The pattern of fMRI activation with STN DBS reveals symptom-specific networks

Abstract

The network changes elicited by deep brain stimulation (DBS) in Parkinson’s disease (PD) remain incompletely understood. We used functional MRI (fMRI) during active stimulation of the subthalamic nucleus (STN) to identify brain networks associated with clinical improvement. Forty PD patients (61 y ±8, 40% female) underwent 3-Tesla fMRI during active stimulation. Stimulation-induced changes in brain activity were correlated with outcomes (tremor, rigidity, bradykinesia, and axial instability). Symptom-specific fMRI response maps revealed distributed networks associated with long-term motor improvement, but ROI-based models showed poor out-of-sample performance and limited spatial stability. In contrast, somatomotor network modulation demonstrated symptom-specific coupling preferentially in rigidity improvement (ΔR² = 0.24, p_FDR = 0.01) and bradykinesia (ΔR² = 0.13, p_FDR = 0.07). Normative connectivity explained variance less consistently. These findings support the use of patient-specific fMRI as a tool for studying the neural mechanisms of DBS and highlight the importance of distributed network engagement in mediating therapeutic benefit.

Introduction

Parkinson’s disease (PD) is commonly treated using deep-brain stimulation (DBS) of the subthalamic nucleus (STN)¹. Neuromodulation with DBS is among the fastest-growing procedures in neurosurgery, with over 10,000 patients implanted annually^2,3. STN-DBS improves the cardinal symptoms of PD, including axial instability, bradykinesia, rigidity, and tremor. Despite this, its therapeutic mechanisms are incompletely understood, and the degree of improvement between individuals and symptoms remains variable^4,5. A better understanding of the effects of stimulation on brain-wide activity may allow for personalized therapy and targeted symptom-specific improvement.

Prior work has demonstrated that STN-DBS modulates canonical motor circuitry, including key hubs within the cortico-striato-thalamo-cortical (CSTC) loop⁶. These findings provide valuable insight into the mechanisms underlying basal ganglia neuromodulation; however, they do not explain the variable symptom response in individual patients⁶. Although Parkinson’s disease shares a common underlying pathology, its clinical presentation and progression are highly heterogeneous, with patients exhibiting varying combinations of axial instability, bradykinesia, rigidity, and tremor⁷. Personalization of therapy targeting a patient’s most burdensome symptoms may yield greater improvements in quality of life. Indeed, studies have suggested that clinical features of PD map to different brain networks, forming the basis for symptom-specific circuitopathies⁴. Group-averaged normative healthy data have also shown that symptoms are preferentially improved depending on the precise location of stimulation^8,9. By further leveraging patient-specific brain responses to DBS, therapy could be individualized to unique clinical profiles.

Functional magnetic resonance imaging (fMRI) acquired during stimulation can visualize brain networks acutely engaged by DBS^10–16. In this study, we examined 40 STN-DBS PD patients who underwent 3-Tesla fMRI during active cycling stimulation. We sought to identify networks associated with the clinical improvement of specific symptoms. We propose the use of these patterns of network engagement as potential biomarkers for personalizing therapy.

Results

Clinical and imaging responses to STN-DBS

Subjects displayed akinetic/rigid (48%), postural instability gait difficulty (25%), and tremor-dominant (27%) phenotypes and exhibited a mean ( ± standard deviation) improvement of 42% (±44%) in right-sided motor impairment (MDS-UPDRS III) following STN-DBS. A breakdown of mean improvement by specific motor symptom can be seen in Table 1. The mean VTA centroid location (x = −13.28, y = −13.02, z = −4.17) was 1.44 mm from the previously described “sweet spot” for overall motor improvement (Fig. 1b)⁹. Distance from this sweet spot (median Euclidean spacing of 2.88 mm) did not explain variation in clinical outcomes (R² = 0.007, p = 0.60). As described previously¹⁷, STN-DBS in our cohort acutely evoked significant (p < 0.001) group-level BOLD changes, namely increased BOLD signal in the ipsilateral thalamus and reduced BOLD signal in the ipsilateral sensorimotor cortex (Supplementary Material 1: Supplementary Fig. 2).

Table 1.

Subject demographics

	Mean (±standard deviation)
Age at surgery (years)	61.23 (±8)
Disease duration* (months)	164 (±59)
Duration of stimulation* (months)	23 (±18)
Percentage Improvement
Axial instability	44% (±42%)
Right bradykinesia	29% (±64%)
Right rigidity	48% (±60%)
Right tremor	37% (±68%)

Demographic information and clinical outcomes for all 40 patients who underwent fMRI during active STN-DBS. Improvement represents the percent difference between MDS-UPDRS pre- and postoperatively OFF medication. DBS deep brain stimulation, fMRI functional magnetic resonance imaging, MDS-UPDRS Movement Disorder Society Unified Parkinson Disease Rating Scale, STN subthalamic nucleus.

*Measured at time of fMRI acquisition.

Fig. 1. Experimental design and electrode localizations.

a fMRI scans were acquired in 40 patients with STN-DBS. Patients undergo clinical DBS optimization. Once optimized, they undergo stimulation-block design fMRI. Statistical parametric mapping is performed to create functional response t-maps by comparing stimulation-ON blocks to stimulation-OFF blocks. Functional response t-maps are then correlated with clinical outcomes. b The distribution of all patient electrodes can be seen on sagittal and coronal slices of a standard brain template. An exemplar electrode and corresponding VTA (red) are shown in relation to the STN (orange). The mean location of all left-sided VTAs is shown in blue compared to a previously published⁹ motor “sweet spot” (pink) in standard MNI space. DBS deep brain stimulation, fMRI functional magnetic resonance imaging, MNI Montreal Neurological Institute, STN subthalamic nucleus, VTA volume of tissue activated.

Correlation of clinical outcomes with BOLD response

BOLD-efficacy maps, showing the relationship between BOLD response and clinical outcomes across all subjects, revealed significant networks associated with clinical response (p < 0.001). Moreover, these networks differed according to the specific symptom of interest (Fig. 2). Improvements in axial instability were positively correlated with BOLD response within the right cerebellum (lobules I–V), left striatum, and left middle temporal gyrus. Improvements in bradykinesia were negatively correlated with BOLD response in the left precentral gyrus leg representation, right frontal pole, right frontal operculum, and right insula. Rigidity benefits were positively correlated with BOLD response in the left precentral gyrus and negatively correlated with BOLD response in the left supplementary motor area (SMA). Finally, tremor improvement was negatively correlated with right frontal and bilateral temporal pole BOLD response.

Fig. 2. Symptom-specific networks of DBS improvement.

Areas in which DBS-evoked BOLD responses were significantly (p < 0.001) related to symptom-specific clinical outcomes are shown overlaid on axial slices of a standard MNI brain template⁵⁵. Areas in which increased stimulation-dependent BOLD response was correlated with symptom improvement are shown in warm colors. Areas in which decreased BOLD response was correlated with symptom improvement are shown in cool colors. BOLD blood–oxygen-level-dependent, Caud caudate, Cer cerebellum, DBS deep brain stimulation, FOp frontal operculum, FPo frontal pole, Ins insula, MTG middle temporal gyrus, MNI Montreal Neurological Institute, PrGA precentral gyrus arm representation, PrGL precentral gyrus leg representation, SMA supplementary motor area, TFCE threshold-free cluster enhanced, TPo temporal pole.

Nested cross-validation of symptom-specific ROI models

Across symptom domains, nested ROI-based models demonstrated substantial variability in held-out performance and failed to generalize to unseen subjects. Models often explained a moderate proportion of variance within the training folds, consistent with group-level correlations described above (mean R² = 0.26) (Fig. 3a). However, this relationship failed to generalize, with R² values for held-out subjects frequently low (mean R² = 0.02) (Fig. 3a). Dice similarity coefficient demonstrated limited overlap between independently derived ROIs for most outcomes (rigidity = 0.08, tremor = 0.20, axial instability = 0.18), indicating high spatial variability (Fig. 3b). Bradykinesia demonstrated the highest spatial consistency in ROI selection with a Dice similarity coefficient of 0.45 between folds (Fig. 3b).

Fig. 3. Nested ROI selection and cross-validation reveal limited generalizability and spatial instability in outcome prediction.

a Mean predictive performance across nested cross-validation folds (N = 40, 5 folds). For each clinical outcome, symptom-specific regions of interest (ROIs) were independently derived within training data using voxel-wise permutation testing of stimulation-evoked fMRI responses (5000 permutations; TFCE corrected p < 0.05). Mean BOLD signal within each fold-specific ROI was then used to predict clinical improvement using linear regression. Points denote the mean coefficient of determination (R²) across folds for training subjects (blue) and held-out test subjects (red), with connecting lines illustrating the reduction in performance between training and test sets. Although models demonstrated moderate explanatory power within training samples, predictive performance consistently declined in held-out subjects, indicating limited out-of-sample generalization. b Spatial stability of fold-derived ROIs quantified using Dice similarity coefficients. Each point represents the mean Dice overlap across all pairwise fold comparisons, with horizontal lines indicating the observed range (min–max). Low Dice overlap across folds indicates substantial variability in voxel selection, suggesting that symptom-associated ROIs are spatially unstable despite similar within-sample model performance. BOLD blood oxygen level dependent, DBS deep brain stimulation, fMRI functional magnetic resonance imaging, ROI region of interest, TFCE threshold-free cluster enhancement.

Circuit-specific network modeling

Given the instability of data-driven ROI models, we next evaluated whether symptom-specific clinical improvement was better explained by distributed modulation of predefined motor networks. Somatomotor network¹⁸ modulation explained significant incremental variance in rigidity improvement (ΔR² = 0.24, p_FDR = 0.01) and bradykinesia improvement (ΔR² = 0.13, p_FDR = 0.07) (Fig. 4). In contrast, somatomotor network modulation explained minimal variance in gait (ΔR² = 0.05) and tremor improvement (ΔR² = 0.01) (Fig. 4).

Fig. 4. Somatomotor network modeling comparing patient-specific fMRI response to stimulation and normative functional connectivity in STN-DBS.

Incremental variance explained (ΔR²) by somatomotor network¹⁸ modulation for cardinal motor symptoms. Blue circles represent ΔR² derived from patient-specific stimulation-evoked fMRI responses, whereas red circles represent ΔR² derived from normative functional connectivity seeded from each subject’s VTA. Horizontal lines connect paired estimates for each symptom, illustrating differences between modeling approaches. Patient-specific somatomotor network modulation explained significant variance in rigidity and bradykinesia improvement and outperformed normative connectivity. Normative connectivity explained a modest proportion of variance in axial instability improvement but did not survive correction for multiple comparisons. Statistical significance was assessed using permutation testing (5000 permutations) with FDR correction applied across outcomes. DBS deep brain stimulation, FDR false discovery rate, fMRI functional magnetic resonance imaging, STN subthalamic nucleus, VTA volume of tissue activated.

Differentiating patient-specific response to stimulation from normative functional connectivity and electrode position

Normative functional connectivity-efficacy maps based on the correlation between left VTA (i.e., electrode position) and rsfMRI BOLD fluctuations over time in healthy individuals were generated. The network-level analysis was repeated to assess whether patient-specific acute response to stimulation provides additional information to explain outcome variance beyond what is provided by normative connectivity. Within the somatomotor network, normative connectivity explained a modest proportion of variance in gait improvement (ΔR² = 0.16, p_permutation = 0.04), although this effect did not survive correction for multiple comparisons (p_FDR = 0.17) (Fig. 4). Smaller effects were observed for bradykinesia (ΔR² = 0.09), rigidity (ΔR² = 0.06), and tremor (ΔR² = 0.02), none of which reached statistical significance (Fig. 4).

Discussion

We have previously shown that fMRI network activity in response to active STN-DBS contains information about long-term motor improvement and may identify optimal stimulation parameters^19,20. Here, using a unique patient-specific fMRI technique during active STN-DBS, we build on this observation to dissect this pattern into symptom-specific components. Acute stimulation-evoked BOLD responses revealed distributed networks associated with long-term improvement in each of the four cardinal motor symptoms of Parkinson’s disease. However, when region-of-interest (ROI) models derived from these networks were evaluated using nested cross-validation, predictive performance declined substantially in held-out subjects, and the spatial locations of fold-derived ROIs were highly variable. These findings suggest that although group-level symptom-efficacy maps capture biologically meaningful associations, focal data-driven ROIs derived from modest cohorts are spatially unstable and do not support robust single-subject prediction.

The resulting decline in held-out performance suggests that the predictive signal in these data is more distributed and heterogeneous than can be captured by small sets of focal ROIs. Consequently, network-level modeling using predefined circuits may provide a more stable and biologically meaningful framework for interpreting DBS-evoked activity. Engagement of the CSTC circuit during therapeutic STN-DBS is thought to underlie motor improvement^{6,10,16,19,21–23}. We found that modulation of the somatomotor network explained significant variance in rigidity improvement and showed an association with bradykinesia improvement during STN-DBS, whereas tremor and axial instability demonstrated comparatively weak associations. These findings suggest that stimulation-induced modulation of the distributed somatomotor system is symptom-specific, and the degree of engagement may explain outcome variance.

Importantly, the networks identified in this study included both traditional motor regions and areas not classically associated with primary motor function. Specifically, axial symptoms correlate with BOLD responses in the cerebellum and striatum (cortico-basal ganglia-cerebellar network)²⁴; bradykinesia involves the precentral gyrus and frontal pole (striato-frontal connectivity)²⁵; rigidity is linked to the precentral gyrus and SMA (long loop reflex)²⁶; tremor shows association with temporal and frontal poles (striato-frontal connectivity and complex perceptual processing)^25–27. These findings are consistent with evidence that the Parkinsonian brain recruits additional compensatory circuits during movement, which may be captured by stimulation-evoked fMRI responses. The heterogeneity of these networks also suggests that individual symptom improvement may depend on the extent to which DBS engages distributed circuits beyond the canonical motor pathway²⁸.

While the location of stimulation within the STN may be predictive of outcome, a large meta-analysis based on structural imaging from 1391 patients found that stimulation volumes within 2 mm of each other did not significantly affect the extent of motor improvement²⁹. Indeed, in our cohort, in which there was close proximity of stimulation volumes (median Euclidean spacing of 2.88 mm), the variation in clinical outcomes was not explained by stimulation location. In addition, we explored whether normative functional connectivity to the stimulation location could explain variance in symptom-specific outcomes beyond that provided by patient-specific BOLD response. Normative connectivity explained a modest proportion of variance in axial instability improvement but did not survive correction for multiple comparisons and showed less consistent associations for the remaining motor symptoms in our cohort of 40 subjects. These findings suggest that individual brain response to stimulation seen with patient-specific fMRI can capture inter-patient differences not accounted for when using structural imaging or normative functional connectivity in small study populations. Such differences may reflect individual variation in structural connectivity, disease-related network reorganization, or patient-specific physiological responses to stimulation that cannot be inferred from normative connectomes. Nevertheless, the prevalence and ease of acquiring structural imaging make normative connectivity and electrode localization methods well-suited for large-scale multicenter studies; these large studies perform similarly to the patient-specific fMRI data presented here^30–33.

A central finding in fMRI studies during active DBS is that changes to stimulation (e.g., cycling the device on/off) will manifest in rapid fMRI responses, preceding changes in symptoms which may take hours/days to reappear^{11,19,20,34–37}. These studies have further shown an association between acute stimulation-dependent activity and long-term clinical outcome, an association robust enough to build predictive models of outcome^11,19,20,34. Here, we suggest that patient-specific fMRI affords the granularity to reveal network modulation unique to individual symptoms in PD, in advance of the physical manifestation of symptoms. The brevity (6.5 min) of the scan and intermittent active left stimulation meant all subjects tolerated the acquisition with minimal symptomatic artifact or discomfort. Perhaps symptoms of tremor or rigidity may have emerged if more extended periods of non-stimulation occurred, and could have a confounding effect on the results. Early biomarkers of long-term outcomes are of particular importance within clinical constraints, where there is often a considerable delay between setting changes and visible symptom response.

The association between fMRI response and clinical outcome depicted here supports the growing role for patient-specific fMRI in personalizing DBS therapy. Immediate and objective imaging-based biomarkers of treatment efficacy may inform the selection of stimulation parameters^19,20, a time-consuming process based on clinical observation^38,39. Such biomarkers could be used to develop semi-automated programming techniques, which will become necessary as novel DBS devices (e.g., directional leads) introduce an ever-increasing number of parameter combinations that cannot be thoroughly evaluated by conventional programming algorithms^38,40. Furthermore, the ability to precisely manipulate specific symptoms will allow tailoring of therapy for individual patients. This is particularly important in PD patients, who encompass different phenotypes with symptoms that are hard to assess or have a delayed response to therapy⁴¹. Nevertheless, key questions regarding validity, reproducibility, and feasibility require additional prospective assessment to answer.

Several features limit the generalizability of this work. This is mainly a product of the scarcity and uniqueness of the imaging data acquired during active deep-brain stimulation. Fortunately, ongoing improvements in MRI compatibility (i.e., MRI conditional models at 3 T) and a growing body of literature on safety will facilitate future studies that minimize these limitations. Importantly, fMRI was acquired only during unilateral stimulation. Although this highlights the laterality of response and avoids obscuring artifact from DBS hardware, it necessitates assessment of only right-sided outcomes – particularly in the case of axial instability—and does not take into account the contribution from right-sided stimulation. We recognize the potential confound that this imposes, understanding that unilateral stimulation also contributes to 10–20% of ipsilateral improvement^42,43. We assume the majority of the contralateral improvement is the result of the stimulation captured with imaging. Clinical outcomes were also obtained at the time of postoperative optimization, reflecting the stable therapeutic state achieved through iterative programming and medication adjustment rather than early postoperative fluctuations. It does not represent a strictly contemporaneous measurement with the fMRI session and could introduce additional variability into observed associations. Future studies should prospectively evaluate the accuracy of predefined fMRI models with additional patients to improve the power of the analyses and prediction.

In summary, stimulation-evoked fMRI responses during STN-DBS reveal distributed brain networks associated with symptom-specific motor improvement. However, focal ROIs are spatially unstable, suggesting that DBS outcome signals are better understood at the level of distributed circuits rather than isolated cortical hotspots. Network-level modeling demonstrates that modulation of the somatomotor system is particularly associated with rigidity improvement and that patient-specific stimulation-evoked physiology captures clinically relevant variance beyond that explained by normative connectivity. These findings support the use of patient-specific fMRI as a tool for studying the neural mechanisms of DBS and highlight the importance of distributed network engagement in mediating therapeutic benefit.

Methods

Subjects

Forty subjects, meeting inclusion criteria, were enrolled from a larger prospective fMRI cohort (n = 67) that has been described previously (#NCT03153670, 2017-06-01)¹⁹. Subjects were all-comers, reflective of institutional clinical selection practices and not preselected based on symptomatology. Following institutional research ethics board approval (ID: #14-8255) and in accordance with the Declaration of Helsinki, we enrolled and obtained informed consent from PD patients who had undergone DBS surgery at our institution and who had: (i) DBS targeting the STN; (ii) optimized stimulation settings following standard-of-care clinical programming³⁸; (iii) suitable clinical follow-up (complete baseline and postoperative symptom severity assessments); and (iv) unilateral left stimulation during the fMRI acquisition, were considered. Unilateral stimulation was used to disentangle the laterality of fMRI responses (i.e., whether observed fMRI responses were ipsilateral or contralateral to the side of stimulation) and to mitigate the effect of the susceptibility artifact created by the right-sided subgaleal coil of DBS extension wire¹⁷. From the original 67-subject cohort, 9 individuals were excluded because their fMRI acquisition preceded clinical programming, 5 were excluded due to a different DBS target, 10 were excluded because the stimulation parameters used during fMRI differed from the clinically determined optimal parameters (item ii, above), and 3 were excluded due to no follow-up at the time of the fMRI (Supplementary Material 1: Supplementary Fig. 1). Clinical outcomes were recorded off medications at the clinic visit closest to the MRI scan (mean interval: 4.5 months (±5.5 months)). DBS contact location remained unchanged between these points for each subject. The four cardinal motor symptom outcomes were chosen a priori and measured using the following items from the Movement Disorder Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) III: (i) axial instability (items 3.9–3.14); (ii) tremor (items 3.15–3.18); (iii) rigidity (item 3.3); and (iv) bradykinesia (items 3.1, 3.2, 3.4–3.8). For each symptom (except for axial instability, for which the scale does not capture laterality), only right-sided outcomes were used in order to match the left-unilateral stimulation design.

Electrode localization

For visualization, subjects underwent electrode localization as previously described⁴⁴. High-resolution pre- and postoperative T1-weighted 3-dimensional spoiled gradient echo (3D-SPGR) MRI scans were acquired for each patient at either 1.5 or 3-T (GE HealthCare, Waukesha, WI, USA). Other acquisition parameters include voxel size = 1 × 1 × 1 mm³, repetition time (TR) = 9.0/11.9 ms, echo time (TE) = 3.7/5.0 ms, flip angle = 12°/20°). Electrode localization utilized Lead-DBS v3.0 software (https://www.lead-dbs.org/)⁴⁵. Pre- and postoperative images were linearly registered using SPM12 (https://www.fil.ion.ucl.ac.uk/spm/software/spm12/)⁴⁶ and non-linearly normalized to a Montreal Neurological Institute (MNI) template brain (ICBM 2009b NLIN asymmetric) using ANTs SyN and subcortical refinement (http://stnava.github.io/ANTs/)⁴⁷. To localize DBS electrodes, the semiautomated trajectory reconstruction function in Lead-DBS was used and manually refined as necessary by an experienced user (A.L.) and confirmed by a second (B.S.). Volumes of tissue activated (VTAs) were estimated using previously published methods^48,49. The mean VTA location was determined from the VTA centroid of each subject and compared to a well-validated, previously published “sweet spot” (MNI: x = −12.68 mm, y = −13.53 mm, z = −5.38 mm)⁹ (Fig. 1b).

fMRI acquisition and single-subject analysis

For each subject (n = 40; Table 1; Fig. 1), a 6.5-min 3-Tesla blood-oxygen-level-dependent (BOLD) fMRI session was obtained (GE HealthCare, Waukesha, WI, USA) using a 30-s DBS-ON/OFF cycling paradigm as previously described¹⁹. Other acquisition parameters include voxel size = 3.75 × 3.75 × 4 mm³, TR = 2.15 ms, TE = 30 ms, and readout bandwidth = 62.5 kHz. This design was selected to balance statistical sensitivity with patient tolerability and safety considerations for implanted DBS systems. Previous work from our group demonstrated that this paradigm reliably captures stimulation-evoked BOLD responses and can differentiate optimal from non-optimal stimulation settings¹⁹. Moreover, the short acquisition duration allows multiple stimulation conditions to be tested within a clinically feasible imaging session while minimizing patient fatigue and motion artifacts. Safety monitoring followed previously published protocols¹⁷, including close clinical monitoring throughout the scan, pre-, and post-scan impedance measurements of DBS contacts, and immediate post-scan neurological examination. No adverse events or device malfunctions occurred during the study.

Images were acquired during unilateral left-sided stimulation using patients’ clinically optimized (resulting in the greatest clinical benefit) stimulation settings. Patients were instructed to be at rest with eyes closed. Scans were performed off medication; patients received their final medication dose the night preceding scan acquisition (8–12 h medication withdrawal). Preprocessing of images consisted of slice-time correction, rigid-body registration to a simultaneously acquired T1 image, non-linear registration to MNI space, and spatial smoothing¹⁹. Statistical parametric maps (functional response T-maps) were estimated from the preprocessed fMRI data using the designed 30-s DBS-ON/OFF paradigm with motion regression in SPM12 (Fig. 1a). Data acquired during the first 30 s of each fMRI scan were discarded to establish a steady state.

Group analysis

To identify functional networks contributing to individual symptom effects, a group-level random effects analysis was performed using SPM12. Significant hemodynamic changes between ON and OFF stimulation conditions were assessed using the obtained whole-brain t-maps. Outcomes were reported as a percentage improvement from baseline. This approach normalizes symptom improvement relative to each patient’s baseline disease severity. Only right-sided scores were used for non-axial items to isolate the clinical effects of left-sided stimulation. Despite this, clinical improvement likely corresponds to contributions from both ipsilateral and contralateral stimulation and requires further investigation to parcellate^50,51.

Nested cross-validation and ROI stability analysis

To evaluate whether BOLD response patterns could explain outcome variance, we implemented a nested 5-fold cross-validation analysis. Within each training set, voxel-wise symptom-efficacy maps were generated by correlating stimulation-evoked BOLD response with clinical outcome using permutation testing (n = 5000) with threshold-free cluster enhancement (TFCE)^52,53. The candidate regions of interest (ROI) were then defined from the training-derived maps by identifying clusters of maximal t-values and expanded using a fixed 5 mm radius dilation constrained to TFCE-significant voxels. Mean stimulation-dependent BOLD response within the selected ROIs was extracted for each subject and used to fit linear models predicting symptom-specific clinical improvement. Model performance was quantified separately for training and testing datasets, providing an estimate of in-sample and out-of-sample predictive performance.

To evaluate the spatial stability of ROIs, ROI overlap between cross-validation folds was quantified using the Dice similarity coefficient. Dice scores were computed pairwise across all fold combinations for each outcome.

Network-level PCA analysis

To evaluate whether distributed motor-network activity better explained variance in symptom-specific outcomes, we performed a hypothesis-driven network-level analysis using a predefined cortical somatomotor network¹⁸. For each subject, voxel-wise stimulation-evoked BOLD responses within this network were extracted. Dimensionality reduction was performed using principal component analysis (PCA). Because of the limited cohort size, analyses were restricted to the first two principal components, capturing spatially distributed patterns of cortical response that covary across individuals rather than localized activation peaks. Clinical associations were evaluated using linear regression models relating principal component scores to clinical outcomes while controlling for age, disease duration, and baseline symptom severity. Incremental variance explained (ΔR²), therefore, reflects the additional explanatory value of stimulation-evoked network activity beyond these clinical covariates. To quantify the contribution of network modulation beyond clinical covariates, we computed the incremental variance explained (ΔR²) between a covariate-only model and a full model that includes covariates and PCA-derived network components. Statistical significance of ΔR² was assessed using non-parametric permutation testing (n = 5000) and false discovery rate (FDR) correction.

Differentiating patient-specific response to stimulation from normative functional connectivity and electrode position

Prior to adequate safety testing, fMRI investigations into the effects of DBS relied on using a resting-state functional connectome derived from separate healthy subjects or PD patients without DBS (normative functional connectivity)⁵⁴. The unique fMRI dataset presented here capitalizes on patient-specific functional imaging and provides insight into activity changes driven by stimulation in individuals. To compare our patient-specific dataset with normative equivalents, functional connectivity estimates were computed using the disease-matched, Parkinson’s Progression Markers Initiative (PPMI) functional connectome (n = 74) (https://www.ppmi-info.org/). Time series were sampled from the PPMI functional connectome for each of our 40 subjects’ left-sided VTA and correlated with every other voxel in the connectome. Correlations were then expressed as average R values and Fisher z-transformed. For each subject, Fisher z-transformed connectivity values within the predefined somatomotor network were extracted. The same PCA-based dimensionality reduction procedure used for the patient-specific BOLD analysis was then applied to these connectivity maps to obtain principal components describing dominant patterns of normative somatomotor connectivity across subjects. Just as above, incremental variance explained (ΔR²) was assessed using permutation testing with FDR correction across symptom domains.

Statistical analysis

All neuroimaging analyses were performed using the FMRIB Software Library (FSL, version 6.0; Oxford Center for Functional MRI of the Brain, Oxford, UK). Image manipulation, masking, and cluster-based operations were conducted using standard FSL command-line utilities.

Network-level modeling, dimensionality reduction, and permutation analyses were performed in Python (version 3.11) and R (version 4.3; R Foundation for Statistical Computing, Vienna, Austria).

Author contributions

B.S. contributed to the project conception and design, data collection, statistical design and execution, and paper writing and review. A.L. contributed to the project conception and design, and paper review. J.G., A.B., A.A., J.Q., and R.M. contributed to the project conception and design, data collection, statistical design and execution, and paper review. G.A. contributed to data collection. M.A. and G.E. contributed to statistical design and execution. A.V., C.C., and C.S. contributed to the paper review and critique. R.M., G.I., S.K., A.F., and A.L. contributed to project conception and design, execution, and manuscript review and critique. All authors have reviewed and approved the contents of the submission and meet authorship requirements.

Data availability

Data can be made available upon request.

Code availability

Code for analysis and figure creation can be made available upon request.

Competing interests

A.M.L. is the scientific director of Functional Neuromodulation Ltd. and a consultant for Medtronic, Abbott, Boston Scientific, and Insightec. AF received honoraria and/or research support from Abbott, AbbVie, Boston Scientific, Ceregate, Inbrain, and Medtronic. SKK received consulting fees from Medtronic. GMI received consulting and advisory fees from Medtronic, Livanova, and Synergia. A.B., R.M., and A.M.L. have intellectual property related to this paper.

Supplementary information

The online version contains supplementary material available at 10.1038/s41531-026-01368-x.