Predictors of motor outcome with pallidal stimulation for Parkinson’s disease from the CSP468 cohort

Abstract

Deep brain stimulation (DBS) of the globus pallidus interna (GPi) is an effective treatment for medication‑refractory Parkinson’s disease (PD), though outcomes vary widely. The influence of lead location relative to the GPi therapeutic sweetspot has not previously been examined within a multivariable regression framework. This study evaluated the predictive utility of volume-of-tissue-activated (VTA)–sweetspot overlap on postoperative UPDRS‑III outcomes in the CSP #468 cohort receiving bilateral GPi‑DBS, alongside demographic/clinical predictors, using two independently derived sweetspot models. CSP #468 was a multicenter randomized clinical trial with blinded 6‑month outcomes and robust collection of associated features of PD. A prior publication generated a cross validated 6‑month GPi-sweetspot from this dataset. An independent single‑surgeon cohort was used to construct a separate sweetspot uninfluenced by CSP outcomes. Multivariable modeling was completed using backward‑selection linear regression with Bonferroni correction. Models were checked for assumptions, data leakage, and overfitting. Each sweetspot-multivariable model’s ability to predict outcomes in the opposite cohort was assessed. Both sweetspots localized to the primary motor GPi. The independent sweetspot multivariable model included VTA‑sweetspot overlap and levodopa response (R²_Adj = 0.19) and remained robust under leave-one-out cross validation. The CSP sweetspot multivariable model also predicted outcomes in the independent cohort (p = 0.004), demonstrating meaningful external validity.

Introduction

Deep brain stimulation (DBS) of the Subthalamic Nucleus (STN) and Globus Pallidus interna (GPi) has been FDA approved for the treatment of Parkinson’s disease (PD) since 2002¹. Despite well-established efficacy, considerable interindividual variability in postoperative motor outcomes remains. Identifying clinical, demographic, and surgical factors that meaningfully account for this variability is critical for optimizing patient selection, surgical planning and postoperative management.

Among proposed predictors, levodopa responsiveness (LDR) has consistently been associated with short-term DBS outcomes^2–6, though its predictive value over longer time points has been challenged⁷. Other factors—including age^4,6,8–12, age at disease onset^4,6,8–13, cognition, mood^2,7,11, target selection^14,15, intraoperative variables^9,16,17, baseline axial symptom burden¹⁸, and lead location^19–22—have also been examined, predominantly within the STN DBS literature. In contrast, relatively few studies have systematically evaluated predictors of motor outcome following GPi-DBS. Randomized controlled trials have demonstrated comparable motor efficacy between GPi and STN-DBS, with GPi-DBS often favored for patients with cognitive vulnerability or severe and brittle dyskinesias^23–25. Existing GPi focused predictive studies have largely reaffirmed the importance of LDR and baseline tremor burden^18,26 or explored advanced imaging metrics, including resting state functional connectivity (rsfMRI)²⁷ and dopamine transporter availability²⁸. However, these studies were generally limited by single center designs, modest sample sizes, and the absence of anatomic lead location measures within multivariable frameworks.

The present analysis sought to address these gaps by evaluating the clinical relevance of GPi-sweetspot targeting on 6-month postoperative off medication, on stimulation motor outcomes, while accounting for a broad range of preoperative demographic and clinical variables. Two independent GPi DBS datasets were leveraged for this purpose. The primary dataset was derived from the randomized, blinded, multi-institutional CSP #468 trial, which enrolled patients with PD treated with either STN or GPi-DBS between 2002 and 2005 and includes prospectively collected clinical variables alongside a previously validated GPi-sweetspot model²⁹.

To mitigate circularity inherent in applying a sweetspot model to the same cohort from which it was derived, a second, independent single center GPi-DBS cohort (2014–2024; follow up 3–29 months) was used to generate a separate sweetspot model. The combination of a large, multicenter randomized dataset and an independently derived anatomic target provided a robust framework to evaluate the relative contributions of lead location and clinical factors to GPi-DBS motor outcomes.

Results

CSP Cohort Characteristics

The patients in the CSP #468 test cohort were predominantly male, ranging in age of onset from 40 to 81 with an average duration of PD of 11.2 years (Table 1). When using an age of onset of less than 50 years old to define young-onset PD (YOPD), 47.7% of the patients met this criterion. The average LDR was 48.0%, ranging from -21.0 to 90.0%. Negative LDR were seen in a small proportion of the study population, attributable to beginning of dose worsening, end-of-dose rebound, delayed or partial response, or dose failure from erratic levodopa absorption³⁰. Additionally, dyskinesias, hypotension, and nausea induced by levodopa intake can interfere with the exam. Average 6-month postoperative percent change in UPDRS-III was 29.5%, with a range of -44 to 82% (standard deviation [SD] 26.4%). The mean sweetspot score from the CSP GPi-Sweetspot (through leave-one-out cross validation [LOOCV]) for this cohort was 0.84 with a range of 0 to 2.6.

Table 1.

Demographic, Baseline, and Outcome Measures of the CSP #468 Cohort (n = 86)

Patient Variables	Mean (SD) or % (N)	Range
Demographic Measures
Age of onset (years)	61.7 (9.4)	(40.0, 81.0)
Disease Duration (years)	11.2 (5.5)	(2.0, 26.0)
Young Age Onset (<50 years of age)	47.7% (n = 41)
Sex:
Male	84.9% (n = 73)
Female	15.1% (n = 13)
Body Mass Index (kg/m2)	26.8 (5.1)	(18.0, 41.0)
Baseline Measures
NEJM Dopa Equivalents	1246.0 (501.4)	(300.0, 3150.0)
Levodopa Response	48.0% (21.3%)	(-21.0%, 90.0%)
Tremor Off Medication	6.4 (4.7)	(0, 20.0)
Baseline Axial Score On Medication	4.4 (3.4)	(0, 19.0)
S&E Off Medication	52.9 (20.1)	(10.0, 80.0)
Patient Diary:
Hours OFF	11.1 (5.8)	(0, 26.0)
Hours ON without Dyskinesia	13.7 (6.2)	(0, 27.0)
Hours ON with Dyskinesia	8.7 (6.8)	(0, 28.0)
Cognition: DRS Total Score DRS Memory Subscore Clock Score	10.5 (2.7) 9.9 (2.7) 9.0 (1.4)	(5.0, 15.0) (4.0, 13.0) (4.0, 10.0)
Mood:
BDI	11.3 (7.8)	(0, 37.0)
STAI State Anxiety	37.5 (10.5)	(20.0, 68.0)
STAI Trait Anxiety	37.2 (10.1)	(20.0, 62.0)
Off Medication UPDRS-III Baseline	41.8 (10.3)	(16.0, 75.0)
CSP GPi-Sweetspot SSS (LOOCV)	0.84 (0.8)	(0, 2.6)
Independent GPi-Sweetspot SSS	0.41 (0.45)	(0, 1.3)
Outcome Measures
Off medication UPDRS-III % Change (6-months postop)	29.5% (26.4%)	(-44.0%, 82.0%)

SD Standard Deviation.

NEJM New England Journal of Medicine.

S&E Schwab and England Activities of Daily Living Scale.

DRS Dementia Rating Scale-2, Age Adjusted.

BDI Beck Depression Inventory.

STAI State Trait Anxiety Inventory.

UPDRS Unified Parkinson’s Disease Rating Scale.

GPi Globus Pallidus Internus.

SSS Sweetspot Score.

LOOCV Leave-one-out cross validation.

Comparison With Excluded CSP Subjects

When comparing demographics and outcome measures to the excluded CSP subjects due to lack of imaged lead locations (n = 52), there was a higher disease duration in the excluded group (average: 12.9 years, p = 0.03). There was no significant difference between the ratio of male to female subjects (p = 0.77), age of onset (p = 0.2), LDR (p = 0.73), proportion of YOPD (p = 0.4), and 6-month postoperative UPDRS-III outcome (p = 0.77).

Independent Cohort Characteristics

The Independent cohort (n = 51) was predominantly male, ranging in age from 47 to 77, with an average UPDRS-III percent change of 62.7%, with a range of -19 to 100% (SD 24.7%) (Table 2). When testing the Independent GPi-Sweetspot model on the CSP cohort, the mean sweetspot score was 0.41 with a range of 0 to 1.3.

Table 2.

Demographic and Outcome Measures of the Independent Cohort (n = 51)

Patient Variables	Mean (SD) or % (N)	Range
Demographic Measures
Age of onset (years)	64.1 (7.7)	(47.0, 77.0)
Sex:
Male	59.8% (n = 30)
Female	41.2% (n = 21)
Baseline Measures
Levodopa Response	53.4% (22.3%)	(-17.1%, 94.1%)
Off medication UPDRS-III Baseline	46.1 (14.0)	(25.0, 79.0)
Outcome Measures
Off medication UPDRS-III % Change (3-29 months postop)	62.7% (24.7%)	(-19.0%, 100.0%)

SD Standard Deviation.

UPDRS Unified Parkinson’s Disease Rating Scale.

Univariable and Multivariable Analyses

There were 12 variables that reached significance at the p < 0.25 threshold for univariable analysis. (Table 3) Baseline dependency evaluated between baseline UPDRS-III score and 6-month UPDRS-III percent change showed no significant correlation (R = 0.09,p = 0.42). Assumptions for multiple linear regression were met. (Supplementary Table 1, Supplementary Fig. 1)

Table 3.

Univariate Regression between Predictors and CSP #468 UPDRS-III 6-Month Percent Change Sorted by Absolute Value of Standardized Regression Coefficient (n = 86)

Patient Variables	Standardized Regression Coefficient (R)	p-value
Levodopa Response	0.37	<0.01
CSP GPi Sweetspot SSS	0.36	<0.01
Independent GPi Sweetspot SSS	0.30	<0.01
Baseline Axial Score on Medication	-0.28	<0.01
Young Age of Onset	0.23	0.03
Hours off	0.22	0.04
Tremor off Medication	0.21	0.05
NEJM Dopa Equivalent	0.18	0.10
Age of onset	-0.17	0.13
Sex	0.16	0.15
STAI State Anxiety	-0.14	0.19
BMI	-0.13	0.24
Clock Score	-0.12	0.27
STAI Trait Anxiety	-0.11	0.30
Hours on with Dyskinesia	-0.10	0.38
Duration of Disease	0.09	0.39
Baseline Off Medication UPDRS-III	0.09	0.42
S&E score off	-0.07	0.54
BDI	-0.06	0.61
DRS Memory (Age-adjusted)	0.05	0.64
DRS Total (Age-Adjusted)	0.02	0.88
Hours on without Dyskinesia	-0.01	0.95

GPi Globus Pallidus Internus

SSS Sweetspot Score

NEJM New England Journal of Medicine

STAI State Trait Anxiety Inventory

BMI Body mass index

UPDRS Unified Parkinson’s Disease Rating Scale

S&E Schwab and England Activities of Daily Living Scale

BDI Beck Depression Inventory

DRS Dementia Rating Scale-2, Age Adjusted

Backward selection for both multivariable models identified three consistent predictors of 6-month postop UPDRS-III outcome: VTA-Sweetspot overlap (Sweetspot Score [SSS]), LDR, and hours in the off state (non-Bonferroni corrected CSP sweetspot derived multivariable model [CSP-MM]: R = 0.52, Adjusted R² = 0.27; non-Bonferroni corrected Independent sweetspot derived multivariable model [I-MM]: R = 0.49, Adjusted R² = 0.24). Baseline tremor score approached but did not reach statistical significance in the models (B* _{CSP -MM} = 0.19;p = 0.051; B* _{I -MM} = 0.18;p = 0.057). After Bonferroni correction (adjusted alpha = 0.017), only VTA-Sweetspot overlap and LDR remained in the final models. (Tables 4 and 5)

Table 4.

Multiple Linear Regression Results for CSP #468 6-Month Postoperative UPDRS-III % Change with the CSP GPi-Sweetspot Ordered by Level of Importance (n = 86)

Predictor Variable	Standardized Estimate	Estimate	Estimate 95% Confidence Interval	Standard Error	t-value	p-value
Levodopa Response	0.30	0.38	0.13–0.62	0.12	3.1	0.003
CSP Cohort VTA-CSP GPi-Sweetspot Overlap (LOOCV SSS)	0.30	10.46	3.6–17.4	3.5	3.0	0.003

Adjusted R² = 0.21, RMSE 23.6.

GPi Globus Pallidus Internus.

LOOCV Leave-one-out cross validation.

SSS Sweetspot Score.

Table 5.

Multiple Linear Regression Results for CSP #468 6-Month Postoperative UPDRS-III % Change with the Independent GPi-Sweetspot Ordered by Level of Importance (n = 86)

Predictor Variable	Standardized Estimate	Estimate	Estimate 95% Confidence Interval	Standard Error	t-value	p-value
Levodopa Response	0.33	0.41	0.16 – 0.65	0.12	3.31	0.001
CSP Cohort VTA-Independent GPi-Sweetspot Overlap (SSS)	0.25	14.9	3.2– 26.5	5.8	2.5	0.015

Adjusted R² = 0.19, RMSE 24.0.

GPi Globus Pallidus Internus.

SSS Sweetspot Score.

Model Performance and Validation

(Figure 1B, D) CSP VTA overlap with the independently derived sweetspot significantly predicted 6-month UPDRS-III outcomes within the CSP cohort (R = 0.30, p = 0.005). When incorporated into a multivariable framework, the I-MM achieved an R of 0.44 and an adjusted R² of 0.19, yielding the following predictive equation:

$$6-\mathrm{month}\%\mathrm{Change\; in\; UPDRS}-\mathrm{III}={14.9}^{*}\mathrm{Independent\; GPi\; SSS}+{0.41}^{*}\mathrm{LDR}+3.8$$

Fig. 1. Sweetspot models and example patient performances in multivariable models.

A Posterior view of left CSP GPi-Sweestpot model (purple) with GPi (green) and GPe (blue) overlaid. To the right, VTAs of a patient with 78% improvement (yellow) and a patient with 32% worsening (neon blue) in 6-month postop UPDRS-III outcome are overlaid to visualize overlap or lack thereof with the sweetspot model. B Actual vs Predicted 6-month UPDRS-III outcome based on the CSP multivariable model. The highlighted circles represent the performance of the patient with improvement (yellow) and worsening (neon blue) in the model. C Posterior view of left Independent GPi-Sweestpot model (red) with GPi (green) and GPe (blue) overlaid. To the right, the same VTAs of patients with improvement (yellow) and worsening (neon blue) are overlaid to visualize overlap with the sweetspot model. D Actual vs Predicted 6-month UPDRS-III outcome based on the Independent multivariable model. The highlighted circles represent the performance of the patient with improvement (yellow) and worsening (neon blue) in the model.

The CSP-MM demonstrated similar performance (R = 0.47, adjusted R² = 0.21), explaining approximately 21% of the variance in the CSP cohort’s 6-month outcomes. Its corresponding predictive equation was:

$$6-\mathrm{month}\%\mathrm{Change\; in\; UPDRS}-\mathrm{III}={10.46}^{*}\mathrm{CSP\; GPi\; SSS}+{0.38}^{*}\mathrm{LDR}+2.54$$

LOOCV of both models resulted in modest reductions in R² while remaining statistically significant (CSP MM: R = 0.41, R² = 0.17, p < 0.001; I MM: R = 0.38, R² = 0.14, p < 0.001)(Supplementary Fig. 2).

Notably, the CSP-MM maintained its predictive strength in the independent dataset (R = 0.40), providing compelling evidence of external predictive validity (R² = 0.16, p = 0.004).

Discussion

This study evaluated whether overlap between VTA and a GPi sweetspot remains a significant predictor of 6-month postoperative motor improvement when modeled alongside a comprehensive set of preoperative clinical and demographic variables. To accomplish this, we quantified sweetspot overlap with CSP cohort VTAs on the CSP GPi‑sweetspot model. To address the inherent risk of circularity, an entirely independent GPi-sweetspot was generated from a separate cohort and applied to the CSP dataset. Notably, both sweetspot models localized to the primary motor territory of the GPi and, across all multivariable iterations, VTA overlap with the sweetspot emerged as a consistent predictor of motor outcome.

The 6-month postoperative timepoint was selected based on prior evidence demonstrating that the majority of DBS-related motor improvement occurs within the first 3–6 months, after which gains plateau. This pattern was replicated in the independent cohort. (Supplementary Fig. 3) By this interval, stimulation settings and medication regimens are generally stabilized, while the confounding effects of disease progression, particularly axial and non-motor symptoms, remain limited^31,32. Accordingly, 6 month outcomes have been utilized in randomized trials and meta analyses assessing DBS efficacy^33–35.

Previous multivariable studies of GPi-DBS outcomes have been constrained by small sample sizes, single center designs, and limited incorporation of anatomic targeting metrics. Early work identified LDR as a positive predictor of motor improvement, a finding replicated across cohorts at varying follow up durations^18,26,28. The present study extends this literature by integrating stimulation anatomy, quantified as SSS, into a multivariable framework applied to a large, randomized, multicenter cohort. In doing so, it demonstrates that lead location contributes significantly to postoperative motor outcomes across populations.

A key finding is that the Independent GPi-Sweetspot and CSP-MM retained predictive validity when applied to their respective external cohort, despite differences in institution, surgical era, and follow up duration. This external validation suggests the identified GPi sweetspot represents a biologically meaningful therapeutic zone rather than a cohort specific artifact, supporting its utility in both surgical targeting and postoperative programming.

Advances in neuroimaging, normalization to standardized spaces, and lead localization have enabled population level analyses of stimulation anatomy^36–39. While similar approaches have yielded symptom specific sweetspots within the STN^40,41, fewer studies have applied these techniques to the GPi^4,27,42,43. Using these methodological advances, the present analysis identified a consistent GPi-sweetspot located within the posterior third of the GPi, extending toward the GPi–GPe lamina and medial GPe (Fig. 1A, C). This region aligns with established models of GPi functional organization derived from primate physiology, lesion studies, and electrophysiology. Prior work on primates by Baron et al. has demonstrated that intervention within the sensorimotor territory of the GPi preferentially improves akinesia and bradykinesia, while stimulation outside this region, particularly dorsally, may exacerbate dyskinesias⁴⁴. Similarly, ventral GPi targeting has been proposed to alleviate dyskinesias, and anteromedial stimulation has been associated with reductions in levodopa equivalent daily dose^43,45. Structural connectivity and histological studies have further refined GPi subdivisions, placing the identified sweetspot within the primary motor GPi²². This is consistent with electrophysiological studies demonstrating phase–amplitude coupling between the GPi and motor cortex, with pathological high-beta coherence suppressed by GPi-DBS^24,46. Normative rsfMRI analyses by Horn et al. similarly found that stronger anticorrelation between M1 and the stimulation site predicted superior UPDRS-III outcomes⁴. The convergence of these independent lines of evidence provides mechanistic support for the anatomic specificity of the sweetspots identified in this study. Clinically, these sweetspots appear capable of modulating a broad spectrum of motor symptoms with notable efficacy across populations, offering a valuable target for both initial surgical planning and during postoperative management as a target during reprogramming, particularly with the advent of directional leads.

A few variables in our study that were not significant in predicting UPDRS-III % change at 6-months also warrant discussion, including hours in the off state, baseline tremor burden, age-adjusted DRS and YOPD.

Hours in the off state was able to survive backwards selection, however, did not pass Bonferroni correction. Patients’ ability to define an off state and record this at 30 min intervals is dependent on their ability to recognize this as different from the on-medication state and thus could be a surrogate for a robust LDR. However, a univariate regression was conducted and demonstrated that the hours off variable was in fact independent from LDR (R = 0.01, p = 0.9). While this was not included in the final model, future analyses should consider reassessing its predictive value on UPDRS-III outcomes at various time points.

Similarly tremor burden was positively associated with outcome, though did not survive backwards selection in either multivariable model. Prior work has suggested that tremor-dominant PD phenotypes experience greater DBS benefit¹⁸. Tremor is heavily weighted in the UPDRS-III due to multiple questions/items addressing it and the possibility of multiple limbs and the head being involved. Thus, a tremor predominant patient may have a high baseline score and the postoperative resolution of tremor (which is common) results in a larger percent change in the postoperative outcome. As such, patients with tremor-dominant PD often demonstrate relatively robust improvement following DBS. This variable should also continue to be evaluated in future studies of UPDRS-III and subtype outcomes.

Dementia is typically used as a selection criterion for DBS eligibility, with those having more severe dementia leading to poorer outcomes. Patients who underwent DBS in the CSP #468 trial, and specifically our GPi cohort included patients with dementia, with an age-adjusted DRS total range of 5–15, memory subscore range of 4–13, and clock score range of 4–10. It is well established that dementia does not degrade the motor outcomes of DBS but rather affects the quality of life and adverse events, such as postoperative confusion. Our study is consistent with this.

YOPD has been inconsistently described in the literature, with some groups opting for <50 years of age as the cutoff, while others choose <40 years. While this may contribute to conflicting results on young-onset’s effect on non-motor outcomes postoperatively^12,13, it seems YOPD remains a non-predictor of motor outcome. This was seen in a study in the STN population by Park et al¹⁰. This model shows this pattern holds true on a larger, multicenter GPi cohort.

Despite the breadth of variables included in the CSP dataset, the final models only explained approximately 19–27% of variance in motor outcomes. It begs the question as to what additional factors can be responsible for 73–81% of the variance. This result is not limited to the present study, as other studies, primarily of the STN, have had similar success. STN studies utilizing normative datasets or patient specific diffusion tensor imaging (DTI)/rsfMRI have yielded similar R² within the 20–30% range, with one study by Horn et al reaching up to 44% when demographic predictors were included^4,47,48. Although the patient’s own DTI or rsfMRI data would theoretically explain interindividual variability, the resolution of a single patient’s data is poor. This limitation in rsfMRI has been overcome in normal subjects through prolonged scanning. Greene et al utilized the Midnight Scan Club (MSC) dataset with scan times of over 100 min to investigate functional network specificity, integration within the subcortex, individual variability and similarity of functional organization across subjects⁴⁹. They identified subcortical structures with limited interindividual variability, such as the ventral intermediate nucleus of the thalamus (VIM), and structures with high variability across subjects, such as the globus pallidus. They posit that this may be a reason for the variability in outcomes seen with GPi stimulation as compared to the more consistent response to thalamic outcomes on tremor reduction.

Our study comes with limitations. First, a subset of CSP #468 subjects was excluded due to inadequate imaging for lead localization, though excluded patients were comparable to the included cohort in demographics and outcomes. Of note, the proportion of YOPD was high in the CSP #468 cohort, and, though it did not survive backwards selection, the final models may have a subtle inclination towards patients with YOPD. Furthermore, while 6-month postop outcomes were used to mitigate the effects of disease progression on the analysis, it would be beneficial to assess the long-term predictive value of later time point GPi-sweetspots. These analyses are currently underway. Along the same lines, future works on the GPi-sweetspot from other groups and institutions would be informative in assessing the stability of its anatomy across surgeon practices and a wider patient population.

In conclusion, the CSP #468 robust, multicenter database contains a wealth of prospectively collected information including blinded evaluations in patients with PD who underwent randomization to GPi or STN-DBS. Alongside LDR, lead placement within the primary motor GPi sweetspot is a major consistent predictor of 6-month UPDRS-III outcomes and represents a valuable target for surgical planning, programming, and future predictive modeling.

Methods

CSP #468 Subject Enrollment and Imaging Protocol

The enrollment criteria for the CSP #468 cohort, study design, and study procedures have been reported previously and are only described briefly here¹⁴. The CSP #468 study was a randomized, controlled, blinded, clinical trial that recruited patients from seven Veterans Affairs and six United States university medical centers between May 2002 and October 2005 and prospectively randomized patients with PD to either STN or GPi-DBS (Institutional review boards [IRBs] approved at the following institutions: University of California at Los Angeles Los Angeles, California, United States, 90095; VA Medical Center, San Francisco San Francisco, California, United States, 94121; University of California at San Francisco San Francisco, California, United States, 94,143; VA Greater Los Angeles Healthcare System, West LA West Los Angeles, California, United States, 90073; VA Medical Center, Portland Portland, Oregon, United States, 97201; Oregon Health Sciences University Portland, Oregon, United States, 97239; VA Medical Center, Philadelphia Philadelphia, Pennsylvania, United States, 19104; Pennsylvania Hospital Philadelphia, Pennsylvania, United States, 19107; Baylor College of Medicine Houston, Texas, United States, 77030; Michael E. DeBakey VA Medical Center (152) Houston, Texas, United States, 77030; Hunter Holmes McGuire VA Medical Center Richmond, Virginia, United States, 23249; Virginia Commonwealth University Richmond, Virginia, United States, 23298; VA Puget Sound Health Care System, Seattle Seattle, Washington, United States, 98108)¹⁴. Patients were considered eligible if they had idiopathic PD with a Hoehn and Yahr scale score stage 2 or greater while not receiving dopaminergic medication, demonstrated LDR, had three or more hours daily with poor symptom control or poor motor function (e.g., motor fluctuations and/or dyskinesia) despite optimal pharmacologic treatment, stable dopaminergic therapy for one month, and age>21 years. Importantly, LDR was not strictly defined and there was no cutoff value for enrollment. This initial cohort consisted of 299 patients.

Targeting images consisted of T1 with and without gadolinium and T2 coronal slab MRI images merged with a thin slice CT of the patient in the stereotactic frame. Exact parameters were not specified. As outlined in D’Souza et al., postoperative MRI or CT was acquired between 6 and 12 weeks after surgery, adhering to Medtronic’s standard MRI protocol²⁹. Lead location was determined on postoperative T1 images with small slice thickness or preoperative T1 merged with postoperative thin slice CT scans. Exact parameters were not specified.

Of the initial 299 patients, 266 presented for 6-month follow-up, and 180 had currently available and assessable imaging of bilateral leads, crucial for assessment of lead location on outcome. Eighty-six of these patients were randomized to bilateral GPi-DBS.

Collection of baseline and demographic measures of the CSP #468 cohort

At baseline, the motor subscale of the UPDRS-III was assessed by blinded evaluators after overnight cessation of medication as well as on medication⁵⁰. Multiple imputation based on ordinal regression was used to impute missing item information that went into calculating the UPDRS-III scores. The Hoehn and Yahr scale of disability, the Schwab and England scale of activities of daily living; subscales of the UPDRS, including part I (mentation, behavior, and mood), part II (activities of daily living), and part IV (complications of therapy); and assessments of mood and cognition were collected. Patients were trained in the use of motor diaries, recording on medication, motor function with dyskinesia that interfered with movement, off medication symptoms or asleep at half-hour intervals for 2 days. Medications were converted to dopa equivalents using the New England Journal of Medicine (NEJM) formula⁵¹. Basic demographic information, such as age of onset, gender, duration of disease, and body mass index (BMI) was recorded. Additionally, the baseline age-adjusted dementia rating scale (DRS) score and subcomponent DRS memory, clock score, Beck Depression Inventory (BDI), state-trait anxiety inventory (STAI), Schwab and England Activities of Daily living score in the off-medication state (S&E), and UPDRS-II defined axial score were collected. Multiple imputation by predictive mean matching was used to impute missing information for the clock score (n = 1). As part of the CSP #468 protocol, blinded off-medication UPDRS-III assessments were conducted at 6 and 24 months postoperatively. The percent change at the 6-month visit was selected as the primary outcome measure to capture the maximal surgical effect while reducing the influence of disease progression or intervening non-Parkinson related conditions.

Independent Cohort

A second, convenience cohort of subjects was used in this study for creation of an independent sweetspot model and serve as an external validation cohort for the final CSP derived multivariable model. This cohort of patients underwent bilateral GPi-DBS by a single surgeon (KH) at Virginia Commonwealth University Health, Department of Neurosurgery (IRB HM12846) and McGuire Veterans Affairs Medical Center (IRB 1572662) from January 1, 2014 to May 1, 2024 (n = 51). Patients were selected for GPi-DBS based on presence of breakthrough symptoms, on/off fluctuations, and/or dyskinesias. Patients were also selected based on a > 30% LDR, however patients with <30% LDR who could not tolerate a therapeutic dosage due to nausea, hypotension, or had gastrointestinal motility issues limiting absorption were also considered for surgery. Preoperative exclusion criteria included DRS-2 < 123 or significant medical contraindications. Depression or anxiety that was under treatment and considered stable was not considered a contraindication. The majority of such patients underwent unilateral or bilateral GPi-DBS during this time period, and only those with bilateral GPi-DBS were included in the creation of this dataset. Patients who had STN/VIM implantation or unilateral GPi-DBS were also excluded.

Operative targeting was based on volumetric T1 with and without gadolinium, volumetric FLAIR, and FGATIR images when it became available. Additionally, preoperative thin cut CT was employed as the reference image to facilitate merging the intraoperative O-arm images with the MRI data set. Surgery entailed microelectrode recording (MER) and awake testing prior to 2021, and the option of MER/awake or asleep from 2021 till 2023. Subsequently, all surgeries were performed with FGATIR image-based placement under general anesthesia, without a decline in outcomes. Intraoperative O-arm assessment of microelectrode and lead placement was performed throughout the study period.

Written informed consent was obtained from each patient prior to collection and analysis of imaging and outcome data for this project. Patients’ preoperative, off medication UPDRS-III symptom severity was assessed, and again at follow up in the off medication, on stimulation state (follow up range: 3-29 months). No subjects had repeated follow up included in this analysis. The percent change was used as the outcome for the subsequent sweetspot generation and validation analyses.

Assessment of lead location on postoperative outcome

To examine whether stimulation location within the GPi-sweetspot influenced postoperative motor outcome, we calculated the degree of overlap between each patients’ volume-of-tissue-activated (VTA) and optimal therapeutic region (sweetspot). Two independent sweetspot models were generated: (1) a CSP derived GPi-Sweetspot and (2) an Independent cohort derived GPi-Sweetspot.

Sweetspot Generation

The preprocessing steps and creation of a sweetspot have been previously described in detail in prior works^29,36. Briefly, for the Independent GPi-Sweetspot, preoperative T1w, FGATIR and postoperative CT scans were nonlinearly registered to the ICBM Montreal Neurologic Institute 152 nonlinear 2009b template^37,38 (MNI) via LeadDBS v3 preprocessing pipeline. This was followed by LeadDBS’ lead localization, reconstruction, and visualization in the MNI brain space. For the CSP dataset with imaging from the early 2000’s, these preprocessing steps were conducted in Cranial Vault Explorer with preoperative T1 images and postoperative CT scans which registered the patients into MNI space³⁹. For both datasets, the patient VTAs in MNI space were generated using postop follow up stimulation parameters via the SimBio/FieldTrip algorithm⁵². Each VTA was attributed to the corresponding patient’s follow up postoperative percent change. The LeadDBS Sweetspot Explorer was then utilized to create the sweetspot model with the Reich et al. methodology (assigning t-scores to each voxel of the sweetspot based on the voxel’s average outcome compared to the full cohort’s average outcome)⁵³. Specific sweetspot generation parameters are listed in the supplemental materials. (Supplementary Table 2) The sweetspot model was rigorously cross validated through 3, 5, 7, and 10 k-fold testing prior to final LOOCV. In the LOOCV step, one patient was excluded during model creation and VTA overlap on the sweetspot model was assessed. Given the large sample size, the exclusion of individual patients did not meaningfully alter the overall model structure.

Sweetspot scores

SSS quantify the degree to which a subject’s VTA overlaps with a given sweetspot model. For each patient, SSS was calculated during the sweetspot cross validation step by averaging the t‑values of all sweetspot voxels overlapped by that patient’s VTA. For the CSP cohort, SSS derived from the CSP GPi‑sweetspot model were obtained through LOOCV, ensuring that each patient’s SSS was generated from a model that excluded their own outcome data. SSS derived from the Independent GPi‑sweetspot model were generated from CSP VTA overlap with the Independent GPi‑sweetspot. These scores were entered as predictors in the following multiple linear regression analyses.

Statistical Analysis

Descriptive statistics were calculated for all demographic, baseline, and outcome variables. Normality was assessed using Shapiro–Wilk tests. Demographic and outcome differences between included and excluded CSP subjects were evaluated using two tailed t-tests or chi-square tests, as appropriate.

Candidate clinical and demographic predictors were selected a priori based on literature review. Variables demonstrating univariable association with the outcome at p < 0.25 without multiple comparison correction were retained for multiple linear regression to minimize premature exclusion of potentially informative predictors^54,55. Regression assumptions were verified prior to model construction with variance inflation factors (VIFs), Durbin-Watson test, residual plots and P-P plots. (Supplementary Table 1, Supplementary Fig. 1)

Two multiple linear regression models were generated to predict the primary outcome, 6-month percent change in UPDRS-III: one incorporating CSP sweetspot-derived SSS and another incorporating independent sweetspot derived SSS. Backward selection (p < 0.05) was applied prior to Bonferroni correction, with adjusted significance thresholds determined by the number of retained predictors^54,55. Standardized regression coefficients (B*) were reported. IBM SPSS Statistics v31.0.0.0(117) was used for the primary analyses. Final models underwent LOOCV to assess overfitting and residual circularity in RStudio (Version 2025.09.2 + 418).

To evaluate external validity, the final CSP-MM was applied to the independent cohort to predict postoperative motor outcomes. External validation was conducted in RStudio (Version 2025.09.2 + 418).

Author contributions

Authors Contributions:1.Research project: A. Conception, B. Organization, C. Execution; 2.Statistical Analysis: A. Design, B. Execution, C. Review and Critique; 3.Manuscript Preparation: A. Writing, B. Review and Critique; SD: 1B, 1C, 2A, 2B, 2C, 3A, 3B; AB: 1B, 1C, 2A, 2B, 2C, 3A, 3B; JC: 2A, 2B, 2C, 3B; HPS: 2A, 2B, 2C, 3B; VS: 1B, 2B, 2C, 3B; NO: 2C, 3A, 3B; OA: 2C, 3A, 3B; JT: 1B, 1C, 3B; PD: 1B, 1C, 3B; BMD: 1B, 1C, 3B; RL: 1C, 3B; PK: 1C, 3B; PL: 1A, 1B, 1C, 3B; KLH: 1A, 1B, 1C, 2A, 2C, 3A, 3B.

Data availability

The data that support the findings of this study were made available from the study sponsor, United States Department of Veterans Affairs (VA), Cooperative Studies Program (CSP). Restrictions apply to the availability of these data, which were made available under a Data Use Agreement (DUA) for this study. Data may be made available with the permission of CSP by contacting [CSP@va.gov].

Competing interests

Dr. Dawant: The CRAVE and Cranial Suite technology described in this article has been licensed by Vanderbilt University to FHC, Inc., which distributes it under the name WayPoint Navigator. Dr. Dawant receives royalties for this license. This aspect of the work was funded by NIH R01NS095291. The other authors do not have a competing interest. The authors declare no other competing financial or non-financial interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41531-026-01312-z.