Predicting optimal deep brain stimulation parameters for Parkinson's disease using functional MRI and machine learning

Abstract

Commonly used for Parkinson's disease (PD), deep brain stimulation (DBS) produces marked clinical benefits when optimized. However, assessing the large number of possible stimulation settings (i.e., programming) requires numerous clinic visits. Here, we examine whether functional magnetic resonance imaging (fMRI) can be used to predict optimal stimulation settings for individual patients. We analyze 3 T fMRI data prospectively acquired as part of an observational trial in 67 PD patients using optimal and non-optimal stimulation settings. Clinically optimal stimulation produces a characteristic fMRI brain response pattern marked by preferential engagement of the motor circuit. Then, we build a machine learning model predicting optimal vs. non-optimal settings using the fMRI patterns of 39 PD patients with a priori clinically optimized DBS (88% accuracy). The model predicts optimal stimulation settings in unseen datasets: a priori clinically optimized and stimulation-naïve PD patients. We propose that fMRI brain responses to DBS stimulation in PD patients could represent an objective biomarker of clinical response. Upon further validation with additional studies, these findings may open the door to functional imaging-assisted DBS programming.

Fig. 1. Experimental design of 3 T fMRI imaging with DBS activation in PD patients.

A DBS patient implanted with bilateral fully internalized and active DBS electrodes targeting the STN. The DBS lead (Medtronic 3387) has four contacts (width = 1·5 mm) spaced 1.5 mm apart. Using the handheld DBS programmer, DBS programming involves titrating the current delivered by adjusting multiple parameters (i.e., electrode contact, voltage, frequency, and pulse-width) in order to provide the best symptom relief. B Coronal T1-weighted image demonstrating a PD patient with fully internalized and active DBS electrodes (blue) implanted in the STN. C fMRI block design paradigm used during 3 T fMRI data acquisition. While the patient laid still in the scanner, unilateral (left) DBS stimulation was cycled ON and OFF every 30 s for six cycles. The DBS ON/OFF cycling was manually synchronized to fMRI acquisition. Each fMRI sequence was acquired at either optimal (green) or non-optimal (red) contacts or voltages. In this example, the four contacts were screened with fMRI; the a priori clinically optimal contact (marked in green) and non-optimal contacts (marked in red) are shown. DBS deep brain stimulation, fMRI functional magnetic resonance imaging, PD Parkinson’s disease.

Fig. 2. Summary of the methods.

(Top row) After DBS surgery, PD patients undergo fMRI with fully implanted and active DBS systems. Contacts or voltages are screened and their associated fMRI patterns are fed into the machine learning model, which classifies the pattern as optimal or non-optimal. (Middle row) Pipeline for fMRI data processing. (Bottom row) Machine learning model is built with a train dataset using linear discriminant analysis and 5-fold cross validation. Then, unseen test datasets can serve as input to the model for validation. fMRI functional magnetic resonance imaging.

Fig. 3. Typical pattern of fMRI changes resulting from different settings.

BOLD response maps associated with left DBS-STN stimulation at multiple DBS lead A Contacts and B voltages for two a priori clinically optimized PD-STN patients. The fMRI BOLD signal changes at the optimal contact (A top row) and voltage (B middle row) are shown. Brain regions with a significant increase (hot colors, positive t-values, DBS-ON > OFF) and decrease (cool colors, negative t-value, DBS-ON < OFF) (p < 0.001, cluster size = 50) in BOLD response were identified. A The optimal contact showed changes in BOLD response in the left (ipsilateral) motor cortex and thalamus, and right (contralateral) cerebellum. We considered the clinically optimal contact as the origin (i.e., 0) and the non-optimal contacts were mapped as a function of distance in mm from the optimal contact. B When using the optimal stimulation contact, decreasing stimulation amplitude from optimal to low (subtherapeutic) voltage stimulation triggered a decrease in magnitude of the BOLD changes but maintained the topographic pattern. High (supratherapeutic) voltages produced a relatively stronger BOLD response in the left (ipsilateral) motor cortex and right (contralateral) cerebellum but was also accompanied by increased BOLD signal in non-motor regions such as the inferior frontal and occipital lobes. The subtherapeutic voltage was defined as 1.5 V below optimal voltage because a reduction of this magnitude yields a change in clinical status for most PD patients. The supratherapeutic voltage was defined as the voltage just below the side effects threshold (i.e., highest tolerated voltage). BOLD blood-oxygen-level-dependent, DBS deep brain stimulation, fMRI functional magnetic resonance imaging, PD Parkinson’s disease, STN subthalamic nucleus.

Fig. 4. Group analysis of fMRI responses to optimal DBS stimulation shows a specific response pattern.

A Distribution of peak t-values overlaid on a standard Montreal Neurological Institute (MNI) brain when the clinically optimal left DBS settings are used (n = 39 total, n = 35 STN-DBS and n = 4 GPI-DBS, train data). Red circles reflect increased BOLD activity (DBS ON > OFF) whereas blue circles indicate decreased BOLD activity (DBS ON < OFF). Left thalamic regions showed high overlap of peak activation t-values (DBS ON > OFF) across subjects and left motor regions showed peak deactivation t-values (DBS ON < OFF) across subjects. B The optimal contact was considered the origin (i.e., 0) and the non-optimal contacts were labeled with distances relative to the optimal contact. When the optimal contact was the most dorsal or ventral, the maximum distance to the furthest contact was 9 mm. Changes in BOLD signal in the ipsilateral primary motor cortex in response to stimulation at the optimal and non-optimal contacts on STN-DBS leads are shown. Absolute values of t-values at the left primary motor cortex ROI (shaded red) were normalized by t-values in the visual and operculum ROIs (y-axis). Mean normalized BOLD activity in the left primary motor cortex at the optimal contact was significantly different from the non-optimal contacts 3–9 mm away from optimal location (inset, n = 20 (optimal), n = 22 (3 mm), n = 13 (6 mm), n = 8 (9 mm), train data contact with at least one non-optimal contact, Table 1, two-sided Wilcoxon rank sum test). C Effects of varying voltage delivered at the optimal contact on BOLD signals are shown. Absolute values of t-values at the left primary motor cortex ROI (shaded red) were normalized by t-values in the contralateral motor cortex ROIs (y-axis). The mean normalized BOLD activity (t-values) in the left primary motor cortex (y-axis) were maximal at the left optimal contact, but not significantly different from non-optimal voltages BOLD activity (n = 19 optimal voltage, n = 15 supra-therapeutic, and n = 16 sub-therapeutic voltage settings, train data voltage (Table 1), two-sided Wilcoxon’s rank sum test). Error bars indicate SEM. Source data are provided as a Source Data file. BOLD blood-oxygen-level-dependent, DBS deep brain stimulation, fMRI functional magnetic resonance imaging, ROI regions-of-interest, STN subthalamic nucleus.

Fig. 5. fMRI responses predict optimal DBS parameters.

Confusion matrices depicting the performance of classifiers trained to identify optimal DBS settings using features from A contact and voltage cohorts, C contact cohort alone, and D voltage cohort alone in an independent test set (n = 9 a priori clinically optimized patients). B Confusion matrix depicting the performance of the classifier trained to identify optimal DBS settings using features from contact and voltage cohorts in an independent test set (n = 9 stimulation naïve patients). E Summary of performance (overall accuracy) for classifiers in A–D. Bars from dataset 1 depict classifier test accuracy on n = 9 a priori clinically optimized patients. Bars from dataset 2 depict classifier test accuracy on n = 9 stimulation naïve patients. Dashed line indicates chance at 50% accuracy. Source data are provided as a Source Data file. DBS deep brain stimulation, fMRI functional magnetic resonance imaging, NOpt non-optimal, Opt optimal.