Summary
Background
Deep brain stimulation (DBS) of the subthalamic nucleus (STN) is highly effective in controlling motor symptoms in patients with Parkinson’s disease. However, correct selection of stimulation parameters is pivotal to treatment success and currently follows a time-consuming and demanding trial-and-error process. We aimed to assess treatment effects of stimulation parameters suggested by a recently published algorithm (StimFit) based on neuroimaging data.
Methods
This double-blind, randomised, crossover, non-inferiority trial was carried out at Charité – Universitätsmedizin, Berlin, Germany, and enrolled patients with Parkinson’s disease treated with directional octopolar electrodes targeted at the STN. All patients had undergone DBS programming according to our centre’s standard of care (SoC) treatment before study recruitment. Based on perioperative imaging data, DBS electrodes were reconstructed and StimFit was applied to suggest optimal stimulation settings. Patients underwent motor assessments using the Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale part III (MDS-UPDRS-III) during OFF-medication and in OFF-stimulation and ON-stimulation states under both conditions, StimFit and SoC parameter settings. Patients were randomly assigned (1:1) to receive either StimFit-programmed DBS first and SoC-programmed DBS second, or SoC-programmed DBS first and StimFit-programmed DBS second. The allocation schedule was generated using a computerised random number generator. Both the rater and patients were masked to the sequence of SoC and StimFit stimulation conditions. All patients who participated in the study were included in the analysis. The primary endpoint of this study was the absolute mean difference between MDS-UPDRS-III scores under StimFit and SoC stimulation, with a non-inferiority margin of 5 points. The study was registered at the German Register for Clinical Trials (DRKS00023115), and is complete.
Findings
Between July 10, 2020, and Oct 28, 2021, 35 patients were enrolled in the study; 18 received StimFit followed by SoC stimulation, and 17 received SoC followed by StimFit stimulation. Mean MDS-UPDRS-III scores improved from 47·3 (SD 17·1) at OFF-stimulation baseline to 24·7 (SD 12·4) and 26·3 (SD 12·4) under SoC and StimFit stimulation, respectively. Mean difference between motor scores was –1·6 (SD 7·1; 95% CI –4·0 to 0·9; superiority test psuperiority=0·20; n=35), establishing non-inferiority of StimFit stimulation at a margin of –5 points (non-inferiority test pnon-inferiority=0·0038). In six patients (17%), initial programming of StimFit settings resulted in acute side-effects and amplitudes were reduced until side-effects disappeared.
Interpretation
Automated data-driven algorithms can predict stimulation parameters that lead to motor symptom control comparable to SoC treatment. This approach could significantly decrease the time necessary to obtain optimal treatment parameters.
Funding
Deutsche Forschungsgemeinschaft through NeuroCure Clinical Research Center and TRR 295.
Introduction
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. Currently, strategies to optimise 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.
Typically, this process is initiated by a monopolar review for contact selection and amplitude adjustment, the two most important parameters for effective DBS.
Variation of stimulation frequency and pulse width usually has limited impact on clinical outcome
and therefore current clinical programming strategies suggest that these parameters should be kept at default values (commonly 130 Hz, 60 μs) during initial programming and refined at later stages according to patients’ symptomatic profile and treatment response.
Nevertheless, this procedure 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. This problem has been aggravated by 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.
Research in contextEvidence before this studyWe searched PubMed for clinical trials published up to July 28, 2022, combining the terms (“deep brain stimulation”, “programming”) AND (“automated” OR “assisted” OR “guided” OR ”software”) including publications with abstracts in English or German. We identified four studies that prospectively assessed the therapeutic benefit of deep brain stimulation (DBS) parameters derived from neuroimaging metrics. All studies were conducted in patients with Parkinson’s disease treated with subthalamic nucleus (STN) DBS, and a double-blind crossover design was used to compare image-guided programming with standard of care (SoC). Reduced programming time and comparable motor symptom control for image-guided DBS was reported in all studies. However, three studies lack statistical power and do not provide an a priori definition of a non-inferiority margin. The margin in the fourth study should be considered too liberal, leading to inconclusive findings of non-inferiority of image-guided DBS on the one hand and statistical superiority of SoC on the other hand. Crucially, all studies used a commercially available software that provides visual feedback of stimulation sites in relation to patient anatomy but does not specify clear optimisation objectives or suggest optimal stimulation parameters. More elaborate automated or data-driven algorithms have been developed but not yet been systematically tested in prospective applications.Added value of this studyTo our knowledge, this is the largest randomised trial assessing the therapeutic benefit of DBS settings derived from neuroimaging metrics. Moreover, this is the first time a data-driven model to suggest optimal stimulation parameters was prospectively applied in a randomised, double-blind, crossover clinical trial. Leveraging the recent advances in the field of DBS neuroimaging, this approach allows identification of individual stimulation parameters, which maximise predicted motor symptom control, while accounting for potential stimulation-induced side-effects in a fully automated manner. In a cohort of 35 patients treated with STN-DBS, we established non-inferiority of motor symptom control under these stimulation settings compared with SoC treatment, with no statistically significant difference between both conditions. Furthermore, we identified differential effects with regard to specific motor symptoms. These findings provide prospective evidence for distinct symptom-specific stimulation targets and suggest that automated DBS parameter selection tailored to patients’ individual symptom profiles could further improve therapeutic benefit.Implications of all the available evidenceIndividualised adaptation of stimulation parameters is essential to optimise DBS treatment benefit. Current optimisation strategies are based on clinical trial-and-error, a process which is not only time-consuming and exhausting for both patients and medical personnel, but can also easily lead to selection of suboptimal stimulation parameters. This problem was amplified with the introduction of directional electrodes and limits the potential benefit of more complex electrode designs. Multiple retrospective studies have linked electrode location and stimulation parameters to clinical outcome. This has shed light onto the anatomical structures involved in therapeutic neuromodulation, but despite the urgent need, translational applications remain sparse. First prospective trials have reported clinical advantages in using anatomical visualisation software to guide DBS parameter selection. The current study expands these findings by providing strong evidence that neuroimaging data can be used to identify beneficial stimulation settings in an automated manner. This could potentially be translated to other diseases and provides much needed prospective validation in the field. Data-driven algorithms could assist future programming strategies and pave the way for more complex electrode designs to further optimise treatment benefit.
,
,
Electrode localisation 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.
,
,
,
This seems especially feasible since electrodes can be reconstructed from routinely acquired perioperative neuroimaging data, allowing for potential implementation in routine clinical practice without the need for additional data acquisition or equipment.
Commercial software that can provide visual feedback of stimulation and electrode location in relation to patients’ individual anatomy is already available to aid clinical programming procedures, and first prospective applications indicate a potential benefit by reducing the time needed for clinical programming.
,
,
,
Despite these advantages, image-guided optimisation 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 that there are more than 1010 combinatorial possibilities to distribute electric current in octopolar electrodes. Second, optimisation objectives are unknown. Decision making is currently based on visualisations of simplified volumetric estimates of neuronal activation (volume of tissue activated, VTA) and their overlap with anatomical regions. However, stimulation targets as well as regions of avoidance are not clearly defined.
Moreover, the VTA as a simplified “bottom-up” biophysical model of neuronal activation should be considered a rather vague metric to estimate the anatomical regions affected by stimulation. It is based on many pre-assumptions, most of which are unknown in the individual patient—eg, fiber diameters and their orientation relative to the electric field, which has shown to impact fiber activation in silico and in vivo.
,
Subsequently, VTA-based predictive modelling approaches have been shown to be vulnerable towards parameter modifications and varying statistical implementations.
Hence, DBS parameter selection in current image-guided programming strategies is driven (and potentially misled) by the programmers’ individual understanding of the interplay between model estimations, patient anatomy, and DBS effects.
StimFit was trained and tested based on a large sample of over 600 different stimulation settings applied in 50 patients with Parkinson’s disease.
,
,
Electrode reconstructions were obtained using the Lead-DBS toolbox (figure 1A) and a predictive model was implemented linking electrode locations and stimulation parameters to corresponding improvements of akinetic-rigid symptoms and tremor as well as probabilities of potential stimulation-induced side-effects (figure 1C).
Model predictions were based on the properties of the electric field generated by each stimulation setting. Opposed to commonly used binarised VTA models, this approach does not assume an all-or-nothing firing behaviour of the axons of passage but instead allows modelling of a probabilistic interrelation between voltage gradients (and directionalities) and clinical outcome in the target region. This way, the model is less constrained by unknown biophysical pre-assumptions and might therefore be more robust, especially in anatomically complex areas, such as the subthalamic region. We retrospectively validated the predictive model using an independent dataset, showing that the trained algorithm could predict differences in stimulation outcome caused by varying stimulation parameters in individual patients. Next, a non-linear optimisation procedure was implemented to identify the setting that would maximise therapeutic effects, while minimising probabilities of stimulation-induced side-effects in a time-efficient manner (figure 1C). It allows the clinician to predefine a maximum side-effect probability, constraining the number of possible solutions (all solutions with side-effect probabilities greater than the predefined threshold are discarded), and to define whether solutions should aim at maximising suppression of tremor or akinetic-rigid symptoms. StimFit was embedded in a graphical user interface for a clear and streamlined use (appendix p 4). The algorithm is supplied as open-source code.
Figure 1Image-based optimisation of DBS parameters
In short, electric fields were simulated for each stimulation setting (see panel A) and electric field properties (vector magnitudes and directionalities) were used as input features for model training. To obtain suggestions for optimal stimulation settings within our study cohort, electrode locations were provided to the StimFit algorithm (input). For each electrode, 200 monopolar stimulation settings were simulated first, and corresponding electric fields were provided to the predictive models to obtain outcome predictions (grid-search, right upper panel). Monopolar solutions with maximum predicted motor improvements were then used as starting points for a gradient descent algorithm (middle lower panel). Within each iteration of the optimiser, stimulation parameters were slightly adjusted, and corresponding electric fields were simulated to predict stimulation outcome. The optimiser aimed at maximising predicted motor improvement, while being constrained by a maximum side-effect probability. This way StimFit explored complex (multipolar) stimulation settings until stopping criteria were fulfilled. For graphical representation the surface plot shows the optimisation procedure of only two contacts (x and y). In this study, however, StimFit optimisation was applied on octopolar electrodes in all patients and the number of contacts could in theory be scaled up, limited by computational power only. The resulting stimulation settings (output) were applied in our crossover design. DBS=deep brain stimulation. MNI=Montreal Neurological Institute. ANTs=Advanced Normalization Tools. PaCER=Precise and Convenient Electrode Reconstruction for Deep Brain Stimulation. DiODe=Directional Orientation Detection.
In the present study, we prospectively assessed the clinical effects of DBS parameter suggestions made by StimFit and compared them with the ones derived during standard of care (SoC) programming strategies.
Results
Figure 3Trial profile
Trial profile depicting patient recruitment and analysis. Reasons for exclusion (total n=16) during screening of medical records (n=7) or recruitment (n=9) were the inability to undergo dopaminergic withdrawal (n=7), severe neuropsychiatric symptoms (n=3), cerebral atrophy (n=2), or cognitive impairment (n=1), as well as surgical complications such as electrode replacement (n=1) and the development of a large cyst near the electrode (n=1). One patient was reported to have died. Of note, all patients who were enrolled have completed the study. SoC=standard of care.
TableDemographic and treatment data
Data are n (%) or mean (SD; range). SoC=standard of care. MDS-UPDRS-III=part III of the Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale. LEDD=levodopa equivalent daily dose.
Figure 4Primary outcome and summary statistics
(A) Violin plots showing total MDS-UPDRS-III scores under OFF (grey), SoC (purple), and StimFit (light blue) stimulation conditions on the left, as well as differences in 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 (total score) and symptom-specific MDS-UPDRS-III subscores. Mean absolute differences between both ON-stimulation conditions are shown together with their 95% CIs: total score: –1·6 (SD 7·1; 95% CI –4·0 to 0·9); akinesia-rigidity score: –0·2 (SD 4·4; 95% CI –1·7 to 1·3; p=0·98; n=35); tremor score: –1·4 (SD 3·3; 95% CI –2·7 to –0·1; p=0·046; n=28); and axial score: –0·2 (SD 2·0; 95% CI –0·9 to 0·5; p=0·67, n=34). The 95% CI of the total score did not include the margin of –5 points, establishing non-inferiority. SoC=standard of care. MDS-UPDRS-III=part III of the Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale.
Figure 5Secondary outcome 1—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.
Figure 6Secondary outcome 2—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. One patient was accidentally unmasked before the first ON-stimulation assessment by looking at the patient programmer and was therefore excluded from the self-assessment analyses. SoC=standard of care.
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 mA [SD 0·16], ranging from 0·3 to 0·7 mA). Delayed onset dyskinesias appeared in three patients (located at the head, shoulder, and leg) under StimFit 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 (head and shoulder) but did not affect motor assessments.
To obtain StimFit settings, computational time for image normalisation and electrode reconstruction using the Lead-DBS toolbox was around 45 min per patient, plus an additional 5 min to visually check the results. Other commercial or non-commercial software packages to reconstruct DBS electrodes could in theory be used as well and corresponding computational and visual inspection times might differ accordingly. Reconstructions were then provided to the StimFit algorithm. In order to predict stimulation outcome, StimFit requires patient-specific E-field templates, which are automatically generated once (taking around 45 min) when processing a patient for the first time and stored in the patient folder for later use. StimFit converged to bilateral suggestions of stimulation parameters within approximately 50 min in all patients. Of note, since SoC treatment was conducted before study recruitment, the exact times spent for clinical parameter optimisation in each patient were not available and a direct comparison between both approaches regarding the programming time necessary to achieve optimal settings was not possible.
Discussion
In line with this, previous long-term SoC treatment had led to reduction of dopaminergic medication (LEDD) by 57% compared with preoperative treatment in our cohort.
However, the therapeutic potential of these technical innovations is currently limited by the small number of parameter combinations that can be explored in clinical routine with trial and error.
,
,
,
prospective applications have remained sparse. Frankemolle and colleaugues used image-based models to identify stimulation settings that would minimise spread of electric current to non-motor regions of the STN and 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 could markedly reduce programming time compared with standard clinical procedures.
,
,
,
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 patients with Parkinson’s disease treated with STN-DBS based on electrode location in a fully automated fashion. This approach would allow the possibilities of the latest electrode designs for complex parameter settings to be exploited without being constrained by the complexity of clinical programming procedures, and would open up avenues for further development of multisegmented DBS electrodes.
Specifically, in Parkinson’s disease, multiple recent publications point towards an anatomical segregation of DBS “sweetspots” for suppression of tremor on the one hand and akinesia and rigidity on the other hand.
This has potential implications for personalised DBS programming procedures since optimal DBS parameters would depend on individual patients’ symptom profiles. Within the StimFit software, this concept is implemented by allowing the optimisation objective (motor symptom control) to be modified to maximise the predicted therapeutic effects on tremor or akinetic-rigid symptoms on a continuous spectrum (appendix p 4). In the present study, however, we did not consider the presence of tremor in StimFit predictions, forcing the model to find optimal settings based on the cardinal features of bradykinesia and rigidity alone. The reason for this choice was that we aimed to evaluate the effects of automated DBS programming based exclusively on neuroimaging data. Although a multimodal approach, integrating—among other information—the patients’ baseline symptomatology (especially the presence of tremor) could further benefit model predictions, this issue was outside the scope of this trial. The choice to exclude tremor from the optimisation objective resulted in suboptimal tremor response. This supports the hypothesis of an anatomical segregation of symptom-specific stimulation sites and emphasises the importance of taking clinical information into account in imaging-based DBS programming procedures.
No statistically significant difference in battery drain between the two stimulation conditions was found 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.
Future algorithms could incorporate estimated energy efficiency as an additional variable to obtain settings with an optimised battery lifecycle.
Of note, three patients experienced severe dyskinesias under StimFit stimulation. Dyskinesias are usually thought of as “on-target” side-effects and could therefore indicate a high efficacy of StimFit DBS. On the other hand, different approaches to medication adjustment might be needed during long-term StimFit stimulation. 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 effects of varying frequencies, pulse shapes, or durations to be accounted for. We further noticed that the variation of stimulation amplitudes suggested by StimFit was relatively small compared with SoC. Initial amplitudes ranged between 2·3 mA 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 with stimulation amplitude and suggests that our modelling approach for side-effects should be optimised further. Furthermore, 5% of the patients who were screened for eligibility were excluded due to structural brain changes or electrode replacement (figure 3). Although this only represents a small fraction of patients, it is important to note that anatomical abnormalities could impact performance of image-based outcome predictions. Finally, StimFit was trained on a large monocentric dataset of monopolar review data. Retraining the model on fine-grained data from different centres could increase its generalisability and account for potential centre-specific or population-specific characteristics.
In conclusion, results of this prospective, randomised, double-blind, crossover trial showed that application of STN-DBS parameters suggested by a data-driven optimisation algorithm in a cohort of 35 patients with Parkinson’s disease led to a significant reduction of motor impairment compared with OFF stimulation, similar to the effects obtained during SoC stimulation. This finding suggests that data-driven strategies that allow for quantitative predictions of stimulation effects, embedded in mathematical optimisation procedures, could govern future programming strategies. Additional longitudinal studies are required to confirm long-term motor benefit and to assess the impact of data-driven DBS programming on quality of life, dopaminergic medication, and programming time.
JR, TAD, AH, and AAK conceptualised the study. Funding was acquired by JR and AAK. JR, JA, JLB, A-PK, G-HS, KF, and PK contributed to data acquisition. JR conducted the statistical analysis and data visualisation. JA and JLB 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 had full access to all the data in the study, had final responsibility for the decision to submit for publication, 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.