Parkinsonian rest tremor can be distinguished from voluntary hand movements based on subthalamic and cortical activity

activity


Introduction
Deep brain stimulation (DBS) is a widely used treatment for advanced Parkinson's disease (PD) (Krauss et al., 2021).While well accepted and effective, it also has significant side-effects (Zarzycki and Domitrz, 2020) most of which result from current spread to structures adjacent to the DBS target (Koeglsperger et al., 2019).Hence, a general strategy for reducing side-effects is to reduce the energy applied in DBS.One way to achieve this while maintaining clinical benefits is to adapt stimulation to the current motor state.This approach is called adaptive DBS (Krauss et al., 2021;Meidahl et al., 2017;Neumann et al., 2019).Various signals can be used to control either the onset or the amplitude of DBS https://doi.org/10.1016/j.clinph.2023.10.0181388-2457/Ó 2023 International Federation of Clinical Neurophysiology.Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).(Marceglia et al., 2021).Currently, the most investigated neural control signal is subthalamic beta band activity (Little et al., 2016(Little et al., , 2013;;Piña-Fuentes et al., 2017;Tinkhauser et al., 2017), but other signals have also been considered, such as cortical gamma band activity (Gilron et al., 2021;Swann et al., 2018) or local evoked potentials (Dale et al., 2022).
Adapting DBS to the current motor state makes sense for treating tremor, particularly, because tremor is highly variable, waxing and waning spontaneously.In case of tremor, one natural strategy for adaptive DBS is to monitor tremor directly by means of peripherals such as accelerometers or electromyography (Cagnan et al., 2017;Cernera et al., 2021;Malekmohammadi et al., 2016).This approach, however, requires additional hardware and thus raises additional security and compliance issues.
An alternative strategy is to use brain signals rather than peripherals.It is known that Parkinsonian rest tremor originates in the brain (Milosevic et al., 2018), involving an extended subcorticocortical network including the basal ganglia, cerebellum, thalamus and motor cortex (Helmich, 2018;Timmermann et al., 2003).Thus, it should be feasible to track tremor by monitoring its neural correlates.Several studies have investigated subcortical correlates by analyzing local field potentials (LFPs) recorded from DBS electrodes and identified numerous tremor-related changes, such as an increase of power at individual tremor frequency (Hirschmann et al., 2013a), a beta power decrease (Qasim et al., 2016;Wang et al., 2005), an increase of low gamma power (Beudel et al., 2015), and a shift of power in the very high frequency range (Hirschmann et al., 2016).While these studies were able to distinguish tremor from tremor-free rest periods, no study has been able to distinguish PD rest tremor from voluntary hand movements so far.When applying adaptive DBS in practice, it would be desirable to make this distinction to reduce unnecessary stimulation.
In this study, we re-analyzed simultaneous recordings of forearm EMG, magnetoencephalography (MEG) and subthalamic nucleus LFP data, obtained from tremor-dominant PD patients, aiming to distinguish between PD rest tremor, self-paced fistclenching, static forearm extension and tremor-free rest periods (referred to as ''quiet" episodes from here on), by means of gradient-boosted tree learning, a popular method in machine learning.We show that these motor states can be distinguished when considering cortical and subthalamic data.

Participants
In this study we used a subset of previously collected data (Hirschmann et al., 2013a).We selected patients with enough data for comparing rest tremor, quiet episodes, and two motor tasks (see below).Recordings from 6 patients met these requirements.We the kept participant identifiers consistent with (Hirschmann et al., 2013a).Overall, we had 11 datasets (5 subjects in Med On and Med Off, one subject in Med Off only; see Table 1).
All patients involved were diagnosed with idiopathic PD, and underwent DBS surgery the day before measurement.They experienced spontaneous waxing and waning of rest tremor during the recordings.Patient details are provided in Table 1.The study was approved by the ethics committee of the Medical Faculty of the Heinrich Heine University Düsseldorf (Study No. 3209).It was carried out in accordance with the Declaration of Helsinki and required written informed consent.

Recordings
Patients were recorded one day after subthalamic nucleus (STN) electrode implantation with leads still externalized.All but one patients were recorded in two sessions: one session OFF oral dopaminergic medication for at least 12 h and one session ON medication; one patient had only the OFF medication session recorded.Subcutaneous apomorphine administration was paused 1.5-2 h before measurements started.Each session contained four parts: rest, motor task 1, rest, motor task 2. Rest periods lasted 5 min.Motor task 1 was static forearm extension (''hold") and motor task 2 was self-paced fist-clenching (''grasp") at approximately 1 Hz (Hirschmann et al., 2013b).Movements were performed with the symptom-dominant body side in five 1-min blocks which were interleaved by 1 min pauses to avoid fatigue.During rest and between the blocks of voluntary movement tremor appeared and disappeared spontaneously.
Local field potentials from the STN, the magnetoencephalogram (MEG; Vectorview, MEGIN) and the surface electromyogram (EMG) of the extensor digitorum communis and flexor digitorum superficialis muscles of both upper limbs were recorded simultaneously.The sampling rate was 2000 Hz.DBS electrodes were connected to the amplifier of the MEG system by externalized, nonferromagnetic leads.Electrode contacts were referenced to the left mastoid and rearranged to a bipolar montage offline by subtracting signals from neighboring contacts.EMG electrodes were referenced to surface electrodes at the muscle tendons.A hardware filter was applied with a pass-band of 0.1-660 Hz.

EMG labeling
The hand performing the voluntary movements showed intermittent rest tremor in all subjects.In order to label the data, EMG data was 10 Hz high-pass filtered, rectified and smoothed with a 100 ms box-filter.Quiet periods were identified semiautomatically.Epochs with both EMG channels (forearm flexor and extensor) deflecting less than their respective 40% quantiles (computed using the data from the entire recording) were labeled as candidate quiet periods automatically and then adapted manually.Tremor and voluntary movements were labeled manually.Epochs with uncertain labels, as well as postural and kinetic tremor were discarded.When possible, ''safety offsets" were included to ensure that different behavioral states were separated by at least 1 s to mitigate the risk of confusing states (Fig. 1A).The amount of data available per motor state is provided in Table 2.Among the epochs with tremor, the proportion of bilateral tremor was 70% (std = 36%) in the medication OFF condition and 61% (std = 40%) in the medication ON condition.

Electrode localization
The placement of DBS electrodes was reconstructed with the Lead-DBS software package; (Horn and Kühn, 2015) from preoperative MR and postoperative CT images (Fig. 1B).This was not possible for two out of six patients due to noisy CT images.

Preprocessing
The MEG sensor data and LFP data were resampled to 256 Hz to increase computation speed.MEG artifacts were identified as time periods when sensor data deflected more than 2.5 times the signal's 95%-trimmed mean in multiple sensors.Spatiotemporal signal space separation (Taulu and Simola, 2006) with a 10 s time window was applied to the MEG data, using the MNE toolbox (Gramfort et al., 2014).LFP artifacts were identified as periods with strong broadband modulation of the signal.

Source reconstruction
For forward modelling, we made use of realistic, single-shell head models based on the individual, T1-weighted MR image (Nolte, 2003).Source reconstruction was performed by means of Linearly Constrained Minimum Variance beamforming, as implemented in the Fieldtrip toolbox (Oostenveld et al., 2011).The beamformer grid contained 567 locations, spread out evenly across the cortical and cerebellar surface.It was aligned to Montreal Neurological Institute space, allowing for grid parcellation into 30 supersets of regions defined in the Automatic Anatomic Labeling atlas (Tzourio-Mazoyer et al., 2002).For the computation of the spatial filter, we used only the quiet periods without artifacts.For beamformer regularization, the lambda parameter was set to 0.1% of the maximum eigenvalue of the covariance matrix computed on the quiet periods.

Feature engineering
For feature engineering, the data were separated into 1 s-long disjoint windows, i.e. the temporal resolution was 1 s.Windows containing MEG or LFP artifacts were discarded.The number of windows per dataset averaged across subjects and medication con-ditions after artifact and uncertain EMG state rejection was 1112 (std = 172), whereas total recording duration (without artifact rejection) was 1638 s (std = 168 s).For each window, we computed the logarithm of Hjorth activity, equal to windowed signal variance (Hjorth, 1970).The logarithm served to bring the data distribution closer to normal.For the source-reconstructed data, we averaged Hjorth activity across all grid points belonging to the same cortical parcel to obtain one feature per cortical area and time window.

Hjorth activity x t
The use of Hjorth activity rather than spectral power or connectivity was motivated by the intention to compare decoding performance across brain areas while keeping the number of features as low as possible to avoid overfitting.Thus, we preferred metrics without frequency-resolution, uniquely associated with a single area.Our analysis choices were guided by the work of Yao et al. who performed a detailed investigation of the optimal choices for automated tremor detection (Yao et al., 2020).The authors tested several features and identified Hjorth activity as one of the most useful features.Further, they found XGBoost to be the best performing machine learning model among several candidates, and 1 s to be a good window size.

Machine learning
To predict the four different motor states from Hjorth activity, we performed 4-label classification on the windowed data using 5-fold, shuffled cross-validation.The XGBoost algorithm was used to perform the classification.XGBoost is an advanced decisiontree-based algorithm, widely used machine learning (Chen and Guestrin, 2016).It often provides better results than linear models and requires relatively small amounts of training data.Since the number of windows differed substantially between behavioral states both within and between subjects (Table 2), we used balanced accuracy, defined as the mean of sensitivity and specificity averaged across classes, as a measure of performance (Kelleher et al., 2020).When training the model, we used oversampling via imbalanced-learn toolbox (Lemaître et al., 2017) to artificially balance classes.The entire python & Matlab pipeline code (including preprocessing, source reconstruction and machine learning) is available at the code repository https://github.com/todorovdi/ContBehFeatExplorer.

LFP channel selection and hyperparameter tuning
For each subject, we performed single-feature classification for each LFP channel separately and selected the channel with best out-of-sample performance.Similarly, hyperparameter tuning was done by sampling combinations of the XGboost parameters max_depth, min_child_weight, subsample and eta and selecting the combination giving the best 5-fold cross-validated balanced accuracy.Neither LFP channel selection nor tuning had a strong effect on performance.All selected LFP channels were located contralateral to movement, except for patient S05.

Distinguishing movements based on STN activity
We investigated how well the different motor states could be distinguished based on subthalamic activity alone.Fig. 2 shows a confusion matrix for each subject and medication state, providing discrimination performance for each pair of movements.STN activity alone turned out to be a rather weak predictor of movement type.In medication OFF (N = 6), the average percentage of correctly classified windows was 62% for quiet (std = 5%), 32% for rest tremor (std = 19%), 34% for hold (std = 16%), and 23% for grasp (std = 27%).The most common misclassification was labeling tremor as hold (mean number of cases = 22%, std = 9%).The least common misclassification was labeling tremor as grasp (mean = 8%, std = 8%).Random shuffling of labels lowered performance, demonstrating that the STN did provide relevant information (Table 3).
Levodopa did not have a strong influence on movement discrimination (Table 3).Balanced accuracy did not differ significantly between medication states (p = 0.7, all vs. all, independent t-test, N OFF = 6, N ON = 5).
In addition to the multiclass problem, we tested how well our approach can discriminate tremor from any other movement, as this classification is relevant for tremor therapy through adaptive DBS.When merging quiet, hold and grasp epochs into a single non-tremor class, we obtained balanced accuracy of 59% in the Med Off state (std = 8%) and 58% in Med On (std = 6%; Table 2).When discriminating merely between rest tremor and quiet, accuracy was 64% in Med Off (std = 4%) and 60% in Med On (std = 7%).This performance is similar to previous papers dealing with tremor vs. quiet (Hirschmann et al., 2017;Yao et al., 2020), taking into account that they reported regular, not balanced accuracy.

Distinguishing movements based on STN and cortical activity
When classifying epochs based on both subthalamic activity and the activity of all cortical parcels, we observed a strong increase in performance (average performance gain: 37%; Table 3 and Fig. 3).In medication OFF (N = 6), the average percentage of correctly classified epochs was 74% for quiet (std = 21%), 58% for rest tremor (std = 36%), 80% for hold (std = 12%), and 81% for grasp (std = 18%).The most common misclassification was labeling tremor as hold (mean = 22%, std = 31%).The least common misclassification was labeling rest tremor as grasp (mean = 1%, std = 1%).As in the STN-only case, random shuffling of labels lowered performance, lowering the number of classes increased performance, and medication had little influence on performance (p = 0.63, ttest; see Table 3).
When the model was trained across rather than within subjects, performance of all-to-all classification using STN LFP and MEG data dropped considerably, to 24% in medication OFF and to 28% in medication ON condition (compare Table 3), suggesting that movement-specific neural signatures vary across patients.When the model was trained on medication OFF data and tested on medication ON data of the same patient, mean balanced accuracy dropped to 31% (std = 12%).ON to OFF generalization was comparably poor (mean balanced accuracy = 23%, std = 4%), suggesting that movement-specific neural signatures vary across medication states as well.

STN-cortex two-feature models
While it would be difficult for an adaptive DBS system to monitor the entire cortex, it is feasible to monitor the STN and individual cortical areas (Gilron et al., 2021;Opri et al., 2020).Thus, we investigated the performance gain achieved by adding the activity of a single cortical area to STN activity as a second feature.Fig. 4 shows a brain map of the performance gain.In most cases, sensorimotor cortical areas added the most information.
Adding a single cortical area was usually enough to cover much of the total gain achievable by adding all cortical areas (Fig. 5).Adding all remaining cortical areas to the STN-cortex two-feature models as additional features led to an average performance increase of only 10% (std 11%) compared to the two-feature model.
The spatial distribution of information was not particularly lateralized.The single most informative cortical area was not always on the contralateral side with respect to movement, and adding all contralateral areas to STN activity did not consistently yield higher accuracy than adding all ipsilateral areas (Fig. 5).In this context, it is important to note that the non-moving hand was trembling in some of the patients and epochs.This likely contributes to the bilateral patterns.

Discussion
In this paper, we demonstrate the feasibility of distinguishing tremor from voluntary movements of the same hand.We found that LFP recordings from the STN are not sufficient for this distinction.It becomes possible, however, when considering additional cortical signals.

Previous studies
Distinguishing different movements in neural recordings has been the aim of several brain-computer interface studies, typically addressing cortical activity, as measured by EEG (Xu et al., 2021) or electrocorticography (ECoG) (Volkova et al., 2019).More recently, the possibility of recording subcortical activity in the context of DBS has facilitated classification studies in basal ganglia and thalamus, most of which deal with movement detection, i.e. the distinction between movement and rest.Loukas and Brown were the first to decode the onset of voluntary movements from subthalamic activity (Loukas and Brown, 2004).Several other studies aimed for tremor detection, due to its relevance for adaptive DBS (Bakstein et al., 2012;Camara et al., 2015;Hirschmann et al., 2017;Shah et al., 2018;Yao et al., 2020).Our group has previously applied Hidden Markov Modelling to STN LFPs, using a superset of the data analyzed here (Hirschmann et al., 2017).Spontaneously occurring PD rest tremor was detected with an average sensitivity of 70% and a specificity of 89%.Yao et al. achieved a sensitivity of 89% and a specificity of 50%, using XGBoost and Kalman filtering of selected input features (Yao et al., 2020).The same group later demonstrated the feasibility of XGBoost tremor detection on chip, facilitating its integration into small, battery-powered neurostimulators (Zhu et al., 2020).Tremor detection has also been demonstrated for thalamic LFPs recorded in ET patients (Tan et al., 2019), including the feasibility of conditioning DBS on these signals under laboratory conditions (He et al., 2021).
In this study, we obtained comparable tremor detection performance for the multi-feature case (STN + cortex).Importantly, this work goes a step further by distinguishing between tremor, tremor-free rest, self-paced fist-clenching and static forearm extension.

Clinical implications
Distinguishing between tremor and voluntary hand movements of the same limb based on brain activity is a challenging task, which has both clinical and basic science implications.From a clinical perspective, mastering this task would allow for better adaptation of DBS to the current situation, as it would reduce unnecessary stimulation occurring when an adaptive DBS system confuses voluntary movement, or dyskinesia, with tremor.Unnecessary stimulation is not only a waste of energy, but might contribute to maladaptive changes in response to DBS (Reich et al., 2016).In addition, an unnecessary sudden onset of stimulation or a sudden amplitude increase may interfere with fine motor control, as both may elicit transient side-effects such as paresthesia.

Neuroscientific implications
From a basic science perspective, it is warranted to understand the neurophysiological differences between voluntary and nonvoluntary movements such as tremor.Tremor and voluntary movement share a common oscillatory signature, characterized by beta desynchronization and gamma synchronization (Beudel et al., 2015;Pfurtscheller et al., 2003;Qasim et al., 2016;Wang et al., 2005).While this pattern is very robust, it signals the presence of any movement rather than what movement is being executed, nor whether the movement is voluntary or not.The latter question is highly interesting for general neuroscience and has been addressed by contrasting tremor and mimicked tremor.Although some differences emerged, there were generally more similarities than differences (Muthuraman et al., 2018(Muthuraman et al., , 2012)).Hence, there is currently no consensus on the neural mechanisms underlying involuntary movement and hence no compelling explanation of why tremor cannot be stopped at will.
Even without a neural correlate of volition, it may be possible to distinguish tremor from other movements based on brain signals.A smart DBS system could simply try to detect oscillatory movements of a certain frequency and assume tremor, as rhythmic movements of this kind are rarely voluntary.This appears to be a feasible approach, given the presumably unique neural correlates of such movements, such as an increase of power at individual tremor frequency (Hirschmann et al., 2013a;Pollok et al., 2004;Timmermann et al., 2003).The characteristic kinematics of tremor and its reflection in brain activity is most likely the reason why were able to distinguish it from other movements here.Note, however, that our approach was based on Hjorth activity, which is not frequency-resolved.Thus, knowing the individual tremor frequency is not necessary for tremor discrimination.

Movement-related information in STN and cortex
One of the key findings of this paper is that cortical signals can dramatically increase tremor discrimination.This observation tallies with recent studies in PD patients decoding grip force from STN and primary motor cortex, the latter recorded through intraoperative ECoG (Merk et al., 2022;Peterson et al., 2023).Both studies achieved much better predictions when using cortical signals, as compared to subthalamic signals.The authors restrained from drawing conclusions about the information content of cortex versus STN because ECoG signals have a better signal-to-noise ration than recordings made with a DBS electrode, which might explain the difference in performance.In our case, however, the invasive deep brain recording is generally considered to have better signal-to-noise ratio than the noninvasive MEG, suggesting that the STN might not be the optimal brain area for inferring the kind of movement being performed.In fact, a number of studies indicate that the basal ganglia control movement vigor, i.e. the strength and velocity of a movement, rather than being concerned with movement coordination (Dudman and Krakauer, 2016;Lofredi et al., 2018;Turner and Desmurget, 2010;Yttri and Dudman, 2016).This may imply that STN activity is more useful for predicting grip force (Merk et al., 2022;Shah et al., 2018) than movement type.
While the observations discussed above are in line with the poor decoding of movement type based on STN activity alone, as observed here and in a different analysis of the same data (Hirschmann et al., 2017), it is challenged by recent findings of Golshan and colleagues, who managed to decode different voluntary movements from STN LFPs with accuracies of up to 90% (Golshan et al., 2020(Golshan et al., , 2018)).These studies did not investigate tremor, however, and the approach requires knowing when a movement is initiated.This information, of course, would not be available when applying adaptive DBS in practice.
So as of now, it seems that any smart DBS system that needs know to more about a patient's motor state than whether or not the patient is at rest would likely profit substantially from external information, such as peripherals (Cagnan et al., 2017;Cernera et al., 2021;Malekmohammadi et al., 2016) or electrodes in additional brain areas.If so, how many and which brain areas should be monitored?Our results suggest that a few areas suffice for distinguishing tremor from voluntary movement and between different voluntary movements.Unsurprisingly, primary and premotor areas were found to be most useful, although other areas con- tributed information as well.Thus, a system monitoring both STN and motor cortex appears promising for real-time motor state discrimination in PD.
In fact, such systems have already been tested both in-clinic and at home.Using an implantable DBS system capable of recording neural activity and a combination of deep brain electrodes and cortical strip electrodes placed on primary motor cortex, Gilron et al. found that conditioning STN DBS on cortical gamma oscillations increased the time spent in the Med On/Stim On state without experiencing dyskinesia (Gilron et al., 2021).This is a crucial finding, since reducing stimulation-induced side effects is a key promise of adaptive DBS.Using a similar approach, Opri et al. demonstrated that conditioning thalamic DBS on motorcortical low-frequency activity is as effective for suppressing essential tremor as standard DBS, despite delivering less energy (Opri et al., 2020).

Generalizability of movement decoding
To be useful in practice, a decoding approach needs to work equally well in different situations, e.g.before and after intake of anti-Parkinsonian medication.In agreement with (Golshan et al., 2020), we obtained similar decoding performance in medication On and Off.This needs to be tested further, however, in light of recent findings linking decoding performance to motor impairment (Merk et al., 2022;Peterson et al., 2023), which is closely linked to dopamine availability.Although it was possible to achieve decent decoding in both medication states, each state required a different decoder.While this is clearly suboptimal for clinical application, it may turn out to be a minor hurdle, as PD patients rarely reach a complete OFF state when medicated properly.In practice, an ON-state decoder might suffice for most patients.Ideally, a decoder would not only generalize to different dopamine levels but also to different patients.As in almost all previous works on this topic (but see Hirschmann et al., 2022), our approach required within-patient training, i.e. a decoder working in one patient does not necessarily work in another.Tuning the decoder to a new patient requires individual, electrophysiological data from the DBS target and cortex.While this seemed infeasible a few years ago, the advent of new DBS systems capable of recording brain signals in addition to delivering stimulation has made this scenario more realistic (Jimenez-Shahed, 2021).Note, however, that a general decoder would still be preferable from both a practical point of view (no tuning necessary) and a basic science perspective because only general decoders allow for conclusions on the general mechanism of tremor.

Limitations
One obvious limitation of this work is that we did not detect tremor in real-time nor condition DBS on tremor.In addition, the performance achieved here is not enough for clinical application, although much better than chance, as evidenced by the decline caused by shuffling class labels randomly.Moreover, due to a limited amount of data, parts of the training data were close in time to parts of the test data, which might cause overfitting in the presence of temporal autocorrelation.In this scenario, the classifier groups epochs based on their closeness in time rather than their movement class.This potential confound can be mitigated in larger datasets by requiring e.g. a certain distance in time between all train and test epochs.
Further, the cortical signal was recorded with MEG rather than ECoG.While MEG has much worse spatial resolution, it offers whole-brain coverage.This allows us to exclude movement artifacts as the true source of information because artifacts have a different topography than the information maps provided here (Kandemir et al., 2020).While we found most information in sensorimotor regions, DBS hardware-related artifacts are centered on the adapter for the externalized lead (right parietal and right temporal regions) whereas movement-related MEG artifacts are strongest close to the moving limb.
Finally, we included only men in this re-analysis because they happened to fulfill the inclusion criteria accidentally (spontaneous waxing and waning of tremor in the limb used in the motor task and presence of both hold and grasp movements in the recordings).Hence, we cannot be sure that our findings are transferable to females.Yet, we see no reason for assuming sex differences.

Conclusions
PD rest tremor can be distinguished from voluntary hand movements when considering subthalamic and cortical signals.This distinction is possible regardless of the dopamine level.

Fig. 1 .
Fig. 1.Data labelling and electrode placement.A. Labeling of behavioral states in forearm EMG, state transition examples.B. Localization of deep brain stimulation (DBS) electrodes for four out of six patients.For two other patients (S02 and S03) the CT image quality was not sufficient for localization.

Fig. 2 .
Fig. 2. Distinguishing movements based on subthalamic activity.Confusion matrices for Med Off and On per subject, averaged across cross-validation folds.The color bar shows the overall maximum and minimum, the maximum of all off-diagonal entries and the minimum of all diagonal entries, and the session-average chance level.

Fig. 3 .
Fig. 3. Distinguishing movements based on subthalamic and cortical activity.Confusion matrices for Med On and Med Off per subject, averaged across cross-validation folds.The color bar shows the overall maximum, the maximum of all off-diagonal entries and the minimum of all diagonal entries, and the session-average chance level.

Fig. 4 .
Fig. 4. Gain of using cortical activity in addition subthalamic nucleus activity.The balanced accuracy achieved with subthalamic activity alone is provided in the headings (''LFP perf").Colors code the improvement (in %) obtained by adding the activity of an individual cortical brain area as a second feature.LFP = local field potential.

Fig. 5 .
Fig.5.Most informative cortical areas and lateralization of information.Bar height indicates the performance difference with respect to the STN-only case, with original (blue) and shuffled class labels (red).Uppermost bar: best combination of one STN feature and one cortical feature.Second bar: best single cortical feature (STN not used).The lower three bars illustrate the gain of adding all cortical areas (both hemispheres), only the ipsilateral areas (with respect to the moving hand) and only the contralateral areas, respectively.LFP = local field potential; STN = subthalamic nucleus.

Table 2
Amount of data per motor state in seconds and balanced accuracy (bacc) in percent, for the feature set consisting of Hjorth activity of the best local field potential (LFP) channel and all cortical areas.The last column identifies the hand that was used to define the behavioral states.The amount refers to data used for classification, i.e. after discarding artifacts.N/A = not available.

Table 3
Balanced accuracy, group results.''Shuffled" performance is 90% percentile of balanced accuracies obtained from 100 permutations of class labels.STN = subthalamic nucleus; LFP = local field potential.