Deep brain stimulation of symptom-specific networks in Parkinson's disease

Abstract

Deep Brain Stimulation can improve tremor, bradykinesia, rigidity, and axial symptoms in patients with Parkinson's disease. Potentially, improving each symptom may require stimulation of different white matter tracts. Here, we study a large cohort of patients (N = 237 from five centers) to identify tracts associated with improvements in each of the four symptom domains. Tremor improvements were associated with stimulation of tracts connected to primary motor cortex and cerebellum. In contrast, axial symptoms are associated with stimulation of tracts connected to the supplementary motor cortex and brainstem. Bradykinesia and rigidity improvements are associated with the stimulation of tracts connected to the supplementary motor and premotor cortices, respectively. We introduce an algorithm that uses these symptom-response tracts to suggest optimal stimulation parameters for DBS based on individual patient's symptom profiles. Application of the algorithm illustrates that our symptom-tract library may bear potential in personalizing stimulation treatment based on the symptoms that are most burdensome in an individual patient.

Fig. 1. Electrode placement.

Active contacts visualized on a coronal slice of the cortex separately for each of the three subcohorts of the discovery cohort (total N = 129, left) and the three validation cohorts (N = 93, 10, and 5, respectively, right). Please note that orientations here refer to Superior (S), Posterior (P), and Lateral (L).

Fig. 2. Symptom-network library.

Views (A)–(C) from medial. A symptom-response tracts shown in sagittal view from medial and magnified at the level of the STN (orange, insets, one rotated by 180 degrees, i.e., shown from lateral view). Symptom-response tracts follow a rostrocaudal gradient with tremor most occipital, followed by bradykinesia, axial symptoms, and rigidity. All shown tracts significantly correlated with symptom improvements after correcting for multiple comparisons (α < 0.05) using a two-tailed correlation analysis test. Note that tracts are in proximity to one another, making it possible to modulate all of them with a single well-placed electrode (matching clinical experience). B Symptom-response tracts visualized separately at the STN level with the other tracts grayed out for spatial comparison. Insets represent circular and 10-fold cross-validation results for each symptom tract. C Segregation of symptoms within indirect pathway streamlines between STN and pallidum, following a similar rostrocaudal gradient. D Cortical origins of hyperdirect projections. Streamlines associated with tremor improvements originated in primary motor cortex, whereas the ones associated with improvements in hypokinetic symptoms originated from premotor regions in a more interspersed fashion. I = Inferior, A = Anterior, L = Lateral, P = Posterior.

Fig. 3. Anatomical considerations of circuits associated with improvements of tremor and axial symptoms.

As opposed to the other figures, tracts in this figure are not thresholded at significance after FDR correction but include a broader set of tracts to appreciate the broader distribution of symptoms across streamlines (lower threshold). A Tremor tracts included projections from the cerebellar nuclei to thalamus as well as the cortical projections from primary motor cortex to STN, matching current pathophysiological models of tremor. B Tracts associated with axial symptoms included a brainstem connection to the pedunculopontine nucleus region. C Segregating axial symptoms into gait vs. all other (axial) items revealed that this connection was driven by gait (and not by other axial symptoms). D Comparison to the projection site with a matching slice from a histological atlas published by Coulombe and colleagues at z = +5.08 mm (panel adapted under the Creative Commons Attribution (CC-BY) license from Coulombe et al., 2021 Frontiers in Neuroanatomy). A = Anterior, L = Lateral, P = Posterior.

Fig. 4. Network Blending.

A Two example patients’ stimulation volumes are shown alongside the optimal streamlines associated with symptom-response tracts. These two patient examples illustrate both extremes of the model estimation: one, where the absolute error value of model estimate is low and the other, where the absolute error is higher. To derive group level statistics, we employ the multi tract model across the four symptoms, and this process led to four scores, each coding for one symptom. These were linearly weighted by the symptoms prevalent in each patient (since, for instance, a patient with severe tremor would profit more from modulating the tremor streamlines) and averaged, leading to a weighted-average score that was converted to UPDRS-III improvements based on the training data. These estimated improvements significantly correlated with actual improvements when analyzed via a two-tailed correlation analysis (R = 0.33 p = 0.00016, mean absolute error: 17.87%, RMSE: 0.22, R² = 0.08). B Stimulation volume of the same two patients shown alongside the optimal streamlines associated with global UPDRS-III improvements. These fiber scores (0.54, 0.20) were transformed to estimated values of global UPDRS-III improvements based on the training data within the 10-fold cross-validation process. These estimated improvements significantly correlated with empirical improvements, when analyzed through a two-tailed correlation analysis (R = 0.28, p = 0.01, mean absolute error: 18.11%, RMSE = 0.22, R² = 0.05). The shaded area in the correlation plot signifies the 95% confidence interval on the slope of the line. L = Lateral, A = Anterior, I = Inferior. Source data available as source data file.

Fig. 5. Retrospective validation on long-term clinical outcome data.

A The fiber distribution of the original model as shown in previous figures, B fiber distribution when recalculating the same model on the independent test dataset (N = 93). C Estimation of UPDRS-III improvements in the test set (R = 0.37, p = 0.0006, R² = 0.07, RMSE = 0.22, MAE = 17.16), based on the original symptom response model, using two-tailed correlation analysis. The shaded area of the correlation plot signifies 95% confidence interval on the slope of the line. L = Lateral, A = Anterior, I = Inferior. Source data available as source data file.

Fig. 6. Retrospective validation on monopolar review dataset.

A The left panel illustrates a raincloud plot where each data point represents a Spearman’s correlation coefficient between estimated and empirical UPDRS-III improvements for settings in one of the 20 electrodes. A one sided t-test is significant, illustrating that Spearman’s rho is positive across most electrodes (T = 4.15, p = 5.3e-04, Average R = 0.41 ± 0.44). All correlation plots are shown in Fig. S32. The right panel gives four representative examples. A red eclipse is used to represent the stimulation contact that renders the highest improvement for a given electrode, while the contact chosen by the model is marked with a blue eclipse, corresponding stimulation fields are shown for the example electrodes. B, C To assess symptom-specificity of the model, the analysis was repeated, this time maximally weighting either bradykinesia or rigidity symptoms, respectively. Correlations across settings in the 20 electrodes were almost all positive when the model was used to estimate improvements in the correct symptom, but significantly dropped when used to estimate improvements in the respective other symptom. In each panel, two representative examples of correct vs. incorrect symptom pairings are given. In both (B) and (C), the T value is derived from a paired t-test between the Spearman’s rho for estimated improvements in the correct symptom (for instance, when the model, trained on bradykinesia improvement estimated empirical bradykinesia improvement, T = 3.2987, p = 0.0045) vs estimated improvements in the incorrect symptom (for instance, when the model trained on bradykinesia improvements estimated rigidity improvement, T = 2.5484, p = 0.02). The shaded area of the correlation plots signifies the 95% confidence interval on the slope of the line. Source data available as source data file.

Fig. 7. Hypothetical future use of symptom-tract model.

A DBS Today. A well-placed, standard omnidirectional Medtronic 3389 electrode is shown with a single stimulation volume that equally covers all symptom-specific tracts. B DBS in the future. A hypothetical future concept with a modern electrode (Boston Scientific Cartesia X electrode with 15 directional and one omnidirectional contact) is shown. With some devices, it is possible to steer multiple stimulation volumes toward individual tracts. In our example, one stimulation could target tremor streamlines (potentially with a high frequency of ~180 Hz). A second volume would focus on the axial/gait streamlines connecting to the PPN region (potentially with a low frequency of ~25 Hz). L = Lateral, A = Anterior, I = Inferior.

Fig. 8. Methods for calculating and cross-validating the multi-tract model / symptom network library.

A An example fiber tract from the pathway atlas is shown (red dashed line). For each E-field that it passes, the peak magnitude is recorded and correlated with symptom improvements. For instance, the example tract was strongly activated by E-fields that led to improvements in bradykinesia (blue scatter plot), leading to general UPDRS-III improvement (gray scatter plot). The tract is tagged by five Spearman rank correlation coefficients, one for each symptom domain, and one for global motor improvement. B This process is repeated across all fiber tracts in the pathway atlas to create the symptom network library. Tracts can be filtered and visualized based on the correlation coefficients (and significance values) they obtained. C The single tract model (coding for global motor improvements) is cross-validated by estimating motor improvements in left-out patients based on their activation of the tract. D A more elaborate symptom-specific tract model repeats this procedure four times (for each symptom tract) and weight estimates by baseline scores of each symptom.

Fig. 9. Methodological overview of the Cleartune algorithm.

In this example, a novel (hypothetical) patient is treated based on Cleartune, who appears to be tremor-dominant. Cleartune performs network blending (Fig. 2) taking into consideration the patient’s symptom profile, which directly leads to strong weights for tremor and rigidity tracts. After localization of DBS electrodes for the new patient, E-fields for various permutations and combinations of the contacts are simulated using a surrogate optimizer, and their impact on each symptom network is calculated. These impacts are weighted by the symptom profile, leading to an estimated improvement score for each solution. Finally, the solution leading to the best estimate is selected and reported to the clinician, who may consider programming the solution into the DBS pulse generator.