Automated Deep Brain Stimulation programming based on electrode location – a randomized, cross-over trial using a data-driven algorithm

INTRODUCTION

Deep Brain Stimulation (DBS) of the subthalamic nucleus (STN) is a well-established treatment option for advanced Parkinson’s Disease (PD), improving motor symptoms, quality of life and allowing to reduce dopaminergic medication.^1-3 Yet, treatment success depends on the correct selection of stimulation parameters, which includes adaptation of amplitude, stimulation frequency, pulse-width, and the relative distribution of electric current across contacts.⁴ Currently, strategies to optimize these parameters are exclusively based on clinical testing and require highly trained medical personnel to iteratively adjust DBS settings in response to therapeutic or adverse effects.⁵ Typically, this process is initiated by a monopolar review for contact selection and amplitude adjustment, the two most important parameters for effective DBS.⁶ Other parameters are often refined at later stages according to patients’ symptomatic profile and treatment response. This procedure, however, is highly time consuming and impeded by multiple factors including a delayed response to parameter adjustments, symptom fluctuations and patient fatigue. Hence, only a fraction of the vast number of parameter combinations can be evaluated in this manner imposing the risk of selecting suboptimal settings even despite multiple thorough and time-consuming programming sessions. This problem has been aggravated with the introduction of directional electrodes, which allow for a more flexible shaping of the electric field but come at the cost of further inflating the number of possible stimulation settings.⁷

To utilize the full therapeutic potential of modern DBS systems, data-driven algorithms could guide DBS programming by suggesting a subset of stimulation parameters.^8,9 Electrode localization represents a promising input feature for such an approach since numerous studies have established a link to therapeutic or adverse DBS-effects across various stimulation targets and diseases.^10-15 This seems especially feasible, since electrodes can be reconstructed from routinely acquired perioperative neuroimaging data allowing for a potential implementation in clinical routine without the need of additional data acquisition or equipment.^16-19 Commercial software which can provide visual feedback for neurologists in relation to patients’ individual anatomy to aid clinical programming procedures is already available and first prospective applications indicate a potential benefit by reducing the time needed for clinical programming.^20-23 Despite these advantages image-guided optimization of DBS parameters remains challenging for two reasons. First, to derive optimal stimulation settings iterative adjustments of DBS parameters need to be conducted manually within the software. Although this allows for a much faster probing of different settings, this approach still does not solve the initial problem considering >10¹⁰ combinatorial possibilities to distribute electric current in octopolar electrodes. Second, optimization objectives are unknown. Decision making is currently based on visualizations of simplified volumetric estimates of neuronal activation (VTA) and their overlap with anatomical regions. However, stimulation targets as well as regions of avoidance are not clearly defined and many therapeutic as well as adverse effects are likely being mediated by a complex interplay of concurrent modulation of multiple white matter bundles.²⁴ Hence, decisions are driven (and potentially mislead) by the programmers’ individual understanding of the interplay between anatomy and function.

Aiming at overcoming those limitations we recently developed and retrospectively validated StimFit, a data-driven algorithm capable of suggesting optimal stimulation parameters in PD patients treated with STN-DBS based on electrode location in a fully automated fashion.²⁵ The model was trained on a large set of standardized monopolar review data to predict therapeutic effects, as well as side-effect probabilities of different stimulation settings. An optimization algorithm was then implemented to find the stimulation setting which would maximize the therapeutic benefit while being constrained by a user-defined side-effect threshold. In the present randomized double-blind cross-over study we now prospectively assess the clinical effects of suggestions made by StimFit to the ones derived during standard of care (SoC) programming strategies.

Methods

StimFit setting

Electrodes were reconstructed and normalized to Montreal Neurological Institute (MNI) coordinate system based on multimodal preoperative Magnetic Resonance Imaging (MRI) and postoperative Computed Tomography (CT) using the default Lead-DBS pipe-line as described in the original publication of the algorithm.^17,25 Algorithmic suggestions of DBS parameters were then computed using StimFit. Briefly, the model predicts motor improvements and side-effect probabilities of stimulation settings based on electrode reconstructions and simulations of the electric field in the target region. To identify which contact selection and stimulation amplitude would lead to the maximal therapeutic benefit an optimizer iterated through different stimulation settings until the model converged to a final solution (Figure 1). To constrain stimulation parameters StimFit required to specify a maximum side-effect probability, which was set to 20% (50% within the first four patients). Since the occurrence of tremor varied across patients, tremor was excluded from the prediction, i.e., the model optimized DBS parameters exclusively based on akinetic-rigid symptoms. Computational time to obtain DBS parameter suggestions for both hemispheres was approximately 50 minutes on a local workstation. The resulting stimulation parameters will be referred to as StimFit settings.

• Download figure
• Open in new tab

Figure 1: Image-based optimization of DBS parameters.

A & B: Perioperative neuroimaging data (preoperative MRI and postoperative CT) is processed using the Lead-DBS pipeline to reconstruct DBS electrodes in MNI-space. C: Based on electric field simulations StimFit predicts clinical effects of different stimulation settings. A nonlinear optimization algorithm iteratively explores the predicted outcomes to converge to a final optimal solution (StimFit setting); MNI = Montreal Neurological Institute; MRI = magnetic resonance imaging; CT = computed tomography

Study protocol

Examinations were performed within one day after overnight withdrawal of dopaminergic medication according to the study protocol (Figure 2). In the preparation phase contact impedances were measured to confirm integrity of the DBS system and to estimate energy efficiency of stimulation settings (see suppl. material). StimFit settings were activated and assigned to a stimulation program on the pulse generator. In case any acute permanent side-effects occurred, stimulation amplitudes were reduced until side-effects disappeared. Stimulation was then switched off and motor impairment was evaluated after a 45-minute wash-out period using part III of the Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS-III). Both SoC and StimFit ON-stimulation conditions were then evaluated after 45-minutes wash-in periods in randomized order. All ratings were carried out by the same rater (JR). Finally, patients were asked to guess the randomization sequence and to self-assess the effects of both stimulation conditions.

• Download figure
• Open in new tab

Figure 2: Study protocol

Patients underwent study examinations after overnight withdrawal of dopaminergic medication. In the preparation phase contact impedances were measured and StimFit settings were programmed and activated for a short duration to detect potential acute stimulation-induced side-effects. Amplitudes were reduced in those cases until side-effects disappeared. MDS-UPDRS-III scores were assessed in OFF-stimulation conditions after a wash-out period of 45 minutes. Afterwards ON-stimulation assessments were performed under both, StimFit and SoC stimulation, again after a 45-minutes wash-in. Rater and patients were blinded to the sequence of both conditions which were randomized in a 1:1 ratio. After motor assessments were completed, patients were asked to self-assess both stimulation conditions on a VAS from “0 = very unsatisfactory” to “100 = very satisfactory” and to guess the correct order of both stimulation conditions. VAS= visual analogue scale; SoC= standard of care; MDS-UPDRS-III = part III of the Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale

RESULTS

Medical records of 86 PD patients who underwent bilateral STN-DBS surgery at Charité between Apr 1, 2018 and Apr 30, 2021 were screened for eligibility criteria. Out of the remaining 67 potential study participants, 52 patients were contacted for recruitment until the intended sample size of n = 35 (32 Boston Vercise™ Cartesia, 3 Medtronic SenSight™) was achieved (Supp. Fig. 2). Study visits were conducted between Oct 23, 2020 and Oct 28, 2021. All patients who participated in the study were included in the analysis. The first four patients were reinvited for study participation after a final patch to the StimFit software was applied and data from their first study visits were excluded from the analysis. Patient demographics and treatment data as well as electrode localizations are summarized in the supplementary material (Suppl.Table 1, Suppl. Fig. 3). OFF-stimulation MDS-UPDRS-III scores improved from 47.3 ± 17.1 to 24.7 ± 12.4 (48 %) and 26.3 ± 12.4 (43 %) under SoC and StimFit stimulation, respectively (Fig. 3A). Mean difference between motor scores was -1.6 ± 7.1 (95% CI: [-4.0, 0.9], p = 0.20, n = 35) establishing non-inferiority at a margin of five points (p = 0.0038, Fig. 3B). 17 patients (49 %) had better motor scores under StimFit stimulation compared to SoC.

• Download figure
• Open in new tab

Figure 3: Primary endpoint and summary statistics

A: Violin plots showing total MDS-UPDRS-III scores under OFF (grey), SoC (purple) and StimFit (blue) stimulation conditions on the left as well as differences of motor scores between StimFit and SoC stimulation on the right (green). Mean and 95% CIs are displayed at each plot. B: Summary statistics of the primary endpoint and symptom-specific MDS-UPDRS-III sub-scores. Mean absolute differences between both ON-stimulation conditions are shown together with their 95% CIs. The CI of the total score did not include the margin of five points, establishing non-inferiority. SoC= standard of care; CI = confidence interval; MDS-UPDRS-III = part III of the Unified Parkinson’s Disease Rating Scale

Analysis of MDS-UPDRS-III sub-scores revealed an improvement for akinesia-rigidity of 46 % and 43 %, tremor 70 % and 62 %, and axial score 36 % and 35 % under SoC and StimFit stimulation, respectively (Fig. 4). This resulted in following mean differences in sub-scores between both stimulation conditions (Fig. 3B): Akinesia-rigidity score: -0.2 ± 4.4 (95% CI: [-1.7, 1.3], p =0.98, n = 35), tremor score: -1.4 ± 3.3, (95% CI: [-2.7, -0.1], p = 0.046, n= 28) and axial score: -0.2 ± 2.0 (95% CI: [-0.9, 0.5], p = 0.67, n = 34).

• Download figure
• Open in new tab

Figure 4: Secondary endpoint I: Symptom-specific motor effects

Violin plots showing symptom specific motor scores for akinetic-rigid (A) and axial (B) symptoms, as well as tremor (C). Each panel depicts scores under OFF (grey), SoC (purple) and StimFit (light blue) stimulation conditions on the left and differences of motor scores between StimFit and SoC stimulation on the right (green). Mean and 95% CIs are displayed at each plot. SoC= standard of care; CI = confidence interval

Patients’ self-assessments of the overall stimulation effects from 0 = “very unsatisfactory” to 100 = “very satisfactory” showed significant differences in favor of SoC, with SoC settings on average scoring 74 ± 19 points as compared to 55 ± 24 under StimFit stimulation (mean difference: 19 ± 28, 95% CI: [9, 29], p < 0.001, n = 34, Fig. 5). StimFit settings were rated superior to SoC by eight (24 %) and equal ratings were given by three patients (9 %). 19 patients (56 %) correctly guessed the order of stimulation conditions, which was not significantly above chance level (p = 0.50). One patient was accidentally unblinded before the first ON-stimulation assessment by looking at the patient programmer and was therefore excluded from the self-assessment analyses. The patient was not excluded from the remaining analyses since the sequence was not revealed to the rater before the study visit was completed.

• Download figure
• Open in new tab

Figure 5: Secondary outcome II: Self-assessments

Violin plots showing results of patient ratings of both SoC (purple) and StimFit (light blue) stimulation conditions on the left as well as differences between ratings on the right (green). Mean and 95% CIs are displayed at each plot. SoC= standard of care; CI = confidence interval

In six patients (17 %) initial programming of StimFit settings (during the preparation phase) resulted in acute side-effects (two muscle contractions, two dysarthria, two vertigo). Stimulation amplitudes were reduced until side-effects disappeared (mean reduction 0.41 ± 0.16 mA ranging from 0.3 to 0.7 mA). Delayed onset dyskinesias appeared in three patients under Stim-Fit stimulation. In two cases these were rated as severe and potentially interfered with motor ratings. Dyskinesias were also observed in two cases under SoC stimulation but did not affect motor assessments.

LEDDs were reduced by 57 % under SoC treatment compared to pre-surgical medication and SoC settings had remained unchanged for 6.7 ± 5.8 months prior to study participation. Mean stimulation amplitudes were 2.6 ± 1.1 mA and an average of 3.2 ± 1.4 contacts were active. Bipolar settings were applied on four electrodes. One patient was treated with interleaving stimulation bilaterally. Mean pulse-widths and frequencies were 60 ± 3 μs and 134 ± 25 Hz. Mean stimulation amplitude suggested by StimFit was 2.4 ± 0.3 mA distributed across 5.4 ± 2.1 active contacts. Battery drain was estimated to be 57 ± 29 μA in SoC compared to 50 ± 21 μA in StimFit settings (p = 0.5, n = 32). All impedance values were within normal ranges. The parameters of both stimulation conditions are provided in the supplementary material for each subject and in a summarized form (Supplementary Tables 2 & 3).

Since SoC treatment was conducted before study recruitment the exact times spent for parameter optimization in each patient were not available. However, during 20 ± 11 months since DBS surgery, patients received DBS-related treatment at specialized in-and out-patient facilities on 28 ± 16 days, including thorough monopolar reviews and OFF-medication assessments. The StimFit algorithm converged to a final solution within ∼50 minutes in all patients and did not require manual steps.

DISCUSSION

In this prospective double-blind cross-over randomized non-inferiority trial bilateral STN-DBS was applied in 35 PD patients using stimulation parameters suggested by an automated algorithm (StimFit) based on electrode locations. StimFit stimulation reduced motor symptoms with a statistically non-significant mean difference of 1.6 points compared to SoC on the MDS-UPDRS-III scale, establishing non-inferiority within a predefined margin of five points. Importantly, both stimulation conditions resulted in significant motor improvements of 48 % and 43 % compared to OFF-stimulation baseline confirming effective STN-DBS and sufficient wash-out/wash-in time between conditions in this patient group.^1-3 In line with this, previous long-term SoC treatment had led to reduction of dopaminergic medication (LEDD) by 57% compared to preoperative treatment in our cohort.

A strength of this study was the prospective double-blind cross-over design and the fact that it used for the first time a data-driven algorithm capable of suggesting optimal stimulation parameters in PD patients treated with STN-DBS based on electrode location in a fully automated fashion. A limitation was the short observational period, which precluded assessment of long-term motor effects and quality of life. Moreover, only cathodal contact configurations are suggested since the algorithm has not been tested for bipolar or interleaving settings. In addition, the algorithm was trained and validated on data with a fixed pulse-width of 60 μs and stimulation frequency of 130 Hz. It does not allow to account for effects of varying frequencies, pulse-shapes or -durations. We further noticed that the variation of stimulation amplitudes suggested by StimFit was relatively small compared to SoC. Initial amplitudes ranged between 2.3 and 2.5 mA across the cohort at a side-effect threshold of 20 %. This might indicate that variance in electrode location and active contact configuration only had limited impact on side-effect predictions compared to stimulation amplitude and suggests that our modeling approach for side-effects should be optimized further.

Ever more complex DBS electrode designs call for guided programming by integrating biomarkers to identify beneficial combinations of stimulation parameters. Computational models of different complexities are available linking the relationship between electrode location, stimulation parameters and clinical outcome.^32-38 Due to its simplicity and low computational demands the binarized approach of modeling the volume of activated tissue (VTA) in respect to patients’ brain anatomy has become a popular method and was recently implemented in commercially available software.¹⁹ However, “bottom-up” biophysical models of neuronal activation are based on many pre-assumptions, most of which are unknown in the individual patient, e.g., fiber diameters and their orientation relative to the electric field, which has shown to impact fiber activation in silico and in vivo.^39,40 Further, direct axonal activation is the only mechanism of action considered by the model neglecting potential stimulation effects on dendrites, glial cells, intracellular cascades, transmitter depletion and network dynamics.⁴¹ Consequently, the VTA should be considered a rather vague metric to estimate the anatomical regions affected by stimulation. To complicate things even more, exact target regions, as well as regions of avoidance are still a matter of debate and were not defined within currently available software for image-guided DBS. This requires physicians to adjust DBS parameters in respect to visible anatomical structures without a clear optimization objective leaving room for different solutions based on the anatomical intuition and training of the programmer. Finally, this in silico optimization needs to be performed manually demanding time and resources. With StimFit we aimed at overcoming some of these limitations. First, the underlying prediction model is based on the properties of the electric field in the vicinity of the electrode and does not require assumptions about mechanisms of action of DBS. Predictions made by this biophysically naïve approach could proof to be more robust across different behavioral effects since they are independent of the underlying neurophysiological mechanisms.

Second, the model was trained and cross-validated on a large sample of monopolar review data allowing to predict and quantify stimulation effects. Predictions were tested on an independent test dataset before, third, the model was embedded in a mathematical optimization framework to identify optimal stimulation settings in a time-efficient manner and without need of any manual interactions.

While many studies in the field of DBS neuroimaging have applied out-of-sample testing to retrospectively validate their results ^10,11,13-15, prospective applications have remained sparse. Frankemolle et al. used image-based models to identify stimulation settings which would minimize spread of electric current to nonmotor regions of the STN.⁴² Applied in ten patients, these settings led to a better performance in a working memory task compared to standard of care stimulation and they concluded that cognitive decline associated with STN-DBS could be avoided by using model-based stimulation parameters. Other prospective studies concluded that software-assisted programming was significantly faster compared to standard clinical procedures.^20-22 In three studies, model-based parameters achieved therapeutic benefit comparable to clinical programming after parameters were re-adjusted based on behavioral feedback. No adjustments were made in a study by Waldthaler et al. and non-inferiority was reported at a margin of 20 % improvement relative to OFF-stimulation baseline. However, keeping in mind that larger controlled DBS trials have reported mean motor improvements ranging from 24 % to 49 %, this margin should be considered too liberal.⁴³

Interestingly, in our cohort akinetic-rigid and axial sub-scores showed similar improvements under both stimulation conditions, while tremor responded significantly less to StimFit stimulation. Our understanding of anatomical target structures involved in therapeutic neuromodulation is currently undergoing a paradigm shift from a disease-to a symptom-centric view.⁴⁴ More specifically, in PD, multiple recent publications point toward an anatomical segregation of DBS “sweetspots” for suppression of tremor on the one hand and akinesia and rigidity on the other hand.¹³ Moreover, the dentato-rubro-thalamic tract might be a common target structure for the suppression of tremor across different pathologies.⁴⁵ This has potential implications for personalized DBS programming procedures since optimal DBS parameters would depend on individual patients’ symptom profiles. Within the StimFit software this concept is now implemented by allowing to modify the optimization objective (motor symptom control) to maximize the predicted therapeutic effects of tremor or akinetic-rigid symptoms on a continuous spectrum. In this present study, however, we did not yet include tremor subitems from StimFit predictions, forcing the model to find settings which would maximize improvement of akinetic-rigid symptoms, exclusively. Seeing this being reflected in our sub-score analysis provides prospective evidence to support the assumption of symptom-specific stimulation sites and underlines the importance of taking baseline symptomatology into account in automated DBS programming procedures. However, further prospective studies are necessary to proof that optimization procedures adapted to patients’ symptom profile will yield higher clinical benefit.

Despite non-inferiority of StimFit stimulation being established for objective motor assessments, subjective patient ratings showed significant differences favoring SoC. To investigate potential reasons for this discrepancy we explored the characteristics of patients which clearly preferred SoC ratings over StimFit. We found that the presence of hyperkinetic symptoms under StimFit stimulation was associated with lower VAS ratings compared to SoC. More specifically, patients who experienced suboptimal tremor control or strong dyskinesias under Stim-Fit stimulation strongly favored SoC settings. This led us to conclude that hyperkinetic symptoms might have excessively contributed to VAS ratings due to the explicitly observable patient feedback. An optimization procedure tailored to patients’ symptomatic profile might therefore also improve their subjective perception of treatment benefit.

Stimulation settings differed between StimFit and SoC with respect to numbers of contacts activated with more multi-contact configurations suggested by the algorithm. Clinical optimization strategies usually start by conducting a monopolar review to assess therapeutic windows of stimulation amplitudes for individual contacts or contact levels. Based on these information, more complex multi-contact configurations which might provide superior benefit will then be probed. This process is carried out analogously by StimFit which starts the optimization procedure by predicting motor and side-effects at different amplitudes to identify favorable monopolar solutions which are then used as starting points for a gradient descent algorithm to explore potentially superior multi-cathode solutions. The number of iterations in this optimization procedure depends on the solver’s stopping criteria, such as the changes in predicted benefit in previous iterations. Clinical programming strategies also need to apply certain stopping criteria, but due to the limited number of iterations those need to be much more liberal, so that further adjustments are often only made when therapeutic outcome is clearly suboptimal, or patients are not satisfied with treatment effects. In many cases this results in monopolar, or pseudo-monopolar stimulation settings (equal distribution of electric current across all contacts at one segmented level), which were chosen for SoC in 46 % of the cases in this cohort. StimFit, however, predicted superior clinical outcome for multi-contact configurations in 93 % of the cases. Overall, StimFit solutions show a more gradual distribution of electric current across contacts compared to SoC. Such settings could not be derived from clinical trial-and-error programming but might bear potential therapeutic advantages, underlining the importance of automated optimization strategies in modern DBS devices. Importantly, battery drain was similar for both stimulation settings in our study. This was in contrast to a reduction in estimated battery drain that has been reported for stimulation settings obtained from anatomy-guided programming.²³ Future algorithms could incorporate estimated energy efficiency as an additional variable to obtain settings with an optimized battery life cycle.

In conclusion, results of this prospective randomized double-blind cross-over trial showed that application of STN-DBS parameters suggested by a data-driven optimization algorithm in a cohort of 35 PD patients led to a significant reduction of motor impairment compared to OFF stimulation, similar to the effects obtained during SoC stimulation. This suggests that data-driven strategies which allow for quantitative predictions of stimulation effects, embedded in mathematical optimization procedures, could govern future programming strategies. This could not only reduce programming time and resources but might pave the way for novel electrode designs to further optimize treatment benefit.

CONTRIBUTORS

JR, TAD, AH and AAK conceptualized the study. Funding was acquired by JR and AAK. JR, JA, JB, APK, GHS and PK contributed to data acquisition. JR conducted the statistical analysis and data visualization. JA and JB verified the data and reviewed the analysis. TAD and AAK contributed to data interpretation. AAK supervised and administered the study. JR wrote the first draft and all other authors reviewed and commented on the report. All authors approved the final published version and agree to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

DISCLOSURES

JR was supported by the Einstein Center for Neurosciences. TAD was supported by the Cologne Clinician Scientist Program (CCSP) / Faculty of Medicine / University of Cologne funded by the German Research Foundation (DFG, FI 773/15-1). AH was supported by the German Research Foundation (Deutsche Forschungsgemeinschaft, Emmy Noether Stipend 410169619 and 424778381 – TRR 295), Deutsches Zentrum für Luft-und Raumfahrt (DynaSti grant within the EU Joint Program Neurodegenerative Disease Research, JPND), the National Institutes of Health (2R01 MH113929) as well as the Foundation for OCD Research (FFOR). AAK was supported by the German Research Foundation (Deutsche Forschungsgemeinschaft, 424778381 – TRR 295) and the Lundbeck Foundation and declares that she is on the advisory board of Boston Scientific and Medtronic, and has received honoraria from Boston Scientific, Medtronic, Zambon and Stadapharm.

FUNDING

This study was funded by the NeuroCure Clinical Research Center (Germany’s Excellence Cluster – EXC-2049 – 390688087) and by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Project ID 4247788381 -TRR 295 Grant.

DATA AVAILABILITY

Behavioral data as well as data on stimulation settings are provided in the supplementary material. Other data supporting the findings of this study are available upon reasonable request from the corresponding author but will not be made publicly available due to their containing information that could compromise the privacy of study participants.