department of Neurosurgery, University Hospital Zurich, Zurich 8091, Switzerland.

Anina Belvedere, Department of Epileptology, Swiss Epilepsy Center Klinik Lengg, Zurich 8008, Switzerland.

Tena Dubcek, Department of Epileptology, Swiss Epilepsy Center Klinik Lengg, Zurich 8008, Switzerland; Institute for Neuroscience, ETH Zurich, Zurich 8057, Switzerland.

Rafael Polania, Department of Health Science and Technology, ETH Zurich, Zurich 8092, Switzerland; Department of Epileptology, Swiss Epilepsy Center Klinik Lengg, Zurich 8008, Switzerland.

Lukas Imbach, Department of Epileptology, Swiss Epilepsy Center Klinik Lengg, Zurich 8008, Switzerland; Institute for Neuroscience, ETH Zurich, Zurich 8057, Switzerland.

Data availability

According to the subjects’ consent, which did not include a provision stating that individual raw data can be made accessible, the data supporting the findings of this study are not included. The code that supports the findings of this study is available from the corresponding author, upon reasonable request.

Funding

This work was funded by a grant awarded to R.P. and L.I. from the Swiss National Science Foundation (SNSF 197766).

Competing interests

L.I. received honorary compensation from Medtronic for delivering invited talks. The other authors report no competing interests.

Supplementary material

Supplementary material is available at Brain online.

References

  • 1. Chen  Z, Brodie  MJ, Liew  D, Kwan  P. Treatment outcomes in patients with newly diagnosed epilepsy treated with established and new antiepileptic drugs: A 30-year longitudinal cohort study. JAMA Neurol. 2018;75:279–286. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2. Kwan  P, Brodie  MJ. Early identification of refractory epilepsy. N Engl J Med. 2000;342:314–319. [DOI] [PubMed] [Google Scholar]
  • 3. Toprani  S, Durand  DM. Mechanisms of neurostimulation for epilepsy. Epilepsy Curr. 2023;23:298–302. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4. Markert  MS, Fisher  RS. Neuromodulation—Science and practice in epilepsy: Vagus nerve stimulation, thalamic deep brain stimulation, and responsive neurostimulation. Expert Rev Neurother. 2019;19:17–29. [DOI] [PubMed] [Google Scholar]
  • 5. Abouelleil  M, Deshpande  N, Ali  R. Emerging trends in neuromodulation for treatment of drug-resistant epilepsy. Front Pain Res. 2022;3:839463. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6. Fisher  R, Salanova  V, Witt  T, et al.  Electrical stimulation of the anterior nucleus of thalamus for treatment of refractory epilepsy. Epilepsia. 2010;51:899–908. [DOI] [PubMed] [Google Scholar]
  • 7. Peltola  J, Colon  AJ, Pimentel  J, et al.  Deep brain stimulation of the anterior nucleus of the thalamus in drug-resistant epilepsy in the MORE multicenter patient registry. Neurology. 2023;100:e1852–e1865. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8. Touma  L, Dansereau  B, Chan  AY, et al.  Neurostimulation in people with drug-resistant epilepsy: Systematic review and meta-analysis from the ILAE surgical therapies commission. Epilepsia. 2022;63:1314–1329. [DOI] [PubMed] [Google Scholar]
  • 9. Kaufmann  E, Peltola  J, Colon  AJ, et al.  Long-term evaluation of anterior thalamic deep brain stimulation for epilepsy in the European MORE registry. Epilepsia. 2024;65:2438–2458. [DOI] [PubMed] [Google Scholar]
  • 10. Laxpati  NG, Kasoff  WS, Gross  RE. Deep brain stimulation for the treatment of epilepsy: Circuits, targets, and trials. Neurotherapeutics. 2014;11:508–526. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11. Salanova  V, Sperling  MR, Gross  RE, et al.  The SANTÉ study at 10 years of follow-up: Effectiveness, safety, and sudden unexpected death in epilepsy. Epilepsia. 2021;62:1306–1317. [DOI] [PubMed] [Google Scholar]
  • 12. Nair  DR, Laxer  KD, Weber  PB, et al.  Nine-year prospective efficacy and safety of brain-responsive neurostimulation for focal epilepsy. Neurology. 2020;95:e1244–e1256. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13. Piper  RJ, Richardson  RM, Worrell  G, et al.  Towards network-guided neuromodulation for epilepsy. Brain. 2022;145:3347–3362. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14. Fisher  RS. Deep brain stimulation of thalamus for epilepsy. Neurobiol Dis. 2023;179:. [DOI] [PubMed] [Google Scholar]
  • 15. Anderson  DN, Charlebois  CM, Smith  EH, et al.  Closed-loop stimulation in periods with less epileptiform activity drives improved epilepsy outcomes. Brain. 2024;147:521–531. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16. Trevelyan  AJ, Marks  VS, Graham  RT, Denison  T, Jackson  A, Smith  EH. On brain stimulation in epilepsy. Brain. 2025;148:746–752. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17. Bergey  GK, Morrell  MJ, Mizrahi  EM, et al.  Long-term treatment with responsive brain stimulation in adults with refractory partial seizures. Neurology. 2015;84:810–817. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18. Salanova  V, Witt  T, Worth  R, et al.  Long-term efficacy and safety of thalamic stimulation for drug-resistant partial epilepsy. Neurology. 2015;84:1017–1025. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19. Liu  DF, Chen  YC, Zhu  GY, et al.  Effects of anterior thalamic nuclei stimulation on gene expression in a rat model of temporal lobe epilepsy. Acta Neurol Belg. 2020;120:1361–1370. [DOI] [PubMed] [Google Scholar]
  • 20. Chen  YC, Zhu  GY, Wang  X, et al.  Deep brain stimulation of the anterior nucleus of the thalamus reverses the gene expression of cytokines and their receptors as well as neuronal degeneration in epileptic rats. Brain Res. 2017;1657:304–311. [DOI] [PubMed] [Google Scholar]
  • 21. Hoppe  C, Elger  CE. Epilepsie: Anfallstagebücher im klinischen Alltag und in der Forschung. Aktuelle Neurol. 2016;43:493–500. [Google Scholar]
  • 22. Blachut  B, Hoppe  C, Surges  R, Stahl  J, Elger  CE, Helmstaedter  C. Counting seizures: The primary outcome measure in epileptology from the patients’ perspective. Seizure. 2015;29:97–103. [DOI] [PubMed] [Google Scholar]
  • 23. Elger  CE, Hoppe  C. Diagnostic challenges in epilepsy: Seizure under-reporting and seizure detection. Lancet Neurol. 2018;17:279–288. [DOI] [PubMed] [Google Scholar]
  • 24. Rao  VR, Rolston  JD. Unearthing the mechanisms of responsive neurostimulation for epilepsy. Commun Med (Lond). 2023;3:166. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25. Ryvlin  P, Rheims  S, Hirsch  LJ, Sokolov  A, Jehi  L. Neuromodulation in epilepsy: State-of-the-art approved therapies. Lancet Neurol. 2021;20:1038–1047. [DOI] [PubMed] [Google Scholar]
  • 26. Johnson  KA, Dosenbach  NUF, Gordon  EM, et al.  Proceedings of the 11th annual deep brain stimulation think tank: Pushing the forefront of neuromodulation with functional network mapping, biomarkers for adaptive DBS, bioethical dilemmas, AI-guided neuromodulation, and translational advancements. Front Hum Neurosci. 2024;18:1320806. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27. Medtronic PLC . Medtronic achieves CE Mark approval for BrainSenseTM Adaptive deep brain stimulation and Electrode Identifier a groundbreaking advance in personalized sensing-enabled care for people with Parkinsons through innovative brain-computer interface technology. Medtronic News. [Press Release] Published 13 Jan 2025. Accessed 13 Jan 2025. https://news.medtronic.com/
  • 28. Cummins  DD, Kochanski  RB, Gilron  R, et al.  Chronic sensing of subthalamic local field potentials: Comparison of first and second generation implantable bidirectional systems within a single subject. Front Neurosci. 2021;15:725797. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29. Bouthour  W, Mégevand  P, Donoghue  J, Lüscher  C, Birbaumer  N, Krack  P. Biomarkers for closed-loop deep brain stimulation in Parkinson disease and beyond. Nat Rev Neurol. 2019;15:343–352. [DOI] [PubMed] [Google Scholar]
  • 30. Zrenner  C, Ziemann  U. Closed-loop brain stimulation. Biol Psychiatry. 2024;95:545–552. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31. Wang  S, Zhu  G, Shi  L, et al.  Closed-loop adaptive deep brain stimulation in Parkinson’s disease: Procedures to achieve it and future perspectives. J Parkinsons Dis. 2023;13:453–471. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32. Quigg  M. Circadian rhythms: Interactions with seizures and epilepsy. Epilepsy Res. 2000;42:43–55. [DOI] [PubMed] [Google Scholar]
  • 33. Leguia  MG, Andrzejak  RG, Rummel  C, et al.  Seizure cycles in focal epilepsy. JAMA Neurol. 2021;78:454–463. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34. Durazzo  TS, Spencer  SS, Duckrow  RB, Novotny  EJ, Spencer  DD, Zaveri  HP. Temporal distributions of seizure occurrence from various epileptogenic regions. Neurology. 2008;70:1265–1271. [DOI] [PubMed] [Google Scholar]
  • 35. Baud  MO, Proix  T, Rao  VR, Schindler  K. Chance and risk in epilepsy. Curr Opin Neurol. 2020;33:163–172. [DOI] [PubMed] [Google Scholar]
  • 36. Baud  MO, Kleen  JK, Mirro  EA, et al.  Multi-day rhythms modulate seizure risk in epilepsy. Nat Commun. 2018;9:88. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37. Niu  R, Guo  X, Wang  J, Yang  X. The hidden rhythms of epilepsy: Exploring biological clocks and epileptic seizure dynamics. Acta Epileptol. 2025;7:1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38. Daley  JT, DeWolfe  JL. Sleep, circadian rhythms, and epilepsy. Curr Treat Options Neurol. 2018;20:47. [DOI] [PubMed] [Google Scholar]
  • 39. Karoly  PJ, Rao  VR, Gregg  NM, et al.  Cycles in epilepsy. Nat Rev Neurol. 2021;17:267–284. [DOI] [PubMed] [Google Scholar]
  • 40. Liu  M, Ding  J, Wang  X. The interaction between circadian rhythm and epilepsy. Acta Epileptol. 2022;4:28. [Google Scholar]
  • 41. Karoly  PJ, Ung  H, Grayden  DB, et al.  The circadian profile of epilepsy improves seizure forecasting. Brain. 2017;140:2169–2182. [DOI] [PubMed] [Google Scholar]
  • 42. Zhong  D, Luo  S, Zheng  L, Zhang  Y, Jin  R. Epilepsy occurrence and circadian rhythm: A bibliometrics study and visualization analysis via CiteSpace. Front Neurol. 2020;11:984. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43. Karoly  PJ, Goldenholz  DM, Freestone  DR, et al.  Circadian and circaseptan rhythms in human epilepsy: A retrospective cohort study. Lancet Neurol. 2018;17:977–985. [DOI] [PubMed] [Google Scholar]
  • 44. Baud  MO, Rao  VR. Gauging seizure risk. Neurology. 2018;91:967–973. [DOI] [PubMed] [Google Scholar]
  • 45. Ilyas  A, Alamoudi  OA, Riley  KO, Pati  S. Pro-ictal state in human temporal lobe epilepsy. NEJM Evid. 2023;2:EVIDoa2200187. [DOI] [PubMed] [Google Scholar]
  • 46. Wiebe  S, Gotman  J. Understanding heightened seizure risk: The pro-ictal state. NEJM Evid. 2023;2:EVIDe2300004. [DOI] [PubMed] [Google Scholar]
  • 47. Worrell  GA, Kremen  V. Chapter 9—Next-generation brain sensing, stimulation, and adaptive control devices for epilepsy. In: Rao  VR, ed. Neurostimulation for epilepsy. Academic Press; 2023:215–227. [Google Scholar]
  • 48. Chiang  S, Khambhati  AN, Wang  ET, Vannucci  M, Chang  EF, Rao  VR. Evidence of state-dependence in the effectiveness of responsive neurostimulation for seizure modulation. Brain Stimul. 2021;14:366–375. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 49. Gregg  NM, Sladky  V, Nejedly  P, et al.  Thalamic deep brain stimulation modulates cycles of seizure risk in epilepsy. Sci Rep. 2021;11:24250. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 50. Aiello  G, Ledergerber  D, Dubcek  T, et al.  Functional network dynamics between the anterior thalamus and the cortex in deep brain stimulation for epilepsy. Brain. 2023;146:4717–4735. [DOI] [PubMed] [Google Scholar]
  • 51. Sanger  ZT, Henry  TR, Park  MC, Darrow  D, McGovern  RA, Netoff  TI. Neural signal data collection and analysis of Percept™ PC BrainSense recordings for thalamic stimulation in epilepsy. J Neural Eng. 2024;21:012001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 52. Thenaisie  Y, Palmisano  C, Canessa  A, et al.  Towards adaptive deep brain stimulation: Clinical and technical notes on a novel commercial device for chronic brain sensing. J Neural Eng. 2021;18:042002. [DOI] [PubMed] [Google Scholar]
  • 53. Shirhatti  V, Borthakur  A, Ray  S. Effect of reference scheme on power and phase of the local field potential. Neural Comput. 2016;28:882–913. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 54. Bernabei  JM, Litt  B, Cajigas  I. Thalamic stereo-EEG in epilepsy surgery: Where do we stand?  Brain. 2023;146:2663–2665. [DOI] [PubMed] [Google Scholar]
  • 55. Aiello  G, Soper  D, Hadar  PN, et al. Thalamic connectivity-based biomarkers for neuromodulation in patients with refractory epilepsies. medRxiv. [Preprint] doi: 10.1101/2025.02.05.25320539 [DOI]
  • 56. Firtinidou  A, Stieglitz  L, Imbach  L. Deep brain stimulation of the anterior nucleus of the thalamus reduces the risk for status epilepticus in focal drug-resistant epilepsy. Seizure Eur J Epilepsy. 2025;126:71–75. [DOI] [PubMed] [Google Scholar]
  • 57. Perperoglou  A, Sauerbrei  W, Abrahamowicz  M, Schmid  M. A review of spline function procedures in R. BMC Med Res Methodol. 2019;19:46. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 58. R Core Team . R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria; 2022. https://www.R-project.org/
  • 59. Oostenveld  R, Fries  P, Maris  E, Schoffelen  JM. FieldTrip: Open source software for advanced analysis of MEG, EEG, and invasive electrophysiological data. Comput Intell Neurosci. 2011;2011:156869. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

awag120_Supplementary_Data

Data Availability Statement

According to the subjects’ consent, which did not include a provision stating that individual raw data can be made accessible, the data supporting the findings of this study are not included. The code that supports the findings of this study is available from the corresponding author, upon reasonable request.


Articles from Brain are provided here courtesy of Oxford University Press

RESOURCES