Tractography-Based Modeling Explains Treatment Outcomes in Patients Undergoing Deep Brain Stimulation for Obsessive Compulsive Disorder

Abstract

Background:

Deep brain stimulation is an established and expanding therapy for treatment-refractory obsessive-compulsive disorder (OCD). Li et al. postulated that a white matter circuit providing hyperdirect input from dorsal cingulate and ventrolateral prefrontal regions to the subthalamic nucleus (STN) could be an effective neuromodulatory target.

Methods:

We tested this concept by attempting to retrospectively explain through predictive modeling the ranks of clinical improvement as measured by Yale-Brown Obsessive Compulsive Scale (Y-BOCS) in ten OCD patients who underwent DBS to the ventral ALIC with subsequent programming uninformed by the putative target tract.

Results:

Rank predictions were carried out using the tract model by a team who was completely uninvolved in DBS planning and programming. Predicted Y-BOCS improvement ranks significantly correlated with actual Y-BOCS improvement ranks at 6-months follow-up (r = 0.75, p = 0.013). Predicted score improvements correlated with actual Y-BOCS score improvements (r = 0.72, p = 0.018).

Conclusion:

Here, we provide data in a first-of-its kind report suggesting that normative tractography-based modeling can blindly predict treatment response in DBS for OCD.

Introduction

Neuropsychiatric disorders such as obsessive-compulsive disorder (OCD) involve dysfunction across broad networks regulating cognitive and affective domains.(1) The goal of neuromodulatory therapies such as deep brain stimulation (DBS) is increasingly being viewed as engaging and restoring balance within these dysfunctional networks.(2) Connectomic imaging strategies, especially using white matter tractography, are becoming increasingly reliable methods for identifying these networks to inform surgical planning and DBS programming.(2, 3) Li et al. recently described a white matter connectomic target theorized to serve as a nexus connecting multiple existing striato-capsular grey matter targets in DBS for OCD.(4) By analyzing tractographic data from four separate cohorts (n = 50), they showed that a fiber bundle providing hyperdirect cortical input from dorsal anterior cingulate and ventrolateral prefrontal cortices to the subthalamic nucleus (STN) through a specific subregion of the anterior limb of internal capsule (ALIC) could be used to cross-estimate clinical outcomes between cohorts of patients undergoing ALIC, STN, or nucleus accumbens DBS.(4)

Since the original publication, several reports have accumulated direct support for this “OCD response tract” in equally retrospective fashion in other targets, such as the bed nucleus of stria terminalis (BNST), ventral capsule/ventral striatum (VC/VS) and internal pallidum (GPi)(1–4) (for a review, see(1)). However, prospective or blinded replication has not been attempted, thus far. Here, we tested the OCD response tract model by attempting to blindly explain through predictive modeling ranks of clinical improvement in individuals who underwent DBS for OCD targeting the ventral ALIC (vALIC) based on stimulation-derived anatomical overlap with the published connectomic target.

Methods and Materials

Participants

We included consecutive patients who underwent deep brain stimulation for severe OCD to the vALIC at BCM from Sep 2018 – July 2021 and had at least 6 months of post-operative follow-up. Two patients undergoing bilateral vALIC DBS for OCD during this time period were excluded for not having at least 6 months of follow-up, and one was excluded due to having concomitant bilateral globus pallidus internus (GPi) electrodes implanted to treat comorbid Tourette Syndrome. All patients fulfilled accepted criteria for surgical intervention in OCD, including OCD as the main diagnosis, severity (Yale Brown Obsessive Compulsive Scale(5) (YBOCS)>28), chronicity (>5 years with OCD), drug refractoriness (3 SSRI trials, clomipramine, augmentation), and refractoriness from cognitive-behavioral therapy (minimum of 20 sessions of expert exposure-response prevention). This study was approved by the institutional review board at BCM (IRB number: H-43183).

Surgical Procedure

All patients underwent DBS electrode (n = 9 bilateral Medtronic 3387, n = 1 bilateral Medtronic SenSight 1.5 mm spacing, Minneapolis, MN) implantation using robotic stereotactic guidance.(6) Preoperative high-resolution MRI (3D, T1 weighted with and without contrast) and 3D CT were obtained and uploaded to the ROSA planning station (Zimmer Biomet, Warsaw, Indiana, USA). Two trajectories were planned for each hemisphere (one just anterior to the anterior commissure (AC) and one just posterior to AC) and intraoperative awake testing was used to adjudicate the optimal target, as described in a separate recent publication.(7) Briefly, a monopolar review of all contact pairs is performed by an expert programming psychiatrist (W.K.G) to assess for stimulation-related positive affect responses weighed against acute adverse side effects (i.e., paresthesias, anxiety). We hypothesize that achieving a net positive response may predict future improvement in OCD symptoms. Please see Shofty et al. for further details regarding this testing methodology.(7) Four patients were implanted with Summit RC+S generators (available through BRAIN Initiative funded research trial (UH3 NS1005493, NCT 04281134) Activa RC, 2 Percept, and 1 Activa PC (Medtronic, Minneapolis, MN)).

DBS Programming

The DBS system was activated at the initial psychiatry visit approximately 2 weeks following surgery. Programming visits were generally held every few weeks for the first 3 months and then slowly spaced when a stable programming setup was achieved. All patients were programmed by the same psychiatrist (WKG) and assessed for outcome by the same clinical psychologist (EAS). In brief, our programming psychiatrist uses positive valence behavior (improved mood and energy and increased “approach” behavior such as talkativeness) as a guidepost to strive for with a working hypothesis that these acute behaviors predict future OCD improvement.(7) Induction of adverse effects such as hypomanic behavior (impulsivity, reckless behavior, etc.) or weight gain indicates the need to reduce stimulation or reorient the direction of the stimulation field. We consider potentially effective contacts to be those that produce positive valence responses. We then titrate stimulation intensity (using amplitude and pulse width controls) to achieve a balance between improvements in OCD symptoms and mood (by providing enough stimulation) without inducing hypomania (by providing too much stimulation) based on serial programming sessions by the same psychiatrist trained in DBS programming for psychiatric indications. All surgical planning, stimulation optimization, and outcome assessments were done by the BCM group without input from the Li et al. tract.

DBS Electrode Reconstruction

The BCM group sent de-identified pre-operative MRI and post-operative CT scans as well as programming settings for all 10 patients to the group at Brigham & Women’s Hospital (BWH) which had been completely uninvolved in DBS planning and programming. Any clinical information from visits documenting stimulation titration and Y-BOCS evaluations was withheld in a database at BCM until modeling was complete and final rank predictions were returned from the BWH group. DBS electrode localization was performed using Lead-DBS software (http://www.lead-dbs.org) with the default parameters.(8) Briefly, postoperative CT and preoperative MRI scan were linearly coregistered and then normalized into ICBM 2009b Nonlinear Asymmetric (“MNI”) template space using Advanced Normalization Tools (ANTs, http://stnava.github.io/ANTs/).(9) Brain shift correction was also applied as implemented in Lead-DBS. Electrodes were then automatically pre-localized using the PaCER algorithm and manually refined by an expert user.(10) Volumes of tissue activated (VTAs) were estimated based on each patient’s individual stimulation parameters using a finite element method (FEM) as described in Horn et al.(8)

Statistical Analysis

The described OCD tract is available within the Lead-DBS software.(4) To build the published atlas, structural connectivity of VTAs was calculated based on a normative whole-brain connectome, which was built upon diffusion-weighted imaging data from 985 subjects from the Human Connectome Project (HCP), as had been done in prior reports.(2, 11) Each fiber tract in the connectome received a T-score (from a mass-univariate two-sample t-test) that contrasted the clinical improvements associated with connected vs. unconnected VTAs. A high T-score indicated that the tract was strongly discriminative between better and less responsive VTAs. The published OCD response tract comprises a bundle of fibers that were found to have the highest absolute (positive or negative) T-scores in cross-cohort (n = 50) validation.(4)

Overlapping novel VTAs (e.g., from the present study) with the fibers of the tract will select a subsection of the fibers (each weighted by a T-score), resulting in an average T-score per novel VTA. The resulting scores were termed “Fiber T-scores” in the original study and represent “gain scores” in statistical terms. Based on patients from the original study, these scores were fit to actual percent YBOCS improvement scores using a linear model (LM).

To predict i) ranks of improvements and ii) actual percent YBOCS improvement scores of the present ten unseen patients, Fiber T-scores were calculated using the exact same method. Ranks of these scores were directly used for the primary outcome (rank prediction). Converting the scores to actual percent YBOCS scores using the model (LM) from above led to the secondary outcome (actual prediction of improvements). This process implies a prediction (estimating mappings from datapoints that the model has not seen) based on various definitions,(12, 13) including ones in the context of DBS.(14) For the purpose of this article, we use this definition when speaking about prediction: A model that was calculated on dataset X is used to estimate variance within an unseen dataset Y. The predicted score reduction ranks were then sent back to BCM where they were compared with actual YBOCS score reduction ranks at 6 months follow-up in Spearman rank correlation.

Furthermore, to validate and redemonstrate the OCD response tract in an additional cohort, the tract was re-calculated using the VTAs of the BCM cohort in a similar fashion to the methodology published in Li et al.(4) (Figure 1)

Figure 1. Blinded Tractography-Based Patient Outcome Prediction in DBS for OCD.

A) OCD response tract as published by Li et al. based on tractographic data from four separate cohorts targeting ALIC, STN, STN/ALIC, and NAcc regions. B) As the primary outcome of this study, the ranks of 10 unseen patients from BCM were blindly predicted by the BWH team using overlap of stimulation sites with the tract shown in panel A. C) As a secondary outcome, the actual Y-BOCS scores were correlated with the predicted Y-BOCS improvements. D) The OCD response tract independently recalculated based on the 10 patients of the BCM vALIC cohort.

Results

Across 10 included patients, the mean age was 37 ± 13 years with a mean 18 ± 6-year history of refractory OCD. (Table 1) The median preoperative Y-BOCS score was 38 indicating extreme severity. Eight of 10 patients had comorbid major depressive disorder, and 3 had additional diagnoses including generalized anxiety disorder. All patients had attempted pharmacologic-behavioral intervention with multiple SSRIs, clomipramine, antipsychotics, and exposure-response prevention therapy. Following implantation and stimulation titration, mean stimulation amplitude was 4.76 ± 0.76 mA and 5 ± 0.57 mA on right and left sided contacts, respectively. In the majority of cases, the 2^nd most ventral contact on the electrodes were ultimately chosen to be most effective following programming and titration (for individual stimulation parameters, see Table 1).

Table 1.

Baseline patient characteristics and stimulation parameters at 6 months follow-up.

Participant	Age	Years with OCD	Preop Y-BOCS	OCD Subtype	Comorbid psychiatric disorders	Psychiatric Medications	R active contact (mA)	L active contact (mA)
1	31	18	39	Intrusive Thoughts - Scrupulosity	MDD	Sertraline, Ziprasidone, Clomipramine, Zolpidem, Clonazepam	C+10- (5.5)	C+1- (5.5)
2	31	17	29	Intrusive Thoughts	MDD	Fluvoxamine, Clomipramine	C+9- (4.7)	C+0–1- (5)
3	36	20	34	Just Right	MDD, GAD	Fluvoxamine, Clomipramine	C+9- (5.5)	C+1- (5.3)
4	31	14	34	Contamination	MDD	Escitalopram, Clomipramine, Risperidone	C+9- (ring)* (5)	C+1- (ring)* (5)
5	73	27	39	Contamination	None	Diazepam, Lithium, Hydroxyzine, Clomipramine	C+9- (4.5)	C+1- (4.5)
6	37	25	40	Just Right	TS, MDD, GAD, PTSD	Fluoxetine, Trazodone, Clomipramine	C+9- (4.7)	C+1- (5.2)
7	31	14	40	Intrusive Thoughts - Harm	MDD, GAD, ADHD, ASD	Fluvoxamine, Clonazepam, Buspirone, Clomipramine	C+10- (4)	1+0- (4.5)
8	31	6	40	Contamination	MDD	Fluvoxamine, Risperidone, Vortioxetine	C+9- (6)	C+1- (6)
9	40	24	38	Contamination	BPII	Paroxetine, Lurasidone, Clonazepam, Clomipramine	C+9- (4.2)	C+1- (4)
10	28	18	32	Contamination	MDD	Sertraline, Fluvoxamine, Lorazepam	C+9- (3.5)	C+0- (5)

Abbreviations: MDD = major depressive disorder, GAD = generalized anxiety disorder, TS = Tourette syndrome, PTSD = post-traumatic stress disorder, ADHD = attention-deficit/hyperactivity disorder, ASD = autism spectrum disorder, BPII = bipolar disorder type II

^*This patient received bilateral SenSight 1.5 mm spaced leads (Medtronic, Minneapolis, MN).

Actual Y-BOCS percent improvement ranged from 2.5–48% at 6 months follow-up (mean ± SD = 22 ± 17%, Table 2). The patients were ranked based on their percent Y-BOCS improvement. Fiberscore values coding for improvement predictions were ranked (Table 2). Ranks of empirical and predicted Y-BOCS scores were compared and showed significant agreement in Spearman rank correlation (r = 0.75, r² = 0.56, p = 0.013, 95% CI [0.20–1.28]).

Table 2. Actual and predicted Y-BOCS score improvement at 6 months follow-up.

The median split is color coded by highest (white) and lowest (grey) responding patients.

Participant	Preop Y-BOCS	6-month Y-BOCS	Predicted Y-BOCS improvement	Actual Y-BOCS improvement	Predicted rank improvement	Actual rank improvement
1	39	20	0.486	0.487	1	1
2	29	22	0.434	0.241	2	5
3	34	25	0.396	0.264	3	4
4	34	24	0.373	0.294	4	3
5	39	21	0.348	0.461	5	2
6	40	39	0.322	0.025	6	10
7	40	32	0.312	0.2	7	6
8	40	34	0.303	0.15	8	7
9	38	37	0.299	0.026	9	9
10	32	31	0.297	0.031	10	8

As a post-hoc secondary outcome, actual Y-BOCS scores from the BCM cohort were also linearly fit to empirical Y-BOCS improvement scores based on the model trained on the N = 50 patients published in Li et al. Empirical and predicted Y-BOCS scores were also significantly correlated (r = 0.72, r² = 0.52, p = 0.018, 95% CI [0.42–3.35]). Additionally, the median split of treatment response was found to be accurately predicted by the tractography-based model. (Table 2) Finally, the re-calculated tract based on the BCM cohort was nearly identical to the previously published OCD tract model. (Figure 1)

Discussion

Our results validate the utility of the OCD response tract proposed by Li et al. in a blinded fashion. This data-driven white matter bundle, which links cortical, striatal, and capsular hubs, was capable of rank-predicting the relative outcomes of patients undergoing DBS to the vALIC regions based on a tractographic atlas generated from patients undergoing DBS to sites including ALIC, STN, and NAcc.(4) Our findings echo the increasing understanding that optimal targeting requires engagement of this network and that tractography likely has a role to play in identifying this structural target.(1, 15) Furthermore, continued optimization of this neuroanatomical model may lead to improved efficacy in DBS for OCD trials as work is done to make this therapy available to more patients.(16)

One of the key strengths of our approach is the blinded prediction between two centers mutually uninvolved in each other’s role in this project, which to our knowledge has not been carried out in this fashion. The concept of postulating published treatment target models based on retrospective data such as the one by Li et al. is a critical first step. Before carrying out prospective validation studies directly, however, an intermediate step could be the concept followed here: One DBS center provides the necessary ingredients to localize the stimulation sites (neuroimaging data and stimulation parameters) but deliberately withholds information about clinical outcomes. Based on this information, the second center then makes blinded (rank) predictions on clinical outcomes, which can be compared to the empirical data by the first center. (Figure 2) These analyses can be readily carried out using public open-source software such as Lead-DBS or others.(8)

Figure 2. Schematic for Collaboration Between Centers in DBS for OCD Prediction Studies.

One center provides imaging, stimulation setting, and clinical outcome data. A second center performs electrode reconstruction-based score predictions using statistical modeling of the overlap between volume of activated tissue and tractographic targets. This method can provide external validation of a given tractographic atlas.

There are limitations of our work. We report a retrospective analysis incorporating a modest sample size of patients who underwent stimulation optimization concurrent with continuous medication and psychotherapy. While this is standard to all DBS for OCD studies, it could have resulted in confounding effects due to the individualization of therapy. Further, the relatively short follow-up time point assessed may be a limitation in the reporting of ultimate treatment effects as has been documented in DBS for OCD studies.(17) At the time we performed these analyses, 6 of the 10 patients had 12-month outcomes available. Two more reached this time point during manuscript review. Average YBOCS reduction improved from 22% to 39%, and responder rate increased from 2/8 (25%) to 6/8 (75%) between 6- and 12-month follow-up. Interestingly, we found that rank scores at 6 and 12 months were significantly correlated (p=0.014), meaning that longer term improvement (at 12 months) was correlated with shorter term improvement (at 6 months). This finding further supports our decision to use 6-month outcomes, as rank improvement at 6 months still seems valid, and power at this time point is higher with greater N. It would be useful in future studies to validate the model’s predictive ability in larger study with longer follow-up. Additionally, tractography-based modeling predicted a relatively smaller dynamic range than the actual response, perhaps due to the constraints of fitting the outcome variable to a linear regression model. Additionally, the practical real-world challenges of managing OCD with the associated life stressors and co-morbidities could affect the Y-BOCS score at any given time point, even with engagement of optimal fiber tracts and connected networks.

In this first of-its-kind study, we have demonstrated the ability to explain the variance in clinical response to DBS for OCD using normative connectomic modeling. Our findings further validate a white-matter circuit linking prefrontal regions to the STN via a tract in the ALIC that, when effectively modulated, can lead to clinical improvement in severe OCD.