Effective stimulation sites and networks for substantia nigra pars reticulata deep brain stimulation for auditory verbal hallucinations in schizophrenia

Abstract

Deep brain stimulation (DBS) of the substantia nigra pars reticulata (SNpr) is under investigation for managing auditory-verbal hallucinations (AVH) in treatment-resistant schizophrenia (TR-SZ). We assessed acute AVH suppression during initial SNpr-DBS programming in three TR-SZ patients and mapped associated brain network engagement using normative connectomes. One-month post-implantation, monopolar stimulation at each electrode contact was evaluated for its effect on AVH severity. Volumes of tissue activation (VTA) were integrated with normative structural and functional connectivity data to generate individualized network maps. Among 86 VTAs, stimulation volumes associated with greatest AVH relief localized to left anterior-dorsal and right posterior-ventral SNpr. Greater AVH suppression correlated with structural connectivity to sensorimotor cortex and inferior frontal gyral subregions, and functional connectivity to the insula, orbitofrontal cortex, and dorsolateral prefrontal cortex, regions that fit salience attribution, value/belief updating, and top-down control of source/reality monitoring deficits reported in schizophrenia. These preliminary results highlight specific SNpr subregions and circuits linked to acute symptom reduction, supporting the potential of network-targeted DBS for TR-SZ.

Introduction

Schizophrenia is a complex neuropsychiatric disorder with global prevalence of 0.28%, causing significant deterioration in quality of life and significantly reduced life expectancy [1]. About 60 to 80% of patients with schizophrenia experience auditory and verbal hallucinations (AVH) [2]. While antipsychotic drugs are considered first-line therapies, 25–30% of patients with AVH remain refractory to pharmacotherapy [3], resulting in an unmet need for alternative interventions for persons with treatment resistant schizophrenia (TR-SZ).

For the last decade, neuromodulation therapies such as Deep Brain Stimulation (DBS) have shown great promise for many neuropsychiatric disorders, including obsessive-compulsive disorder (OCD) [4–7] and major depressive disorder (MDD) [8–10], among others. Ongoing investigations aimed at improving DBS in these neuropsychiatric disorders have focused on identifying precise and accurate stimulation targets [8, 11, 12], as well as elucidating key neural circuitries that must be engaged for effective DBS. These studies have provided a mechanistic explanation on how DBS intervenes in a larger neural circuitry to achieve clinical benefit and allowed optimization of its effectiveness in future studies through iterative targeting of specific structures and/or brain networks.

In contrast to OCD and MDD, neuromodulatory applications in TR-SZ have been limited. DBS for TR-SZ has been attempted, via anatomical targeting of anterior cingulate cortex, nucleus accumbens, and habenula; albeit with variable outcomes within and between targets limited by small sample sizes. We have recently shown substantial improvement in AVH in a patient with TR-SZ following DBS of the substantia nigra pars reticulata (SNpr) [13]. SNpr was chosen as a target for AVH because we hypothesized that high-frequency stimulation of GABAergic inhibitory SNpr projections to the dorsomedial thalamus may normalize activity in connected cortical regions involved in AVH generation, such as the superior temporal gyrus. Here, we explore the network mechanism of SNpr DBS for AVH in three patients with TR-SZ using normative functional and structural connectomes to investigate which SNpr connections are involved in AVH suppression during acute stimulation.

Materials and methods

Participants

We included three patients with TR-SZ and chronic AVH who had received investigational bilateral electrode implantation (Medtronic 3387 leads) targeting the SNpr (Clinical Trial: #NCT02361554). One patient had an Activa PC Implantable Pulse Generator placed, and two patients had a Percept PC IPG implanted (Medtronic USA). Informed consent was obtained from all subjects. The patients had a diagnosis of SZ as determined by a review of medical records, discussion with referring psychiatrist as well as the Structured Clinical Interview for DSM-V (SCID-V). They were determined to be treatment-resistant for at least one year prior to the screening visit as demonstrated by clinical evidence of persistent auditory hallucinations and/or delusions that had not responded to treatment with three adequate regimens of antipsychotic medication including one failed trial of clozapine. Adequate trials of two different antipsychotic drugs belonging to different classes of at least 12 weeks equivalent to at least 500 mg/day of chlorpromazine within the previous 5 years and a trial of clozapine for at least 12 weeks at a dose of at least 400 mg (or a clozapine level of at least 350 ng/mL) were required. Patients who were unable to tolerate clozapine at this dose or for this duration because of intolerable side effects were also eligible. Patients had at least a score of 6 (severe) on 2 of the 4 Brief Psychiatric Rating Scale or BPRS positive symptoms (conceptual disorganization, grandiosity, hallucinatory behavior and unusual thought content).

One month after electrode implantation each patient underwent a systematic monopolar review stimulating each electrode contact level at escalating amplitudes while reporting their hallucination severity scores based on the Likert Scale (0–10). This acute Likert rating was used as a composite patient-reported AVH severity measure reflecting the overall momentary burden of hallucinations during stimulation, including dimensions such as loudness, distress, and controllability, rather than a single isolated symptom component. Contact testing was performed in a fixed, nonrandomized sequence, moving from ventral to dorsal contacts and from the left to the right electrode (LC0–LC3, then RC0–RC3) in 0.10 V (patient 1) or 0.10 mA (patient 2 and 3) increments until 1.0 V or 1.0 mA then by 0.25 V or mA until side-effects manifested or 5 V or mA were reached if tolerated, rather than in a randomized or counterbalanced order. To reduce potential carryover effects, approximately 1 min was allowed at each voltage step before transitioning to the next stimulation setting, and this wash-in/wash-out interval was kept constant throughout the monopolar review. Each side was tested individually with the contralateral hemisphere de-activated during monopolar review, with a 2 min washout in between. Patients were kept blind to the stimulation condition during testing. The acute AVH effect was transient, based on patient report during blinded programming changes, and persisted only while the given contact/setting was being stimulated.

During monopolar review, stimulation amplitude was escalated systematically until either a clinical effect or a stimulation-related adverse effect was observed. Higher amplitudes (>2.0 V or mA) were sometimes associated with transient paresthesia, dysarthria, or ocular deviation, depending on the active contact and patient, consistent with spread of stimulation to adjacent anatomical structures. When such effects occurred, that level was treated as the safety threshold, and subsequent testing was restricted to lower amplitudes. Accordingly, clinically effective contacts were interpreted within a contact-specific therapeutic window that balanced AVH benefit against acute side effects.

Because SNpr is a novel DBS target for AVH and no target-specific psychiatric programming consensus currently exists, pulse width and frequency were held constant at 60 μs and 130 Hz to reduce parameter heterogeneity during monopolar review and allow more interpretable comparison of contact- and amplitude-dependent effects, since these are the main drivers of outcome in other neuropsychiatric applications. These values were selected as pragmatic, clinically familiar DBS settings based on broader DBS programming experience, including movement disorders applications. However, for patient 3, stimulation was delivered at 125 Hz rather than 130 Hz to allow simultaneous recording with the Medtronic Percept system, as 125 Hz is a sensing-compatible alternative used to reduce recording artifact during active stimulation. Finally, to allow for group-wise analysis, hallucination scores reported for each patient were transformed into z-scores.

Electrode and volume of tissue activation modeling

Bilateral electrodes from three patients were localized and modeled using the Lead-DBS computational pipeline [14]. Postoperative CT sequences were co-registered with preoperative MRI sequences using Statistical Parametric Mapping (SPM) [15]. These images were normalized into a template ICBM MNI152 brain space using Automated Normalization Tools (ANTs) [16]. Based on the metal artifact shown in the CT image, a 3D electrode model was rendered.

Based on stimulation conditions surveyed during the monopolar review for each patient an electric field model was simulated based on Finite-Element-Modeling (FEM)-based algorithm using StimFit / FieldTrip [17]. Binary volumes of tissue activation (VTAs) were generated by thresholding regions where the magnitude of the electric field was above 0.2 V/mm [18–20]. Because VTA size depends on stimulation parameters, particularly amplitude/current, each stimulation setting was modeled separately using a FEM-based electric field approach, such that higher-amplitude settings generated correspondingly larger VTAs.

Mean effect image

In order to characterize how stimulation locations and stimulation volumes in the SNpr target contribute to acute AVH score during monopolar review, Z-scored AVH scores were mapped on each associated VTA and spatially averaged across all patients. This resulted in “Mean Effect Image (MEI)” for the left and right SNpr, where each voxel corresponds to the expected average AVH score after stimulation. To avoid type I error, only voxels occupied by more than 25% of total number of VTAs on each stimulation side were included in the MEI. To precisely localize regions associated with maximum and minimum acute AVH suppression, centroid coordinates representing the top and bottom 5% voxels of MEI were quantified and visualized. Centroid coordinate with top 5% voxels were considered as “sourspot” and bottom 5% voxels as “sweetspot”. All voxel-wise centroid coordinates that were reported following the standard MNI x,y,z coordinate system (mm) such that negative and positive x values denote left and right hemisphere, respectively. SNpr localization was defined using the probabilistic subcortical atlas of Pauli et al., which includes substantia nigra pars reticulata labels in MNI152 2009c nonlinear asymmetric space [21].

Structural connectivity analysis

To explore how structural connections from the stimulated SNpr target are associated with AVH score, each VTA was also seeded into a normative structural connectome composed of 1,600,000 streamlines produced from diffusion MRI of 32 healthy controls as part of the Human Connectome Project (HCP) [22]. After seeding VTAs into the connectome, we generated whole-brain “fiber” streamlines and a structural connectivity (SC) map where each voxel corresponds to a number of fibers passing through the VTAs. Similar to the FC analysis, for each voxel, a multiple linear regression model was fitted by regressing SC across all patients against their respective z-scored AVH scores with patient ID as grouping variables following the equation: $\mathit{AVH}~{\beta }_0+{\beta }_1\mathit{SC}+\mathit{Patient\; ID}$ where AVH denotes z-scored AVH score, SC denotes voxelwise structural connectivity strength, and Patient ID denotes patient identity as a categorical covariate. The t-statistic of the SC term ${\beta }_1$ at each voxel was used to generate the whole-brain T-map.

Fiberfiltering analysis

Similar to structural connectivity analysis, fiberfiltering analysis can be used to estimate how each radiographic streamline “fiber” relative to VTAs is associated with clinical benefits of DBS [23, 24]. For each individual streamline from the normative structural connectome (identical to the one used for the structural connectivity analysis) we tested whether targeting that fiber was more likely to result in AVH score changes. VTAs corresponding to the monopolar review stimulation settings were used to classified whether or not each fiber was stimulated or not by that VTA. Across these two groups – “stimulated” and “unstimulated” – their respective AVH z-scores were compared using a two-sample t-test, resulting in a t-value assigned to each fiber. The magnitude and direction of the t-value indicates the normalized degree of AVH based on whether that fiber is stimulated or not by a VTA, such that a positive t-score is associated with relatively higher AVH score indicating less acute DBS efficacy, while a negative t-score is associated with relatively lower AVH score and more DBS efficacy. Only fibers connected to more than 10% but less than 90% of total number of VTAs were considered to reduce potential type II and type I statistical errors.

For each fiber, after performing a two-sample t-test between AVH z-scores in unconnected and connected VTAs, only fibers assigned to the top 20% and the bottom 20% of t-scores were considered for subsequent analysis. To account for potential laterality effect across hemisphere, identical methods were applied separately for left-sided VTAs and right-sided VTAs.

Functional connectivity analysis

Seed-based functional connectivity was explored by seeding VTAs into a normative fMRI connectome. Whole-brain functional connectivity (FC) maps were derived by calculating voxel-wise correlation coefficients (r) of BOLD signals relative to the VTA. The normative fMRI connectome was created from fMRI sequences of 1000 healthy adults derived as part of the Brain Genomics Superstruct Project (GSP-1000) [25].

For each voxel, a multiple linear regression model was fitted by regressing functional connectivity (FC) strength (r) across all patients against their respective z-scored AVH scores. In this model, patient ID (e.g., 1, 2, 3) corresponding to each VTA was included as a categorial grouping variable to account for potential inter-patient confounding effects between FC and resulting AVH changes such that $\mathit{AVH}~{\beta }_0+{\beta }_1\mathit{FC}+\mathit{Patient\; ID}$ where AVH denotes z-scored AVH score, FC denotes voxelwise functional connectivity strength, and Patient ID denotes patient identity as a categorical covariate. The t-statistic of the FC term β₁ at each voxel was used to generate the whole-brain T-map, where voxels with t-scores greater than 0 indicate regions where an increase in functional connectivity is associated with more AVH (i.e., less suppression), and voxels with t-scores less than 0 indicate regions where an increase in functional connectivity is associated with a less AVH (i.e., more suppression).

AVH prediction modeling and permutation inference

Prediction performance was evaluated separately for structural connectivity (SC) and functional connectivity (FC) analyses using three complementary schemes: naïve fitting (leave-no-patient-out), leave-one-VTA-out cross-validation, and leave-one-patient-out cross-validation. In the naïve analysis, the full dataset was used to construct connectivity-derived features and fit a multiple linear regression model predicting AVH z-score across all stimulation settings. In the leave-one-VTA-out analysis, one stimulation setting was held out at each iteration, the full feature-construction and model-fitting procedure was repeated using the remaining VTAs, and the fitted model was then applied to the held-out stimulation setting to generate a predicted AVH z-score. In the leave-one-patient-out analysis, all stimulation settings from one patient were held out together, the model was trained on the remaining patients, and predicted AVH z-scores were generated for all held-out stimulation settings from the unseen patient. Prediction performance in each setting was quantified as the Spearman correlation between predicted and observed AVH z-scores.

For both SC and FC, prediction was based on ROI-binned similarity features derived within each training iteration. ROIs were defined using the combined cortical Schaefer 200-parcel and subcortical Tian atlas [26]. At each training iteration, voxel-wise T-values were averaged within each ROI to generate an ROI-level mean T-score. ROIs with positive mean T-scores and ROIs with negative mean T-scores were rank-ordered separately and divided into tertiles, yielding six ROI groups corresponding approximately to the top 33%, middle 33%, and bottom 33% of positive ROIs, and the top 33%, middle 33%, and bottom 33% of negative ROIs. For each stimulation setting, correlation was computed between the voxel-wise connectivity pattern within each ROI and the corresponding training-derived voxel-wise T-map restricted to that ROI. These ROI-level similarity values were then averaged within each tertile-defined ROI group to generate six summary features per stimulation setting. These six tertile-based features were then entered into a multiple linear regression model to predict AVH z-score. In the naïve and leave-one-VTA-out analyses, patient identity was additionally included as a categorical covariate. In the leave-one-patient-out analysis, prediction was performed without a patient term because the held-out patient was absent from the training model. This feature-construction and prediction framework was applied analogously to both SC and FC.

To assess statistical significance, empirical permutation testing was performed separately for SC and FC. AVH z-scores were shuffled 1000 times within each patient, thereby preserving the within-patient score distribution while disrupting the association between connectivity-derived features and clinical outcome. For each permutation, the full cross-validated prediction pipeline was rerun and the resulting Spearman correlation between predicted and observed AVH z-scores was recorded to generate an empirical null distribution. Because the a priori hypothesis was that prediction performance would exceed chance in the positive direction, significance was assessed using a one-sided empirical permutation test, with the null hypothesis that the observed Spearman correlation coefficient was not greater than expected under the within-patient permutation null distribution.

Results

Demographics and clinical information

Demographic and clinical information of three patients is listed below in Supplementary Material 1. All patients experienced auditory hallucinations with a primary diagnosis of schizophrenia.

During the monopolar review, patient 1 had total 27 unique stimulation parameters tested (14 left side, 13 right side, 0.5–2 V, 60 μsec, 130 Hz). For patient 2, 41 unique stimulation parameters were tested (24 left side, 17 right side, 0.5–4.9 mA, 60 μsec, 130 Hz). Finally for patient 3, 18 unique stimulation parameters were tested (9 left side, 9 right side, 0.9–2.5 mA, 60 μsec, 125 Hz). This resulted in a total of 86 VTAs (47 left, 39 right). Stimulation parameters used during monopolar review across three patients are further reported in Supplementary Material 2–4.

After all the VTAs and paired AVH Likert scale scores were pooled together, AVH scores underwent Z-score transformation to allow group-level analysis. The raw Likert AVH ratings and Z-scored AVH ratings were highly correlated (r = 0.635, p < 0.0001) shown in Fig. 1A, confirming that the transformation could reliably capture the variance of AVH scores across different stimulation parameters.

Fig. 1. Comparison of AVH Scores from DBS.

A Correlation between raw Likert-Scale AVH scores and AVH Z-scores. B Comparison of AVH Z-scores across the patients. Comparison of AVH Z-scores across hemisphere of stimulation across all patients in the cohort (C) and within individual patients (D–F).

AVH Z-scores were also compared between the three patients to explore if variance was driven by a certain patient. One-way ANOVA test, followed by post-hoc Tukey’s test for pairwise comparison, revealed non-significant differences of scores between patients (F = 0.1739 (2,83), p = 0.84) shown in Fig. 1B. Further, the distribution of AVH scores between each stimulated hemispheres were not statistically different across patients (Fig. 1C), or within each patient (Fig. 1D–F).

At the contacts associated with the greatest reduction in AVH, stimulation-related side effects were still observed across patients. In patient 1, stimulation at the contacts associated with the greatest AVH reduction were frequently accompanied by transient dysarthria and ocular deviation during left-sided stimulation, and ocular deviation during right-sided stimulation. In patient 2, left-sided stimulation was associated with ocular deviation, whereas right-sided stimulation was associated with anxiety and ocular deviation. In patient 3, paresthesia was observed with both left- and right-sided stimulation.

Mapping AVH scores across SNpr stimulation sites

Reconstructed anatomical electrode locations are visualized in Fig. 2 relative to atlas-defined SNpr [21]. To account for variability in stimulation parameters, each stimulation setting was modeled as a separate VTA, such that higher-amplitude/current settings yielded distinct, and typically larger, stimulation volumes. VTA models weighted by z-scored AVH scores were used to create MEI for each hemisphere (Fig. 3). For left-side stimulation, lower than average AVH scores (i.e MEI regions > 0) were seen when the anterior-dorsal portion of SNpr was stimulated, and higher-than-average AVH scores (i.e MEI regions < 0) were associated with stimulation at the posterior-ventral border of SNpr. However, for right-side stimulation the posterior-ventral SNpr was associated with optimal AVH suppression (Fig. 3B, C). Centroid coordinates of “sourspot” and “sweetspot” of MEI are shown in Fig. 3D.

Fig. 2. DBS Electrode Localization and Monopolar Stimulation Space.

A DBS electrodes of three patients are localized and visualized relative to SNpr atlas (Pauli et al., 2018) [21] with 7 T 100 umm ex-vivo MRI backdrop (Edlow et al., 2019). B Stimulation Space in SNpr. During acute monopolar review of each patient, each stimulation parameters was used to model VTA. Stimulation space was defined as area captured by all VTAs produced per patient in both hemispheres shown relative to SNpr (white mesh).

Fig. 3. AVH Mean Effect Image (MEI) and VTA Density Map.

A Voxel-wise density of VTAs is represented as N-Map bilaterally. Mean AVH z-scores attributed to VTAs occupying each voxel are shown as MEI for (B) left and (C) right SNpr. D Centroid of MEI where top (red sphere) and bottom 5% extremes (blue sphere) of voxel intensities are located are mapped relative to SNpr. R Right, L Left, S Superior, I Inferior.

For the right-sided MEI, using stereotactic coordinates in MNI space referenced to the anterior commissure-posterior commissure framework, the centroid coordinate of “sourspot” was located at X = 11.15 mm, Y = −13.97, Z = −9.28 mm encompassing regions with mean z-score of 1.32, and “sweetspot” at X = 11.46 mm, Y = −18.98 mm, Z = −14.58 of regions with mean z-score of −0.95 (Fig. 3D). For the left-sided MEI, the centroid coordinate of “sourspot” was located at X = −12.38 mm, Y = −17.58, Z = −14.84 mm encompassing regions with mean z-score of 1.13, and the “sweetspot” was located at X = −11.70 mm, Y = −12.33 mm, Z = −8.96 of regions with mean z-score of −1.43.

To highlight patient-specific variability in “sweetspot” location, we additionally mapped the centroid coordinates of the VTAs associated with the lowest AVH score in each hemisphere for each patient. In the left hemisphere, the lowest AVH score occurred at X = −10.60 mm, Y = −16.44 mm, Z = −13.96 mm for patient 1; X = −12.02 mm, Y = −13.14 mm, Z = −9.48 mm for patient 2; and X = −13.74 mm, Y = −14.85 mm, Z = −9.61 mm for patient 3. In the right hemisphere, the corresponding coordinates were X = 10.96 mm, Y = −16.99 mm, Z = −10.09 mm for patient 1; X = 10.33 mm, Y = −16.37 mm, Z = −13.41 mm for patient 2; and X = 12.63 mm, Y = −17.64 mm, Z = −13.20 mm for patient 3.

Structural connectivity associated with acute AVH suppression

When seeding VTAs from both hemispheres, SC T-map revealed that lower AVH scores were significantly correlated to increased streamlines (i.e negative T-values) from VTAs reaching right inferior frontal gyrus subregions, sensorimotor cortex, supplementary motor area (Fig. 4A, B). On the other hand, less AVH suppression (i.e positive T-values) was observed when there was increased connectivity of VTAs to left sensorimotor cortex supplementary motor area, rolandic operculum, and temporal regions including inferior temporal gyrus and middle temporal gyrus.

Fig. 4. AVH Structural Connectivity T-Map.

A Cortical rendering of T-Map of whole brain voxel-wise structural connectivity streamlines numbers relative to bilateral VTAs associated with more or less AVH suppression. B T-score distributions across subcortical regions at different axial planes. Positive T-values (red – yellow) are areas associated with more AVH suppression, and negative T-values (blue - green) as areas with less AVH suppression. C Cross-validation and prediction of AVH scores using T-maps with a naïve model (leave-no-patient-out), leave-one-VTA-out cross-validation, and leave-one-patient-out cross-validation. R Right, L Left.

SC T-map could further predict AVH score and was significantly correlated with actual AVH score across patients (r = 0.62, p < 0.001) using the naïve model (leave-no-patient-out cross validation scheme), leave-one-VTA-out cross validation model (r = 0.42, p = 0.006), but not using leave-one-patient-out cross validation model (r = −0.18, p = 0.91) (Fig. 4C).

Streamline fibertracts associated with AVH suppression

To better characterize how distribution of streamline fibers corresponds to AVH suppression, fiberfiltering analysis was adopted. Distribution of fibers predictive of AVH scores were further explored separately for right and left SNpr stimulation. For right SNpr stimulation, only fibers assigned with top and bottom 25% of T-scores were considered. Amongst the total number fibers (516) passing through the right SNpr VTAs, the majority of fibers (98.8% (510)) were occupied by fibers with negative T-scores (−2.26 ± 0.0039) transversing posterior-lateral portion of SNpr.

Right sided fibers associated with higher AVH severity (6.17 ± 0.13) occupied a 1.16% of total number of distributed fibers with, transversing the anterior-medial portion of the right SNpr (Fig. 5A). Similar to the analysis with pooled VTAs above, mean T-scores of fibers were significantly correlated with AVH score (r = 0.75, p = 0.0002) and scores could be predicted during leave-one-vta-out (r = 0.32, p = 0.041) and 5-fold (r = 0.56, p = 0.0004) cross-validation, shown in Fig. 5B.

Fig. 5. Streamline Fibers Associated with AVH from Unilateral VTAs.

Streamlines separately seeded from VTAs produced from (A) right-sided and (C) left-side stimulation. Mean T-scores attributed to each fiber were used to predict AVH scores paired with each VTAs (blue = more suppression, red = less suppression). The performance of the prediction model was tested in naïve condition, leave-one-vta-out, and 5-fold cross-validation scheme for separately for (B and D) left stimulation. R Right, L Left.

For left side stimulation, amongst the 740 total fibers (fibers with top and bottom 25% of t-scores) passing through left SNpr, 64% (472) of fibers assigned negative scores (−3.38 ± 0.011) associated with lower AVH scores transversed through the anterior-medial portion of SNpr (Fig. 5C). On the other hand, posterior SNpr was occupied by fibers assigned positive t-scores (2.72 ± 0.011) occupying 36% of total fibers. Mean T-scores of the fibers were also strongly correlated with AVH scores (r = 0.71, p = 0.0002), and this model performed significantly well during leave-one-vta-out (r = 0.32, p = 0.027) but not during 5-fold (r = 0.18, p = 0.22) cross-validation shown in Fig. 5D.

Functional connectivity associated with AVH suppression

Similar to T-maps generated based on structural connectivity using seed-based streamline counts, T-maps were produced using functional connectivity (FC) measures to further survey potential polysynaptic connections not explained from SC profiles (Fig. 6). Positive (warm colored) voxels in T-maps represent regions were increased in FC relative to VTAs is correlated with higher AVH scores (i.e. less suppression of AVH), and negative (cool colored) voxels with lower AVH scores (i.e. more AVH suppression).

Fig. 6. AVH Functional Connectivity T-Map.

A Cortical rendering of T-Map of whole brain voxel-wise functional connectivity strength relative to bilateral VTAs associated with AVH severity. B T-score distributions across subcortical regions at different axial planes. Positive T-values (red – yellow) are areas associated with more AVH suppression, and negative T-values (blue - green) as areas with less AVH suppression. C Cross-validation and prediction of AVH scores using T-maps with a naïve model (leave-no-patient-out), leave-one-VTA-out cross-validation, and leave-one-patient-out cross-validation. R Right, L Left.

When seeding VTAs from both hemispheres, increased FC to the following cortical regions was associated with lower AVH scores: bilateral occipital cortex, orbitofrontal cortex (OFC), dorsolateral prefrontal cortex (DLPFC) spanning superior frontal gyrus and medial frontal gyrus, middle cingulate cortex, insula, and inferior frontal gyrus subregions. In addition to cortical regions, increased FC to thalamic pulvinar was associated with lower AVH scores.

On the other hand, higher AVH severity (less AVH suppression) was correlated with increased FC strength to bilateral precuneus, subgenual anterior cingulate cortex, and frontal pole. Stronger FC to mesial temporal regions as bilateral hippocampus and amygdala were also correlated with higher AVH scores. In addition to cortical regions, stronger FC to medial portion of thalamic medial dorsal nucleus was associated with higher AVH score.

Finally, FC T-map could further predict AVH score and was significantly correlated with actual AVH score across patients (r = 0.42, p < 0.001) using the naïve model (leave-no-patient-out cross validation scheme), leave-one-VTA-out cross validation model (r = 0.22, p = 0.016), but not using leave-one-patient-out cross validation model (r = −0.19, p = 0.97) (Fig. 6C).

Discussion

In this study, we have identified potential surgical DBS targets within the SNpr that are associated with more or fewer AVH when acutely stimulated during initial programming, suggesting a spatial gradient of anti-hallucination effect in and around the SNpr. Though limited by a focus on only short-term and self-reported responses that may not represent long-term outcomes, these exploratory findings can help to shape the iterative approach to development of ideal targeting for AVH suppression in schizophrenia since the degree of AVH suppression is associated with distinctive stimulation sites within the left and right SNpr and with connectivity to key thalamic subnuclei and prefrontal cortical regions.

The SNpr is one of the major output nuclei of basal ganglia forming the cortico-basal-ganglia-thalamic loop [27]. It receives major inhibitory projections from the striatum, primarily from the caudate [28]. The SNpr consists of a large number of GABAergic neurons sending inhibitory projections to brainstem regions, dorsal midbrain regions such as inferior and superior colliculi, and thalamus [29], including medial-dorsal thalamus (MD).

Changes in MD metabolic activities and volumes have been previously observed in patients with schizophrenia [30, 31]. In addition to SNpr input, MD serves an association hub receiving various sensory input from other brainstem nuclei and other limbic and basal ganglia structures [32]. MD also sends efferent connections to prefrontal regions as such orbitofrontal cortex and other cortical regions extending to premotor, insular, and cingulate cortex [33]. Because MD thus serves as a gateway for subcortical-cortical connections involved in limbic and associative processing, it follows that changes in thalamo-cortical connectivity have been observed in psychosis and schizophrenia [34–38].

These exploratory findings may suggest that SNpr-DBS may exert its effect on AVH by alteration of nigrothalamic signaling that normalizes dysfunctional thalamocortical connectivity. Our connectivity results support this as while higher AVH burden correlated with increased functional connectivity with medial dorsal nucleus increased functional connectivity to thalamic pulvinar was associated with fewer hallucinations. Also, functional connectivity with, thalamic cortical projection targets such as OFC and dorsolateral prefrontal cortex was associated with AVH suppression. Based on previous fMRI mapping studies in AVH, robust evidence supports the activation of language processing and higher-order cognitive areas like inferior frontal gyrus (including auditory cortex), anterior cingulate cortex, and temporal lobe during hallucinations [39]. Our finding that connectivity strength between the SNpr VTA and dorsolateral prefrontal and orbitofrontal subregions regions is associated with fewer AVH supports a mechanism of AVH reduction through normalization of distributed networks of subcortical and cortical regions involved in language processing. While studies reviewing connectivity directly from SNpr are lacking, the caudate, part of the dorsal striatum selectively connected to the SNpr, showed increased functional connectivity to regions as insula and inferior frontal gyrus after clozapine administration in schizophrenia [40, 41].

Interestingly, but not unexpectedly, there was not complete overlap between patterns of functional and structural connectivity seeded by the SNpr VTA. For example, ventromedial prefrontal cortex and dorsolateral prefrontal cortex showed increased functional connectivity with AVH suppression but no structural connectivity. This may be because FC can represent mono- or polysynaptic connectivity whereas SC typically looks at monosynaptic connectivity between regions. Our findings from the FC analysis align with prior research showing that networks involved in AVH include neural substrates not only in auditory perception, but syntax and semantical processing regions involved in language and memory such as hippocampus and amygdala [42]. In a PET imaging study, patients experiencing AVH were reported to have increased activity in regions such as the broca’s area and hippocampus and parahippocampus [43], fitting with our results. Similarly, based on a lesion mapping study reviewing functional networks, auditory hallucination was associated with increased connection to the posterior hippocampal subiculum [44]. Abnormal engagement of hippocampus and its hyperactivity via associated circuitry also seem to play a role in other psychosis beyond AVH from abnormal contextualizing of sensory information [45, 46]. From our study, increased connectivity to these “psychosis” network hubs involving hippocampal and amygdala regions were associated with relatively less suppression and higher AVH burden. This raises the question of whether SNr-DBS acts by modifying networks associated with AVH generation, or instead, by engagement of compensatory networks involving DLPFC and other prefrontal regions that allow “top-down” suppression of AVH.

AVH is driven by dysfunction in distributed network across cortical regions involved in language processing and auditory perception that are likely lateralized in function, potentially explaining the lateralized findings of our structural connectivity analysis. There is evidence that increased auditory hallucination severity has been associated with decreased language lateralization [47–49], such that there is relative hyperactivation of right hemisphere and hypoactivation of the left hemisphere. Complementary to this pattern, our structural connectivity and fiberfiltering results illustrate a strong lateralized effect such that streamlines from right SNpr stimulation were associated with more AVH suppression, and left stimulation with less AVH suppression. Because group-level differences in AVH scores between entire left and right-side stimulation were not significant (Fig. 1C), it is reasonable that engagement of selective streamlines passing through SNpr may drive AVH depending on the hemisphere of stimulation. In AVH and schizophrenia, lateralization of structures such as inferior frontal gyrus [50], supramarginal gyrus [51] and angular gyrus [52, 53] has been demonstrated. Similarly, we found that acute AVH-suppression was associated with increased connectivity between the stimulation VTA and the right inferior frontal gyral subregions.

Taken together, SNpr-DBS may rebalance circuit nodes that support cognitive processes consistently implicated in AVH, including expectation updating, aberrant salience processing, and auditory source monitoring, spanning regions such as the OFC, insula, and DLPFC. In schizophrenia patients with AVH, lateral OFC cortical thickness is reduced relative to non-hallucinating patients and correlates inversely with hallucination severity, implicating this region in disrupted prior belief encoding and failures of expectation updating [54, 55]. For insula, especially the anterior insula, have shown convergent AVH-related structural and functional abnormalities, with abnormal salience network–default mode network coupling linked to the misattribution of internally generated speech as external [42, 56]. Finally, the DLPFC, a key region for top-down control of speech and language processing, exhibits reduced functional connectivity with Broca’s area in AVH. This abnormal coupling may impair the monitoring systems that normally distinguish internally generated speech from external auditory input, consistent with the therapeutic benefit of frontotemporal stimulation in treatment-resistant AVH [57, 58].

Another indication of the lateralized effects of SNpr DBS is that AVH suppression was maximal with stimulation in the left anterior dorsolateral SNpr, which was the opposite for the right SNpr where posterior-ventral stimulation was most effective. The observed hemispheric asymmetry within SNpr should be interpreted cautiously, as the present study was not designed to formally test laterality effects. However, this inverse “sweetspot” and “sourspot” location between right and left SNpr may be attributed to the lateralization of aforementioned circuitry involved in language processing. In addition to the lateralized circuitry across hemisphere, the gradient of clinical outcomes may also be attributed to the topography of ipsilateral connections made from SNpr. Investigations on topographic connections within human SNpr are sparse, but early studies on non-primate models have illustrated overlapping populations of nigrotectal, nigroreticular, and nigrothalamic occupying distinctive zones within SNpr [59, 60]. Following the anterior-posterior axis, nigrothalamic cells projecting to ventroanterior-ventrolateral (VA-VL) and MD thalamus diminish in population. Following the medial-lateral axis within the first anterior part of the SNpr, nigrotectal cells projecting to colliculi regions heavily populate the lateral margin of SNpr. Based on this topography, one may theorize that our “sweetspot” for left SNpr DBS may be at a region with intersecting neuronal populations forming nigrothalamic and nigrotectal connections. Given the significance of these two circuitries in forming the cortico-thalamic-cortical loop, future radiographic studies are strongly encouraged to validate whether stimulating these two pathways achieve sufficient AVH benefits with a larger sample size and with chronic outcomes.

Several limitations exist in this study. First, our analysis was performed based on unvalidated Likert-style clinical ratings from three patients. When performing DBS programming, quick assessments are often needed to test each setting in a timeline the patient can tolerate. Validated psychosis scales cannot be repeated for each of the stimulation parameters tried during the duration of visit that most patients can tolerate. Given the experimental nature of our study with a novel surgical target for AVH in schizophrenia, a restricted number of highly refractory patients were eligible to undergo the procedure and with such a small sample size our results may have limited generalizability. However, through postoperative monopolar review montage, we were able to test numerous stimulation settings and resulting AVH scores. By incorporating each AVH scores paired with stimulating settings as individual datapoints, and pooling these scores together from all three patients, we could maintain the power of our analysis. We also assessed the possibility that the results were confounded by inter-patient differences in AVH severity. Mean AVH scores did not differ significantly across the three patients, arguing against a dominant between-patient bias. In addition, because multiple AVH measurements were obtained from the same individual, all analyses included patient identity as a covariate to account for the non-independence of repeated observations within patients.

Furthermore, due to the nature of monopolar review, the clinical outcomes recorded were based on acute stimulation effect, and may not accurately reflect outcomes after chronic DBS, since the collection of longitudinal data on these subjects is ongoing. In addition, monopolar review was performed in a fixed contact order rather than using randomized or counterbalanced testing, and formal within-session re-testing was not performed for every setting. Accordingly, nonspecific influences such as attention, arousal, expectation, or carryover effects cannot be fully excluded, although a constant wash-in/wash-out interval was used, and clinically effective final settings were reproducible at follow-up. We will further attempt to validate and update our current findings after we have sufficient datapoints during chronic stimulation in a subsequent study.

Another major limitation is that we used structural and functional connectomic data generated from aggregated dMRI and fMRI sequence from group of healthy controls. Schizophrenia and AVH is considered a distributed network disorder, reflected by changes in whole-brain white matter properties derived from dMRI [61–63] or BOLD activation patterns [64, 65]. As a result, normative connectomes may not fully capture disease-specific or patient-specific connectivity patterns relevant to AVH suppression or exacerbation, which may limit the generalizability of our findings. At the same time, similar to prior network-mapping studies in movement disorders [66, 67] and neuropsychiatric disorders [24, 68], we used normative connectivity as a hypothesis-generating approximation to identify candidate circuit-level substrates associated with AVH and SNpr DBS. Future investigations should study SNpr DBS connectivity associated with AVH suppression using patient-specific fMRI or dMRI data.

In summary, we have identified stimulation volumes in and around the SNpr that are associated with acute AVH suppression. Many of the executive control and language processing regions identified in the structural and functional connectivity analyses herein are also implicated in imaging studies of AVH, and our work supports a causal relationship between AVH and these stimulation related regions. Our study may inform future iterative efforts to establish an optimal target in or near the SNpr that maximally suppresses AVH.

Author contributions

MJK and AB contributed equally to this work. MJK contributed to study conception and design, data analysis and interpretation, and drafting of the manuscript. AB contributed to study conception and design, data analysis, and interpretation. YS, ASa, MF, KSC, ASh, and DS contributed to data acquisition and critically revised the manuscript. KAM and NC supervised the study, contributed to study design and data interpretation, and critically revised the manuscript. All authors reviewed and approved the final manuscript.

Funding

KAM discloses support for the research reported in this work from the National Institutes of Health/National Institute of Neurological Disorders and Stroke [5K23NS101096-01A1], the Michael J. Fox Foundation, the Parkinson’s Foundation, UCB Pharma, and the U.S. Food and Drug Administration [U01FD005942]. M.J.K. discloses support from the National Institutes of Health through the University of Pennsylvania Medical Scientist Training Program.

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Code availability

Codes that were written and utilized for analysis in this study are available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

All methods were performed in accordance with the relevant institutional guidelines and regulations. The study was approved by the Johns Hopkins Medicine Institutional Review Board (IRB protocol NA_00042065). Written informed consent was obtained from all participants prior to participation.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41398-026-04298-7.