Patient cohorts, imaging and clinical assessments

Every stage of the research process complied with all relevant ethical regulations, and post hoc analyses performed for the purpose of the present manuscript were approved by the institutional review board of Charité – Universitätsmedizin Berlin (master vote EA2/186/18). Procedures of clinical trials and studies leading to the collection of these data were approved by the institutional review boards at each of the respective data collection sites. They were carried out in accordance with the 1975 Declaration of Helsinki, and all participants signed an informed consent before study participation. Participants received no compensation in exchange for taking part in this research.

Discovery cohort

The present study sought to establish models of optimal focal stimulation sites and streamlines, harnessing a retrospective discovery sample of eight patient cohorts (n = 197) spanning seven international DBS centers (San Francisco, Shanghai, Berlin, Würzburg, Grenoble, London and Pisa/Milan). Each of these patients had been bilaterally implanted with subthalamic DBS for treatment of DYT (n = 70), PD (n = 94), TS (n = 14) or OCD (n = 19). The full sample consisted of two patient cohorts per disease, with the Shanghai center contributing two cohorts (DYT and TS data). Among all available patients with complete neuroimaging and clinical outcome information, no patient was discarded from our analyses. Instead of prospective randomization, we leveraged incidental variability in electrode placement within each disease cohort, which can be presumed to be random. Supplementary Table 1 summarizes the included discovery cohorts, with more detailed patient-wise demographic and clinical information listed in Supplementary Tables 2–5.

Given the exploratory nature of our study, no statistical methods were used to pre-determine sample sizes. As neuropsychiatric applications of STN-DBS are recent and rare, samples, especially in TS and OCD, are limited by the small number of worldwide surgeries. In the TS cohort, we included all globally treated patients undergoing STN stimulation for this condition at the time of analysis. Overall, we were able to include two cohorts per disease group. Our initial assumption of expected effect sizes was based on Li et al.21 and Treu et al.70, with an R of approximately 0.4 for reported correlations between empirical clinical outcomes and estimated gain scores. In view of the natural restrictions in available sample sizes, we calculated a ‘compromise’ type power analysis using G*Power, version 3.1.9.6 (refs. 71,72), to determine the power of our analyses based on the accessible data per disorder to detect the assumed effect. Given a β/α ratio of 1, the available DYT sample (n = 56 in the main cohort) used for the model setup was powered to 0.94, the PD sample (n = 94) to 0.98, the TS sample (n = 14) to 0.77 and the OCD sample (n = 19) to 0.81 for detecting the assumed effect size. To our knowledge, this is the largest transdiagnostic study of its kind.

Retrospective validation cohorts

To further validate streamline models in two exemplary disorders based on out-of-sample data, two additional patient cohorts were integrated. The first consisted of a further cohort of patients with PD from Würzburg receiving STN-DBS (n = 32). The second comprised an additional cohort of patients with OCD, pooled across London, Cologne and Boston centers, treated with DBS of the VC/VS region (n = 35). Crucially, these patients contributed entirely independent data points that had not been used to inform the previous streamline model setup. The only exception was formed by the OCD-DBS cohort from London, in which patients had received a set of electrodes each to both targets (STN and VC/VS, with n = 4 electrodes per patient) that had been activated independently during the original study73. For this cohort, stimulation settings and clinical scores with ‘optimized’ stimulation of both targets combined or of each target separately were available. For model generation within the discovery cohort (with subthalamic focus), stimulation parameters and corresponding Y-BOCS improvement values collected during the ‘STN-DBS-only’ phase were implemented, whereas corresponding information acquired during the ‘VC/VS-DBS-only’ phase was used to inform the retrospective OCD model validation. Supplementary Table 6 summarizes these two additional retrospective cohorts. Patient-specific information is provided in Supplementary Table 7 (PD validation cohort) and Supplementary Table 8 (OCD validation cohorts). Again, none of the available patients with complete neuroimaging and clinical information was excluded from further analysis.

Prospective patient cases

Streamline models for PD and OCD were further prospectively validated by reprogramming DBS settings in a patient with PD and in a patient with OCD from Würzburg and Boston, respectively, guided by the aim of maximized engagement of their stimulation volumes with the corresponding streamline model. Both patients were recruited and investigated within the ongoing clinical service—namely, in the inpatient DBS program at University Hospital Würzburg in the case of the patient with PD or in the DBS program of the psychiatry and neurosurgery departments at Massachusetts General Hospital (MGH) in the case of the patient with OCD.

Finally, a single patient with OCD from São Paulo underwent DBS surgery and programming as informed by the OCD streamline model. This patient was recruited within the regular surgical service of Clínica de Dor e Funcional after classification as a refractory case of OCD that was associated with depression. He underwent evaluation by a neurologist, a psychologist and two functional neurosurgeons before approval of the DBS implantation surgery by the Ethics Board Committee of the State of Rio Grande do Sul. Supplementary Table 11 provides additional details on these three patient cases.

To inform surgical planning and for exclusion of structural abnormalities, all patients received high-resolution multi-spectral structural MRI that had been acquired at 3T field strength. High imaging quality was ensured through visual inspection by a multi-disciplinary team during stereotactic planning, and, in case of movement artifacts, pre-operative acquisitions were repeated under general anesthesia. Intra-operative microelectrode recordings and macrostimulations as well as either post-operative MRI (n = 73) or computed tomography (CT) of the head (n = 188) (Supplementary Tables 1–8 and 11) were acquired to confirm accurate lead placement.

Specifics on electrode models implanted in each cohort used for the model setup are summarized in Supplementary Table 1, and the same information for retrospective model validation cohorts and prospective patient cases is provided in Supplementary Tables 6 and 11, respectively. Stimulation settings and corresponding clinical improvement scores for all cohorts were selected from times of follow-up to which stimulation effects had sufficiently stabilized (Supplementary Tables 1, 6 and 11).

Times of follow-up available for some patients within the n = 58 cohort of patients with DYT from Shanghai were shorter than those of other disease cohorts. In addition, as DYT is a heterogeneous disease of several forms (for example, generalized, segmental and focal somatotopic expressions), pre-operative BFMDRS summary scores in some Shanghai patients were considerably lower than those of patients in the San Francisco cohort. To ascertain stabilized and comparable DBS effects across cohorts, main analyses were, thus, carried out on the DYT sample including a subcohort of Shanghai patients (n = 44), which sufficed to more conservative inclusion criteria (baseline BFMDRS scores ≥5 and follow-up ≥6 months). However, we repeated our results on the complete DYT sample (n = 70) including the full Shanghai cohort (n = 58) to demonstrate stability of effects (Supplementary Fig. 10).

Clinical outcome data and DBS parameters were retrieved from the collecting sites using Microsoft Excel, version 16.70, and imported for analysis using MATLAB R2022b, version 9.13.0.2105380 (MathWorks). Clinical improvement was measured in the form of relative change from pre-operative baseline to post-operative follow-up under DBS ON (or from post-operative OFF to ON DBS conditions in the case of PD) within the primary outcome assessment of each disease cohort: BFMDRS in DYT, UPDRS-III in PD, Y-BOCS in OCD and YGTSS in TS. Blinding was not relevant in the case of the secondary analyses of existing datasets performed here. To mitigate the risk of observer bias, we tested the explanatory value of our models in hold-out data and performed retrospective and prospective validation experiments (see below).

Of note, most statistical results in the present manuscript involve Spearman’s rank correlations, which do not make assumptions about the underlying distribution. In addition, for these results, scatter plots of individual data points are shown. For analyses in which t-tests were calculated, normality and equality of variances were formally tested (and present for all cases).

DBS electrode localization and E-field modeling

DBS electrodes of all patients were localized based on default settings in an advanced, state-of-the-art processing pipeline as implemented in Lead-DBS software, version 3.0 (https://www.lead-dbs.org)74. MATLAB R2022b, version 9.13.0.2105380, was used to apply this Lead-DBS-based analysis stream. In brief, our approach involved linear co-registrations of post-operative head CT or MRI scans to pre-operative T1-weighted images by means of Advanced Normalization Tools (ANTs; http://stnava.github.io/ANTs/)75. Co-registration results were subsequently corrected for potential intra-operative brain shift via an automatized subcortical refinement module (as implemented in Lead-DBS) but also needed to conform to meticulous visual inspection by two expert users (B.H. and N.L.). This latter step led to manual refinement in cases where aberrations were detected.

All pre-operative acquisitions were used for multi-spectral spatial normalization into ICBM 2009b Nonlinear Asymmetric (‘MNI’) template space76 using the Symmetric Normalization (SyN) approach included in ANTs with the ‘effective: low variance + subcortical refinement’ preset in Lead-DBS. This method had outperformed similar approaches for subcortical normalizations (including STN segmentation) across >10,000 nonlinear warps and different normalization techniques in two independent studies, with precision approaching manual expert segmentation62,63. For all analyses and visualizations of results, atlas definitions of the STN were based on the DBS Intrinsic Template (DISTAL) atlas, version 1.1 (ref. 28), a precise subcortical atlas explicitly created for use within Lead-DBS and based on convergent information from multi-modal MRI, histology and structural connectivity.

To maximize registration accuracy further, normalization warp fields were manually refined using the WarpDrive64 toolbox included in Lead-DBS, version 3.0 (ref. 74), wherever mismatches in the registration were clearly visible and with particular attention to the STN as the anatomical structure in focus. Although high registration precision between a template brain and individual brain anatomy is of utmost importance for accurate reconstructions of electrode localizations, the low contrast of basal ganglia structures on typically applied clinical imaging sequences renders the automated registration between an individual and an atlas STN challenging62,77. WarpDrive is conceived as a dedicated but optional module that allows to manually counteract small misalignments after automated normalization has been performed. Supplementary Fig. 11 shows examples of optimized normalization warp fields after manual WarpDrive refinements (ANTs + WarpDrive) in head-to-head comparison to unrefined, direct results of the automated pipeline (ANTs only). Across the entire discovery cohort, displacements of 0–1 mm were applied in n = 296 electrodes, of 1–2 mm in n = 82 electrodes and of >2 mm in n = 16 electrodes.

Subsequently, electrodes were pre-localized using the phantom-validated Precise and Convenient Electrode Reconstruction for Deep Brain Stimulation (PaCER) algorithm61 in the case of post-operative CT. In the case of post-operative MRI, the trajectory search/contact reconstructions (TRAC/CORE) algorithm78 was implemented instead, both as included in Lead-DBS, version 3.0 (ref. 74). The resulting pre-localizations were visually inspected and manually refined by two experienced users (B.H. and N.L.).

Integrating patient-specific active electrode contacts with corresponding stimulation parameters, the E-field as the gradient distribution of electrical potential in space was simulated in native patient space via an adaptation of the SimBio/FieldTrip pipeline (https://www.mrt.uni-jena.de/simbio/ and http://fieldtriptoolbox.org/)79 as implemented in Lead-DBS, version 3.0 (ref. 74). Using a finite element (FEM) approach, a volume conductor model was created on the basis of a four-compartment mesh60, which involves a realistic three-dimensional model of electrodes (metal and insulating electrode aspects) and surrounding anatomy (gray and white matter). Again, gray matter was defined using the DISTAL atlas, version 1.1 (ref. 28). Finally, electrodes and E-fields were transformed into template space based on the (manually optimized) warp fields priorly determined during normalization of pre-operative MRI acquisitions. These steps allowed for visualization and analysis of electrodes and stimulation fields at the group level using the Lead-Group toolbox70 as well as DBS Sweet Spot and Fiber Filtering Explorers74.

Dysfunction mappings at the subthalamic level

Model definition (Fig. 1a). Our group-level approach intended to delineate and compare the organization of disorder-specific stimulation effects across different neuroanatomical levels, namely (1) that of the subthalamic target site (DBS Sweet Spot Mapping) as well as (2) that of fronto-subthalamic pathways and interconnected cortical sites (DBS Fiber Filtering).

In the first part of our analysis stream, DBS Sweet Spot Mapping18 (Fig. 1a) was performed in each disease cohort separately to identify subthalamic voxels linked to optimal stimulation-related improvements within each respective cardinal dysfunction. For this purpose, information from patient-specific E-fields was integrated with corresponding clinical outcome scores. The E-field denotes the first derivative of the estimated voltage distribution administered to voxels in space, exhibiting greater intensity near active electrode contacts and diminishing rapidly as distance increases. On a voxel-by-voxel basis, the E-field magnitude in each voxel encompassed by the E-field volume was denoted for each patient across the cohort. To account for variability in voxels covered across E-fields within each cohort and to circumvent unrepresentative results biased by too few data points, the region of interest (ROI) was limited to brain voxels encompassed by at least 50% of E-fields exceeding a magnitude threshold of 200 V/m. This E-field magnitude corresponds to a commonly assumed estimate of voltage needed to activate axons45,80,81 and has been repeatedly successfully applied in models of DBS effects on anatomy surrounding the active electrode contact across disorders (for example, in refs. 19,82). Nonetheless, sweet spot modeling and corresponding quantitative validations were repeated for a range of different thresholds (that is, 180 V/m, 200 V/m and 220 V/m) to demonstrate robustness of results (Supplementary Fig. 1). For each considered voxel, this procedure resulted in a vector of E-field magnitude values of the length of the respective patient sample.

Iterating through brain voxels encompassed by the group of thresholded E-fields in template space, Spearman’s rank correlations were calculated between the vector of E-field magnitudes and the vector of relative clinical improvements of all patients. This procedure resulted in a map of positive peak voxels associated with beneficial stimulation effects (sweet spot) as well as negative peak voxels related to detrimental stimulation effects (sour spot). The resulting model can be conceived as an optimal map of where E-fields should ideally stimulate the focal anatomy to maximize treatment success within the respective domain of dysfunction. Of note, these correlation coefficients should not be interpreted as significant results due to the mass-univariate (voxel-wise) nature of our analysis. Instead, they were validated by probing model performance in estimating clinical outcomes in a fivefold CV design (see below).

Estimation of outcomes based on the model

Once optimal sweet spot models had been established in each disease cohort, each model was tested for its explanatory value for clinical outcome variance. To do so, magnitudes of individual E-fields were multiplied with the model in a voxel-wise fashion, and results were averaged across voxels. This led to the attribution of one ‘Sweet Spot Score’ per E-field, and scores were finally averaged across bilateral E-fields to achieve one single Sweet Spot Score per patient. Our modeling approach followed the logic that E-fields in which peaks spatially overlapped highly with the sweet spot (receiving a high positive Sweet Spot Score) would be associated with considerable clinical improvements, whereas E-field peaks markedly encompassing the sour spot (high negative Sweet Spot Score) would result in low or negative estimates. To probe the tenability of this hypothesis, we performed in-sample Spearmanʼs correlations (two-sided tests) between Sweet Spot Scores and empirical clinical outcomes across the cohort. More precisely, the model was calculated on the full discovery cohort, and E-field overlap with it was used to estimate clinical outcomes in each discovery patient. Although representing circular outcomes, these in-sample correlations allowed to compare results (1) across disorders and (2) between sweet spot and sweet streamline findings.

To investigate the generalizability of these results, we further tested whether disease-wise models were robust when subjected to a fivefold CV design, where the sweet spot model was built on a subset of four-fifths of the respective disorder’s discovery cohort (training set) in each of the five folds, and results were used to estimate clinical outcome of the remaining (held-out) fifth of patients (test set). Crucially, because data of remaining patients were not used to inform the model, respectively, the CV strategy was unbiased by circularity. Once estimates for all patients had been derived, the sweet spot model accuracy was finally tested by correlating model-based Sweet Spot Scores with empirical outcomes across the disease cohort. In all analyses, P values were derived based on permuted testing building on 5,000 iterations.

Visualization of subthalamic dysfunction mappings

Disease-wise sweet spots were smoothed by a kernel of two at full width at half maximum using Statistical Parametric Mapping (SPM12) software (https://www.fil.ion.ucl.ac.uk/spm/) to visualize the organizational pattern of subthalamic dysfunction mappings across disorders. Smoothed profiles were projected onto the surface of a three-dimensional model of the STN in ICBM 2009b Nonlinear Asymmetric space derived from the DISTAL atlas, version 1.1 (ref. 28), using Surf Ice software, version 1.0.20211006 (https://www.nitrc.org/projects/surfice). Three-dimensional density plot renderings of sweet spots were further generated by plotting R value magnitudes coded by spheres with different sizes and alpha values in space using Lead-DBS, version 3.0 (ref. 74). Namely, the size and alpha value (transparency) of spheres were weighted by the correlation of modulating the coordinate with clinical outcomes, so that higher correlation results were visualized in the form of larger and less transparent spheres. Two-dimensional axial and coronal views of sweet and sour spots were additionally displayed separately for each disorder using 3D Slicer software, version 5.2.1 (https://www.slicer.org/).

Dysfunction mappings at streamline and cortical levelsModel definition (Fig. 1b)

The second part of our analysis stream followed the intention of deriving the topographical organization of dysfunction mappings (1) at the level of hyperdirect fronto-subthalamic streamlines and (2) that of interconnected sites within the frontal cortex.

To understand the relationship between DBS-based modulation of specific streamlines and a given clinical effect, we, thus, harnessed a previously validated structural connectivity analysis, termed DBS Fiber Filtering65 (Fig. 1b), in an adapted form for implementation in (non-binarized) E-fields19. Structural connectivity was primarily defined by a population-based group connectome derived from multi-shell dMRI-based tractography data of 985 healthy participants acquired within the HCP 1,200 subjects release22. Details on the calculation procedure of this connectome are reported in Li et al.21. In brief, computation of a whole-brain connectome for each of the 985 healthy patients was performed using the Lead-Connectome tool, as provided within the Lead-DBS environment, version 2.0 (ref. 60) (https://www.lead-dbs.org/about/lead-connectome/). Normalization of streamlines into template space involved a multi-spectral warp building on T1-weighted, T2-weighted and diffusion-weighted acquisitions through use of ANTs (using the ‘Effective Low Variance’ preset in Lead-DBS, version 2.0 (ref. 60)). In total, 6,000 streamlines were sampled per individual, which were finally aggregated across all 985 HCP participants to form a collective dataset in MNI space (encompassing a total of 6,000,000 fibers).

Although, by design, normative connectivity is unable to fully account for patient-specific anatomical variability, it is optimally suited for ‘broad lens’ insight into the average human brain at particularly high resolution, as precisely aimed at in the present investigation. Although small inter-individual differences in the topography of human fronto-subcortical interconnections exist, at least a general agreement can be presumed (Supplementary Fig. 12). To assess the general topography of cortico-subthalamic interconnections, we seeded streamlines from five seed points that were manually placed along the dorsal convex shape of the STN in standard space (Supplementary Fig. 12a). We normalized streamlines originating from nine random individual HCP subjects and selected streamlines that connected to each of the seed points in individualized connectomes (Supplementary Fig. 12b). This enabled visual comparability between the topography of STN connections. Although, as expected, the general fronto-rostral topography of interconnections between cortex and STN was revealed in all nine individuals (Supplementary Fig. 12c), the results also show individual variance. In this context, we must mention that only an unknown fraction of these differences should be attributed to true anatomical variance. As demonstrated by several authors using test–retest analysis of brains that were scanned multiple times with dMRI, a substantial fraction of individual differences needs to be attributed to noise and distortions in the data, the choice of MRI machine and tractography algorithms83,84.

Again, the streamline modeling procedure was performed in each disease cohort separately. Per disorder-wise cohort, we first isolated the subset of streamlines from the normative connectome that passed in proximity of at least a minimal number of electrodes. These were characterized in the form of streamlines traversing a rather high E-field magnitude (>0.8 V/mm) close to active contacts in more than 0.5% of E-fields within that cohort. Iterating through this subset of streamlines, one streamline at a time, the stimulation impact per E-field on each streamline was estimated by the peak value among E-field magnitudes collected from points along its passage. This resulted in a ‘streamlines by E-field peaks’ matrix in which each entry denoted the peak impact of each E-field on each streamline.

Second, the entries of the ‘streamlines by E-field peaks‘ matrix were Spearman’s rank correlated with clinical improvements across the disease cohort. Following this procedure, each streamline was tagged by an R value coding for the association strength of its modulation with clinical outcome. The resulting streamline profile can be seen as a model of optimal connectivity for maximal clinical improvements, where streamlines with positive weights would be strongly modulated by E-fields of good performers (sweet streamlines) and such with negative weights by E-fields of poor performers (sour streamlines). As these correlation coefficients relied on a mass-univariate approach, streamline profiles were later validated by probing their capability to estimate clinical improvement in data that had not been used to inform the model (see below).

Estimation of outcomes based on the model

To determine how well disease-wise optimal sweet streamline profiles would perform in estimating clinical improvement in single patients, the peaks of their E-fields were overlapped with the streamline model of optimal electrode connectivity. Specifically, streamlines from the model touched by that E-field were first isolated. Iterating through this subset, the R value of each streamline was subsequently multiplied by the peak E-field magnitude to account for the strength of its modulation by this E-field. Moreover, only the peak 5% of these weighted sweet streamline R values were retained to maximally account for the streamlines most impacted on by that E-field. These top 5% of weighted R values were then summed up, so that each E-field was tagged by the ‘weighted peak 5% of Fiber R scores’. Because each patient within the cohorts considered in the present study had been implanted to the STN at both hemispheres, the Fiber R score was finally averaged across bilateral E-fields to result in one single value per patient.

Following the logic of this procedure, E-field peaks displaying high overlaps with beneficial streamlines would receive high clinical scores, whereas those with no (or merely peripheral) overlap would receive low clinical estimates. A subset of the most relevant sweet streamlines was selected for this validation step using an R value threshold set at the top 1% of the cumulative distribution function of R values of all streamlines. This pre-selection step included such streamlines where high modulation had led to high improvement (with high meaningfulness for the model) while discarding potentially less relevant or noisy correlations.

Model validation within the discovery cohort followed a similar strategy as implemented in the case of Sweet Spot Mapping. Specifically, we tested for an association between spatial overlap of a patient’s E-fields with the model (weighted peak 5% of Fiber R scores) with empirical clinical improvements across patients using Spearmanʼs correlation (two-sided test). In doing so, the ability of streamline models to explain in-sample variance could be scrutinized for comparability of results (1) across disorders and (2) with sweet spots. For in-sample analyses, sweet streamline models were derived based on E-fields of all patients within each disease cohort and validated using each patient of the respective sample (circular analysis). Ultimately, all models were subjected to CVs in a fivefold design to investigate the generalizability of their explanatory value in hold-out data. Again, this approach was implemented by randomly splitting each disorder’s discovery patient cohort into five folds. The disease-specific model was built on four-fifths of patients and validated on the remaining fifth. This strategy was repeated five times, so that model-based estimates (weighted peak 5% of Fiber R scores) were obtained for all patients across all folds. These estimates could then be correlated with empirical clinical outcomes to evaluate model validity in a non-circular fashion. Again, P values were calculated using permutation testing based on 5,000 iterations.

Visualization of cortical dysfunction mappings

To elucidate the topographical organization of interconnected fronto-cortical regions, disease-wise sets of sweet streamlines were first converted to voxelized images (streamline density maps). The resulting maps were then smoothed using an 8-mm Gaussian kernel at full width at half maximum as implemented in SPM12 (https://www.fil.ion.ucl.ac.uk/spm/) and projected onto the cortical surface of the MNI template using Surf Ice software, version 1.0.20211006 (https://www.nitrc.org/projects/surfice). Anatomical correlates of disease-wise cortical sites interconnected with sweet streamlines were then defined based on the Johns Hopkins University (JHU) atlas parcellation68.

Quantification of spatial uncertainty

Furthermore, we aimed to quantify and visualize the degree of spatial uncertainty per streamline within disorder-wise dysfunction mappings at the streamline level. For this purpose, the thickness of each streamline was determined by the −log(P) value, meaning that thicker streamlines would be illustrative of lower P values.

Influence of electrode placement

Subsequently, we intended to scrutinize the relative impact of different model inputs. Besides the choice of a normative connectome, DBS Fiber Filtering results are determined mainly by two major sources of variability across patients—namely, (1) by the precise placement of the stimulation volume and (2) by clinical improvements. In three out of the four disorders of interest in the present study (DYT, PD and TS), stereotactic targeting aims at the same site within the dorsolateral aspect of the STN, whereas the OCD target resides more antero-medially. Thus, the partitioning of dysfunction mappings among various disorders could predominantly be driven by the stimulation impact on clinical outcomes and may not rely solely on differential electrode placement.

To investigate this hypothesis, we implemented a total of three data-driven control analyses. First, plain streamline connections seeding from bilateral stimulation volumes were isolated for each disorder. These comprised the entirety of structural connections activated by a bilateral E-field, irrespective of the importance of their modulation for clinical outcome. Among these, only the subset of streamlines shared across disorders was retained and contrasted to disease-specific sweet streamlines. Four-sample and pairwise tests for equality of proportions (two-sided tests) were performed to compare the degree of overlap between them.

Second, we fit a three-dimensional Gaussian distribution to the standard (second-to-lowest) electrode contacts of all patients with a specific disorder, leading to four blurred volumes within the STN. Streamlines were then seeded from each of these Gaussians as ROIs. The partitioning among the resulting disease-wise connectivity profiles was consequently visually compared to the streamline segregation model that had been achieved using DBS Fiber Filtering. In the latter approach, streamlines connected to empirical stimulation volumes of patients had been weighted by stimulation-related outcomes within the four different domains of dysfunction.

Third, each connection within each respective sweet streamline model per disorder was color-coded by a specificity value, which was calculated by dividing its R value by the average of R values that it received across the three remaining disorders. The streamline segregation result of this approach was finally visually compared to that of the connectivity profile that had been established based on ‘conventional’ color-coding as informed by a streamline’s unbiased R value (resulting from our DBS Fiber Filtering analysis).

Model specificity

Besides validation of each model within the respective disorder that it had been calculated on, we were interested in the degree of specificity of disease-wise models in their ability of explaining clinical outcome variance. To demonstrate specificity, we considered the sweet streamline models for each disorder and overlapped E-fields of patients in all remaining three disorders with the model to predict clinical outcomes in a disorder-by-disorder fashion. Details of this cross-prediction approach were equivalent to those of the CV strategy described above. In brief, weighted peak 5% of Fiber R scores were estimated for each patient based on the degree to which their E-fields encompassed the model, and, finally, Spearmanʼs correlations were performed between these estimates and the empirical clinical outcomes across the cohort of patients to test for model accuracy. In the case of specificity of dysfunction mappings, each of the models would show predictive utility uniquely for clinical improvements within the corresponding outcome measure (good fit between estimates and outcomes) but not for those of other clinical scales (poor fit).

Influence of choice of connectome

Furthermore, we aimed to scrutinize the influence imposed by a particular normative resource chosen to inform connectivity in our DBS Fiber Filtering analyses. To do so, we repeated modeling and model validation procedures using five additional connectomes based on otherwise equivalent model parameters. The first such resource consisted of a normative whole-brain connectome, derived from a multi-shell diffusion-weighted imaging dataset at 760-µm isotropic diffusion acquired in vivo in a single healthy participant over a total duration of 18 scanning hours23 (MGH Single Subject 760 µm Connectome; openly available from https://datadryad.org/stash/dataset/doi:10.5061/dryad.nzs7h44q2). Although generalizability of results derived using this connectome to a larger population is naturally limited by its single-patient origin, it lends itself particularly well for detailed anatomical insight and visualization by dint of its unprecedented imaging resolution.

Second, an axonal pathway atlas26 (Basal Ganglia Pathway Atlas; openly available from https://osf.io/mhd4z/) was implemented, which did not rely on tracking of streamlines based on dMRI data and, thus, circumvents some of the most important drawbacks of dMRI-based tractography (such as the possibility of integrating false-positive connections)24. Instead, streamlines included in this tractogram were manually defined by expert anatomists within an advanced augmented reality (holography) framework. Guided by control points, this technique allows for precise localization and reconstruction of basal ganglia anatomy aided by three-dimensional images created from laser beams. Although the expert-characterized nature of this resource ensures a highly accurate representation of empirically existing (true-positive) connections, it is limited by a higher degree of false-negative streamlines (as the focus in its creation by the expert anatomists lies in accuracy at the expense of exhaustiveness).

Third, we employed a custom-made pathway atlas (DBS Tractography Atlas, version 2.1; openly available from https://github.com/netstim/DBS-Tractography-Atlas.git) informed on previously defined pathway atlases, including the DBS Tractography Atlas, version 1 (ref. 27), and the aforementioned Basal Ganglia Pathway Atlas26. It was completed by additional streamline tracking with focus on a comprehensive description of subthalamic interconnections with multiple cortical and subcortical nodes, leading to a finite set of 6,525,876 streamlines. Its creation specifically followed the intention of representing streamlines that had previously lacked delineation in other resources.

To generate this atlas, a first subset of streamlines connecting the STN to different cortical regions was derived via streamline tracking based on the HCP-1,065 diffusion data, which scanned 1,065 young and healthy adults85. These data are openly available within DSI-Studio (https://brain.labsolver.org/hcp_template.html; fiber orientation maps at 1-mm resolution). Using DSI-Studio (https://dsi-studio.labsolver.org/), the STN as defined within the DISTAL atlas, version 1.1 (ref. 28), was specified as an end region, and 1,500 streamlines were tracked from each of nine cortical Brodmann areas (BAs) as ROIs from the digitized Brodmann atlas86. These comprised BA1/2/3 (primary and secondary somatosensory cortex), BA4 (primary motor cortex), BA6 (supplementary motor area), BA10 (fronto-parietal cortex), BA13 (insular cortex), BA24/32 (cingulate cortex), BA25 (subgenual anterior cingulate cortex) and BA45/47 (frontal gyrus). Two further ROIs of the subcortical region—the substantia nigra pars compacta and pars reticulata—were added from the California Institute of Technology reinforcement learning atlas, version 1.1 (CIT168)69.

To enable cortical branching, the angular threshold was set to a range of 60°–90°. Sampling was thresholded at a minimum length of 5 mm to avoid the inclusion of short streamlines and to prioritize long-range cortico-subthalami