Circuit response to neuromodulation characterized with simultaneous deep brain stimulation and precision neuroimaging in humans

Abstract

Deep brain stimulation (DBS) is an effective treatment for Parkinson’s disease (PD) but its neural mechanisms remain poorly understood. A mechanistic understanding requires precise characterization of functional responses to various stimulation conditions within the same individual. Here we use 3-T magnetic resonance imaging (MRI)-compatible DBS and precision imaging to collect extensive data from 14 patients with PD who received DBS. Across five timepoints spanning 1 year, each patient underwent 11.7 hours of functional MRI (fMRI) under seven stimulation conditions (30–172 min per session), 2.2 hours of structural MRI (26 min per session), 1.3 hours of diffusion-weighted MRI (16 min per session) and neurological assessments. Imaging data were also collected from 27 healthy participants. DBS normalizes connectivity in the somatocognitive action network and evokes differential responses in two distinct neurocircuits: the primary motor and globus pallidus circuits. Target cortical functional connectivity predicts clinical outcomes. This densely sampled dataset provides reliable, individually specific functional measures and is shared with the community to accelerate research into DBS mechanisms and improve personalized treatment strategies.

Using a 3-T MRI-compatible deep brain stimulation (DBS) system, Ren at al. densely sampled longitudinal, multimodal neuroimaging data from patients with Parkinson’s disease across multiple stimulation conditions spanning a year, mapping circuit-specific DBS-induced responses.

Main

Deep brain stimulation (DBS) is one of the most effective therapies for treatment-resistant Parkinson’s disease (PD)¹ and holds promise for treating a variety of neurological and psychiatric disorders^2–5. However, DBS—and neuromodulation therapies in general—faces two major challenges. The first is the lack of in-depth understanding of the biological mechanisms of neuromodulation, specifically how stimulation transiently and longitudinally impacts individual patient’s functional brain networks^6–8. This knowledge gap has impeded progress in optimizing DBS for PD and extending its application to other disorders, including depression and Alzheimer’s disease^2–5. The second challenge lies in developing precise, personalized treatment strategies^9,10. Functional brain organization varies substantially across individuals^11–15, therefore neuromodulation treatments must be personalized to achieve optimal effectiveness. For example, DBS requires careful adjustment of stimulation parameters for each patient after electrode implantation surgery, which currently is a trial-and-error procedure. It has been suggested that both invasive and noninvasive neuromodulation treatments would benefit from personalized protocols based on the patient’s specific functional network characteristics^16–19. However, such procedures have not yet been established in clinical practice. To address these challenges, a fundamental step is to accurately characterize functional brain organization at the individual level and measure how the brain responds to different types of stimulation at different stages of the treatment. To date, few studies have focused on individual-level analyses of brain activity during DBS^20,21, which requires extensive data from each individual to obtain reliable functional measures^13,22,23.

Currently, brain activity during neuromodulation is mainly measured using electrophysiological²⁴ and functional magnetic resonance imaging (fMRI) recordings^20,21,25. A downside of electrophysiological studies is that they cannot inform us about the functional responses to stimulation across the whole brain^26,27. Functional MRI fills this gap and recent studies using concurrent DBS–fMRI have begun to uncover how large-scale brain networks are modulated by DBS^28–31 and how this may relate to symptom improvement. However, due to safety concerns, such as heating near the lead electrodes, most DBS–fMRI studies have been carried out in animals or performed in patients using limited MRI field strength, scanning length and sample size, causing many to question the reliability and generalizability of the findings^32–35. The recent development of 3-T MRI-compatible DBS, which incorporates helical coils with nonuniform diameters, has addressed safety challenges related to radiofrequency radiation, overheating, displacement force and cellular damage^36–38. This approach enables safe 3-T DBS–fMRI studies in humans without limitations on scanning duration, thus allowing the acquisition of comprehensive neuroimaging data in individuals under different stimulation conditions.

To gain insights into how the brain responds to different types of stimulation at different stages of DBS treatment, we conducted a systematic, longitudinal and individual-focused fMRI study on 14 patients with PD who underwent subthalamic nucleus (STN) DBS using a 3-T MRI-compatible system (neurostimulator: G106R; quadripolar electrodes: L301C; implanted leads: E202C; Beijing Pins Medical Co., Ltd). All patients were scanned using a 3-T MRI system (Philips Achieva TX whole-body MRI scanner equipped with a 32-channel head coil) under established safety conditions for the implanted hardware. Importantly, we scaled up the data quality and quantity by collecting extensive neuroimaging data in each patient to achieve high reliability in individual-level functional measures. To allow for the investigation of the short-term and long-term effects of DBS, all patients were scanned five times in the span of 1 year, in a presurgical visit and four postsurgical visits. The presurgical visit involved 30 min of fMRI, 26 min of structural MRI (sMRI) using five modalities, 16 min of diffusion-weighted MRI (dMRI) and neurological assessments. The four postsurgical visits consisted of 172 min of fMRI, conducted under 7 different stimulation conditions, along with 26 min of sMRI, 16 min of dMRI and neurological assessments under 4 different DBS stimulation conditions. The extensive and longitudinal data that we acquired allowed us to reliably and comprehensively investigate the effects of STN–DBS on resting-state functional connectivity (RSFC) and immediate DBS-induced responses in patients with PD. Using these individual-specific functional measures, we aimed to identify the functional circuits modulated by STN–DBS and characterize their responses to different stimulation conditions both transiently and longitudinally. Moreover, we explored whether individualized target-cortical RSFC can predict and track motor symptoms after STN–DBS. To foster further research on DBS mechanisms, we have provided to the scientific community an extensive dataset of 14 patients with PD who underwent STN–DBS. To gain insights into how the brain responds to different types of stimulation at different stages of DBS treatment, we conducted a systematic, longitudinal and individual-focused fMRI study on these patients.

Results

Neuroimaging data were acquired in 14 patients during 5 separate visits: 1 month before the DBS implantation surgery as well as 1 month, 3 months, 6 months and 12 months after surgery. For each patient, we collected 2.2 hours of sMRI data using five different modalities, 1.3 hours of dMRI data and 11.7 hours of fMRI data under 7 different stimulation conditions. These conditions include resting-state fMRI (rsfMRI) with no DBS stimulation (‘DBS OFF’) and with continuous stimulation with high frequency (HFS, 130 Hz), low frequency (LFS, 60 Hz) and variable frequency³⁹ (VFS, alternating between 130 Hz and 60 Hz; Methods), as well as block-design fMRI (bdfMRI) with 36-s ON and 24-s OFF cycles under 3 different DBS stimulation frequencies. A detailed overview of the dataset contents is provided in graphic form in Fig. 1. We also collected T1-weighted (T1w) and 18.4 min of rsfMRI data in 27 age-matched healthy control participants for comparison. The demographic, neuropsychological and clinical information for all participants is presented in Supplementary Tables 1 and 2.

Fig. 1. Dataset description.

Fourteen patients with PD underwent assessment 1 month before and 1 month, 3 months, 6 months and 12 months after STN–DBS surgery. In the presurgical visit, 5 modalities of sMRI (T1w imaging, T2w FLAIR, T2w imaging with both coronal and transversal views, QSM and blood vessel-enhanced magnetic resonance angiography (MRA)), dMRI (diffusion-weighted NODDI), 30 min (6 min × 5 runs) of rsfMRI and UPDRS-III assessments from 2 independent neurologists were completed. After surgery, patients underwent sMRI and dMRI scans using identical protocols as presurgery. Then 24 min (6 min × 4 runs) of fMRI were acquired under each of the 7 stimulation conditions (DBS OFF as well as continuous or block-design HFS, LFS and VFS). Postsurgical UPDRS-III assessments were conducted for each of the four stimulation conditions (DBS OFF, continuous HFS, LFS and VFS), by two neurologists, independently. T2w_COR, coronal T2-weighted imaging; T2w_TRA, transversal T2-weighted imaging.

Therapeutic effects of different stimulation protocols

The primary clinical outcomes, as measured by the Unified Parkinson’s Disease Rating Scale Part III (UPDRS-III score)⁴⁰, were evaluated for all 14 patients at multiple timepoints and under different stimulation conditions (Fig. 1). The UPDRS-III scores were assessed independently by two neurologists, except for the rigidity subscale, which was evaluated by a single neurologist. The scores from the two neurologists demonstrated a strong correlation (repeated-measures correlation coefficient (rmcorr) = 0.94, P < 0.001; Fig. 2a) and an excellent intraclass correlation coefficient (ICC) of 0.94, indicating high interrater reliability. Symptom categories, including tremor, bradykinesia, posture stability, gait and postural instability or gait difficulty (PIGD), assessed through aggregates of UPDRS-III subitems, also exhibited excellent interrater reliability (Methods and Supplementary Fig. 1).

Fig. 2. Reliable motor symptom evaluations and clinical outcomes from STN–DBS.

a, Scatter plots and linear fitting lines illustrating the interrater reliability of the UPDRS-III, which was independently scored by two neurologists. Patients were colour coded and individual dots represent the scores from different stimulation frequencies and visits. The line indicates least-square fit regression. Both ICC (0.94) and rmcorr (0.94) indicate high interrater reliability. b, Bar graphs illustrating clinical outcomes, as assessed with the UPDRS-III, in all 14 patients 1 month before surgery and 1 month, 3 months, 6 months and 12 months after the STN–DBS surgery. Scores for each stimulation condition are shown as mean ± s.e.m. DBS significantly improved motor symptoms compared with the DBS OFF status across all postsurgical visits. HFS was significantly more effective than the other stimulation conditions. Detailed statistical results from the LME model are provided in Supplementary Tables 3 and 4. Preop, preoperative; Postop, postoperative. ***P < 0.001, **P < 0.01, *P < 0.05.

Source data

Compared to the DBS OFF state, patients showed significant improvement in motor symptoms when DBS was ON (linear mixed effects (LME) model, stimulation effect, F(1, 179.13) = 134.93, P = 2.2 × 10⁻¹⁶; Fig. 2b) and under all stimulation frequencies (HFS: F(1, 79.23) = 113.33, P = 2×10⁻¹⁶; VFS: F(1, 79.13) = 66.76, P = 4.01 × 10⁻¹²; LFS: F(1, 79.15) = 54.80, P = 1.25×10⁻¹⁰). On average, DBS ON led to an improvement of 51% in motor symptoms across follow-up visits (DBS ON versus DBS OFF: 1 month: 47%; 3 months: 49%; 6 months: 59%; 12 months: 53%). Significant improvements were also observed in all six motor symptom dimensions when DBS was ON compared to OFF (Supplementary Fig. 2 and Supplementary Table 3).

There was a significant effect of stimulation frequencies on the clinical outcome. Specifically, the UPDRS scores were significantly lower during HFS compared to VFS (Fig. 2b; LME, F(1, 79.24) = 15.81, P = 1.54 × 10⁻⁴) and LFS (LME, F(1, 79.30) = 20.73, P = 1.88 × 10⁻⁵). Post hoc tests revealed that this difference was particularly significant at 3 months, 6 months and 12 months (Fig. 2b and Supplementary Table 4). However, no significant difference was observed between VFS and LFS (LME, F(1, 79.84) = 1.25, P = 0.266).

Over time, UPDRS-III scores during the DBS OFF condition exhibited a slight increase 6 months after surgery compared to baseline, although it was statistically nonsignificant (one-way repeated-measures analysis of variance (ANOVA): F(3, 30) = 2.36, P = 0.091). The therapeutic effect did not significantly differ across follow-up visits but showed a slight trend of diminishing over time (HFS: F(1.6, 15.55) = 1.82, P = 0.198; VFS: F(1.7, 17.37) = 1.44, P = 0.263; LFS: F(3, 30) = 0.86, P = 0.475).

Individualized functional networks can be reliably characterized

Unraveling the unique functional network organization of individuals is an essential step toward personalized DBS. Using a 3-T MRI-compatible DBS system allowed us to collect high-quality functional data that met the requirements for individual-level analyses. To mitigate noise in rsfMRI data induced by DBS, we applied a denoising technique, termed background component-based noise correction (bCompCor), during data preprocessing (Methods, Extended Data Fig. 1 and Supplementary Fig. 3). We employed an iterative algorithm to parcellate the cerebral cortex into 152 individualized functional regions in each patient^12,41 (Fig. 3a and Methods). To assess the test–retest reliability of individual-specific parcellations, we divided each patient’s rsfMRI data for each stimulation frequency and each timepoint into two halves, generated a functional parcellation from each half and measured their similarity using the Dice similarity coefficient. Moreover, to examine whether the parcellations can capture individual-specific functional region topography, we calculated the average similarity in parcellation between any two patients. The results showed excellent intrapatient reliability of network parcellations (Dice coefficient = 0.88 ± 0.02), which was significantly higher than interpatient similarity (0.64 ± 0.01; two-tailed paired t(13) = 38.19, P = 9.74 × 10⁻¹⁵; Fig. 3b and Supplementary Fig. 4). Network sizes of personalized, large-scale functional networks remained stable across follow-up timepoints (one-way repeated-measures ANOVA, all P > 0.05; Supplementary Fig. 5). These findings indicate that our parcellations are reliable and capable of detecting individual differences in functional organization. Based on the parcellations, we then constructed individualized functional connectomes by estimating the RSFC matrix of the 152 functional regions (Supplementary Fig. 6). Similar to the parcellations, functional connectomes exhibited high test–retest reliability (intrapatient similarity, Pearson’s r = 0.96 ± 0.01) and carried individual-specific information, because intrapatient similarity was significantly higher than interpatient similarity (Pearson’s r = 0.77 ± 0.03; two-tailed paired t(13) = 23.00, P = 6.46 × 10⁻¹²; Fig. 3c).

Extended Data Fig. 1. The noise correction performance of bCompCor in RSFC of PD patients with DBS.

Seed-based RSFC analyses were conducted to evaluate the noise correction performance of preprocessing pipelines with and without bCompCor in patients with DBS. (a) The first seed was located in the right-hemispheric hand sensorimotor area, with its corresponding contralateral cerebellar region also displayed. (b) RSFC maps derived from data preprocessed without (left) and with (right) bCompCor are shown for three patients with DBS-ON. Black circles denote the seed locations in the cerebral cortex, while green circles highlight the corresponding regions in the contralateral cerebellar hemisphere. (c) Canonical cortico-cerebellar RSFC strength was significantly enhanced following the application of bCompCor (two-tailed Wilcoxon signed-rank test, *P = 0.022). Each dot/line represents an individual patient (n = 14). (d) Similar analyses were performed using a seed in the left-hemispheric angular gyrus within the default-mode network. As shown in the hand ROI analysis, seed-based RSFC maps without and with bCompCor are displayed. Cortico-cerebellar RSFC strength was significantly improved in the default-mode network (two-tailed Wilcoxon signed-rank test, **P = 0.005). Data are presented as mean ± s.e.m., and each dot/line represents a patient (n = 14).

Fig. 3. High intrapatient reliability and interpatient differences in functional parcellations and functional connectomes.

a, For each patient, parcellation of the cerebral cortex into 152 individual-specific brain regions using rsfMRI (Methods). These 152 brain regions were assigned to 1 of the 18 large-scale functional networks, which include 17 canonical networks¹³ and the recently discovered SCAN²². A lateral view of the native left hemisphere for each patient is shown here and the four medial and lateral views of both hemispheres are shown in Supplementary Fig. 4. The group-level atlas is pictured in the bottom right corner. b, The intrapatient reliability for each patient’s functional parcellation, represented by a red dot, calculated by splitting each patient’s data into two halves and comparing the parcellations derived from them. For every patient, the parcellation intrapatient reliability was significantly higher than the parcellation similarity between the patient and all other patients, represented by gray dots, with the average interpatient similarity indicated by a blue dot. c, A functional connectome constructed for each patient. Example connectomes from two patients are shown in Supplementary Fig. 6. The connectomes’ intrapatient reliability was also higher than their interpatient similarity. AMN, action-mode network; Ant, anterior; DMN, default-mode network; Dors attn., dorsal attention; Frontopar, frontoparietal; Lat., lateral; Med., medial; Mem, memory; MTL, medial temporal lobe; Post, posterior; SM, sensorimotor.

Source data

DBS target-cortical connectivity estimated within individuals predicts clinical outcomes

We used presurgical sMRI and postsurgical computed tomography (CT) images to localize the DBS leads and to estimate the volume of tissue activated (VTA) by the stimulation electric field in each patient (Fig. 4a and Methods). To investigate target-cortical RSFC, we employed the STN (VTA) as the target region of interest (ROI) and computed its RSFC to the cerebral cortex, using both the normative rsfMRI data from the control group as well as the individual rsfMRI data of each patient. The group-averaged, normative target-cortical RSFC map (‘Norm.’ in Fig. 4b) was similar to previously reported weighted average maps⁴². The group-averaged, individual-specific RSFC map (‘Indi.’ in Fig. 4b) showed a pattern similar to the normative map (Pearson’s r = 0.70, P_spin < 0.0001; Supplementary Fig. 7), albeit with some marked differences in the central sulcus. Specifically, the individualized RSFC map showed selectively and significantly stronger RSFC in the recently discovered somatocognitive action network (SCAN) compared to the normative RSFC map (Fig. 4b,c; two-tailed paired t(13) = 4.77, P = 0.004, Bonferroni’s correction). The SCAN has been shown to be associated with PD^22,43,44 and is hyperconnected to DBS targets, including the STN, substantia nigra and ventral intermediate nucleus of the thalamus, in patients with PD compared to healthy controls⁴³.

Fig. 4. Electrode localization, VTA estimation and RSFC of stimulation targets.

a, Reconstructed bilateral DBS leads for each patient. Individual VTA (purple), STN (orange), GP pars interna (GPi, green), GP pars externa (GPe, blue) and red nucleus (RN, red) are shown. b, Cortical RSFC maps derived from the STN (VTA) seeds. The ‘Norm.’ and ‘Indi.’ were calculated using the control group’s rsfMRI data and patients’ own presurgical rsfMRI data, respectively. Maroon boundaries highlight the marked differences between the two maps in the SCAN within the central sulcus. c, STN (VTA)-cortical RSFC averaged within the 18 large-scale functional networks, using normative data (hatched bars) and individual patients’ data (solid bars) (n = 14). Data are presented as mean ± s.e.m. The top ten networks with the largest differences in STN (VTA)-cortical connectivity between normative and individual data are shown. There was a significant difference in the SCAN (two-tailed paired t(13) = 4.77, **P = 0.004, Bonferroni’s correction). d, Intrapatient reliability of presurgical STN (VTA)-cortical RSFC maps (red bar) significantly higher than interpatient similarity (blue bar; two-tailed paired t(13) = 3.07, **P = 0.009) and also significantly higher than the similarity between individual patient mapping and normative mapping (orange bar; two-tailed paired t(13) = −4.20, **P = 0.001). The similarity between patient-specific STN (VTA)-cortical RSFC maps and the group-averaged normative RSFC map (orange bar) was significantly lower than interpatient similarity (blue bar; t(13) = −5.19, ***P = 0.0002). Data are presented as mean ± s.e.m. Each dot or line represents a patient (n = 14). e, Predicted changes in UPDRS-III scores using postsurgical, individual-specific STN (VTA)-cortical RSFC across different follow-up visits and stimulation frequencies significantly associated with the observed changes (LME: F(1, 138.58) = 15.92, ***P = 1.06 × 10⁻⁴). The colours represent different patients and each line indicates least-squares fit regression.

Source data

To examine whether STN (VTA)-cortical RSFC is reliable and individual specific, we divided fMRI data into two halves for each patient and calculated the Dice similarity coefficient between any pair of halves from the same or different patients. The RSFC maps from the same patient are reliable (Fig. 4d left, intrapatient similarity = 0.63 ± 0.06) and have significantly higher similarity than RSFC maps between different patients (Fig. 4d middle, interpatient similarity: 0.56 ± 0.05; two-tailed paired t(13) = 3.07, P = 0.009), suggesting that STN (VTA) RSFC is reliable and idiosyncratic. Moreover, the similarity between patient-specific STN (VTA)-cortical RSFC maps and the group-averaged normative RSFC map (Fig. 4d right, similarity = 0.53 ± 0.05) was significantly lower than both intrapatient similarity (two-tailed paired t(13) = −4.20, P = 0.001) and interpatient similarity (t(13) = −5.19, P = 0.0002). This finding suggests that the normative RSFC map may not be sufficiently representative of each patient’s target-cortical connectivity.

After confirming the reliability of our measures, we predicted the changes in UPDRS-III scores for each patient using a leave-one-subject-out crossvalidation (LOSOCV) analysis based on the normative or presurgical individual-specific STN (VTA) RSFC, using a method similar to that of a previous report⁴² (Methods). Using presurgical individual-specific RSFC for the prediction yielded a significant correlation between the predicted and observed changes in motor symptoms (Extended Data Fig. 2a; Spearman’s ρ = 0.573, P = 0.031), whereas using normative RSFC led to a nonsignificant correlation (Extended Data Fig. 2b; Spearman’s ρ = 0.500, P = 0.067). These findings highlight the enhanced capability of predicting clinical outcomes using individualized RSFC compared to normative RSFC. In addition, we observed that the Euclidean distance between the presurgical individualized SCAN-guided optimal target (OT) and the actual stimulation site was significantly associated with clinical outcomes (Methods; Spearman’s ρ = −0.56, P = 0.039; Extended Data Fig. 3), indicating that stimulation sites closer to the SCAN-derived OT yielded greater clinical improvement. Notably, control analyses using group-level RSFC data or literature-derived targets did not show such associations (all P > 0.05), highlighting the potential of individualized functional imaging for guiding DBS contact selection.

Extended Data Fig. 2. Predicting clinical outcomes based on ideal target-cortical connectivity map.

(a) Left: The ideal map was derived by performing clinical outcome-weighted averaging of target-cortical functional connectivity across 14 patients with PD using their presurgical rsfMRI data. Right: DBS-induced improvements in motor symptoms were successfully predicted by the similarity between the presurgical patient-specific target-cortical connectivity map and the ideal connectivity map, as assessed using leave-one-subject-out cross-validation analysis (two-tailed Spearman ρ = 0.573, *P = 0.031, see Methods). Each colored dot represents a patient (n = 14). (b) Left: The ideal map was generated using the same procedure but based on rsfMRI data from healthy controls. Right: Predicted improvements based on the normative connectivity map were modestly but not significantly correlated with observed improvements in motor symptoms (two-tailed Spearman ρ = 0.500, P = 0.067). Shaded areas indicate 95% confidence intervals of the mean estimate.

Source data

Extended Data Fig. 3. Comparisons between pre- and post-operative BOLD signals and cortical connectivity maps of DBS targets.

(a) To compare the pre- and post-operative blood oxygen level-dependent (BOLD) signal fluctuations of DBS targets, we extracted BOLD timeseries from the volume of tissue activated (VTA, pink) of the STN (dark yellow) and cortical somato-cognitive action network (SCAN; maroon) in the pre- and post-operative data of one representative patient (DBS07). The timeseries showed significantly similar between two regions in both pre-operative (Pearson’s r = 0.25, P = 0.0006) and post-operative data (r = 0.23, P = 0.002). (b) Group-averaged cortical resting-state functional connectivity (RSFC) map seeded from the STN (VTA) across pre- and post-operative data, showing only positive connectivity. Two maps are spatially similar (Dice similarity = 0.73; two-tailed Pearson’s r = 0.54, P_spin < 0.0001), particularly around the central sulcus, where SCAN patterns were observed. (c) Pre vs. post subtraction map of positive connectivity, highlighting the top five networks with the greatest RSFC changes after DBS surgery, which include the SCAN (maroon), action mode network (AMN, purple), salience (black), parietal memory (Parietal Mem, blue), and auditory (light purple) networks. Red and blue colors indicate decreased and increased functional connectivity, respectively, in the subtraction maps (pre – post). (d) Bar plot showing the magnitude of positive RSFC change (pre - post) for the top five networks. Data are presented as mean ± s.e.m., and each dot represents a patient (n = 14).

Source data

MRI signals are often compromised near the DBS leads. Thus, it is important to examine whether blood oxygen level-dependent (BOLD) signals around the leads still contain meaningful biological information. Our analyses confirmed that the STN (VTA) preserved biologically meaningful BOLD signal fluctuation⁴⁵ and its cortical RSFC after surgery (Extended Data Fig. 4 and Supplementary Fig. 8; Dice similarity = 0.73; Pearson’s r = 0.54, P_spin < 0.0001 for positive RSFC), consistent with a previous report³⁰. We further found that postsurgical STN (VTA) RSFC maps successfully predicted changes in UPDRS-III scores across different timepoints and stimulation conditions using a similar LOSOCV procedure (Fig. 4e and Methods; LME: F(1, 138.58) = 15.92, P = 1.06 × 10⁻⁴). This finding suggests that STN (VTA) RSFC maps could serve as an objective neuroimaging biomarker for tracking motor symptom changes after DBS.

Extended Data Fig. 4. Association between closeness to individualized functionally-defined SCAN targets and clinical outcomes in DBS patients.

(a) For each patient, the optimal target (OT) was defined as the centroid of the largest subthalamic nucleus (STN) cluster functionally connected to individualized SCAN cortical regions. The actual target (AT) was the center of the patient’s STN volume of tissue activated (VTA). Target distance was calculated as the mean Euclidean distance between OT and AT across hemispheres. This distance was significantly associated with motor improvement (two-tailed Spearman’s ρ = –0.56, P = 0.039), suggesting that DBS closer to functionally-defined SCAN targets yielded better outcomes. (b) Two representative patients are shown. DBS03 showed a strong clinical response (UPDRS-III change rate = 0.818) and short OT–AT distances (left: 2.83 mm; right: 4.00 mm), while DBS06 had a weaker response (0.171) and longer distances (left: 9.17 mm; right: 5.65 mm). (c) Control analyses using group-level RSFC-defined OT or (d) a literature-based STN “sweet spot”¹¹² showed no significant correlation between target distance and outcome (two-tailed Spearman’s correlation, all P > 0.05). Shaded areas indicate 95% confidence intervals of the mean estimate.

Source data

The modulatory effects of STN–DBS on RSFC in the sensorimotor and visual cortices

Previous studies have shown that the sensorimotor and visual cortices exhibit robust RSFC abnormalities in PD^25,46–48 and that STN–DBS modulates their RSFC²⁵. Thus, we assessed the effects of STN–DBS in the sensorimotor and visual cortices (Methods and Extended Data Fig. 5a). In these cortical regions, a total of 29 RSFC edges was significantly modulated by STN–DBS. The modulation effect sizes were averaged across edges of each ROI for visualization in the cerebral cortex (Fig. 5a and Extended Data Fig. 5d; Cohen’s d: sensorimotor: 0.84 ± 0.21, visual: 0.68 ± 0.06).

Extended Data Fig. 5. The estimation of modulation effects of STN-DBS on RSFC.

(a) All candidate regions of interest (ROI) in the sensorimotor (blue) and visual (purple) cortices are shown. The sensorimotor and visual cortices included 27 and 34 regions, respectively. (b) Comparison of resting-state functional connectivity (RSFC) changes before and after STN-DBS identified 29 significantly modulated RSFC edges, illustrated in the Circos plot. Among these, 9 edges (red) were normalized, and 20 edges (blue) were denormalized following STN-DBS. (c) Modulation effect sizes for all RSFC edges derived from the candidate ROIs, measured by Cohen’s d, are visualized in the functional connectome. Red and blue edges represent normalization and denormalization effects, respectively. Black outlines highlight edges with significant modulation effects. (d) Modulation effect size was averaged across RSFC edges within each ROI and visualized on the cortical surface. (e) In the sensorimotor cortex, the normalized ROIs (red) closely aligned with the SCAN. (f) ROIs of the denormalized edges (blue) in the sensorimotor cortex are primarily located within the effector-specific networks.

Source data

Fig. 5. Modulatory effects of STN–DBS on RSFC in the sensorimotor cortices.

a, ROI-based modulation effect size map. Modulation effect sizes were averaged across RSFC edges within each ROI and visualized on the cortical surface. b, ROIs in the sensorimotor cortex exhibiting modulation of specific normalized and denormalized RSFC edges. The normalized ROIs (red) largely overlapped with the SCAN, whereas the ROIs of the denormalized edges (blue) were primarily located within effector-specific networks. c, Bar plots showing a significant enhancement in the normalized edges’ RSFC after STN–DBS, rendering them more similar to those of healthy controls (control versus preoperative: two-tailed independent t(39) = 5.47, ***P < 0.001; DBS ON versus preoperative: two-tailed paired t(13) = 3.91, **P = 0.002; DBS ON versus control: two-tailed independent t(39) = 2.12, *P = 0.041). d, The modulatory effects on normalized RSFC edges exhibiting a frequency-dependent effect (one-way repeated-measures ANOVA: F(2, 39) = 2.92, P = 0.036). HFS induced significantly stronger improvements in RSFC strength compared with VFS (two-tailed paired Student’s t-test, *P = 0.022) and LFS (*P = 0.042). e, Bar plots revealing a significant decrease in RSFC strength in the denormalized edges after STN–DBS, making them more divergent from healthy controls (control versus preoperative: two-tailed independent t(39) = 1.15, P = 0.259; DBS ON versus preoperative: two-tailed paired t(13) = 4.09, ***P = 0.001; DBS ON versus control: two-tailed independent t(39) = 3.99, ***P < 0.001). f, The modulatory effects on denormalized RSFC edges not frequency dependent (LME: F(2, 39) = 0.695, P = 0.508). ***P < 0.001, **P < 0.01, *P < 0.05. NS, not significant. Data are presented as mean ± s.e.m. and each dot represents an individual participant (controls: n = 25; patients: n = 14).

Source data

We found that nine corticocortical RSFC edges became more similar to those of healthy controls after STN–DBS. These ‘normalized edges’ were localized to specific regions within the sensorimotor cortex (Fig. 5b and Extended Data Fig. 5e, red). Notably, the normalized ROIs in the sensorimotor cortex closely aligned with SCAN. Preoperatively, these edges exhibited significantly weaker RSFC in patients with PD compared to healthy controls (Fig. 5c; two-tailed independent t(39) = 5.47, P < 0.001). After STN–DBS, their RSFC strength was significantly enhanced in patients with PD (DBS ON versus preoperative: two-tailed, paired t(13) = 3.91, P = 0.002), becoming more similar to that of healthy controls, although significant differences remained (DBS-ON versus control: two-tailed independent t(39) = 2.12, P = 0.041). The normalization of SCAN-related edges was found to be frequency dependent (Fig. 5d, one-way repeated-measures ANOVA: F(2, 39) = 2.92, P = 0.036). HFS induced significantly stronger improvements in RSFC strength compared to VFS (post hoc one-tailed paired t(13) = 2.24, P = 0.022) and LFS (t(13) = 1.86, P = 0.042), with no significant differences between VFS and LFS (t(13) = 0.17, P = 0.435).

In contrast, DBS-modulated edges that diverged from those of healthy controls after STN–DBS were classified as denormalized RSFC edges. The 20 denormalized edges showed a similar pattern to normalized edges in the medial visual cortex but exhibited a distinct pattern in the sensorimotor cortex, being predominantly localized within effector-specific networks (Fig. 5b and Extended Data Fig. 5f, blue). Seed-based RSFC analyses, with seeds placed in adjacent SCAN and hand effector-specific networks, validated the distinct patterns of normalized and denormalized edges (Extended Data Fig. 6). The normalized and denormalized edges and their effect sizes are also depicted using Circos plots and the functional connectome (Extended Data Fig. 5b,c). After STN–DBS, denormalized edges exhibited a significant reduction in RSFC strength (Fig. 5e; DBS ON versus preoperative: two-tailed paired t(13) = 4.09, P = 0.001) and had lower RSFC compared to that of healthy controls (Fig. 5e; preoperative versus control: two-tailed independent t(39) = 1.15, P = 0.259; DBS ON versus control: t(39) = 3.99, P < 0.001). Unlike in normalized edges, no significant frequency-dependent effects were observed in the denormalized edges (Fig. 5f; F(2, 39) = 0.695, P = 0.508). In addition, neither normalized nor denormalized RSFC edges showed significant time-dependent effects (Supplementary Fig. 9; normalized: one-way repeated-measures ANOVA: F(3, 45) = 0.56, P = 0.643; denormalized: F(3, 45) = 0.07, P = 0.978).

Extended Data Fig. 6. Visualization of DBS modulation effects using seed-based RSFC maps and edge-based RSFC plots.

(a) Normalized edges in the sensorimotor cortex were predominantly located within the SCAN. A representative seed (red dot) in the superior region of the SCAN (purple boundary) was selected to illustrate modulatory effects. Group-averaged seed-based RSFC maps highlight robust connectivity between this seed and the medial visual cortex (black circle) in healthy controls (left). This connectivity is lower in presurgical patient data (middle) but normalized following STN-DBS (right). (b) Edge-based visualization of normalized RSFC. Circos plots depict all normalized RSFC edges across the three groups, showing systematic normalization of RSFC strengths following STN-DBS. (c) Denormalized edges in the sensorimotor cortex were primarily situated within effector-specific networks. A representative seed (red dot) in the hand effector network (blue boundary), adjacent to the SCAN seed, was chosen for visualization. Robust effector-visual connectivity is evident in healthy controls (left) and presurgical data (middle) but is weakened following STN-DBS (right). (d) Edge-based visualization of denormalized RSFC. Circos plots display all denormalized RSFC edges, showing systematic reductions in RSFC strengths following STN-DBS.

Source data

Reliable DBS-induced responses uncovered by block-design stimulation in individuals

To explore the immediate effects induced by STN–DBS, we employed an ON or OFF block-design stimulation paradigm with different stimulation frequencies (HFS, VFS and LFS) during fMRI at each follow-up visit (1 month, 3 months, 6 months and 12 months). DBS-induced response mapping was performed at the individual level across all conditions and timepoints, involving 90–288 min of bdfMRI. To control noise in DBS-induced response mapping, we applied the sparse noise modeling approach to construct a sparse noise dictionary, which was integrated with expected responses into the general linear model analysis (Methods). Individual-specific activation maps revealed substantial interindividual variability in DBS-induced responses (Fig. 6a; see activation peak coordinates in Supplementary Table 5). The group-level results showed activations and deactivations in distinct primary motor cortex (M1) and globus pallidus (GP) circuits identified using RSFC (Extended Data Fig. 7). There were strong deactivations in the M1 circuit, which also involved the posterior putamen and anterior cerebellar lobe, whereas DBS-induced activations were observed in the GP circuit, which also involved the thalamus. These findings are consistent with our previous report based on a subset of the data²¹. We extracted BOLD signal time courses from M1 and GP, because these represented the main nodes of the two circuits, and observed strong correlations or anti-correlations with the modeled canonical hemodynamic response function (Supplementary Fig. 10; M1: Pearson’s r = −0.70, P < 10⁻³⁷; GP: r = 0.78, P < 10⁻⁵³), which conforms with the activations and deactivations observed.

Fig. 6. Reliable DBS-induced responses in individuals.

a, Individual-level and group-averaged DBS-induced response maps with subcortical outlines (putamen, dark purple; GPe, light blue; GPi, green; thalamus, khaki). Block-design fMRI data from different stimulation conditions and follow-up timepoints were concatenated for this individual-level analysis. The group-averaged response map across all patients is shown in the bottom right corner. b, The DBS-induced responses showing high reliability within patients (‘test’ and ‘retest’) in M1 (left: rmcorr = 0.82, ***P < 10⁻³⁶, ICC = 0.82), represented by the blue ROI, and GP (right: rmcorr = 0.93, P < 10⁻⁶⁵, ICC = 0.93), indicated by the red ROI. Patients are colour coded and each dot represents the result of a stimulation condition during a visit. The line indicates the least-square fit regression. c, Similarity of whole-brain response maps within a given patient (red bar) that are significantly higher than those between patients (blue bar; two-tailed paired Student’s t-test, ***P = 2.68 × 10⁻⁹). Data are presented as mean ± s.e.m. and each dot represents a patient (n = 14).

Source data

Extended Data Fig. 7. Distinct M1 and GP circuits.

(a) The group-averaged DBS-induced response map shows deactivations primarily in the primary motor cortex (M1), posterior putamen, and anterior cerebellar lobe, as well as activations in the globus pallidus (GP) and thalamus. (b) To examine whether these regions are organized into distinct intrinsic circuits, seed-based resting-state functional connectivity (RSFC) maps were compared using paired t-tests between adjacent seeds from regions with activations and deactivations. Two seeds in the right and left posterior putamen (white circles, MNI coordinate = [31, -13, 7] and [-29, -15, 11]) were selected to represent the M1 circuit. Similarly, two seeds in the right and left GPi (green circles, MNI coordinate = [21, -5, 3] and [-19, -9, 3]) were selected to represent the GP circuit. The resulting contrast RSFC map shows a similar pattern with the DBS-induced response map, particularly in the M1 and anterior cerebellar lobe.

Source data

To evaluate the reliability of DBS-induced responses, we performed test–retest analyses in M1 and GP by splitting the 24-min data for each stimulation frequency at each timepoint into two halves. A comparison of the responses between the two halves revealed that they are highly reliable at the individual level (Fig. 6b; M1: rmcorr = 0.82, P < 10⁻³⁶, ICC = 0.83; GP: rmcorr = 0.93, P < 10⁻⁶⁵, ICC = 0.93). In addition, to assess the reliability and individual specificity of whole-brain DBS-induced response maps, we aggregated the data halves across all timepoints and stimulation conditions for each patient and calculated Pearson’s correlation coefficients between response maps derived from the data halves from the same patient or different patients (Supplementary Fig. 11). Response maps show high intrapatient similarity (Fig. 6c; Pearson’s r = 0.701 ± 0.121) that is significantly higher than interpatient similarity (r = 0.286 ± 0.064; two-tailed paired t(13) = 14.21, P = 2.68 × 10⁻⁹). These results demonstrate the reliable and idiosyncratic responses to STN–DBS.

Distinct frequency and time effects on M1 and GP circuits

To investigate the effects of stimulation frequency and time on immediate DBS-induced responses, we analyzed group-level activations across different DBS stimulation frequencies (Fig. 7a) and follow-up visits (Fig. 7b). Although different stimulation frequencies of STN–DBS evoked similar activation maps, DBS-induced activations within the GP circuit showed a strong frequency-dependent effect (Fig. 7c; one-way repeated-measures ANOVA: F(2, 26) = 41.54, P = 8.02 × 10⁻⁹). HFS evoked significantly larger responses compared to VFS (two-tailed paired t(13) = 5.27, P = 1.52 × 10⁻⁴) and LFS (t(13) = 7.59, P = 3.98 × 10⁻⁶). In addition, VFS evoked significantly larger responses than LFS (t(13) = 5.04, P = 2.26 × 10⁻⁴). In contrast, the M1 circuit showed no significant differences in responses to different frequencies (one-way repeated-measures ANOVA: F(2, 26) = 0.78, P = 0.468). Over time, DBS-induced activations within the GP circuit remained stable (Fig. 7d; F(3, 30) = 1.29, P = 0.296), whereas deactivations within the M1 circuit progressively increased (one-way repeated-measures ANOVA: F(3, 30) = 9.82, P = 1.13 × 10⁻⁴). These time-dependent and frequency-dependent effects remained robust when timepoints and frequencies were decoupled and analyzed separately (Supplementary Fig. 12). Thus, STN–DBS evoked distinct frequency-dependent and time-dependent effects on the M1 and GP circuits.

Fig. 7. Distinct time-dependent and frequency-dependent effects on two circuits.

a, Group-level activation maps induced by HFS (130 Hz, top), VFS (alternating between 130 Hz and 60 Hz, middle) and LFS (60 Hz, bottom). b, Group-level activation maps at 1 month, 3 months, 6 months and 12 months after DBS implantation. In both a and b, anatomical boundaries of key subcortical structures are overlaid for reference: putamen (dark purple), GPe (light blue), GPi (green) and thalamus (khaki). c, DBS-induced activations within the GP circuit (left) showing a strong frequency-dependent effect (one-way repeated-measures ANOVA: F(2, 26) = 41.54, ***P < 0.001), whereas the M1 circuit (right) showed no significant differences in responses to different stimulation frequencies (F(2, 26) = 0.78, P = 0.468). d, In contrast, DBS-induced deactivations within the M1 circuit (bottom) progressively enhanced over time (Fig. 7d; one-way repeated-measures ANOVA: F(3, 30) = 9.82, ***P < 0.001), whereas activations within the GP circuit (top) remained consistent (F(3, 30) = 1.29, P = 0.296). Data are presented as mean ± s.e.m. and each dot represents a patient (n = 14).

Source data

Importantly, further analyses revealed that SCAN in M1 exhibited significantly stronger deactivation than the effectors in response to DBS (Extended Data Fig. 8; one-way repeated-measures ANOVA: F(3, 39) = 11.36, P = 1.71 × 10⁻⁵), indicating preferential engagement of SCAN in DBS-induced network modulation. Moreover, neither SCAN nor effectors showed significant frequency-dependent effects (Extended Data Fig. 9a,b: SCAN: F(2, 26) = 0.88, P = 0.427; effectors: F(2, 26) = 0.49, P = 0.616), but both exhibited significant time-dependent effects (Extended Data Fig. 9c,d: SCAN: F(3, 30) = 6.47, P = 0.002; effectors: F(3, 30) = 6.29, P = 0.002). Meanwhile, SCAN exhibited significantly higher fractional anisotropy (FA) than effectors (Extended Data Fig. 10; 2 mm beneath cortex; one-way repeated-measures ANOVA: F(3, 33) = 77.95, P < 10⁻¹⁴) in preoperative PD data, in line with the findings in healthy participants²². In addition, postoperative FA in both SCAN and effector networks progressively declined over time (one-way repeated-measures ANOVA: SCAN: F(3, 21) = 4.42, P = 0.015; effectors: F(3, 21) = 4.65, P = 0.012).

Extended Data Fig. 8. DBS-induced cortical responses across SCAN and effector networks.

(a) The group-level DBS-induced activation map (DBS ON vs. OFF) is overlaid with boundaries of the SCAN (maroon) and Effector networks, with subnetworks corresponding to the foot (green), hand (cyan), and mouth (orange). (b) The SCAN network exhibited significantly stronger deactivation compared to Effector networks (one-way repeated measures ANOVA: F(3, 39) = 11.36, P = 1.71×10⁻⁵). Post-hoc tests revealed that this difference was significant for all Effector networks (foot: *P = 0.026; hand: **P = 0.002; mouth: ***P < 0.0001; Bonferroni correction). Additionally, deactivation in the foot network was significantly stronger than that in the mouth network (**P = 0.006; Bonferroni correction). Data are presented as mean ± s.e.m., and each dot represents a patient (n = 14).

Source data

Extended Data Fig. 9. The time and frequency-dependent effects on SCAN and Effector.

(a) Group-level activation cortical maps induced by high-frequency stimulation (HFS, 130 Hz, top), variable frequency stimulation (VFS, alternating between 130 Hz and 60 Hz, middle), and low-frequency stimulation (LFS, 60 Hz, bottom). The SCAN (maroon) and the effector network, comprising foot (green), hand (cyan), and mouth (orange) subregions, are outlined for illustration. (b) Group-level activation cortical maps at 1 month, 3 months, 6 months, and 12 months after DBS implantation. (c) DBS-induced activations within the SCAN network and effector network showed no significant differences in responses to different stimulation frequencies (SCAN: one-way repeated measures ANOVA: F(2, 26) = 0.88, P = 0.427; Effector: F(2, 26) = 0.49, ***P = 0.616). (d) In contrast, DBS-induced deactivations within the SCAN network progressively enhanced over time (one-way repeated measures ANOVA: F(3, 30) = 6.47, **P = 0.002), and the deactivation within the effector network also significantly changed (Effector: F(3, 30) = 6.29, **P = 0.002). n.s. = not significant. Data are presented as mean ± s.e.m., and each dot represents a patient (n = 14).

Source data

Extended Data Fig. 10. Fractional anisotropy of the SCAN and Effector networks before and after DBS.

(a) The SCAN network exhibited significantly higher fractional anisotropy (FA) compared to effector networks (one-way repeated measures ANOVA: F(3, 33) = 77.95, ***P < 0.001). Post-hoc tests revealed that this difference was significant between SCAN and each effector network (hand, foot, mouth: all ***P < 0.001; Bonferroni correction). Additionally, FA in the hand network was significantly higher than that in the mouth network (***P < 0.001; Bonferroni correction). (b) FA progressively declined over time in both the SCAN (left; one-way repeated measures ANOVA: F(3, 21) = 4.42, *P = 0.015) and effector network (right; F(3, 21) = 4.65, *P = 0.012). Data are presented as mean ± s.e.m., and each dot represents a patient (n = 14).

Source data

Discussion

To investigate the effects of DBS, we used a 3-T MRI-compatible DBS system (neurostimulator: G106R; quadripolar electrodes: L301C; implanted leads: E202C; Beijing Pins Medical Co., Ltd) to collect extensive imaging data from patients with PD undergoing stimulation. This dataset enabled us to comprehensively characterize the functional organization of an individual patient’s brain and the responses to different stimulations during different stages. Our data indicate that target-cortical RSFC differs between patients and healthy controls and clinical outcomes after STN–DBS can be better predicted by patient-specific RSFC. Moreover, STN–DBS selectively normalizes RSFC within the recently identified SCAN while denormalizing RSFC in effector-specific networks. Importantly, we identified two distinct corticosubcortical circuits modulated by STN–DBS, each exhibiting specific frequency-dependent and time-dependent effects of stimulation. We are making this dataset openly available as a resource for the neuroscience community.

Individual-specific functional measures can be reliably obtained during DBS

DBS–fMRI studies often face challenges in achieving reliable data quality, particularly at the individual level. These challenges primarily arise from two factors. First, the limited scanning duration per individual hinders data reliability. Precision functional mapping (PFM), which involves extensive collection and analysis of fMRI data for individual patients, has advanced significantly in recent years. PFM has demonstrated the ability to produce highly reliable functional maps at the individual level, leading to critical discoveries in the functional organization of the human brain^12,13,22,44 and functional abnormalities associated with brain disorders^23,49. However, in DBS–fMRI research, scanning duration remains constrained due to the aforementioned safety considerations. Technical advancements in 3-T MRI-compatible DBS systems enabled us to collect extensive multimodal MRI data, including rsfMRI and bdfMRI, thereby revealing highly reliable functional parcellation, RSFC and DBS-induced responses at the individual level. Second, the DBS system introduces additional electromagnetic noise in fMRI^50,51, reducing data reliability and potentially leading to misinterpretation of findings. Consistent with previous reports, we observed widespread noise artifacts across the entire fMRI field of view (FOV; Supplementary Fig. 3c). To address these issues, we developed the bCompCor method for rsfMRI denoising and the sparse noise modeling approach for bdfMRI denoising, achieving test–retest reliability exceeding 88% and 64%, respectively. In summary, by integrating PFM with advanced noise-correction algorithms, we achieved highly reliable mapping of DBS effects at the individual level, paving the way for more precise investigations into brain function and dysfunction.

Individual-specific investigation of patients undergoing DBS

Descriptions of an individual’s functional brain organization is a prerequisite for personalized treatment^23,52–54. In our analyses of STN (VTA) connectivity, cortical parcellation and DBS-induced response mapping, we observed substantial individual differences in functional organization and in the effects of DBS on brain activity. The necessity of individual-level analysis is particularly evident in patients implanted with DBS systems^55,56. Electrode location and stimulation parameters differ from patient to patient (Supplementary Table 6), leading to naturally greater variability of RSFC in patients with DBS compared to other populations. In this view, precise characterization of each patient’s brain networks is key to improving DBS treatment^57,58. For example, connectivity between stimulation sites and cortical networks is related to clinical outcomes^59,60. However, directly measuring connectivity to the DBS stimulation site in patients is challenging because of the aforementioned obstacles. As a workaround, previous studies have used normative connectivity derived from healthy populations to approximate DBS patient connectivity^29,42. However, in the current study, we found that normative STN (VTA) connectivity significantly differed from patient-specific connectivity, particularly in SCAN, a critical network in PD⁴³. Patient-specific STN (VTA) connectivity had superior predictive value for clinical outcomes and showed promise in tracking individual symptom progression, in line with a previous report⁶¹. This has important implications for clinical practice^30,42, because current DBS programming requires trial-and-error procedures by physicians, often leading to suboptimal patient management. The discovery of an effective neuroimaging biomarker, such as the one proposed here, may allow for a more systematic approach to DBS programming that is tailored to each patient¹⁸. Together, these findings emphasize the necessity for direct and accurate functional measurements in DBS patients.

SCAN-related RSFC is normalized by STN–DBS

Previous evidence has demonstrated abnormal RSFC in the sensorimotor and visual cortices in PD^25,46,47. However, until recently the sensorimotor cortex has been regarded as having a homogeneous organization, which may have led to paradoxical findings in the literature. Here leveraging PFM and fine-grained individualized functional parcellation, we revealed distinct modulatory effects of DBS on RSFC in the sensorimotor cortex—specifically, denormalization in the effector-specific networks and normalization in SCAN.

RSFC edges of the effector and visual cortices were weakened and denormalized by STN–DBS, aligned with findings from an independent study showing reduced sensorimotor–visual RSFC strength in the STN–DBS ON state compared to the OFF state²⁵. Notably, in patients with PD, sensorimotor–visual RSFC strength is negatively associated with motor symptom severity^25,48. Surprisingly, however, Zhang et al. found that RSFC strength decreased whereas motor symptoms were alleviated after STN–DBS²⁵, contradicting their observed negative association. This paradox may suggest that the relationship between effector–visual RSFC and motor symptoms is correlative rather than causal. Although RSFC changes might not directly underlie DBS-induced motor symptom improvements, the denormalization of the visually related RSFC could be linked to worsened ophthalmological symptoms—such as blurry vision and hallucinations—which occur in approximately 44.8% of patients with PD after STN–DBS, which is attributed to the reduction in medication after DBS⁶². Unfortunately, detailed visual symptoms were not assessed in the current dataset but warrant future investigation.

SCAN, rather than effector-specific networks, may represent the core macroscale network modulated by STN–DBS. The newly recognized SCAN is posited to play a critical role in movement coordination and multisensory integration^22,44, with its dysfunction closely aligning with the multifaceted symptoms of PD⁴³ and other movement disorders⁶³. Growing evidence underscores the pivotal role of SCAN in the neural mechanisms of STN–DBS for PD. In a recent study⁴³, we demonstrated that target-cortical RSFC exhibits SCAN-specific alterations in patients with PD and that SCAN links a number of neuromodulation targets used for PD treatment. Moreover, STN–DBS was found to normalize not only corticosubcortical SCAN RSFC, as shown previously⁴³, but also corticocortical SCAN RSFC as shown in the study.

SCAN–visual connectivity may contribute to multisensory integration, particularly visuomotor coordination²², with its decrease in connectivity potentially being linked to impairments in visuomotor coupling in patients with PD⁶⁴. The normalization of this connectivity could underlie the improved multisensory integration observed in patients treated with STN–DBS⁶⁵. Together, the evidence suggests that SCAN may play a central pathophysiological role in PD and may be a promising target for optimizing DBS or developing new neuromodulation therapies.

Frequency-dependent and time-dependent effects on DBS-induced responses

In this study, we identified reliable responses to STN–DBS immediately after stimulation onset. DBS-induced deactivations in the M1, posterior putamen and anterior cerebellar lobe emerged progressively over time, but were stable across stimulation frequencies. The observation of M1 deactivation is aligned with the well-documented, cortical β-band suppression and reduced STN-cortical synchronization after STN–DBS in PD^66,67, given that BOLD signal changes are related to β-band oscillations in the motor cortex^68,69. These deactivations were minimal 1 month after implantation but became pronounced at the 12-month follow-up. This gradual enhancement likely reflects neuroplastic changes in the circuit driven by long-term, continuous stimulation. Although DBS-induced responses in M1 have been reported previously, the findings have been inconsistent across studies^20,70,71. Such variability may stem from differences in postoperative scanning timelines and limited scanning durations per patient, which potentially led to unreliable measurements. This underscores the need for precision imaging over a long-term period that also includes various stimulation conditions^23,53.

DBS-induced activations were primarily observed in the GP and thalamus. The activations may result from the increased blood flow and glucose metabolism in the subcortical regions induced by STN–DBS, as shown in positron emission tomography studies^72,73. Moreover, the DBS-induced activations demonstrated stimulation frequency dependency, with activation strength increasing progressively from LFS to VFS and peaking at HFS. These frequency-dependent effects, together with SCAN-related RSFC findings, may, in part, explain why various symptoms respond differently to DBS frequencies. For instance, LFS and VFS appear more effective for addressing gait dysfunction, which often emerges in patients after prolonged HFS treatment^39,74,75. Similarly, speech impairments seem to improve with LFS rather than HFS⁷⁶. Our DBS dataset, which includes neuroimaging data related to HFS, LFS and VFS, is ideal for revealing neurological mechanisms underlying different treatment strategies and finding an effective DBS protocol for patient-specific management.

Conclusion

The primary obstacle in conducting neuroimaging studies in DBS-implanted patients is the safety concern associated with the DBS electrodes, which hinders the collection of adequate amounts of data per individual. The current resource marks a crucial advancement in addressing this challenge. The primary purpose of this work is to describe the DBS dataset that we are releasing, now available as a public resource for neuroscientists to further investigate DBS’s mechanisms of action. We demonstrated the quality of the dataset and presented several new findings stemming from it. We have shown that both rsfMRI and bdfMRI are valuable for measuring functional changes under DBS. Given that DBS is relevant to a number of brain disorders, the experimental paradigm outlined in this study could be used to explore personalized treatments for a range of neurological conditions.

Methods

Participants

Fourteen patients with the akinetic-rigid dominant form of idiopathic PD were recruited from Tiantan Hospital, Beijing, Peking Union Medical College Hospital, Beijing and Qilu Hospital, Jinan, China. The inclusion criteria were as follows: (1) age between 18 years and 75 years; (2) Mini-Mental State Examination (MMSE)⁷⁷ score higher than 24; (3) Hoehn and Yahr scale higher than stage 2 in the medication ‘OFF’ state; (4) PD duration longer than 5 years; (5) established positive response to dopaminergic medication, with at least a 30% improvement in the motor symptom section (Part III) of the UPDRS⁴⁰ with levodopa; and (6) ability to provide informed consent as determined by preoperative neuropsychological assessment. Exclusion criteria were as follows: (1) ineligibility for DBS, for example, known inability to undergo anesthesia; (2) history of hydrocephalus, brain atrophy, cerebral infarction or cerebrovascular diseases; (3) inability to follow verbal instructions; (4) history of any other severe chronic condition that may confound treatment effects or interpretation of the data; and (5) presence of MRI contraindications or inability to complete MRI. Fourteen patients (five women and nine men; age 54.9 ± 7.7 years; age range 40–67 years; Supplementary Table 1) were included in the study: eleven patients with a full set of data and three patients with incomplete data (DBS01 was unavailable after the 1-month postsurgical visit, DBS03 was unavailable after the 3-month postsurgical visit and DBS08 was unavailable for the 1-month postsurgical visit only). No adverse events were reported during the study. This project was approved by the ethics committees of Tiantan Hospital (no. QX2016-009-02, 21 July 2016), Peking Union Medical College Hospital (no. HS2016094, 21 September 2016) and Qilu Hospital (no. 2016008, 28 August 2016), with ClinicalTrails.gov identifier NCT02937727. Written informed consent was obtained from all participants.

We recruited 28 healthy control participants who were age matched to the patient group (14 women and 14 men; age 56.0 ± 6.8 years; age range 45–68 years; Supplementary Table 2). Similar to the patient group, they were excluded if they had any of the following criteria: (1) history of hydrocephalus, brain atrophy, cerebral infarction or cerebrovascular diseases; (2) inability to follow verbal instructions; (3) history of any other severe chronic condition that might confound interpretation of the data; and (4) presence of MRI contraindications. Twenty-seven participants were finally included in the study; one participant was excluded due to discomfort in the MRI scanner.

DBS surgery

Each patient underwent standard frame-based stereotaxic DBS implantation surgery at one of the aforementioned hospitals. The targets for DBS were the bilateral STNs, which were localized using presurgical structural MR scans, intraoperative electrophysiological recordings and motor symptom improvement after stimulation during the surgery. For each patient, two quadripolar DBS electrodes (model L301C, Pins Medical Co., Ltd) were implanted bilaterally into the STN. A low field potential sensing-enabled neurostimulator (G106R, Beijing Pins Medical Co., Ltd) was connected to the leads (model E202C, Pins Medical Co., Ltd) during a single operation. Each component of the MRI-compatible DBS system received approval from the National Medical Products Administration of China, formerly known as the China Food and Drug Administration, with the following registration nos.: G106R IPG: 20223120085; L301C leads: 20223120086; and E202C extension cables: 20223120084. The registration information can be found from the National Medical Products Administration website https://www.nmpa.gov.cn/. For the DBS system, stimulation can be turned on or off during MRI and either monopolar or bipolar stimulation is permitted when the stimulation is on.

DBS parameters were initially programmed 2 weeks after surgery by two experienced neurologists following standardized clinical protocols^78–80. The main stimulation settings included frequency, pulse width, amplitude and active electrode contacts, all optimized individually based on motor symptom improvement and side-effect profile^78–80. At each postsurgical visit, two neurologists were involved in the management of each patient’s DBS system. Stimulation parameters and contact selections were reviewed and adjusted as needed based on standardized motor assessments (for example, UPDRS-III), patient feedback and observed side effects. The optimized DBS programming, which led to the best motor symptom improvement, was achieved by choosing positive and negative contacts, searching for stimulation frequency, amplitude and pulse width.

MRI safety for MRI-compatible DBS system

Before MRI use in humans, MRI safety was systematically assessed for physical effects of prolonged 3-T scanning on the implanted DBS system, including displacement force, torque, vibration and heating effect^{36–38,81–84}. This assessment addressed major risks such as lead displacement, heating and unintended radiofrequency-induced stimulation, as described in our prior study²¹. Specifically, the maximal observed values—0.17 N for displacement force, 8 mN m for torque and 52.4 m s⁻² for vibration—were all well below the safety thresholds defined by the American Society for Testing and Materials^81,82, accounting for only 46%, 36% and 8% of the respective limits. Heating effect, evaluated using a phantom model and fluoroptic probes, revealed an average temperature increase of just 0.4 °C near the electrode—far below the 1 °C safety limit⁸⁴ recommended by the Health Protection Agency. In addition, a histological analysis of brain tissue in rodents after 25 min of 3-T MRI⁸³ showed no evidence of cellular damage, supporting the overall safety of the DBS system under these scanning conditions.

Data acquisition

Patients came in for data acquisition during five visits, including one presurgical visit and four postsurgical follow-up visits. The presurgical visit took place approximately 1 month before the DBS surgery and the postsurgical visits took place 1 month, 3 months, 6 months and 12 months after surgery. Data acquisition comprised MRI, neurological assessments and CT scanning. Out of the initial cohort, 11 patients completed all data acquisition, whereas 3 patients had incomplete data because they were unable to attend all postsurgical visits (DBS01 and DBS08 missed the 1-month follow-up and DBS03 the 3-month follow-up). Control participants came for one visit. Data acquisition comprised one T1-weighted MRI run and three BOLD rsfMRI runs.

MRI protocols

All MRI data were acquired with a 3-T Philips Achieva TX whole-body MRI scanner equipped with a 32-channel head coil. Seven different MRI sequences, five for structural imaging (magnetization-prepared rapid gradient echo (MPRAGE), fluid-attenuated inversion recovery (FLAIR), turbo spin-echo (TSE), quantitative susceptibility mapping (QSM) and simultaneous noncontrast angiography and intraplaque hemorrhage (SNAP)), one for functional imaging (BOLD) and one for diffusion-weighted imaging (neurite orientation and dispersion density imaging (NODDI)), were employed in the study.

MPRAGE: T1w sagittal images were acquired for 4 min 14 s using a MPRAGE sequence (FOV = 256 × 256, pixel spacing = 1 mm × 1 mm, 180 slices, slice thickness = 2 mm, spacing between slices = 1 mm, repetition time (TR) = 7.52 ms, echo time (TE) = 3.70 ms, flip angle = 8°).

FLAIR: T2w sagittal images were acquired for 6 min 30 s using a FLAIR sequence (FOV = 384 × 384, pixel spacing = 0.602 mm × 0.602 mm, 120 slices, slice thickness = 1.5 mm, spacing between slices = 1.5 mm, TR = 5,000 ms, TE = 340.29 ms, TI = 1,650 ms, flip angle = 90°).

TSE: T2w images were acquired with a 2 min 56 s TSE sequence, with FOV = 512 × 512, pixel spacing = 0.391 mm × 0.391 mm, 45 slices, slice thickness = 2 mm, spacing between slices = 2 mm, TR = 2,595 ms, TE = 90 ms, flip angle = 90°. One transversal TSE imaging and one coronal TSE imaging were performed and the center plane was set to cover the STNs completely.

QSM: magnetic susceptibility images were acquired with a 3 min 55 s transversal QSM sequence, with FOV = 512 × 512, pixel spacing = 0.449 mm × 0.449 mm, 173 slices, slice thickness = 1.5 mm, spacing between slices = 0.75 mm, TR = 30.253 ms, flip angle = 12°.

Magnetic resonance angiography: images were acquired with a 5 min 8 s transversal SNAP sequence, with FOV = 400 × 400, pixel spacing = 0.602 mm × 0.602 mm, 150 slices, slice thickness = 0.8 mm, spacing between slices = 0.4 mm, TR = 10.025 ms, TE = 5.64 ms, flip angle = 11°.

BOLD: functional images were acquired with a 6 min 14 s transversal gradient-echo echo-planar imaging sequence, with FOV = 256 × 256, pixel spacing = 2.875 mm × 2.875 mm, 37 slices, slice thickness = 3.5 mm, spacing between slices = 4 mm, TR = 2,000 ms, TE = 30 ms, flip angle = 90°, 184 frames.

NODDI: diffusion-weighted images were acquired with a 7 min 47 s transversal NODDI sequence, with FOV = 128 × 128, pixel spacing = 1.75 mm × 1.75 mm, 50 slices, slice thickness = 2.8 mm, spacing between slices = 2.8 mm, TR = 5499.745 ms, TE = 85.069 ms, flip angle = 90°, 81 images with 3 b values (b, diffusion weighting factor, s mm⁻²; 1 image with b = 0, 40 images with b = 1,000 and 40 images with b = 2,000).

CT image acquisition

CT images were acquired with a uCT 760 (United Imaging) scanner 1 month after surgery. A head helical sequence, with FOV = 512 × 512, pixel spacing = 0.449 mm × 0.449 mm, 204 slices and slice thickness = 0.625 mm, was used.

Neurological assessments

The primary outcome measure of patients’ motor function was obtained using the UPDRS-III⁴⁰. The assessment processes were video recorded. Two neurologists subsequently scored each subitem of the UPDRS-III according to the recorded videos. The rigidity subitems were scored by one neurologist on-site. All assessments were carried out when patients were in a medication OFF state for at least 12 h. UPDRS-III subitems were aggregated into several subcategories⁸⁵. Tremor scores were the sum of items 20a–e and 21a–b. Rigidity scores were the sum of items 22a–e. Bradykinesia scores were the sum of items 23a–b, 24a–b, 25a–b and 26a–b. Gait scores were directly from item 29. Posture scores were the sum of items 27, 28 and 30. Cardinal sign scores were the sum of tremor, rigidity and bradykinesia scores. PIGD scores were the sums of gait and posture scores. The MMSE⁷⁷ and 39-item Parkinson’s Disease Questionnaire were administered to all patients before surgery.

Stimulation conditions

We acquired BOLD fMRI and UPDRS-III data under several DBS frequencies: low frequency (60 Hz), high frequency (130 Hz) and variable frequency (alternating between 60 Hz and 130 Hz)^75,86. A neurologist optimized the DBS settings with regard to contacts selection, stimulation amplitude and pulse width. A total of seven different stimulation conditions was investigated during postsurgical fMRI and four of them were included in postsurgical UPDRS-III assessment. A washout period of over 1 hour was implemented between each condition. The order of conditions was pseudorandomized. The conditions are described in more detail below.

• Condition 1 (C1): acquisition without STN stimulation, that is, presurgical status or postsurgical DBS OFF status.

• Condition 2 (C2): acquisition with continuous 60-Hz stimulation to the STN.

• Condition 3 (C3): acquisition with continuous 130-Hz stimulation to the STN.

• Condition 4 (C4): acquisition with continuous VFS to the STN, that is, stimulation frequency was periodically alternated between 60 Hz and 130 Hz every 2 s.

• Condition 5 (C5): block-design acquisition using 60-Hz stimulation to the STN, that is, 36 s of 60-Hz stimulation interleaved with 24 s of DBS OFF status.

• Condition 6 (C6): block-design acquisition using 130-Hz stimulation to the STN, that is, 36 s of 130-Hz stimulation interleaved with 24 s of DBS OFF status.

• Condition 7 (C7): block-design acquisition using VFS to the STN, that is, 36 s of VFS (stimulation frequency was periodically alternated between 60 Hz and 130 Hz every 2 s) interleaved with 24 s of DBS OFF status.

Conditions 2–4 mimicked conventional therapeutic strategies with different stimulation frequencies, which are critical for revealing frequency-related modulation effects. Conditions 5–7 employed block designs to investigate the immediate modulatory effects of different stimulation frequencies. The motor symptoms were assessed using the UPDRS-III at least 1 hour after each stimulation condition switch to wash out the immediate after-effects of stimulation. As it is difficult to interpret UPDRS-III measurement during block-design stimulation (conditions 5–7), only DBS OFF state (condition 1) and continuous stimulation with different frequencies (conditions 1–4) were accompanied by UPDRS-III assessments.

Data preprocessing

DBS electrode localization and the VTA estimation

The locations of the electrode contacts were identified based on presurgical T1w MPRAGE MR scans and postsurgical CT images using a protocol similar to that previously described^87,88. Specifically, the postsurgical CT image was linearly co-registered to the presurgical T1w images using SPM12 (Wellcome Department of Cognitive Neurology, London, UK)⁸⁹. The CT and presurgical T1w images of each patient were registered to a common space, the Montreal Neurological Institute (MNI) ICBM152 template, using the advanced normalization tools⁹⁰. DBS electrode contacts were then semi-automatically identified from the normalized CT images. Finally, DBS electrodes from all 14 patients and several subcortical nuclei were reconstructed within the MNI space using Lead-DBS software⁸⁸.

The VTA estimation largely followed the procedure described previously^91–93. Specifically, a tetrahedral volume mesh was generated based on the surface mesh of the DBS contacts and subcortical nuclei using the Iso2Mesh toolbox of the Lead-DBS software. Regions were assumed to be filled with electrode materials, gray matter or white matter. Conductivities of 0.33 S m⁻¹ and 0.14 S m⁻¹ were assigned to gray and white matter, respectively. For the platinum or iridium contacts and insulated parts of the electrodes, values of 108 S m⁻¹ and 10,216 S m⁻¹ were used, respectively. Based on the volume conductor model, the potential distribution resulting from DBS was simulated using the integration of the FieldTrip-SimBio pipeline⁹⁴. The voltage applied to the active electrode contacts was introduced as a boundary condition. Subsequently, the gradient of the potential distribution was calculated by derivation of the finite element method solution. Due to the first-order finite element method approach that was used, the resulting gradient is piecewise continuous. The gradient was thresholded for magnitudes above a commonly used value of 0.2 V mm⁻¹ to define the extent and shape of the STN (VTA). The MNI coordinates of STN (VTA) centers have been recorded for each individual in Supplementary Table 5.

Resting-state fMRI preprocessing

Resting-state fMRI, that is, BOLD fMRI data acquired under conditions 1–4, was preprocessed using a previously published pipeline^12,15,16,95, which includes: (1) slice-timing correction (SPM12; Wellcome Department of Cognitive Neurology, London, UK); (2) rigid-body correction for head motion (FSL6.0; Oxford Centre for Functional MRI of the Brain, UK); (3) linear detrending and band-pass filtering (0.01–0.08 Hz); and (4) regressing nuisance variables, including the global signal, the six parameters obtained through rigid-body head motion correction and their first temporal derivatives, as well as the component-based confounds (see below for a detailed description).

T1w images were preprocessed to reconstruct individual cortical surface using FreeSurfer (http://surfer.nmr.mgh.harvard.edu/) v6.0.0. The reconstructed cortical surfaces were registered to the FreeSurfer cortical surface template (fsaverage6). The individual anatomical images were also nonlinearly aligned to the MNI152 template. The anatomical and functional images were affinely aligned using boundary-based registration from the FsFast software package (http://surfer.nmr.mgh.harvard.edu/fswiki/FsFast). The preprocessed functional images were projected to the surface and volumetric template. A 6-mm full-width half-maximum (FWHM) smoothing kernel was then applied to the fMRI data in both the surface and the volumetric space.

Background component-based noise correction

The rsfMRI data showed strong noise distributions not only within but also outside the brain (Supplementary Fig. 3c), which may have resulted from DBS artifacts and head motion. To accurately extract noise distributions and avoid removing neural signals, we employed a noise correction method, bCompCor (Supplementary Fig. 3), adapted from anatomical CompCor⁹⁶. Specifically, we applied a principal component analysis to the background of the FOV, in which BOLD signals are deterministically not modulated by neural activity. The noise ROI was generated by using an FOV mask that excluded a slightly enlarged brain mask (dilated by 1 voxel). The principal component analysis was applied to extract the top ten principal components of the BOLD signals within the noise ROI, which explained a cumulative 85.5% of the variance in the noise ROI (Supplementary Fig. 3b). These components were regressed out alongside other nuisance variables. To demonstrate the denoising effect, we performed seed-based RSFC analyses in preprocessed data with and without the regression of these components (Extended Data Fig. 1). The bCompCor has also been incorporated in our preprocessing pipeline DeepPrep⁹⁷.

Data analyses

Therapeutic effects of DBS

We used LME models to analyze the therapeutic effects of DBS, with changes in motor symptoms, as assessed by UPDRS-III total scores, as the dependent variable. The independent variables included stimulation conditions (OFF, HFS, VFS and LFS), follow-up visits (1 month, 3 months, 6 months and 12 months) and their interactions, with random intercepts for individual participants. For comparisons between the ON and OFF states, we considered HFS, VFS and LFS as the ON state and constructed the LME model using the complete dataset. In addition, for pairwise comparisons between different stimulation frequencies, such as HFS versus LFS, the LME models were constructed using data corresponding to the specific pairs of stimulation frequencies. To analyze the time effects on therapeutic effects, a one-way repeated-measures ANOVA was performed. In addition to analyzing the total score of the UPDRS-III, we also compared various motor symptom dimensions, such as tremor and PIGD, between the ON and OFF states. All analyses were conducted using R v.4.3.1. The LME models were fitted using the function ‘lmer’ from the ‘lme4’ package (v1.1.35.5) and the ANOVA was conducted using the ‘lmertest’ package (v3.1.3), which uses a Sattherwaite approximation for degrees of freedom for ANOVA. Post hoc Tukey’s comparisons were performed using the estimated marginal means from the ‘emmeans’ package (v1.10.5). One-way repeated-measures ANOVA was conducted using the function from the ‘rstatix’ package (v0.7.2), with Mauchly’s test for sphericity and a Greenhouse–Geisser correction.

Individualized functional parcellation

Each patient’s cerebral cortex was individually parcellated into 152 regions (76 regions for the left and right hemispheres, respectively) based on their rsfMRI data, using an iterative clustering approach similar to that described in our previous studies^12,16,41,95. Briefly, a population-level parcellation atlas generated by a k-means algorithm was projected to each patient’s space and then the boundaries of these parcels were refined using an iterative approach. The first iteration of the process represented the population-level parcellation. During subsequent iteration steps, the contribution of the group-level atlas was weighted according to prior knowledge of between-participant variability and signal-to-noise ratio distribution. The last iteration represented the individual-level parcellation. Accordingly, the strategy was able to balance information coming from population-level and individual-level data. In addition, we assigned the 152 regions to 1 of 18 large-scale functional networks, which included 17 canonical networks¹³ and SCAN²², according to the overlapping surface area of each region with each network. These regions were delineated by black boundaries and colour coded to represent the 18 networks.

RSFC estimation

Time courses of preprocessed rsfMRI signals were averaged across voxels within ROIs. ROIs were spherical regions centered around empirical functional locations, STN (VTA) or regions from individual parcellations. Then, a RSFC map was generated by calculating Pearson’s correlation coefficients between the averaged ROI time courses and time courses from whole-brain signals. RSFC maps were then Fisher z-transformed. Moreover, RSFC was estimated between each pair of the 152 parcellated regions and comprised a 152 × 152 connectivity matrix (connectome) for each participant. Group-level RSFC maps were calculated by averaging maps from all individuals. Spin-based permutation tests, also termed spin tests, were used to assess the statistical significance of the similarity between two STN (VTA)-cortical RSFC maps (Fig. 4b, Extended Data Fig. 4 and Supplementary Fig. 8), while accounting for spatial autocorrelation^98–100. The null distribution was generated by 10,000 random, surface-based rotations of one of the RSFC maps for comparisons.

Ideal target-cortical connectivity map

The estimation of the ideal target-cortical connectivity map largely followed a procedure described previously⁴². For each patient’s STN (VTA), STN (VTA)-cortical RSFC maps were generated using presurgical patient-specific rsfMRI data, normative rsfMRI data (Extended Data Fig. 2) and postsurgical patient-specific data across various follow-up visits and stimulation frequencies (Fig. 4e). These RSFC maps were weighted by the corresponding UPDRS-III improvement rates (score change/baseline score) and averaged to create weighted average maps. The RSFC maps were then correlated with UPDRS-III outcomes for each vertex, yielding R maps. Following the previously reported approach⁴², the R maps were used to mask the weighted average map. The resulting masked map was deemed the ideal STN (VTA) or target connectivity profile, designed to guide motor symptom prediction in the LOSOCV analysis.

LOSOCV

For the prediction of clinical outcomes at baseline (1 month before surgery) (Extended Data Fig. 2), presurgical patient-specific and normative STN (VTA) connectivity maps were used. LOSOCV was applied: connectivity maps from 13 patients’ STNs (VTAs) were used to generate the ideal target-cortical connectivity map, which was then employed to predict the clinical outcome for the remaining patient. Clinical outcome estimates were derived from the similarity, measured using Pearson’s correlation, between the ideal map and the STN (VTA) connectivity map of the remaining patient. This procedure was repeated 14× to estimate clinical outcomes for all patients and Spearman’s correlation was performed between observed and predicted outcomes. For postsurgical clinical outcome prediction (Fig. 4e), a similar approach was used, but each patient’s STN (VTA) corresponded to the patient-specific data across all stimulation conditions and follow-up visits. To prevent information leakage, all data from the patient whose outcomes were being predicted, across all stimulation conditions and follow-up visits, were excluded when generating the ideal map. An LME model was used to test the association between observed and predicted clinical outcomes. The dependent variable was the observed clinical outcomes and the independent variable was the predicted scores of different timepoints and stimulation conditions, with random intercepts for individual patients. Models were also fitted using the ‘lmertest’ package.

Individualized OT definition and distance analysis

For each patient, an individualized OT was identified within the STN as the centroid of the largest cluster functionally connected to the cortical SCAN regions, based on preoperative fMRI data. The actual target (AT) was defined as the centroid of the patient’s STN VTA. The target distance was calculated as the average Euclidean distance between the OT and the AT across both hemispheres. To assess the behavioral relevance of this spatial relationship, we computed Spearman’s correlations between OT–AT distances and clinical improvement rates (UPDRS-III score change rate). For comparison, control analyses were performed using OT coordinates derived from group-level RSFC maps and from a previously reported STN sweet spot location¹⁰¹.

Estimation of DBS’s modulatory effects on RSFC

As shown in our previous study⁴⁸, we demonstrated consistent abnormalities in RSFC primarily within the sensorimotor and visual cortices in PD, with these abnormalities significantly correlating with both motor and nonmotor symptoms. Building on this, we investigated DBS-induced modulatory effects on RSFC within the two cortices, which included 61 ROIs from the 152-region parcellation (Extended Data Fig. 5a). First, we identified RSFC edges that were significantly modulated by DBS. These were defined as edges exhibiting both (1) significant differences between patients with PD (C1) and healthy controls and (2) significant differences between DBS-ON conditions (C2–C4) and the presurgical condition (C1). The modulation effect size of these identified edges was estimated using Cohen’s d, based on the contrast between DBS ON and presurgical conditions (C2–C4 versus C1). To visualize the modulatory effect, the effect sizes were shown in the functional connectome (Extended Data Fig. 5c) and averaged across edges for each ROI and presented as a modulation effect map (Fig. 5a and Extended Data Fig. 5d). Next, the DBS-modulated edges were categorized into two groups: normalized edges, where RSFC strength became more similar to that of healthy controls, and denormalized edges, where RSFC strength diverged from that of healthy controls after DBS modulation. The RSFC strengths of edges in each category were averaged for each participant and included in subsequent statistical analyses. To illustrate these modulation effects, we performed seed-based RSFC analyses for two representative ROIs: the superior SCAN region for normalized RSFC and the effector-specific network (hand) for denormalized RSFC (Extended Data Fig. 6). Given three stimulation frequencies (HSF, LFS and VSF) and four follow-up visits (1 month, 3 months, 6 months and 12 months), we assessed the impact of stimulation frequency on modulatory effects within each group (Fig. 5) and tracked longitudinal changes in modulatory effects over time (Supplementary Fig. 9). In addition we utilized Circos (http://circos.ca/) to visually represent DBS-modulated edges.

DBS-induced response mapping using block-design fMRI

Immediate DBS-induced responses were estimated by contrasting the ‘ON’ and ‘OFF’ periods from block-design conditions 5–7. First, fMRI data were processed using a previously published pipeline designed to control DBS-induced noise during fMRI²¹. Briefly, processing steps included: (1) postsurgical fMRI data preprocessing using slice-timing correction and rigid-body head motion correction, linear co-registration to the presurgical structural image (FreeSurfer), nonlinear registration to MNI space, spatial smoothing with 6-mm full-width half-maximum, intensity normalization and highpass filtering (0.01 Hz); and (2) noise component estimation and removal using an iterative sparse noise modeling technique. Specifically, a nuisance mask, including white matter and ventricles, was defined according to probability maps. Noise components were then identified as a sparse noise dictionary from the signals in the nuisance mask using an iterative K-SVD algorithm. The sparse noise dictionary and expected DBS-induced response signals were combined. DBS-induced responses were estimated using a general linear model and restricted maximum likelihood approach, in which the noise term was prewhitened by an autoregression with an optimized autocorrelation parameter set. To examine functional activation at the individual level, β-contrast maps from each condition and each follow-up visit were analyzed using a fixed-effects general linear model for each patient. Moreover, to examine functional activation at the group level, β-contrast maps from the same condition and/or the same follow-up visit were analyzed using a random-effects model across patients. To compare DBS-induced responses across different conditions and follow-up timepoints in the M1 and GP circuits, β-values were averaged within the M1 and putamen ROIs within the M1 circuit and GP pars interna and thalamus ROIs within the GP circuit. The M1 ROI was identified as the sensorimotor cortical region with the strongest DBS-induced deactivation in the group-level map, using a threshold of z < −9. The putamen, GP and thalamus ROIs were extracted using anatomical segmentation from the HybraPD Atlas¹⁰².

Diffusion data processing and analysis

The diffusion data were preprocessed using performed using QSIPrep¹⁰³ (v1.0.2), which is based on Nipype¹⁰⁴ (v1.9.1). Standard preprocessing steps were applied to the diffusion data, including dwidenoise, N4 bias field correction, head motion correction and eddy current correction. The processed diffusion-weighted imaging data were then resampled to each participant’s anterior commissure–posterior commissure line space at 2-mm isotropic resolution. To compute fractional anisotropy (FA), the diffusion signal was modeled using the free water elimination approach implemented in the Diffusion Imaging in Python library¹⁰⁵ (v1.11.0). FA values for each motor network were computed by following a previously established approach²², by extracting the mean FA in the white matter located under the cortex. Specifically, we extended the line connecting corresponding vertices on the pial and gray–white surfaces by 2 mm into the white matter to generate a 2-mm under-surface. The pial and gray–white surfaces were derived from each participant’s preoperative T1 image using FreeSurfer. Using the ribbon-constrained mapping method in Workbench, FA values were mapped between the gray–white surface and 2-mm under-surface. FA values were then averaged across the surface vertices within SCAN and effector (hand, mouth and foot) networks. A repeated-measures ANOVA was employed to assess differences in FA across motor networks and evaluate potential postoperative changes.

Assessments of reliability and individual differences

To evaluate the reliability of neurological assessments, we compared the UPDRS-III total scores and five subscores assessed by two neurologists, using two metrics: the ICC and the rmcorr^106,107. For individualized rsfMRI parcellation, reliability was assessed by halving the rsfMRI data into two equivalent halves and comparing the parcellation assignments derived from each half using the Dice similarity coefficient for each patient. Interindividual similarity in parcellation was evaluated by comparing the assignments between pairs of patients using the Dice coefficient. Similarly, reliability and interindividual similarity in functional connectomes were assessed using Pearson’s correlation coefficient. For the presurgical STN (VTA)-cortical RSFC maps, reliability and interindividual similarity were quantified using the Dice coefficient, while the similarity between patient-specific and normative STN (VTA)-cortical RSFC maps was also examined. To assess the reliability of DBS-induced response mapping, each patient’s fMRI data from conditions 5–7 were divided into two halves, each consisting of 12 min of scanning. Functional activation within two representative ROIs, M1 and GP pars interna, was calculated for each half and compared using ICC and rmcorr. In addition, Pearson’s correlation coefficient was used to evaluate the reliability and interindividual differences in whole-brain activity maps.

Reporting summary

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

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41593-026-02228-w.

Author contributions

Conception: H.L. and L.L. Design: J.R., C.J., W.Z., L.L. and H.L. Data acquisition, analysis and interpretation: J.R., C.J., W.Z., L.D., L.S., F.Z., S.L., C.H., Y.Y., X.F., J.H., Y. Long, D.L., Y.G., Y. Liu, S.X., F.M., J.Z., D.W. and H.L. Paper writing and revision: J.R., L.D., W.Z., D.W., L.L. and H.L.

Peer review

Peer review information

Nature Neuroscience thanks Andreas Horn and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Data availability

Anonymous individual patient data, including all brain imaging modalities—sMRI, rsfMRI, bdfMRI, diffusion tensor imaging and CT—demographic data and neurological assessments, are available in the Science Data Bank repository: 10.57760/sciencedb.26302. The preprocessed imaging data will also be available on the repository. Source data are provided with this paper.

Code availability

All fMRI data were preprocessed using a cloud medical image data-processing software, pBFS cloud, at https://app.neuralgalaxy.cn/research/. Code specific to analyses can be found at https://github.com/pBFSLab/open-DBS. The bCompCor code can be found in the GitHub repository and has also been directly implemented in our open-access preprocessing pipeline, DeepPrep⁹⁷, available at https://deepprep.readthedocs.io/en/latest/. Software packages incorporated into the above code for data analysis include: Python v3.7, https://www.python.org; MATLAB R2020b, https://www.mathworks.com/; R v4.3.1, https://www.r-project.org; Connectome Workbench v1.5, http://www.humanconnectome.org/software/connectome-workbench.html; FreeSurfer v6.0.0, https://surfer.nmr.mgh.harvard.edu/; FSL v6.0, https://fsl.fmrib.ox.ac.uk/fsl/fslwiki; ANTs v0.3.8 https://github.com/ANTsX/ANTs; and Lead-DBS v2.0, https://www.lead-dbs.org/.

Competing interests

H.L. is the chief scientist of Neural Galaxy Inc. C.J. serves on the scientific advisory board for Beijing Pins Medical Co., Ltd. C.J. is also listed as an inventor in issued patents and patent applications for the deep brain stimulator used in this work. The other authors declare no competing interests.

Extended data

is available for this paper at 10.1038/s41593-026-02228-w.

Supplementary information

The online version contains supplementary material available at 10.1038/s41593-026-02228-w.