Toward a unified connectomic target for deep brain stimulation in obsessive-compulsive disorder

Multiple surgical targets have been proposed for treating obsessive-compulsive disorder (OCD) with deep brain stimulation (DBS). However, different targets may modulate the same neural network responsible for clinical improvement. Here we analyzed data from four cohorts of OCD patients (N = 50) that underwent DBS to the anterior limb of the internal capsule (ALIC), the nucleus accumbens (NAcc) or the subthalamic nucleus (STN). Fiber tracts that were predominantly connected to electrodes in good or poor DBS responders were isolated from a normative structural connectome and assigned a predictive value. Strikingly, the same fiber bundle was related to treatment response when independently analyzing two large training cohorts that targeted either ALIC or STN. This discriminative tract is a subsection of the ALIC and connects frontal regions (such as the dorsal anterior cingulate, dACC, and ventral prefrontal, vlPFC, cortices to the STN). When informing the tract solely based on one cohort (e.g. ALIC), clinical improvements in the other (e.g. STN) could be significantly predicted, and vice versa. Finally, clinical improvements of eight patients from a third center with electrodes in the NAcc and six patients from a fourth center in which electrodes had been implanted in both STN and ALIC were significantly predicted based on this novel tract-based DBS target. Results suggest a functional role of a limbic hyperdirect pathway that projects from dACC and vlPFC to anteriomedial STN. Obsessive-compulsive symptoms seem to be tractable by modulating the specific bundle isolated here. Our results show that connectivity-derived improvement models can inform clinical improvement across DBS targets, surgeons and centers. The identified tract is now three-dimensionally defined in stereotactic standard space and will be made openly available.

Obsessive-compulsive disorder is a debilitating disease with a life-time prevalence of around 2.3% 1 . Treatment of severe cases by deep brain stimulation (DBS) to the ALIC has been approved by the FDA (Humanitarian Device Exemption) in 2009 2 . A variety of other targets have been proposed, however, including the STN 3,4 , the NAcc 5-7 , ventral capsule/ventral striatum (VC/VS) 8 , inferior thalamic peduncle (ITP) 9,10 , bed nucleus of the stria terminalis (BNST) 11 , anteromedial globus pallidus interna (amGPi) 12 , superolateral branch of the medial forebrain bundle (slMFB) 13 and medial dorsal and ventral anterior nuclei of the thalamus (MD/vANT) 14 (for an overview of targets see 15 ). A recent prospective clinical trial implanted four electrodes per patient with one pair in the STN and one in the ALIC 16 .
In parallel, DBS has experienced a conceptual paradigm-shift away from focal stimulation of specific brain nuclei (such as the subthalamic nucleus or globus pallidus in Parkinson's Disease; PD) toward modulating distributed brain networks (such as the motor basal-ganglia cortical cerebellar loop in PD) 13,[17][18][19][20][21][22][23] . While this concept of modulating white-matter tracts (instead of grey matter nuclei) is certainly not new (and anterior capsulotomy was introduced already in the ~1950ies by Talairach and Leksell 24 ), novel MRI technologies such as diffusion-weighted imaging based tractography have been used in functional neurosurgery in order to more deliberately target white-matter tracts 21 . In this translational development, the Coenen and Mayberg groups should be explicitly mentioned, among others, for pioneering and rapidly translating the use of tractography to functional surgery since around 2009 13,[18][19][20]25,26 .
Thus, it could be possible that, of the multiple targets proposed, some -or most -may in fact modulate the same brain network to alleviate symptoms. Such a concept has been proposed in the past by Schlaepfer and colleagues for the case of treatment-refractory depression 27 . According to their concept, the supero-lateral branch of the medial forebrain bundle may connect most if not all surgical targets that were proposed for treatment of depression (e.g. subgenual cortex, ALIC, NAcc, habenula). Thus, in theory, the tract itself could be a surgical targetand could be modulated in a similar way when targeting various points along its anatomical course. Accordingly, already, Coenen and colleagues surgically implanted electrodes in two OCD patients, targeting a tract instead of a localized target 13 .
The tract connected the ventral tegmental area and the prefrontal cortex and authors referred to it as the superolateral branch of the medial forebrain bundle.
Other invasive therapies, such as cingulotomy and capsulotomy also aimed at disrupting connectivity from frontal regions by lesioning white matter bundles 28 . It could recently be shown that such tract-or network-based concepts may be used to predict clinical improvements across DBS centers and surgeons for the case of Parkinson's Disease 29 .
Based on modern neuroimaging methods and high-resolution connectomic datasets, connectivity of DBS electrodes to specific cortical regions was associated with stronger therapeutic effects in various diseases treated with this surgical procedure [29][30][31][32][33] .
For the case of OCD, Baldermann and colleagues recently demonstrated that structural connectivity from DBS electrodes to medial and lateral prefrontal cortices were associated with stronger symptom alleviation 30 . Crucially, they were also able to identify a specific subsection of the ALIC that was highly associated with symptom improvements after one year of DBS. Of note, connectivity to this fiber tract was able to predict ~40 % of the variance in clinical outcome in out-of-sample data. The bundle was described to connect to both the medial dorsal nucleus of the thalamus and to the anterior part of the STN (which both have received substantial attention in the context of OCD). The STN itself is a prominent target for DBS of various diseases including PD 34 , dystonia 35 , OCD 36 and Tourette's Syndrome 37 . The small nucleus receives wide-spread direct afferents from most parts of the prefrontal cortex and is involved in motor, associative and limbic processing 38 . Due to these spatially organized cortico-subthalamic projections, the nucleus has various functional zones that largely follow the organization of the frontal cortex, i.e. the sensorimotor parts of the STN are situated posterior and followed by pre-/oculomotor-, associative and limbic domains in anteromedial direction.
Consequently, the anterior (associative/limbic) parts of the STN have been targeted by DBS for OCD 36 ; these same anterior subregions were exclusively connected to the tract-target identified by Baldermann et al. in ALIC-DBS patients 30 . Following up on this, our present study aimed at testing whether the same tract could be associated with good clinical outcome in a cohort treated with STN-DBS. We retrospectively analyzed two cohorts of DBS patients that were treated with either STN-DBS or ALIC-DBS in order to test our hypothesis, that the same tract could potentially predict clinical improvement in STN-DBS as well as ALIC-DBS. In this attempt, we identified a common tract that already became apparent when analyzing either cohort alone. After calculating the tract exclusively based on data of one cohort (e.g. ALIC), we cross-predicted outcome in the other cohort (e.g. STN), and vice versa. We then tested predictive utility of this tract in two additional cohorts from a third and fourth center. Finally, we set the resulting tract target into the larger context of OCD-DBS literature and tested, whether it could be used to explain outcomes of reported clinical studies with different surgical targets.

Results
Two cohorts (Cologne; ALIC target; N = 22; and Grenoble; STN target; N = 14, two electrodes in each patient) formed a training and cross-validation sample in which the tract target was identified and validated. Each of the two cohorts were first analyzed independently, then used to cross-predict outcome in patients from the other one. The main part of our analyses focuses on these two cohorts. As further validation of results, two additional test-cohorts were included (Madrid: two electrodes in each patient targeting Nucleus Accumbens (NAcc); London: four electrodes in each patient targeting both ALIC and STN).
Patients in all cohorts were of similar age with a similar Y-BOCS score at baseline and comparable Y-BOCS improvement scores (Table 1). In the first test cohort (Madrid; NAcc target; N = 8), improvement scores were taken after activating each of the four electrode contact pairs for 3 months, respectively (following the clinical protocol as described in 39 ).
This resulted in a total of 32 data points. In the second test cohort (London; both ALIC and STN target; N = 6, four electrodes in each patient), stimulation parameters resulted from an optimized phase following parameter optimization.  Connectivity analysis results seeding from electrodes of the two training cohorts (Cologne and Grenoble) based on the N = 985 HCP normative connectome are shown in Figure 2.
The overall connectivity of electrodes to other areas in the brain (without weighing for clinical improvement) was strikingly different between the two cohorts ( Figure 2, top row). This is hardly surprising since it mainly reflects the overall structural connectivity profiles of the two DBS targets and the STN as a widely connected basal ganglia entry point and the ALIC as a white matter structure are differently connected in the brain. However, when tracts were weighted by their ability to discriminate between good and poor responders (using the Fiber T-score method described below), a bilateral positively discriminative tract to the medial prefrontal cortex emerged in each cohort even when cohorts were analyzed independently ( Figure 2, middle row). The degree of lead connectivity to this tract correlated with clinical improvement (R = 0.63 at p < 0.001 in the ALIC cohort and R = 0.77 at p < 0.001 in the STN cohort; Figure 2, bottom row). Of note, these correlations are somewhat circular and meant to describe the degree of how well discriminative tracts could explain the same sample of patients on which they had been built. More interestingly, in the next step, the tract was calculated exclusively on data from the STN cohort and then used to explain outcome in the ALIC cohort (R = 0.50 at p = 0.009) and vice versa (R = 0.49 at p = 0.041) ( Figure 3).
Crucially, some VTAs of the ALIC cohort resided entirely below the identified tract and thus received a Fiber T-score of (near) zero (also see blue example patient in Figure 3, bottom right). The same holds true when either calculating the tract based on the STN cohort ( Figure   3) or the ALIC cohort itself ( Figure 2). To further investigate this matter, two-sample t-tests between improvements of patients with near zero scores (Fiber T-scores below 50) and the remaining patients with VTAs covering the tract well (scores above 50) were calculated. This showed that electrodes that reached the tract well resulted in significantly better clinical improvement (T = 6.0 at p < 10 -5 when the tract was calculated on the ALIC cohort, Figure   2, and T = 3.7 at p < 0.005 when it was calculated on the STN cohort, Figure 3). In the next step, the analysis was performed on the two cohorts combined. Again, the same tract emerged, now even more clearly (Figure 4, top). Thus, bundles were highlighted, that were predominantly connected with VTAs of patients from both cohorts with good or poor improvement, respectively. The resulting positive discriminative tract traversed slightly dorsal to the group of electrodes of the ALIC-cohort and coursed centrally or slightly ventral to the electrodes of the STN cohort. This tract was then used to predict outcome in two completely independent test-cohorts of patients that underwent surgery in a third and fourth center (Madrid & London; Figure 4, bottom). While the surgical target of the Madrid cohort was the Nucleus Accumbens, electrode placement is comparable to the ALIC / Cologne cohort ( Figure 1). Here, improvements were taken for each contact pair during a three-month interval, leading to 32 data points ( Figure  The tract target identified here may potentially "unify" some aspects of the STN and ALIC/NAcc targets for OCD. Thus, in a final analysis, we aimed at setting it into context with other DBS targets that were used in OCD-DBS, before. To do so, we converted literaturebased targets into MNI space 44 and set them into relation with the tract target (see Figure   5, Table 2  Given the potential clinical importance of the identified tract, we characterized its anatomical properties using additional views relative to anatomical landmarks (Figures 6 & S3) as well as in comparison to anatomical dissection results ( Figure S4). Anatomically, the tract is a subpart of the well-characterized ALIC that connects areas of the prefrontal cortex with the subthalamic nucleus and MD nucleus of the thalamus 45,46 . Anatomical validity of the isolated tract was discussed with four anatomists (see acknowledgement section). In the motor domain, the "hyperdirect pathway", i.e., a direct connection from the frontal cortex to the subthalamic nucleus, has been well established 47,48 , functionally, but the STN is known to receive widespread and direct input from most areas of the prefrontal cortex 45

Discussion
We analyzed data from four cohorts of OCD patients with different DBS targets using a connectomic approach. Strikingly, the same optimal tract target emerged when separately Here, we extend anatomical definition of the same circuit and show that it emerges based on data from multiple stimulation sites. The subthalamic nucleus receives afferents from a large portion of the prefrontal cortex by hyperdirect pathways that are known to traverse within the internal capsule 45,57 . Recently, such an input to the STN from prefrontal regions was electrophysiologically described in humans 58 . In rodents, lesions to such a "limbic hyperdirect pathway" led to diminished discriminative accuracy and increased perseveration 59 . One classical cortical region which was described as an origin of limbic hyperdirect input is the dACC 22 Because of the high amount of false-positive connections in diffusion MRI based connectomes 43,62 , we repeated the analysis using an atlas of predefined anatomical tracts 22 . Here, the hyperdirect pathway connecting dACC to the STN was isolated as the only of five bundles in the ALIC that were included in the atlas (Figures S1-2). Thus, hyperdirect cortical input from dACC to STN could be an anatomical and functional substrate of the identified bundle. In this context, it is crucial to note that the atlas by nature cannot represent each and every white-matter bundle that exists in the ALIC / STN region and shows "gaps" in between the included bundles ( Figures S1-2). Thus, while normative connectomes include a large number of false-positive fibers, the atlas may instead be prone to false-negative connections, since some tracts are simply not included. For instance, it is known that the STN receives direct input from other areas of the prefrontal cortex such as the ventrolateral prefrontal cortex 63  in standard stereotactic space.
A highly similar pathway that already served as a tract-target in a small case-series of OCD patients 13 also traversed within the ALIC but has instead been referred to as the superolateral branch of the medial forebrain bundle (sl-MFB) 49 . The original anatomical definition of the medial forebrain bundle (which was first defined in the rat) suggests a more ventral route connecting the ventral tegmental area to the olfactory cortex while bypassing the red nucleus medially 46 . In other words, the anatomical definition of the medial forebrain bundle does not traverse within the ALIC. This mismatch between the surgical target (sl-MFB) and anatomical literature (mfb) has recently been confirmed by the original authors of the surgical target and they now additionally referred to it with "vtaPP" (for ventral tegmental area projection pathway) 64 . This potentially misleading nomenclature of the surgical sl-MFB target has suggested that results in two previous OCD studies would be conflicting, while anatomically, their results agreed. Both studies favored a similarly defined tract within the ALIC, which was referred to as sl-MFB in one study 65 and as anterior thalamic radiation in the second 30 . To readers, this suggested conflicting results while they were actually confirmatory (based on the location of both tracts within the ALIC). Thus, we welcome the recent steps taken to move away from calling the surgical target sl-MFB toward calling it "vta-PP" 64 . This said, our interpretation of the identified tract differs. Our findings reveal a tract connecting frontal areas with the STN (cf. Figure S3 C & results from the basal ganglia pathway atlas, Figure S1-2). Thus, we attribute the tract to afferents of the STN (limbic hyperdirect pathway) as opposed to efferents of the ventral tegmental area implied by the term "vtaPP" 64 .
This interpretation could be further supported by combined analyses of dMRI and tracing methods in nonhuman primates as well as human subjects, which were used to segregate prefrontal fibers passing through the internal capsule 66 . Fibers that originated from ventrolateral prefrontal cortices (areas 45 and 47) were shown to terminate in the medial part of the STN and the MD nucleus of the thalamus -precisely corresponding to the tract described here. Alternatively -or additionally -the hyperdirect pathway projecting from dACC to the STN may be functionally involved in mediating treatment outcome. As mentioned, a strong additional hint for this latter hypothesis is that lesions to the dACC itself have beneficiary effects on OCD 61 .

Toward symptom-specific circuitopathies
Based on our results, two testable hypotheses with implications above and beyond OCD could be proposed. First, different surgical targets may reduce the same symptoms equally well -potentially by modulating the same tract or network. Second, in addition, they may modulate not only one (shared) network but other networks that are not shared, resulting in different changes across other behavioral domains. This can be seen by widely different connectivity profiles of the targets (Figure 2 top row) and differential effects of STN vs. ALIC stimulation on depressive / cognitive functions described by Tyagi et al. 16 . Thus, one may speculate that networks are symptom-specific (and not disease-specific). When modulated, these networks or tracts seem to not ameliorate a specific disease but rather specific symptoms present in the disease.
Potentially, DBS surgery in the (distant) future could involve detailed preoperative phenotyping to establish a broad patient-specific symptom score. Based on databases of clinical improvements along affected symptom axes, a mix of networks that would be modulated to alleviate each patient's specific symptom profile could be identified. Such concepts are still mostly speculation but may be investigated by future studies. This said, we must emphasize that the present study investigated data on a group level and utilized connectivity from individuals without OCD. As mentioned by others in the very context, we could not agree more that surgical decision making for DBS should not be based on such aggregated normative data, alone 49 . Further studies are required to determine whether individual patient connectivity or generic connectome data (or both) could assist with optimizations in surgical targeting or DBS programming by determining crossing sites of symptom networks for specific patients.

Limitations
Several limitations apply for the current work. First and foremost, the retrospective character of the study is not ideal to compare and study effects of clinical outcome which is why we kept clinical information to a minimum and instead referred to the underlying clinical studies.
Second, we used normative connectome data instead of patient-specific diffusion-weighted MRI data (which is unavailable for most of the patients included). This poses dramatic limitations since such data cannot be representative of patient-specific anatomical variations.
Still, we argue that some aspects about general pathophysiological mechanisms may still be investigated using normative data. Use of normative connectomes has been introduced in other clinical domains where patient-specific MRI data was unavailable, such as stroke [76][77][78] or transcranial magnetic stimulation 79 . In DBS, the technique has been applied before and -as in the present study -has led to models that could be used to predict improvement in out-of-sample data 29,30,80 . In addition to the practical advantage of being applicable in cases where patient-specific data is lacking, normative data also has the theoretical advantage of much better data quality. In the present case, a connectome dataset was derived from a high N of 985 subjects scanned under research conditions by one of the best methodological groups in the world 81 . It may be logistically challenging to acquire data of such quality in a clinical routine setting (e.g. pre-operatively) in individual patients but will be feasible in specialized centers. Tractography based DBS targets pointed to coordinates that were sometimes >2 mm when repeating analyses on test-retest scans of the same subject 82 . However, patient-specific connectivity can never be reconstructed when using normative connectomes. Thus, normative connectomes will likely not embody the final solution to the connectomic surgery framework and will be challenged by advances in MRI technology and algorithm developments. Potentially, as a step in-between, using combined information from normative and patient-specific connectomes could embody a promising strategy that should be explored, in the future.
Inaccuracies in lead localization may result from the approach of warping electrodes into common space as done here. To minimize this issue, we used a modern neuroimaging pipeline that has been scientifically validated in numerous studies and involved advanced concepts such as brain shift correction 83 , multispectral normalization, subcortical refinement 83 and phantom-validated electrode localizations 84 . The normalization strategy that was applied was found to automatically segment the STN as precisely as manual expert segmentations 85  anatomists (see Acknowledgements). Fourth, we replicated findings based on an atlas that is based on predefined anatomical tracts (see supplementary material). The tract described in the present study matches results from this atlas ( Figure S1-2) and dissection studies ( Figure S4). However, the potential that the tract represents a false positive result may not be completely ruled out given the fundamental limitations of dMRI-based tractography 43,62 .

Conclusions
Four main conclusions may be drawn from the present study. First, we show that the overall connectivity profiles of STN-and ALIC-DBS electrodes project to largely different areas in the brain. Second, data in each target alone singled out the same fiber tract that was associated with long-term improvement of OCD symptoms when modulated either at the level of the STN or the ALIC. Third, we demonstrated that it is possible to cross-predict clinical improvement of OCD patients across DBS target sites (ALIC / STN) and centers (Cologne / Grenoble). Finally, we confirm results by predicting outcome in two additional cohorts from different centers (Madrid / London) and set results into context of published reports.

Patient Cohorts and Imaging
Fifty OCD patients from four centers were retrospectively enrolled in this study, among them twenty-two patients from University Grenoble and London cohorts. In case of the London cohort, this followed a four-step clinical trial (2x3 months blinded stimulation at one target followed by 6 months of stimulation at both targets, the last three months using clinically optimized parameters. For details see 16 ).
In the Madrid cohort, each of the four contact pairs were activated for three months, with a one month wash-out period between trials and a three month sham period. In our analysis, this leads to 32 data points (i.e. stimulation-based outcomes). Patients' demographic details are provided in Table 1

DBS Lead Localization and VTA Estimation
DBS electrodes were localized using Lead-DBS software (http://www.lead-dbs.org) as described in 86

Connectivity Analysis
Structural connectivity between VTAs and all other brain areas was calculated based on a normative connectome as similarly done in previous work 29,30,41,44,80,83 . Specifically, a wholebrain connectome based on state-of-the-art multi-shell diffusion-weighted imaging data from 985 subjects of the Human Connectome Project (HCP) 1200 subjects data release 81 was calculated in each patient using Lead-Connectome. Whole-brain fiber tracts were then normalized into standard space using a multispectral warp based on T1-weighted, T2weighted, and diffusion-weighted acquisitions using ANTs (using the same "Effective Low Variance" preset implemented in Lead-DBS). In each subject, a total of 6,000 fibers were sampled and aggregated to a joint dataset in standard space, resulting in a set of 6,000,000 fibers across 985 HCP subjects. For each of these tracts, a "Fiber T-score" was assigned by associating the fiber tract's connectivity to VTAs across patients with clinical outcome ( Figure   7). Specifically, (mass-univariate) two-sample t-tests between clinical outcomes in connected and unconnected VTAs were performed for all 6,000,000 tracts. Needless to say, these T-scores were not meant to result in significant results but instead formed a model that could be used for out-of-sample predictions in other DBS cohorts. T-values from these tests and could be positive or negative (since two-sided t-tests were performed). A high absolute T-value meant that the fiber was strongly discriminative or predictive for clinical outcome. For instance, a tract that was connected exclusively to VTAs in good responders (and not to VTAs of poor responders) would receive a high positive score. In return, a patient would most likely show more pronounced clinical benefit, if her/his VTA was strongly connected to many fibers with high positive T-values but not to many with negative scores.
This analysis made it possible to assign aggregated "Fiber T-scores" to each VTA in subsequent prediction analyses.
To account for the fact that larger VTAs would potentially automatically receive higher Fiber T-scores, these were divided by the stimulation amplitude throughout the manuscript. Finally, Monte-Carlo random permutations (× 1000) were conducted to obtain p-values, except for two-sample t-tests. This procedure is free from assumptions about the distributions (e.g. Student-T for R-values) which are typically violated in small sample sizes 91 . Scatterplots are visualized with 95% confidence bounds (gray or light-red areas).

Data and code availability
The DBS MRI datasets generated during and analyzed during the current study are not publicly available due to data privacy regulations of patient data but are available from the corresponding author on reasonable request. The resulting tract atlas and code used to analyze the dataset is openly available within Lead-DBS /-Connectome software (https://github.com/leaddbs/leaddbs).    Literature based DBS target analysis (cf.  Figure 5) For each publication (Table 2), pre-and postoperative average Y-BOCS scores were extracted to calculate the average percent change in Y-BOCS (difference between avg. preop-and postop-scores divided by avg. preop scores). As in all "meta-analysis" situations, due to differences in study design and reported data, we had to decide which exact values to use in some cases (below). identified here. To do so, stereotactic coordinates were converted to MNI space using a novel probabilistic method 44 . A sphere of radius 3 mm was introduced at this site and heavily smoothed with a sigma of 6 mm (to allow for a weighted/distance measure with the tract).
Weighted overlap values between these smoothed volumes and the T-values of the tract-target were multiplied to derive a literature-based Fiber-T-score which was correlated with average improvement scores.