Neurofeedback-enabled beta power control with a fully implanted DBS system in patients with Parkinson’s disease

(cid:1) Subthalamic oscillations can be streamed for neurofeedback. (cid:1) Parkinson patients can control subthalamic oscillations using neurofeedback and improve this control with training time. (cid:1) DBS electrode-guided neurofeedback is effective even with concurrent stimulation and medication.


Introduction
Neurofeedback (NF) has been acknowledged as a novel approach in the management of neurological and psychiatric disor-ders (Sitaram et al., 2017;Hampson, Ruiz, and Ushiba 2020).It relies on the real-time extraction of relevant features from neuronal activity, that are presented to the subject in real time, who can then develop endogenous techniques to self-regulate this ongoing brain activity.In fact, NF has already been explored for Parkinson's disease (PD) by numerous studies using electrophysiological and functional imaging methods (Anil et al., 2021).Most of the electrophysiological NF studies used oscillations of electrical potential in beta frequency  as a feedback signal.These oscillations are pronounced in the local field potential (LFP) of the subthalamic nucleus (STN) and internal globus pallidus (GPi) of PD patients (Brown et al., 2001;Cassidy et al., 2002;Brown and Williams, 2005;Hammond, Bergman, and Brown 2007), even though they are not exclusively localised in the basal ganglia, as coupled oscillatory signals are also found in the cortex (Williams et al., 2002;Brown and Williams, 2005;Hammond, Bergman, and Brown, 2007).
Beta oscillations can be modulated by PD treatment: Levodopa (Brown et al., 2001;Levy et al., 2002;Doyle et al., 2005;Alonso-Frech et al., 2006) and STN deep brain stimulation (DBS) (Brown et al., 2004;Wingeier et al., 2006;Kühn et al., 2008;Bronte-Stewart et al., 2009;Eusebio et al., 2011) each induce a modality-specific decrease of beta frequency activity in the STN (Muthuraman et al., 2018).Moreover, adaptive and conventional DBS have differential effects on the temporal dynamics of beta activity, with only adaptive DBS selectively truncating long beta bursts (Tinkhauser et al., 2017), i.e. temporary increases of beta oscillation amplitudes.The decline of beta power in the STN induced by levodopa correlates with a decrease in the bradykinesia-rigidity subscore of the Unified Parkinson's Disease Rating Scale (UPDRS) of the contralateral side (Kühn et al., 2006b;Ray et al., 2008;Kühn et al., 2009).A correlation of STN DBS induced beta power decrease and motor improvement has also been demonstrated (Kühn et al., 2008;Oswal et al., 2016).Conversely, there is also evidence that external synchronisation of STN LFP with stimulation at 20 Hz disturbs previously intact motor ability in PD patients (Chen et al., 2007).
The connection between beta oscillations in the STN and movement is extensively characterised as well.Without dopaminergic medication, voluntary movements lead to a reduction of LFP beta frequency oscillations (Cassidy et al., 2002).Particularly, in a study investigating finger tapping movements, intrinsically triggered movements showed a continuous decline of beta power during movement whereas movements triggered through extrinsic signals showed a short-term decrease dependent on the tap onset (Bichsel et al., 2018).Interestingly, the decline of beta frequency activity starts prior to the movement (Cassidy et al., 2002;Kühn et al., 2004;Doyle et al., 2005).This premovement decline is magnified and starts sooner through levodopa administration and correlates with motor ability (Kühn et al., 2004;Doyle et al., 2005).Due to these results, it was suggested that this decline in beta activity is a correlate of movement preparation (Kühn et al., 2004;Doyle et al., 2005).Moreover, a correlation between decreased betaactivity and motor planning as well as execution have also been extensively described (Oswal et al., 2012;Tan et al., 2015).Supporting this theory, kinesthetic motor imagery without movement causes a decrease in beta power comparable to performed movements (Kühn et al., 2006a).
Having multiple relationships to PD, beta oscillations are suggested as a possible biomarker (Little and Brown, 2012).Indeed, beta oscillations were successfully used as a feedback signal for adaptive DBS in PD patients, thereby achieving greater motor improvement while also relieving battery strain compared to continuous DBS (Little et al., 2013).More recently, high beta oscillations as well as phase-amplitude coupling between high beta and high-frequency oscillations have been identified as promising biomarkers for future investigations in the intriguing field of adaptive DBS for PD (Neumann et al., 2023;Bockova et al., 2024).Interestingly, electroencephalography (EEG) NF training targeting beta bursts in the sensorimotor cortex showed a significant decline of beta bursts and faster motor response in healthy subjects (He et al., 2020a).Furthermore, there are NF studies in PD patients providing signals through invasive methods.A successful modulation of beta oscillation with a completely implanted electrocorticogra-phy (ECoG) electrode over the sensorimotor areas has been demonstrated (Khanna et al., 2017).Most of the NF studies applying invasive methods, though, make use of the electrodes implanted for DBS to provide information about the current state of beta oscillations in the STN.One study showed a significant suppression of beta frequency oscillations in the STN intraoperatively (Fukuma et al., 2018).Moreover, three studies were conducted in a postoperative setting making use of the time window between the two operations of DBS electrode implantation and neurostimulator implantation, when the DBS leads were externalised and therefore available for recording (He et al., 2019(He et al., , 2020b;;Bichsel et al., 2021).One study showed a partly significant control over the appearance of beta bursts in the STN (He et al., 2019).Another study of the same group confirmed the significant capability of NF to diminish beta bursts in the STN leading to an improvement in a motor task (He et al., 2020b).Moreover, we previously showed that NF using externalised DBS electrodes enabled patients with PD to rapidly gain beta-regulatory control, which in turn quickened movements in a hand pro-and supination task (Bichsel et al., 2021).
The experimental setup using externalised leads was valuable to demonstrate the efficacy of DBS NF, but is not suitable for every-day use as the leads ultimately must be internalised.Here, to our knowledge for the first time, we report on NF training with an entirely internalised DBS system to replicate the results of DBS NF studies using externalised DBS electrodes and provide evidence that NF is feasible in a setting similar to daily life.This prespecified study was a priori designed and powered.

Setting and participants
The participants consisted of seven PD patients and one patient (highly conspicuous family history) with familial PD (five male and three female patients, mean age +/-standard error of mean 68.5 +/-2.38 years) receiving DBS electrode implantation into the STN.They were numbered with their patient identification (PID) number.A sample size of 6 patients was calculated a priori for finding a significant difference between a non-regulating and NFregulating setting at significance level of a = 0.05/2 and statistical power of 0.8 (Bichsel et al., 2022).The estimate of the effect size was derived from the proof-of-principle study (Bichsel et al., 2021) with externalised electrodes using the same primary endpoint (effect size = 1.94 for beta-power estimates during downregulation vs. rest in the last neurofeedback session).DBS electrodes were implanted into the STN at the Department of Neurosurgery of the University Hospital Zurich.The recruitment for this study took place postoperatively.The NF trials were conducted on three separate days at the University Hospital Zurich or Klinik Lengg between day 2 and 35 after electrode implantation.Here we report on the outcome of the first day, for which the study has been a priori designed and powered.Medication or stimulation load were not affected by study participation and determined according to the treating neurologists.The post-operative rehabilitation procedure did not differ from that of non-participating patients.The details of the participants are summarised in Table 1.

Surgery and implanted materials
DBS electrode implantation corresponded to the surgical procedure as previously described (Bichsel et al., 2021).Leads of the model B33005 SenSight TM (Medtronic, Minneapolis, Minnesota, USA) or 3389 (Medtronic, Minneapolis, Minnesota, USA) were implanted into the STN.The leads were connected via the extensions of the model B34000 SenSight TM (Medtronic, Minneapolis, Minnesota, USA) or 37,086 (Medtronic, Minneapolis, Minnesota, USA) to the neurostimulator Percept TM PC B35200 (Medtronic, Minneapolis, Minnesota, USA) during the same surgical procedure or in a second operation on a later date.Information about the DBS system is summarised in Table 2.

Feedback setting
The participants' task during the NF trials was to gain control over pathological LFP beta oscillations while receiving visual feedback of their own current beta power in the STN.The fully implanted DBS system provided this information wirelessly over the communicator (model 8880 T2, Medtronic, Minneapolis, Minnesota, USA) to the tablet clinician programmer (CT900D, Medtronic, Minneapolis, Minnesota, USA) with the clinician programmer software application (A610 3.0, Medtronic, Minneapolis, Minnesota, USA) visualising the data through BrainSense TM Technology (Medtronic, Minneapolis, Minnesota, USA).
The NF tasks were executed only on the symptom dominant side, therefore the information about the current beta power was only obtained in the STN contralateral to the symptom dominant side.In one case (PID 3), the recording occurred on the STN ipsilateral to the symptom dominant side, as the contralateral STN was not suitable for NF training due to excessive artifacts.
In order to provide visual feedback, a frequency range of locally elevated beta oscillatory activity, referred to as beta peak, was determined in a sitting position at rest using the BrainSense TM Survey function on the Medtronic tablet (which takes a real time 30 s averaged fast Fourier transform while stimulation is turned off) and selected manually in the BrainSense TM Setup function as the frequency range of beta peaks shows interindividual variability (Kühn et al., 2009;Darcy et al., 2022).The selected STN side and peak frequency for each patient is listed in Table 2.For PID 6 and 7, peak frequencies in the alpha frequency (8-12 Hz) were selected as no local maximum within the beta frequency was visually identified.It has been hypothesised that when no low beta peak (in the frequency range of 13-20 Hz) is available, it can appear in the high Demographic and clinical data of participants.ON and OFF International Parkinson and Movement Disorder Society (MDS) À UPDRS III scores were the result of a preoperative test (Goetz et al., 2008).Levodopa equivalent dose (LED) was calculated using conversion factors from (Tomlinson et al., 2010;Schade, Mollenhauer, and Trenkwalder, 2020).Benserazid and Carbidopa were not considered for calculation (Nyholm and Jost, 2021).Lacking information on the LED of Madopar DR Ò (Roche Pharma (Schweiz) AG, Basel, Switzerland), its LED was assumed to be equivalent to the one of Madopar HBS Ò (F.Hoffmann-La Roche, Basel, Switzerland) multiplied with the increased bioavailability of approximately 40% (Gasser et al., 1998).Madopar HBS Ò was deemed being of controlled release character (Grahnén et al., 1992) based on (Erni and Held, 1987) alpha frequency (Darcy et al., 2022).According to the BrainSense TM Technology Summary document, +/-2.5 Hz around the manually determined beta peak was used for power calculation.The power of the beta frequency range selected in such way will herein be referred to as beta power.
The bipolar electrode configuration for LFP recording was selected in the BrainSense TM Setup function.The two electrodes adjacent to the stimulation electrode (suggested by the program as 'active therapy') were chosen for recording (see Table 2).As most of the patients were receiving DBS according to their neurorehabilitation plan, simultaneous stimulation and recording was performed.
Visual feedback was provided by the BrainSense TM Streaming function of the BrainSense TM Technology on the tablet Medtronic clinician programmer.Thereby, BrainSense TM Technology determined the beta power at a rate of 2 Hz for the selected beta frequency range based on the LFP data (recorded at 250 Hz) and displayed it in a continuous graphical form with time on the xaxis and beta power on the y-axis, as explained in the BrainSense TM Technology Summary document.The default setting of the Brain-Sense TM Streaming function was not changed.This function displayed two graphs of beta power, one showing the activity of the last approximately 6 s, the other showing the entire streaming session with the last approximately 6 s highlighted in a brighter blue colour.Patients were advised to focus on the brightly highlighted part of the graph displaying the entire streaming session, as this was deemed to provide the best visual information for NF.With time on the x-axis, the scale changed with the highlighted part getting compressed with the passage of time (cf.2.3.3.experimental procedure for the strategy to ensure homogeneous visual feedback).

Tasks
The main tasks were constructed identical to our previous study (Bichsel et al., 2021).The NF trials consisted of three different tasks: baseline, downregulation and upregulation.In the baseline task, baseline activity in the selected beta frequency range was recorded.Participants did not get any visual feedback, were requested to avoid intentionally regulating beta power and refrain from movements.For down-and upregulation tasks, patients received visual feedback.In downregulation tasks, patients were instructed to pay attention to their current beta power and try to lower the curve, i.e. the beta power, by imagining a certain situation, whereas in upregulation tasks, the goal was to drive up the curve.Movements during these tasks were to be avoided by the patients as well.Participants were given an initial strategy to achieve downregulation by imagining asymptomatic movements or situations, and upregulation by imagining situations in which they were afflicted by PD symptoms.
The upregulation part served as the control task for the clinically important downregulation part.This bidirectional NF design has been suggested as a possible control method, as it can expose erroneously achieved NF (Sorger et al., 2019) and was also utilised in two previous DBS NF studies (Fukuma et al., 2018;Bichsel et al., 2021).
Furthermore, two blocks without NF were integrated.These blocks each consisted of the three tasks baseline, downregulation and upregulation as well, but the patients did not receive visual feedback for any of these tasks and were asked to regulate the beta power solely by the application of mental strategies.One non-NF block was located before the NF tasks (pre-NF) to record the initial regulatory ability before NF training and one non-NF block after the NF tasks (transfer) to demonstrate how the regulatory ability is modified by the NF training.

Experimental procedure
Each session consisted of five blocks: one pre-NF block, three NF blocks, and one transfer block.Both NF and non-NF blocks were made up of two runs, each consisting of the tasks baseline, downregulation, and upregulation, with each task lasting 60 s.The order of the tasks in the first run of the block was baseline, downregulation, upregulation, whereas the second run had the sequence baseline, upregulation, downregulation.One BrainSense TM Streaming recording was congruent with one run and therefore had a length of 180 s.All the tasks were performed in sitting position.The experimental procedure is summarised in Table 3.As changes in the scaling were especially pronounced in the first minute, a session always started with a baseline task during which no visual neurofeedback was provided.For the last two minutes, the effect of the scaling x-axis was much less pronounced.Still, we alternated the sequence of downregulation and upregulation for every run such as to correct for any possible confounding effect.The alternating order of down-and upregulation was also important to balance out a carry-over effect (He et al., 2020b) as well as the concentration level.
Initial mental strategies to be employed in the pre-NF block as well as the first half of the NF1 block were provided.For downregulation, participants had to imagine smooth fist forming/opening of the hand contralateral to the recorded STN.The upregulation strategy consisted of a specific self-experienced symptomatic situation.To visualise the neurophysiological mechanisms, a movement-induced beta decline was shown to the patients after the second run.From the second half of the NF1 block onwards, the mental strategy was not predefined, but patients were reminded to develop a strategy imagining situations without Table 3 Experimental procedure.A session was defined as the entirety of measurements conducted for one patient on a single day.A session consisted of 5 blocks from which the first and the last block were non-NF blocks.(see comment) Each block consisted of 2 runs, of which each contained the tasks baseline, downregulation and upregulation.The sequences of tasks were different in the first and second run of a block.symptoms for downregulation and situations with symptoms for upregulation.

Block
After the runs in the NF blocks, patients were asked to rate their impression of control they had for the down-and upregulation tasks on a scale from 0 (no ability to control beta power) to 10 (perfect ability to control beta power), herein referred to as agency.

Data and statistical analysis
BrainSense TM Technology stored the session data in a JSON file.The JSON file was then imported to MATLAB (R2021a Update 4, The MathWorks, Inc., Natick, Massachusetts, USA) for analysis, statistical evaluation and graphical illustration.Analysis concerning the difference of beta power between tasks/blocks and the course of beta power during a run was conducted with data stored under the struct 'BrainSenseLfp'.For the visualisation of the course of beta power during runs (Fig. 4), the 'BrainSenseLfp' data was normalised by the median beta power value of the baseline task of the same run.The median normalised value of all patients was then determined for every time point of a run, further analysed and graphically illustrated.
To calculate the beta power of the task (Figs. 2, 3, 7, 8), the median beta power of each task was determined, and these values were then averaged for the two same tasks of one block (e.g. 2 downregulation tasks in NF1 block).Subsequently, these downregulation values were normalised by the equivalently processed baseline values (Figs. 2, 7, 8A) and respective upregulation values by said downregulation values (Figs. 3, 8B) of the same block.Statistical analysis in Figs. 2 and 3 was conducted with such normalised data.The choice of statistical tests was based on the a priori stated tests for the primary endpoint, our previous study as well as the specific research question and assumptions underlying each analysis.We conducted a paired-sample t-test for evaluating the difference between tasks (with Shapiro-Wilk test for normality testing where relevant) while the difference between blocks was determined by Wilcoxon signed rank test.Furthermore, linear mixed effects analysis of the relationship between the beta power and the neurofeedback time (pre-NF, NF1, NF2, NF3) was carried out.In order to assess the effect of neurofeedback downregulation time on the beta power, we tested the full model with the effect in question as well as random effect for intercept with subjects as grouping variable against a rudimentary fixed intercept model only.We used a theoretical likelihood test to assess the significance of the random effect.
As two patients were lacking a beta peak after visual inspection and therefore the NF tasks were conducted with a peak located in the alpha band, separate plot visualisations and statistical analyses were carried out excluding these patients (Supplementary Figures 1, 2).
In Fig. 7A and B, the normalised downregulation values of pre-NF, NF1, NF2, NF3 and transfer were averaged for each person.
Influence of number of days since electrode implantation (Fig. 7) and agency (Fig. 8) were investigated for linear relationship.Hereby, agency values of each task were averaged for one block.
The data struct 'BrainSenseTimeDomain' containing the LFP data measured at 250 Hz was used when conducting spectral analysis of power (Figs. 1, 5, 6).The basic spectral analysis algorithm of MATLAB, containing fast Fourier transformation, was applied in all baseline tasks for peak analysis (Fig. 5), and in baseline and downregulation of all NF tasks for comparison of the power spectrum of such two tasks (Fig. 6).The resulting spectral data were averaged for baseline and for downregulation for each patient.Smoothing of the mean spectral data was achieved by calculating the moving average of 250 data points.Local maxima were determined within the frequency range of 13-35 Hz.  for Pharmaceuticals for Human Use (ICH) -Good Clinical Practice (GCP) guidelines and local regulations.For the recruitment to the study, an informed written consent was obtained from the patients.This study was conducted in compliance with applicable data protection regulations.Data protection was ensured by encrypting personalised data at acquisition.The unencrypting key as well as unencrypted data were always held at the University Hospital Zurich.Only involved researchers, subjected to professional secrecy, had access to the unencrypted data when and to the extent necessary for the study conduct.In other cases, data evaluation was conducted with encrypted data.The encrypted data can be made available upon reasonable request.Funding for the conduction of this study was provided by the Dr. Wilhelm Hurka Stiftung.There is no conflict of interest to be declared for this study.

Results
As an exemplary course of measurement, the power spectrum across frequencies against time is graphically illustrated in Fig. 1.
Two exemplary runs of two different patients were depicted in a power spectrum over time.The power spectrum for the frequencies between 8 and 35 Hz was visualised for the entirety of the runs of 3 min.Different tasks within a run are separated with a thin black line.

Effect of downregulation compared to baseline
The median beta power values of the downregulation tasks were compared to baseline data of the same block after normalisation.As shown in Fig. 2, the median values of the downregulation tasks were below 1 (i.e.below baseline beta power) in all blocks (median of normalised beta power values for pre-NF, NF1, NF2, NF3, transfer: 0.8991, 0.8170, 0.8358, 0.6885, 0.8372).In the NF3 task, a median beta power decrease of 31.15% in downregulations tasks was quantified.There was a significant difference of downregulation to baseline tasks in all blocks (paired-sample t-test) except for NF1, where an outlier influences the result (p-value of pre-NF, NF1, NF2, NF3, transfer: 0.0439, 0.1475, 0.0026, 0.00013, 0.0043).Based on the Shapiro-Wilk test for the primary endpoint (NF3 downregulation vs. rest), with a test statistic of 0.88633 and a p-value of 0.18602, the test failed to reject the null hypothesis.This suggests that there is insufficient evidence to conclude that the data significantly deviate from a normal distribution at a typical significance level (alpha = 0.05).Therefore, the data may be reasonably assumed to be normally distributed.From pre-NF to NF3, the median beta power value of downregulation tasks decreased significantly within minutes of learning (Wilcoxon signed rank test p-value: 0.0234).The median downregulation beta power was lower in transfer compared to pre-NF, however not marking a significant difference (Wilcoxon signed rank test p-value: 0.4609).In the linear mixed effects analysis, the fixed effect of neurofeedback time was significant (p-value: 0.0069) and the introduction of the random effect for intercept with subject as grouping variable significantly improved the model (p-value: 0.0238).Similar results were achieved by excluding the 2 patients without a beta peak (median of normalised beta power values for pre-NF, NF1, NF2, NF3, transfer: 0.9230, 0.8170, 0.7743, 0.6885, 0.8372; pairedsample t-test p-value of pre-NF, NF1, NF2, NF3, transfer: 0.0938, 0.3020, 0.0043, 0.0023, 0.0096), thus showing a significant improvement when comparing the normalised downregulation tasks of pre-NF and transfer blocks (Wilcoxon signed rank test p-value: 0.0313; Supplementary Figure 1).

Course of beta power in a run
The median beta power was calculated across all patients for every time point in the NF3 runs and is depicted in Fig. 4. It is visible that, in the first few seconds of all tasks the beta power from the earlier task tended to spill over.In NF3, 1 to 7.5 s elapsed until the median beta power (black graph in Fig. 4) crossed the median of the median beta power values (red line in Fig. 4), calculated for each task, for the first time.

Spectral analysis
All baseline tasks were spectrally analysed to identify the beta peak (Fig. 5).There was no clear beta peak within the beta frequency definition of 13-35 Hz for patients with PID 6 and 7.In all other cases, the beta peak defined as a local maximum tended to be lower than the manually chosen peak (mean difference +/standard error of mean À2.40 +/-0.28Hz).The peaks seemed to gather closely in our cohort (mean peak frequency +/-standard error of mean 22.66 +/-0.28Hz).
The comparison of the spectral analysis of the NF baseline and downregulation tasks showed individual characteristics (Fig. 6).It is visible, that there were patients, where only the frequency around the feedback frequency changed (PID 1, 5, 7), one patient, where there was a broad, continuous decrease (PID 3), patients with a decrease at the feedback frequency and a further decrease in another, lower frequency range (PID 0, 2, 4) and even an example of power decrease at a different frequency (PID 6).

Beta power as a function of time since electrode implantation
The dependency of NF regulation ability on number of days since electrode implantation was investigated in our study (Fig. 7).
The decrease of power achieved in downregulation compared to the cumulative number of days after electrode implantation showed a tendency of a linear relationship (r = À0.6442,p = 0.0847; Fig. 7A).
The same correlation was examined for the relationship of stimulation and beta power decrease through downregulation (Fig. 7B).NF sessions took place in the postoperative rehabilitation phase in which the stimulation was slowly increased over the course of days.The magnitude of the stimulation current did not influence the downregulation ability in the postoperative phase, as it did not show a significant linear relationship (r = À0.4944p = 0.2130).
The NF3 downregulation ability (Fig. 7C) did not hold a significant linear relationship with the number of days after electrode implantation (NF3: r = À0.1281,p = 0.7625).
The learning effect (Fig. 7D) defined as the quotient of beta power during NF3 downregulation tasks and NF1 downregulation tasks revealed no significant linear relationship with the number of days after electrode implantation (NF3/NF1: r = 0.2092, p = 0.6190).

Correlation of patients' agency and regulation ability
The communicated agency was compared with the effective ability to regulate the beta power (Fig. 8).There was neither a significant linear correlation for downregulation (r = À0.1958,p = 0.3592) nor for upregulation (r = 0.0641, p = 0.7660).

Key findings
As the main finding of this study, we were able to show, to our knowledge for the first time, that NF downregulation using a fully implanted DBS system led to a significant reduction of pathological beta oscillations compared to a non-regulation condition.A significant improvement of downregulation ability within minutes of learning was observed (primary endpoint of the study).Although the difference in beta power between upregulation and downregulation tasks remained non-significant, a significant improvement of bidirectional NF control from NF1 to NF3, and, therewith, a significant effect of NF on the regulation of beta oscillations within minutes of learning was demonstrated.This study confirms the results obtained in postoperative NF studies with externalised DBS leads (He et al., 2019(He et al., , 2020b;;Bichsel et al., 2021), while, however, relying on an internalised system with a lower signal quality, different signal processing, a cognitively more challenging neurofeedback visualisation, as well as a longer latency and ongoing stimulation.

Rapid neurofeedback control over pathological deep brain oscillations using a fully implanted system
As summarised in the Introduction, NF is a possible treatment of PD in addition to the conventional therapies, resulting in several studies in this field.So far, invasive NF was conducted via an implanted ECoG system providing information of the sensorimotor cortex (Khanna et al., 2017) as well as via partially implanted DBS leads (Fukuma et al., 2018;He et al., 2019He et al., , 2020b;;Bichsel et al., 2021).Herein, we showed that NF regulation is possible with an entirely implanted DBS system and therefore demonstrated its suitability for daily application.This is important as there may be differences in signal quality/resolution and sampling frequency as well as real-time signal processing.
In this study, we provided evidence for a highly significant beta power decrease during the NF downregulation tasks, which is in line with the findings of the other STN DBS NF studies using externalised leads (Fukuma et al., 2018;He et al., 2019He et al., , 2020b;;Bichsel et al., 2021).It was possible to recreate our previous results (Bichsel et al., 2021) in a similar experimental protocol but using a fully implanted DBS system.

The effect of NF on pathological deep brain oscillations in comparison to current treatments
The extent of beta power decrease is also of considerable importance.In the NF3 block, the participants showed a median decrease of 31.15% of beta power in the downregulation tasks compared to the baseline tasks, the greatest value being 41.85%.For comparison, other studies described a beta power decrease of 58.9% (Kühn et al., 2006b) and 54.3% (Ray et al., 2008) with levodopa and 45.8% with DBS (Eusebio et al., 2011) without any additional therapy.The result of NF regulation in this study is thus remarkable, as it was achieved under pharmacological therapy and DBS application (note that DBS was most frequently not in the optimal stimulation range, as the study took place during the rehabilitation phase, where DBS was increased slowly over days).

On bidirectional neurofeedback control
Even though the difference of upregulation and downregulation tasks improved significantly from NF1 to NF3, the difference stayed non-significant in absolute terms.The increased challenge of upregulation compared to downregulation was also seen in two other studies implementing upregulation tasks (Fukuma et al., 2018;Bichsel et al., 2021).In PD, beta-power is often highest during rest and can be reduced by various means such as motor and cognitive tasks (Cassidy et al., 2002;Kühn et al., 2004;Doyle et al., 2005;Kühn et al., 2006a).Along this line, we have found pronounced beta-activity during rest and diminished beta-activity during downregulation as well as upregulation, since the strategies employed comprised cognitive tasks such as motor imagery.Indeed, the initial upregulation strategies during NF1 had a tendency to be even more efficient at reducing beta-activity than the initial downregulation strategies.Yet, the bidirectional feedback design, an established control method in neurofeedback research, revealed a gradual and significant divergence from NF1 until NF3, with upregulation strategies resulting in higher betaactivity values in 6 out of 8 patients.Nevertheless, these findings indicate that downregulation strategies using imagery of asymptomatic movements work so well, as they suit the property of movement preparation and motor imagery while the upregulation strategy, imagining symptomatic movements or situations, might not trigger an increase of beta power.In Fukuma et al. (Fukuma et al., 2018), where the majority of the participants used strategies other than motor imagery, similar difficulties in upregulation occurred.Therefore, the application of motor regulation strategies might not be the sole explanation of the more challenging nature of upregulation.Ultimately, it might be more difficult to increase an already pathologically increased signal as compared to reducing it.Nonetheless, our findings showing successful bidirectional control of beta power were important to exclude erroneously achieved NF effects (Sorger et al., 2019).

Concerning the choice of the neurofeedback parameter
In the spectral analysis of all baseline tasks, it became visible that not all patients have a beta peak.It was shown that, when withheld from medication, beta peaks are present in the majority of PD patients, whereas beta peak prevalence in the low beta frequency band (defined as 13-20 Hz) decreases to 50% by application of medication (Little and Brown, 2012;Darcy et al., 2022).Interestingly, Darcy et al. (Darcy et al., 2022) stated a 100% prevalence of beta peaks in tremor-dominant PD patients withheld from medication in at least one hemisphere, whereas in our study, with all but one patient receiving medication, the 2 patients missing a beta peak on the investigated hemisphere were tremor-dominant PD patients.
The lack of a beta peak in these two tremor-dominant patients can be discussed from three aspects: Firstly, the frequency definition of the pathological oscillations is inconsistent among studies as some of them include lower frequencies than others.Darcy et al. discussed that the classification of the frequency bands as traditional EEG bands may not suit the parkinsonian electrophysiology, the latter showing a dopamine induced reduction of power in alpha and low beta frequency (Darcy et al., 2022).Furthermore, peaks in the high alpha frequency decrease more strongly by the administration of medication if there is no local maximum in the low beta frequency, and therefore, it has been hypothesised that the low beta peak might also appear in the high alpha frequency (Darcy et al., 2022).In line with these results, it can be observed in Fig. 6, that in 4 out of 6 patients showing a beta peak, the reduction in beta power with NF extended into the alpha frequency.Also, the 2 patients without an identifiable beta peak showed a reduction in power of alpha frequencies.Moreover, participants with NF signals in the alpha frequency achieved similar results as patients with feedback markers in the beta frequency.The absence of improvement of the results by excluding the 2 patients lacking the beta peak is another indication that the strict adherence to the beta band may not be of much benefit.The inclusion of PID7 was based on several considerations: data availability, clinical relevance since beta-peaks are not always observed, as well as the chance to investigate modulatory capacity in another frequency band.The inclusion of the patient does not alter the significance of the primary endpoint (cf.Results).It remains to be explored if the localisation of the feedback parameter in the alpha frequency induces a reduction of symptoms through NF.
Based on the aforementioned electrophysiological mismatch and finding tremor worsening with NF targeting beta burst, (He et al., 2020b) queried the effectiveness of NF targeting the LFP beta frequency in terms of tremor improvement.This suggestion cannot be evaluated with this study, as the relationship of clinical improvement with NF was not investigated.
Lastly, for one patient lacking a beta peak (PID 6), power decreases different from the feedback frequency were observed in Fig. 6.One power decrease was prominentely pronounced in the frequency range of approximately 14 to 24 Hz.This can be interpreted as having missed a possible beta peak, which unveiled itself during the NF training.However, part of this power decreasing band was included in the feedback frequency as well.Furthermore, in Figs. 5 and 6, the beta peaks determined in the spectral analysis tended to have a lower frequency than the manually selected peaks.The largest effect of downregulation tended to be at lower frequencies as well.Therefore, it may be useful to adapt the feedback frequency after a few NF regulation tasks to a frequency that is regulated more effectively or that better correlates with clinical improvement.

Towards DBS-neurofeedback in an every-day setting
Previous NF studies with PD patients were restricted to the time window between the two operations while the DBS leads were externalised (He et al., 2019(He et al., , 2020b;;Bichsel et al., 2021), whereas, in this study, the time point of the NF sessions could be selected freely after full implantation of the DBS system.Thereby, downregulation ability may improve with an increasing number of days since the electrode implantation, as we observed a tendency of linear decline of downregulation task beta power in function of the number of postoperative days, as shown in Fig. 7A.The increasing stimulation throughout the days after the electrode implantation did not have a clear effect on the downregulation ability, as there was no significant correlation of stimulation amplitude and downregulation ability, which can be observed in Fig. 7B.Nevertheless, the observation from Fig. 7A does not mean that NF sessions should be carried out at later stages of the postoperative phase, as the learning effect, calculated as the quotient of NF3 and NF1 downregulation beta power, was not influenced by the number of days since electrode implantation.Given the range of the postoperative measurement day (2-30 days), we cannot exclude a varying contribution of potential perilesional surgical effect or perielectrode oedema on the results.Further studies might demonstrate clarifying results when the tasks are conducted at later stages of the rehabilitation phase.
The agency stated by the patients at the end of a NF run failed to show a correlation with the actual regulation result.As the selfassessment of regulation ability was not accurate, it might be difficult for patients to efficiently pursue effective regulation strategies in an entirely independent NF training.Therefore, it could be beneficial for patients to start NF trials in a supervised setting with instructions and advices.

Limitations and outlook
As previously shown (Bichsel et al., 2021), the ability to downregulate beta power via NF and an improvement of this ability within a short time span was demonstrated in this study as well.With the analysis conducted in this study, no conclusion can be drawn for an effect over multiple days.As this study demonstrated the feasibility of NF control with an entirely implanted DBS system, it will be possible to investigate NF with this experimental setup over a much larger time span in the future.
In this study, muscle contractions were not controlled with electromyography (EMG) electrodes.This could represent a possible disruptive factor, as voluntary movements induce a decrease in beta frequency activity (Cassidy et al., 2002).However, movements were prohibited as a regulation strategy and the examinators visually controlled for overt movements.We waived an EMG control, as it was shown in previous studies, that patients did not tend to use muscle contraction in NF settings and that the beta activity decrease of NF was not induced erroneously through movement (Fukuma et al., 2018;He et al., 2019He et al., , 2020b)).
As, within a run, the transition to the next task followed without a pause, the experimental results were affected by a spillover effect.Analysing the course of runs, the spill-over effect varied between 1 and 7.5 s.A carry-over effect was described in the work of (He et al., 2020b).Excluding the spill-over effect, the beta power regulation could reveal to be even more effective.However, it can still be stated that the NF effect of modulating beta power is deployed relatively quickly, within seconds.
Only the STN of the symptom-dominant hemisphere was investigated in our experimental design and setup.A simultaneous streaming from both sides or simultaneous recording while streaming and visualising the symptom-dominant STN was unfortunately not supported by the hard-and software.Future studies might want to consider investigating both hemispheres sequentially, interleaved or even provide bilateral neurofeedback.
The lack of beta peaks in some patients can be discussed as a limiting factor, as NF can possibly not be provided in all PD patients.However, by selecting a peak in lower frequencies, the modulatory ability of the oscillatory activity in the STN was similar to when the selected peak was in the beta frequency.This can also be seen in the similarity of results when the patients lacking a beta peak were excluded (Supplementary Figures 1, 2).The clinical effect of NF in the alpha frequency band, though, remains to be investigated.
The aim of this study was to provide evidence of NF feasibility with an entirely implanted DBS system and therewith the suitability of NF in an every-day setting.The effect of NF in a situation of symptom exacerbation, which was suggested as a possible utilisation of NF (Bichsel et al., 2021), was not evaluated in this study, as most of the patients were oligosymptomatic.Furthermore, motor behavioural outcome was not investigated in this study, whereby other NF studies provide evidence of clinical improvement (He et al., 2020b;Bichsel et al., 2021).Thus, the current study lacked specific clinical or behavioural data to definitively classify the modulated oscillations as pathological or physiological.Our assertion regarding the modulation of pathological subthalamic oscillations stems from a substantial body of existing research in the field (cf.Introduction) as well as our initial study with externalised electrodes (Bichsel et al., 2021), where we observed a quickening of movements upon neurofeedback downregulation.Yet, future studies should try to include clinical or behavioural data to strengthen this claim.Ultimately, by showing the feasibility of NF with an entirely implanted DBS system, long-term studies as well as testing in symptom exacerbating situations become feasible and are of great interest.

Conclusion
We demonstrated successful and significant regulation of STN beta band power through NF using an entirely implanted DBS system in patients suffering from PD. Patients learned to control pathological deep brain oscillations within a few minutes of NF training.Of note is the high amount of beta power regulation (median decrease of 31.15%,maximum decrease of 41.85% in the last NF block) achieved using NF despite concurrent treatment with medication and DBS.As NF was conducted with an entirely implanted DBS system, the evidence presented in this study encourages the transfer of DBS-NF to the every-day setting and the conduction of long-term NF studies.

Fig. 3 .
Fig. 3. Beta power of upregulation tasks compared with downregulation tasks.Median beta power value of the upregulation block normalised by the median downregulation beta power value of the same block visualised as a box-and scatterplot.The black circles of the scatter plot each show the normalised value of one patient.The red line of the boxplot is the median value of all patients in one block, the blue box shows the interquartile range, and the whiskers extend from the nearest edge of the interquartile range box to the maximum or minimum value within the range of 1.5x interquartile range.The line at y = 1 corresponds to the baseline.The p-values marked with y were calculated with the paired-sample t-test, comparing normalised up-and downregulation tasks from the same block.The p-value marked with à was calculated with Wilcoxon signed rank test comparing normalised upregulation tasks from 2 different blocks, connected with a line in the figure.

Fig. 4 .
Fig. 4. Median beta power in the time course of runs.The median beta power values of NF3 runs of all patients normalised by the median value of the associated baseline tasks are displayed for each time point in NF3 runs as a black graph.The red line shows the median value of the displayed graph in each task.A) shows runs in the order of baseline, downregulation, upregulation, B) shows runs in the order of baseline, upregulation, downregulation.

Fig. 5 .
Fig. 5. Spectral analysis of all baseline tasks.The yellow line represents the average of the power spectrum of all baseline tasks.The red line represents the frequency, which was manually chosen as peak frequency for feedback in BrainSense TM Setup and the green star shows the beta peak, which manifested as a local maximum in the retrospective analysis.The black lines enclose the beta frequency band defined as 13-35 Hz.

Fig. 6 .
Fig.6.Spectral analysis of NF baseline and downregulation tasks.The yellow line shows the average of the power spectrum of baseline tasks during NF blocks.The blue line displays the average of the power spectrum of downregulation tasks during NF blocks.The red line represents the frequency, which was manually chosen for visual feedback.The black lines enclose the beta frequency band defined as 13-35 Hz.

Fig. 7 .
Fig. 7. Downregulation depending on time since implantation and stimulation amplitude.The downregulation data was normalised by median baseline values.A & B) The black circles show the average of the normalised median values of all downregulation tasks for each patient in function of the number of days after electrode implantation (A) or the stimulation amplitude during respective NF sessions (B).C) The black circles show the normalised median values of NF3 downregulation tasks of each patient in function of the number of days after electrode implantation.D) The black circles depict the normalised median values of NF3 downregulation tasks divided by the respective values of the NF1 downregulation tasks for each patient to show the learning effect in downregulation.The r-and p-values of linear correlation were calculated in all schematic representations.

Fig. 8 .
Fig. 8. Regulatory capacity in function of agency.Median value of NF downregulation tasks normalised by baseline tasks (A) and NF upregulation tasks normalised by downregulation tasks (B) in dependence on the agency stated by the patients.The r-and p-values of linear correlation are shown.

Table 1
. The LED values were rounded two digits before the decimal point.