Decision letter: Concurrent decoding of distinct neurophysiological fingerprints of tremor and bradykinesia in Parkinson’s disease

Abstract

Full text Figures and data Side by side Abstract Editor's evaluation Introduction Results Discussion Materials and methods Data availability References Decision letter Author response Article and author information Metrics Abstract Parkinson’s disease (PD) is characterized by distinct motor phenomena that are expressed asynchronously. Understanding the neurophysiological correlates of these motor states could facilitate monitoring of disease progression and allow improved assessments of therapeutic efficacy, as well as enable optimal closed-loop neuromodulation. We examined neural activity in the basal ganglia and cortex of 31 subjects with PD during a quantitative motor task to decode tremor and bradykinesia – two cardinal motor signs of PD – and relatively asymptomatic periods of behavior. Support vector regression analysis of microelectrode and electrocorticography recordings revealed that tremor and bradykinesia had nearly opposite neural signatures, while effective motor control displayed unique, differentiating features. The neurophysiological signatures of these motor states depended on the signal type and location. Cortical decoding generally outperformed subcortical decoding. Within the subthalamic nucleus (STN), tremor and bradykinesia were better decoded from distinct subregions. These results demonstrate how to leverage neurophysiology to more precisely treat PD. Editor's evaluation This important study advances our understanding of Parkinson's by identifying micro and macro scale signatures linked to critical symptoms (e.g., tremor and slowness of movement), and effective motor control. The evidence supporting the conclusions is solid, and leverages a rich dataset obtained during naturalistic movement. The work will be of interest to neuroscientists, neurologists, and biomedical engineers. https://doi.org/10.7554/eLife.84135.sa0 Decision letter eLife's review process Introduction Parkinson’s disease (PD) is a common and complex neurodegenerative disorder characterized by the dynamic expression of particular motor features such as tremor and bradykinesia (Armstrong and Okun, 2020; Parkinson, 2002). These distinct motor signs are expressed variably across patients and may respond differently to dopamine replacement therapy; their differential expression is often used to classify patients into phenotypic subtypes (Koller, 1986; Sethi, 2008). Despite this heterogeneity, both of these motor features (and both tremor-dominant and non-tremor-dominant patient subtypes) respond to high-frequency deep brain stimulation (DBS) applied to the subthalamic nucleus (STN) (Katz et al., 2015; Limousin et al., 1998). DBS delivered in a closed-loop fashion (i.e., in response to neurophysiological biomarkers) has shown promising therapeutic potential primarily toward alleviating bradykinesia (Little et al., 2016a; Little et al., 2016b), but current efforts focusing on β frequency oscillations (15–30 Hz) have been shown to inadequately treat or worsen tremor in some cases (Piña-Fuentes et al., 2020; Velisar et al., 2019). Thus, tremor may be better signaled by different components within the local field potential (LFP) spectrum, and closed-loop DBS could benefit from a clearer understanding of the neurophysiological biomarkers that differentiate these motor signs from each other, and from more optimal motor performance in the absence of these impairments. To this point, STN LFP recordings from patients with different PD subtypes have revealed distinct patterns of oscillatory activity (Telkes et al., 2018). In addition to spectral variability, specific stimulation sites within the STN have been associated with the preferential reduction of individual motor signs (Akram et al., 2017). Moreover, these STN sites were associated with specific patterns of anatomical connectivity with cortical structures (Haynes and Haber, 2013). Much like how overlapping subdivisions of basal-ganglia-cortical circuits have been found to encode separate aspects of movement (Mosher et al., 2021; Neumann et al., 2018), separate motor features may be mediated by different sub-circuits involving the STN and sensorimotor cortex (Gibson et al., 2021). In order to better reveal the functional and anatomical substrates of distinct PD motor states, we enlisted patients with PD undergoing awake DBS electrode implantation to perform a continuous visual-motor task that allowed rigorous, concurrent measurement of different motor metrics while we acquired STN (micro- and macroelectrode) and cortical (electrocorticography [ECoG]) recordings. Prior studies have not attempted to simultaneously decode different aspects of disease expression, contrast these measures with symptom-free performance, and examine disease expression on the short timescales relevant to that varying expression. While our group has previously demonstrated the ability to decode global PD motor dysfunction from STN recordings on short timescales (Ahn et al., 2020; Sanderson et al., 2020), we focus here on individual motor features and their specific neurophysiological manifestations. Specifically, we trained machine learning models to directly decode tremor or slowness from neural recordings to reveal the spectral and anatomical fingerprints of these cardinal motor features of PD. Results Motor behavior during the target tracking task Twenty-seven patients with PD undergoing STN DBS implantation and 17 age-matched controls performed a visual-motor task in which they followed an on-screen target with a cursor controlled by either a joystick or a stylus and tablet (Figure 1A). Twenty-three patients (and 12 control subjects) performed a version of the task with fixed patterns of target movement, while four patients (and five control subjects) performed a version with randomly generated target paths. Each patient performed 1–4 sessions of the task during the procedure for a total of 69 sessions, while control subjects each performed 1 session extra-operatively for a total of 17 sessions. Tremor amplitude and cursor speed – task metrics calculated to reflect the expression of tremor and bradykinesia – were quantified from the cursor traces. These behavioral metric data were then averaged into 7 s non-overlapping epochs. To compare metrics across subject populations while considering epochs, trials, and sessions as repeated measurements within individuals, linear mixed models (LMMs) were used (see Materials and methods). The resulting metric distributions for PD vs. control subjects demonstrated increased tremor for PD patients (n=6498 epochs across 44 subjects, PD vs. control, LMM coefficient/LMM β=0.337, Z=2.169, p=0.030), but only a trend for decreased speed (PD vs. control, LMM β=−0.592, Z=−1.194, p=0.232) (Figure 1B and C). Figure 1 Download asset Open asset Tremor and movement speed calculated from fixed- and random-pattern intraoperative visual-motor tasks. (A) Left: Schematic of task target (green) and cursor (gray) traces from a single trial of the fixed- (top) or random- (bottom) pattern task. Center-top: Bandpass filtered cursor traces from a task trial. Ca refers to the amplitude of the analytic signal (a) of the cursor trace (C). Center-bottom: Lowpass filtered cursor traces from a task trial. Right-top: One-dimensional projection of bandpass filtered traces (black), with tremor amplitude measured from the envelope (orange). Right-bottom: Cursor speed measured from lowpass filtered traces (black). Figure adapted from Figure 1 of Ahn et al., 2020. (B, C) Distributions of 7 s tremor amplitude (top) and cursor speed (bottom) epochs for control subject and Parkinson's disease (PD) patient populations in the fixed-pattern (B) (n=5375 epochs across 35 subjects) and random-pattern (C) (n=1123 epochs across 9 subjects) task. ° – degrees of visual angle. (D, E) Task-based tremor amplitude (D) and slowness (E) corresponded to UPDRS measures of tremor (D) or bradykinesia (E) (n=24 subjects). ρ=Spearman correlation statistic. Tremor distributions compiled across task versions revealed that while subjects with PD spent a substantial fraction of time without tremor, they also exhibited a large range of tremor expression not present in control subjects (Figure 1B and C, top). On the other hand, the two task versions generated different movement speed distributions (Figure 1B and C, bottom). While the fixed-pattern version of the task elicited a bimodal distribution (reflecting slower turns and faster straight path segments) in control subjects, the random-pattern version elicited a single peak corresponding to the fixed target speed used in that task. Nonetheless, in the random-pattern version, the PD cursor speed distribution was shifted to the left (i.e., slower) relative to control subjects (n=1123 epochs across 9 subjects, PD vs. control, LMM β=−1.004, Z=−2.210, p=0.027). Cursor speed was converted to ‘slowness’ in order to control for target trajectory/speed variability by normalizing each session’s distribution of cursor speed to its minimum and maximum values (0: highest speed, 1: lowest speed). To determine whether these two motor metrics reflected distinct components of PD motor dysfunction, each patient’s metric distribution medians were correlated against their UPDRS III motor subscores. Indeed, tremor amplitude positively correlated with the resting tremor subscore (ρ=0.607, p=0.001, n=24 PD subjects, Spearman correlation) (Figure 1D) and slowness positively correlated with the hand open/close subscore (a subscore used in part to assess bradykinesia) (ρ=0.406, p=0.035, Spearman correlation) (Figure 1E). However, the opposite correlations were not significant (tremor amplitude – UPDRS hand open/close: ρ=0.022, p=0.446; slowness – UPDRS resting tremor: ρ=0.120, p=0.219). In addition, action/postural tremor subscores trended toward a positive correlation with tremor amplitude (ρ=0.354, p=0.057), while resting and action/postural tremor subscores were positively correlated within subjects (ρ=0.602, p=0.002). Although postural tremor can correlate with resting tremor when patients with PD are measured by the UPDRS for the former and the Washington Heights-Inwood Genetic Study of Essential Tremor (WHIGET) Rating Scale for the latter, resting tremor is thought to be more specific to the PD pathophysiology (Louis et al., 2001). Therefore, tremor and movement speed were considered to reflect two key aspects of PD motor dysfunction. Tremor and slowness were distinct and opposing symptomatic states Relative to each other, tremor and slowness typically did not co-occur but rather were inversely expressed in time (n=27 subjects, tremor × slowness, LMM β=−0.584, Z=−19.351, p=2.00*10–83) (Figure 2A and B). To understand whether this anti-correlation may have been due in part to motor features manifesting on different timescales, autocorrelograms were computed for each metric with 100 ms epochs. Here, tremor was typically expressed continuously for longer periods (autocorrelogram full-width half-maximum [FWHM], 0.898 s) as opposed to slowness (0.297 s) (Figure 2C). Using this FWHM as the minimum, we calculated periods of time where metrics were sustained above control levels (i.e., symptomatic periods) across subjects. The median duration of symptomatic tremor episodes was 2.000 s, and slowness episodes lasted for 0.500 s (Figure 2D). With these differing timescales and anti-correlated presence, tremor and slowness appeared to represent distinct symptomatic states. Figure 2 Download asset Open asset Tremor and slowness represented two non-overlapping motor states with differing timescales. (A) Examples from three individual subjects of cursor (solid lines) and target (translucent lines) traces (top row) and calculated motor metrics (bottom three rows) within single trials. Periods of increased expression of individual motor metrics are highlighted by their respective color. (B) (Left) Scatter plot of all cursor speed and tremor measurements in 7 s epochs across subjects. (Right) Histogram of subject-wide behavioral Spearman correlation with tremor and slowness metrics (n=27 subjects). (C) Autocorrelograms of symptomatic (tremor, slowness) and non-symptomatic (effective motor control) metrics. Colored vertical dashed lines indicate full-width half-maximum (FWHM) for each metric. Top-left inset depicts a zoomed-in window of the autocorrelogram. (D) Histogram of sustained motor metric period duration (i.e., symptomatic state duration) across subjects. Solid lines indicate gamma distribution fit to each motor metric state histogram, while dashed vertical lines indicate the median symptomatic state length for each metric. Because PD produces a fluctuating motor deficit such that there can be moments of normal-appearing motor behavior (Mazzoni et al., 2007), we labeled epochs without motor dysfunction as ‘effective’ motor states. Specifically, epochs with lower tremor and/or higher movement speeds were assigned values closer to 1 while more symptomatic epochs (high tremor and/or slower movement speeds) were assigned values closer to 0. Compared to other metrics, effective motor control was expressed on longer timescales (FWHM = 3.784 s, median state length = 7.900 s) (Figure 2A, C and D). Tremor and slowness had distinct representations within the STN A total of 203 microelectrode and 176 macroelectrode recordings (microelectrode tips and macroelectrode contacts separated by 3 mm on the same electrodes) were acquired from the STN as patients performed the task. To assess whether tremor or slowness could be decoded from these recordings, spectral estimates of power from 3 to 400 Hz were obtained using a wavelet convolution. Narrowband power estimates were grouped into six broad frequency bands (θ/α,β,γlow,γmid,γhigh,hfo) with 7 sub-bands each, for a total of 42 neural ‘features’ per 7 s epoch (Ahn et al., 2020). Neural decoding models (support vector regression [SVR] with a linear kernel and 100-fold cross-validation) were trained directly on the epoch’s average metric (tremor or slowness values averaged within each epoch), and their performance was assessed with squared Pearson’s r (r2) between observed and decoded metrics. Across each subject’s best-performing microelectrode recordings (MER), tremor decoding performance (r2=0.232±0.200) was superior to slowness decoding (r2=0.125±0.108) (n=203 MER models, tremor v. slowness, LMM β=0.051, Z=5.477, p=4.33*10–8). No such difference was observed across macroelectrode recordings (n=176 macroelectrode recordings models, tremor v. slowness, r2=0.209±0.174 v. r2=0.198±0.147, LMM β=−0.005, Z=−0.496, p=0.620). To determine if tremor and slowness had distinct neurophysiological signatures, SVR model feature weights were aggregated for each metric. To understand which spectral features were used consistently across models, feature weights were compared to null distributions generated from models where motor metric values were shuffled with respect to the corresponding spectral features. Microelectrode tremor decoding models positively weighted low-frequency features (θ,α,β; 4–21 Hz; p<0.001, permutation test, see Materials and methods). Hfo (275–375 Hz) weights were also positively associated with tremor decoding (p<0.014, permutation test). In contrast, macroelectrode tremor decoding models negatively weighed β power (14–41 Hz; p<0.026, permutation test) while positively weighing γ/hfo activity (60–375 Hz; p<0.011, permutation test). In other words, optimal macroelectrode tremor decoding relied on decreased β power and increased γ/hfo power. For slowness, microelectrode decoding models had negative θ, γlow , and hfo weights (5–12 Hz, 33–56 Hz, 200–375 Hz) (p<0.012, permutation test). Macroelectrode decoding models positively weighted β frequencies (12–30 Hz; p<0.006, permutation test) along with negative γ/hfo weights (33–375 Hz; p<0.001, permutation test). Tremor and slowness model features differed when compared directly, with hfo frequencies (225–375 Hz) being elevated during tremor in both micro/macroelectrode recordings (Figure 3A). Overall, just as tremor and slowness represented two distinct, anti-correlated symptomatic states of PD, tremor and slowness decoding models from the STN revealed distinguishable patterns of underlying neural activity. Figure 3 Download asset Open asset Subthalamic tremor decoding models emphasized lower frequencies whereas slowness models emphasized higher frequencies. (A) Average tremor decoding and slowness model coefficients for all subthalamic nucleus (STN) microelectrode (left) (n=203 microelectrode recordings, 27 subjects) and macroelectrode (right) recordings (n=176 macroelectrode recordings, 27 subjects). Solid lines indicate average weights, with positive/negative values reflecting a positive or negative relationship with the metric. Error bars indicate s.e.m. across subjects. Black lines (top) represented contiguous spectral features that significantly differed between tremor and slowness decoding models. (B) Average model coefficients for effective motor control and tremor for all STN microelectrode (left) and macroelectrode (right) recordings. (C) Average model coefficients for effective motor control and slowness for all STN microelectrode (left) and macroelectrode (right) recordings. However, in order to rule out the possibility that the alternating patterns of relevant neural decoding features simply reflected the anti-correlated nature of tremor and slowness, we tested whether decoding models trained for tremor could accurately decode slowness. When directly comparing tremor and slowness decoding performance on tremor-trained models, slowness decoding was inferior for both microelectrode (tremor v. slowness decoding, r2=0.232±0.197 v. 0.002±0.001; LMM β=0.101, Z=11.242, p=2.54*10–29) and macroelectrode (tremor v. slowness decoding, r2=0.205±0.172 v. 0.002±0.001; LMM β=0.086, Z=10.294, p=7.53*10–25) recordings. If decoding features for tremor and slowness were simply inverted, applying models for decoding another metric would result in significantly negative r2-values. Thus, the neural features used for individual metric decoding likely reflected a unique spectral state or ‘fingerprint’. To further validate that our approach was able to decode motor dysfunction in a symptom-specific fashion, we examined the relationship between individual tremor expression and tremor decoding performance (as not all patients with PD exhibit tremor). Here, we found that task-based tremor distribution medians positively correlated with individual’s highest decoding performance (n=24 patients, ρ=0.442, p=0.031). While UPDRS measures of tremor did not directly correlate with decoding performance (p<0.878) – likely due to a variety of factors distinguishing standard UPDRS tremor assessment from intraoperative task performance – the dynamic, continuous, low-velocity-biased expression of tremor on the naturalistic task (which itself did correlate with UPDRS resting tremor scores) provided an immediate, ground-truth behavior for patient-specific neurophysiological decoding. Effective motor control had characteristic neural signatures Effective motor control was similarly decoded from both micro- (r2=0.140±0.104) and macroelectrode (r2=0.204±0.097) recordings. Effective motor control decoding was characterized by the absence of β (10–28 Hz) power in both micro- and macroelectrode recordings (p<0.006, permutation test), while macroelectrodes also exhibited positive γ power weights (30–175 Hz; p<0.020, permutation test). Power in γlow frequencies (30–48 Hz) in particular was significantly increased during effective motor control relative to both tremor and slowness decoding models (p<0.006, permutation test) (Figure 3B–C, right). In total, STN activity contained specific features that distinguished symptomatic from non-symptomatic motor states. Tremor was characterized by lower frequencies (θ/α) in microelectrodes, slowness by β frequencies in macroelectrodes, and effective motor performance was uniquely characterized by γlow frequencies from both recording types. Full-spectrum neural decoding outperformed beta-band decoding To directly test whether each behavior model used neural features across the spectrum, we compared the relative ability of full-spectrum and canonical band (β, 12–30 Hz) models. Full-spectrum decoding had significantly greater performance for macroelectrode (full vs. beta-only decoding, LMM β=0.018–0.035, Z=2.388–3.949, p<0.017) and microelectrode (full vs. beta-only decoding, LMM β=0.014–0.017, Z=2.241–3.154, p<0.025) for all three metrics, with the exception of microelectrode-tremor decoding (full vs. beta-only decoding, LMM β=0.014, Z=1.548, p=0.122). Optimal subthalamic tremor decoding sites were dorsolateral to optimal slowness decoding sites across patients To investigate whether tremor and slowness were more optimally decoded from distinct areas within the STN, recording sites for each session were reconstructed using subject-specific neuroimaging (peak MER density in MNI space: x=-12,y=-10,z=-6.0) (Figure 4A; peak macroelectrode recording density: x=-12,y=-9,z=-3.0). For each recording site, the corresponding decoding model performance for each metric was plotted (Figure 4B and C). Figure 4 Download asset Open asset Optimal subthalamic tremor decoding sites were dorsolateral to optimal slowness decoding sites. (A) Recording density of stationary microelectrode recordings across patients (n=182 microelectrode recording sites, 25 subjects) and task sessions overlaid on an MNI reference volume (approximate outline of the subthalamic nucleus [STN] in bolded black, zona incerta outlined above, substantia nigra outlined below). L: left. y-value corresponds to coronal slice in MNI space. (B) Tremor decoding model r2-values for stationary microelectrode recordings. (C) Slowness decoding model r2-values for stationary microelectrode recordings. (D) Difference in tremor vs. slowness decoding r2-values for stationary microelectrode recordings. Warmer colors indicate voxels where tremor decoding was superior, whereas cooler colors indicate where slowness decoding was superior. (E) Recording density of moving microelectrode recordings across all patients and task sessions overlaid on an MNI reference volume. (F) Tremor decoding model r2-values for high-density STN survey recordings. (G) Slowness decoding model r2-values for high-density STN survey recordings. (H) Difference in tremor vs. slowness decoding r2-values for high-density STN survey recordings. r2-Values depicted here are site-specific r2-values generated from the whole-STN model applied to individual depth recordings. Warmer colors indicate voxels where tremor decoding was superior, whereas cooler colors indicate where slowness decoding was superior. We then compared the voxel-wise relative performance between tremor and slowness throughout all recorded STN voxels by using a modified 3D t-test with spatially based permutation shuffling (see Materials and methods). Tremor was better decoded in recordings from dorsolateral STN (n=182 MER sites, x=-14.0,y=-13.0,z=-5.0; Z=2.116,p=0.017), whereas slowness was better decoded from recordings in central/ventromedial STN (x=-12.0,y=-14.0,z=-6.0; Z=1.911,p=0.028) (Figure 4D). Optimal locations for tremor and slowness decoding were not found to differ significantly by macroelectrode location (n=176 macroelectrode recording sites, p>0.05). Moreover, the locus of optimal effective motor control decoding was not observed to differ from those of tremor or speed using either micro- and macroelectrode recordings (p>0.05). Nevertheless, we found that that differences in metric decoding were not only related to the frequencies present, but also to an electrode’s location within the STN, as assessed over the entire study PD population. Optimal subthalamic tremor decoding sites were dorsolateral to optimal slowness decoding sites within individual patients To verify the spatial relationship of optimal tremor and slowness decoding within patients, five additional right-handed patients (70.0±8.9 years of age; 2F, 3M; UPDRS III: 45.2±9.5) underwent a modified version of the random-pattern task. Rather than acquiring recordings from a stationary site, here we surveyed the entire length of the STN by systematically moving the electrodes between task trials in small, discrete steps using automatic, computer-driven microdrive control (see Materials and methods, High-density STN survey). SVR models for tremor and slowness were then calculated by incorporating recording data across all sites/trials within a single trajectory. Although decoding performance of models derived from multi-site data exhibited a trend of lower performance than models trained on single-site recordings (tremor: r2=0.232±0.200 v. 0.073±0.052; slowness: r2=0.125±0.108 v. 0.061±0.079, effective motor control: r2=0.140±0.104 v. 0.043±0.058), these differences were not significant (n=203 stationary MER sites, n=17 moving MER trajectories, moving v. stationary data, LMM β=−0.075 to –0.075, Z=−0.945 to –1.291, p=0.197–0.344). Despite less data at each recording site, whole-STN models demonstrated above-chance decoding performance for all three metrics (tremor: 8/17 trajectories, slowness: 6/17, effective motor control: 6/17). Recording sites along each trajectory were reconstructed using imaging (Figure 4E), and site-specific metric decoding r2-values were calculated by applying the whole-STN SVR model to individual site recordings (Figure 4F and G) (see Materials and methods). Decoding performance was then compared across patients (Figure 4H). Across these recordings, tremor was optimally decoded at (x=−13.0,y=−13.0,z=−5.0; Z=1.911,p=0.028), while slowness was optimally decoded at (x=−12.0,y=−15.0,z=−8.0; Z=1.937,p=0.026). Within individual subjects, tremor was again found to be decoded dorsolaterally to slowness. Cortical recordings also revealed distinct representations of tremor, slowness, and effective motor control Ten subjects additionally had ECoG recordings from sensorimotor cortex (motor cortex: n=16 contacts, somatosensory cortex: n=15, see Materials and methods). SVR models for metric decoding were similarly trained on ECoG signals. ECoG decoding performance did not differ between tremor or slowness (n=85 ECoG recordings across 27 sessions and 10 subjects, tremor: r2=0.323±0.153, slowness: r2=0.314±0.143, tremor v. slowness, LMM β=0.010, Z=0.578, p=0.563). To understand which spectral features contributed to cortical motor metric decoding, SVR model weights were aggregated across all patients and recordings and compared to metric-shuffled models. When compared directly, cortical tremor and slowness models had opposing relationships in α/β (8–40 Hz, p<0.027, permutation test), γmid (45–125 Hz, p<0.023, permutation test), and γhigh (150–225 Hz, p<0.015, permutation test) frequency bands (Figure 5A). Tremor models additionally had positive weights associated with θ frequencies (5–7 Hz, p=0.003, permutation test). Altogether, although cortical signals supported equivalent decoding performance for tremor or slowness, decoding features were nonetheless distinct. On the other hand, effective motor control decoding performance (r2=0.469±0.112) was lower than tremor (effective motor control v. tremor, LMM β=−0.067, Z=−3.975, p=7.04*10–5) and slowness (effective motor control v. slowness, LMM β=−0.057, Z=−3.397, p=0.001). Nevertheless, effective motor control was represented in cortical decoding models by γhigh frequencies. These γhigh features additionally appeared to differentiate effective motor control models from both tremor and slowness models (125–175 Hz, p<0.026, permutation test) (Figure 5B–C). In addition, α/β (8–30 Hz, p<0.001, permutation test) and γlow (45–75 Hz, p<0.010, permutation test) frequencies exhibited an opposing relationship between effective motor control and slowness, much like the interaction between tremor and slowness models (Figure 5C). Taken together, although at different γ frequencies, both STN (γlow) and sensorimotor cortex (γhigh) exhibited features specific to effective motor control. And similarly to STN recordings, ECoG full-spectrum decoding was superior for tremor and effective motor control (full vs. beta-only decoding, LMM β=0.017–0.024, Z=2.224–2.577, p<0.026), but equivalent for slowness (full vs. beta-only decoding, LMM β=−0.013, Z=−1.531, p=0.126). Figure 5 Download asset Open asset Cortical tremor and slowness decoding models exhibited opposing weights for multiple frequency bands, and co-expressed specific features with subthalamic recordings. (A) Average cortical tremor and slowness decoding model coefficients for every recording along sensorimotor cortex (n=85 electrocorticography [ECoG] recordings, 10 subjects). Colored lines indicate average weights, with positive/negative values reflecting a positive or negative relationship with the metric. Error bars indicate s.e.m. across subjects. Black lines (top) represented contiguous spectral features that significantly differed between tremor and slowness decoding models. (B) Average model coefficients for effective motor control and tremor. (C) Average model coefficients for effective motor control and slowness. (D) Average subthalamic nucleus [STN]-cortical coherence tremor and slowness decoding model coefficients for every pairwise recording along sensorimotor cortex and macro contacts within the STN (n=85 ECoG recordings, 10 subjects). (E) Average coherence model for effective motor control and tremor. (F) Average model coefficients for effective motor control and slowness. Finally, we analyzed whether motor features were selectively represented in different