This study uses data from 104 patients with Parkinson’s disease who underwent bilateral implantation of deep brain stimulation electrodes (Medtronic 3,389, Medtronic B33 series, or Abbott 6,172), targeting the STN. In total, MRI data from 104 patients were processed, with 87 patients also having corresponding clinical review information and at least one hemisphere with documented stimulation settings. This yielded 174 implantation instances meeting all inclusion criteria for direct comparison.

All participants provided written informed consent to participate in the study upon enrolment. The study was approved by Ethics Committee of the General University Hospital in Prague (case number 59/18) and conducted in alignment with the Declaration of Helsinki. The data were sourced retrospectively from records of the iTEMPO department of the Neurology clinic, General University Hospital in Prague. For this dataset, the (average ± standard deviation) age at PD onset was 45 ± 8.6 (n = 64), and the age at DBS surgery was 56.2 ± 8.8 (n = 87). The surgeries for the DBS implants were performed between 2007 and 2023. A discrepancy between the numbers of patients with a PD onset record compared to the DBS surgery record is noted, due to incomplete reporting for some patients (hospital transfers).

We incorporated clinical evaluations of electrode contacts, done during initial stimulation setup at a clinical visit after implantation. The records indicated that the clinical review was performed in four contact group configurations per hemisphere: (i) individually activated edge contacts, (ii) separately tested middle non-directional contacts, or two groups of three radially arranged directional contacts in the case of directional leads. These groups were stimulated across incremental current levels ranging from 0.5 mA to 4.0 mA in 0.5 mA steps, with assessments focused on clinical improvement in rigidity, akinesia, and tremor, as well as thresholds for adverse effects.

Statistical analyses were conducted using the SciPy and Pingouin libraries, with Bonferroni-corrected Wilcoxon signed-rank tests for pairwise comparisons.

Pre-operative (3 T) and post-operative (1.5 T) MRI images (Nifti format) were imported into the Lead-DBS toolbox (Horn and Kühn, 2015), for atlas co-registration (using the SPM method; Friston et al., 2007), normalisation (using the ANTs method; Avants et al., 2008 with SyN nonlinear transform and mutual information metric), subcortical brainshift correction (Schönecker et al., 2009), and electrode reconstruction performed manually in Lead-DBS. The reconstructions were performed and validated by an expert neurologist.

All subsequent volume and coordinate processing was done after conversion back to the native patient space for each subject.

The proposed method, outlined below, consists of two steps: (i) contact selection based on the spatial configuration of the STN and the stimulation lead, and (ii) current selection, based on modelling of the Volume of Tissue Activated (VTA). In the second step, two variants of the method were evaluated: (a) geometry-based method only, and (b) a variant considering the clinical review with test stimulation in addition to the geometry-based method. The method is available as a standalone python-based GUI tool (Supplementary Figure S2; Mikroulis, 2025).

The aforementioned method is implemented as a Python-based GUI (Supplementary Figure S2) app (Mikroulis, 2025) and provides the recommended settings in CSV format. These results include the contact selections (for up to three directional contacts in Medtronic B33-series and Abbott 617x leads, by default), the current estimates for the 200 mV/mm median electric field threshold in the motor STN, and the corresponding overlaps with the STN subregions as fractions. Additionally, to facilitate exploration of the space of possible parameter settings by the user, the VTAs for four more currents (0.8, 1.6, 2.4, 3.2 mA) can be optionally calculated (for a total of up to six current—VTA pairs). The resulting parameters, including subregion overlaps, are interpolated using a cubic spline fit over the range of 0.5 to 4 mA for the three subregions of the STN and presented in an interactive window. An additional assisting tool highlights a range of current values to optimise motor STN coverage against VTA outside the motor STN subregion, limbic STN coverage, and associative STN coverage, using a selectable minimum percentage of the corresponding harmonic mean range (over the entire current range).

The contact and current settings, suggested by the proposed method, were compared against the clinical DBS settings used by patients. As most patients were followed longitudinally, often with multiple adjustments over time, only the earliest complete set of stimulation parameters was analysed to minimise confounding effects from disease progression, electrode drift, or current increases. For patients programmed using voltage-controlled stimulation, current values were derived using the recorded electrode impedance. In cases involving interleaved stimulation, the effective current was calculated as a weighted average based on the duty cycle duration, pulse duration, and frequency of each setting.

The contact selection between the two methods (algorithm-based and clinical settings) was compared with a Jaccard index. For each implanted electrode, the cross-method Jaccard index was calculated from the ratio of coinciding contacts selected by both methods over the set of all selected contacts by either method. For directional contacts, all the subcontacts were considered for this calculation.

Motor symptom improvement was estimated using a subset of patients (n = 50) who underwent UPDRS-III evaluation of DBS efficacy within 3 years post-surgery. Of these, a smaller group (n = 18) had their first UPDRS-III assessment conducted during the same visit in which DBS settings were first recorded (6–12 months postoperatively), while the remaining patients had delayed assessments occurring 1 to 3 years after surgery. For each patient, the relative motor improvement was calculated using paired UPDRS-III scores with stimulation turned on and off during the same visit:

Two related variables were used to predict the improvement: (i) normalised motor-STN coverage as a measure of contact selection effectiveness, and (ii) current (mA) as a measure of stimulation strength affecting the size of the VTA.

A random forest regression model was trained using a maximum tree depth of seven splits, as performance plateaued beyond depth six and showed no further improvement at depths eight or nine. To reduce variance and stabilise predictions, the model was configured with 5,000 estimators.

In total, the algorithm was evaluated on 174 implantations from 87 patients. Figure 1 shows examples of the resulting stimulation settings, visualized using VTA approximation in Lead-DBS (Figures 1A,B), or directly using the thresholded electric field data from OSS-DBS (Figures 1C,D, see 3D models in the Supplementary material).

The contact selection was assessed in comparison to the clinical settings, in two different scenarios: (a) without any clinical information, and (b) using clinical review data (weighted at 50%). The contact selection overlapped in approximately 80% of the cases (Figures 2A,B). To quantify the overlaps, for each implant, a Jaccard index was calculated between the expert-selected contacts and the algorithm-selected contacts (expressed as the intersection of the two contact sets over their union). This quantification showed slightly increased partial overlaps with the expert contact selections when the clinical review information was used (Figure 2C).

The VTAs generated with the algorithm-selected contacts and current outperformed the manual expert settings in terms of motor STN coverage (Hedges’ g > 0.94, Wilcoxon p < 5e-13) and VTA containment within the motor STN (g > 0.46, p < 2e-10; Figures 3A,C—detailed breakdown presented in Supplementary Table S1). Although the algorithm-selected currents were similar to the expert settings (Supplementary Table S1), we evaluated the relative efficacy of the contact selection by normalising the calculated VTA overlap with the motor-STN subregion (Figure 3B), and the portion of the calculated VTA within motor-STN boundaries (Figure 3D), dividing with the algorithm-selected currents, and compared them with the corresponding motor STN and VTA fractions when using the expert-selected settings—also normalised by the current indicated in the expert settings. For both comparisons we observed an improvement in terms of the normalised overlaps over the expert settings. The difference was particularly pronounced for the motor STN coverage (Hedges’ g > 0.79, p < 2e-20), and smaller for the VTA containment within the motor STN (g > 0.07, p < 0.03). The inclusion of clinical review data had a minor influence on these measures (motor STN coverage: g = 0.17, VTA containment: g = 0.19).

To evaluate the potential benefit of the calculated settings on patient outcomes, we compiled a subset of 50 patients with a recorded DBS evaluation using UPDRS-III in the first 3 years after surgery. An improvement ratio was calculated from the difference of UPDRS-III scores with DBS stimulation on vs. DBS stimulation off, normalised by the DBS-off score. To estimate the predicted improvement associated with the calculated settings, we trained a random forest regressor using two predictors: current-normalised motor STN coverage (reflecting contact selection effectiveness), and the stimulation current applied.

The regressor predicted the observed improvements with MSE = 0.0095, R² = 0.692 (Supplementary Figure S1A). Predicted improvement scores were comparable between clinical (median: 39.5%) and algorithm-derived settings (36.2% without clinical review data—39.0% with clinical review data, n = 174; see Supplementary Figure S1B). This predicted similarity between clinical and algorithm-derived settings was consistent regardless of whether clinical review data were incorporated, with Hedges’ g ranging from 0.05 to 0.08 and Wilcoxon p-values between 0.096 and 1 (Supplementary Figure S1B). We noted, however, a minor (g = 0.12) but statistically significant difference (p = 0.004) between the predicted improvements of the two algorithm-derived setting sets, with the incorporation of the clinical review information yielding slightly increased predicted improvements, closer to the expert-selected clinical settings.

Contact selection performance was evaluated by normalising the overlap between the calculated VTA and the motor STN by the stimulation current, yielding a measure of motor STN coverage per mA. The algorithm-derived settings demonstrated a small but statistically significant increase in current-normalised motor STN activation compared to clinically selected contacts. This suggests that the proposed method identifies more effective contact configurations––at a given implant position––for selectively targeting the motor subregion of the STN.

Similarly, we calculated the motor-STN and VTA overlap as a fraction of the VTA, also normalised with the current. Comparing this variable between manually selected clinical settings and the calculated settings showed that a larger portion of the VTA (per current unit) is contained within the target region (motor STN) when using the calculated settings compared to the manually adjusted settings. This significantly reduces electric field leakage in the milieu, by containing more of the high-intensity area within the target region.

Taken together, the differences in these two variables show better selective activation of the target area (motor STN in this case) when using the calculated contact selection.

Exact matches occurred in only nine implantation instances without clinical review input and in 10 instances when clinical review data were included, due to the procedural difference in contact selection between the two methods. The algorithm, with the settings used in this study, selects up to three sub-contacts out of a maximum of eight (in the case of directional electrodes). The clinicians, however, predominantly select entire (sub-)contact groups, matching the clinical review resolution. Thus, the only perfectly matching cases are limited to non-directional electrodes where up to two (out of four) contacts are selected by both methods.

The algorithm’s current selection strategy aims to achieve a median electric field intensity in the target structure equal to the VTA threshold, optimising stimulation precision. This approach yields similar or better absolute coverage of the motor STN compared to manually adjusted clinical settings, and current-normalised analyses reveal greater efficiency per unit of current. To mitigate potential under-estimation or over-estimation of the current with the automated method and allow further fine-tuning, we provide an additional tool to facilitate exploration of different currents and their effects on STN subregion coverage.

Most recently proposed methods for initial DBS settings optimisation aim to improve the patient outcome directly (Boutet et al., 2021; Qiu et al., 2024; Roediger et al., 2022) or approximate a desired volume of tissue activation (Connolly et al., 2021), using statistics from large patient data pools. By contrast, our method aims to optimise the stimulation of brain structures (STN in this instance) modulating the symptoms of the disease. Among data-driven methods, our approach most closely aligns with the StimFit toolbox (Roediger et al., 2022), which also leverages patient-specific MRI data for automated DBS programming.

Our predicted improvement, predicted to be similar to expert settings based on retrospective analysis, is also comparable with the non-inferiority to standard care achieved by the StimFit solution (Roediger et al., 2023). However, our method introduces a few advantages: it integrates clinical review data when available, explicitly targets the motor subregion of the STN to optimise motor symptom control and operates without the need for pre-trained statistical models––substantially reducing computational and data acquisition demands.

Another closely related anatomy-guided approach is the sweet spot-guided algorithm proposed by Nordenström et al. (2022), which uses Lead-DBS for lead reconstruction and VTA overlap with an empirical sweet spot (derived from prior cohorts) to suggest lead levels, contacts, and effect thresholds for rigidity reduction. While this method provides interpretable suggestions via a Matlab/Lead-DBS-based GUI, it relies on approximations like spherical VTA models for speed, which may affect precision. Our pipeline builds on similar foundations but opts for direct targeting of the motor STN centroid from atlas-based parcellation, relying on OSS-DBS for higher-resolution electric field modelling, and incorporating multi-symptom clinical reviews, enabling further patient-specific fine-tuning without predefined sweet spots.

Other state-of-the-art approaches, such as Bayesian optimisation and particle swarm optimisation, have been proposed to efficiently explore the stimulation parameter space. Bayesian optimisation has been applied to reduce the number of required trials by iteratively minimising rigidity based on prior clinical responses (Louie et al., 2021), while particle swarm optimisation method was designed to address complex electrode configurations with segmented leads, and optimise stimulation amplitudes (Peña et al., 2018).

However, these techniques are primarily outcome-driven and do not currently provide structural substrate-centric metrics or predicted improvement values, limiting the feasibility of direct comparison with our anatomy-based method. Similarly, multimodal biomarker-guided methods, like that of Shah et al. (2023), combine resting/movement-state LFPs with imaging to predict clinical efficacy and side-effect thresholds, achieving ~90% accuracy in optimal contact selection but requiring intraoperative electrophysiological data not routinely available in our workflow. Connectomic-based programming, as in Hines et al. (2024), optimises pathway recruitment using detailed axonal models, yielding 43.5% UPDRS-III improvements, but requires advanced computational resources compared to our simulation-focused approach. fMRI-driven algorithms, such as Santyr et al. (2024), demonstrate non-inferior motor outcomes to standard care in a blinded trial but rely on functional scans, contrasting our emphasis on routine anatomical MRI for broad clinical accessibility.

The method relies on pre-operative and post-operative MRI data. Since the brain structures are directly derived from the pre-operative images and the electrode position is derived from the post-operative scan, the output is sensitive to the time delay between pre- and post-operative inputs. Increased time intervals between the pre-operative MRI scans and the post-operative scans may introduce errors both in electrode localisation relative to the target structures, due to random electrode drifts, as well as changes in the localisation of the brain structures themselves, as a result of ageing or disease progression. This is expected to progressively reduce the accuracy of the method for long-term parameter adjustments, without error corrections for the MRI inputs.

Our method considers two parameters: contact selection, and current selection. Multiple additional parameters have to be assigned by the clinicians, including frequency of stimulation, pulse width/shape, possible interleaved stimulations, and bipolar settings. Most of these parameters are usually preset to well-documented values, and are only changed later in the decision process, if necessary (Picillo et al., 2016). Accordingly, we prioritised the most time-intensive component of initial DBS programming: contact and current selection.

In the case of current estimation, our method targets the centroid of the motor STN subregion while applying predefined lower and upper bounds to ensure clinical plausibility. Thus, the suggested current values should be interpreted as an informed starting point rather than definitive therapeutic settings.

Additionally, in the present study, the input of the clinical review data was weighed equally with the geometry-based calculation. This may be unsuitable if the clinical review information is highly prioritised, for instance, to ensure a minimum stimulation effectiveness or minimising side effects. In such scenarios, increasing the weighting of clinical inputs may enhance the method’s alignment with clinical priorities and patient-specific therapeutic goals. Furthermore, future prospective studies with standardised side-effect logging could correlate non-motor STN coverage with specific adverse effects.

Finally, this study was retrospective in nature, with predicted motor improvement derived from a subset of patients—limited by digitised records availability—who had early UPDRS-III evaluation records following DBS implantation. Although the predicted outcomes are consistent with those reported by comparable methods (i.e., similar to expert settings), a prospective clinical trial will be necessary to validate the effectiveness of this approach with greater certainty.