Effect of levodopa on electroencephalographic biomarkers of the parkinsonian state

Abstract

The objective of this study was to evaluate proposed electroencephalographic (EEG) biomarkers of Parkinson’s disease (PD) and test their correlation with motor impairment in a new, well-characterized cohort of PD patients and controls. Sixty-four-channel EEG was recorded from 14 patients with rigid-akinetic PD with minimal tremor and from 14 age-matched healthy controls at rest and during voluntary movement. Patients were tested off and on medication during a single session. Recordings were analyzed for phase-amplitude coupling over sensorimotor cortex and for pairwise coherence from all electrode pairs in the recording montage (distributed coherence). Phase-amplitude coupling and distributed coherence were found to be elevated Off compared with On levodopa, and their reduction was correlated with motor improvement. In the Off medication state, phase-amplitude coupling was greater in sensorimotor contacts contralateral to the most affected body part and reduced by voluntary movement. We conclude that phase-amplitude coupling and distributed coherence are cortical biomarkers of the parkinsonian state that are detectable noninvasively and may be useful as objective aids for management of dopaminergic therapy. Several analytic methods may be used for noninvasive measurement of abnormal brain synchronization in PD. Calculation of phase-amplitude coupling requires only a single electrode over motor cortex.

NEW & NOTEWORTHY Several EEG biomarkers of the parkinsonian state have been proposed that are related to abnormal cortical synchronization. We report several new findings in this study: correlations of EEG markers of synchronization with specific motor signs of Parkinson’s disease (PD), and demonstration that one of the EEG markers, phase-amplitude coupling, is more elevated over the more clinically affected brain hemisphere. These findings underscore the potential utility of scalp EEG for objective, noninvasive monitoring of medication state in PD.

INTRODUCTION

In Parkinson’s disease, loss of dopaminergic neurons in the basal ganglia results in circuit abnormalities throughout the basal ganglia-thalamo-cortical network. The akinetic motor signs of the disease are associated with an increase in synchronized beta-band oscillatory activity in this network (Brown 2003; Priori et al. 2004). Several physiological signatures or “biomarkers” of this excessive synchronization have been proposed, including excessive entrainment of basal ganglia neuronal discharge to the motor beta rhythm (Kühn et al. 2005), elevated beta-band coherence within the basal ganglia (DeLong and Wichmann 2007; Weinberger et al. 2009) and between cortical regions (George et al. 2013; Silberstein et al. 2005), high coupling of broadband gamma (50–200 Hz) amplitude to beta phase in motor cortex (phase-amplitude coupling, PAC; illustrated in Fig. 1A) (de Hemptinne et al. 2013), and alterations in the coupling of beta phase to the amplitude of very-high-frequency oscillations in the subthalamic nucleus (López-Azcárate et al. 2010). Successful therapy using either dopaminergic medication or deep brain stimulation (DBS) reduces beta-band synchronization (Jenkinson and Brown 2011). Several biomarkers derived from basal ganglia field potentials appear to correlate with the severity of motor impairment (Kühn et al. 2008; Little et al. 2012; Neumann et al. 2016; van Wijk et al. 2016).

Fig. 1.

Extracting phase-amplitude coupling (PAC) from an electric potential time series. A: illustration of hypothetical signals with and without coupling of low-frequency phase to high-frequency amplitude. B: schematic showing PAC derivation from a single patient’s EEG signal. The raw signal for each electrode is first filtered at both a lower frequency (beta, in our case) and a higher frequency (broadband gamma). The phase of the lower frequency activity and the amplitude envelope of the higher frequency activity are then extracted using a Hilbert transform. The high-frequency amplitude is then binned according to the phase of the lower frequency. From this distribution we compute a modulation index (MI) value, which quantifies the degree to which the broadband gamma amplitude is modulated by the low-frequency oscillation (Tort et al. 2008).

There is great interest in using noninvasive recording methods such as scalp electroencephalography (EEG) to identify the neural signals underlying specific signs and symptoms of Parkinson’s disease (PD). Such biomarkers have potential to aid in diagnosis, rational management of dopaminergic therapy, measurement of disease progression in neuroprotection studies, or as control signals for closed-loop DBS paradigms. Prior EEG work has suggested two potential measures that are excessive in PD and are reduced by levodopa: cortico-cortical coherence (George et al. 2013; Silberstein et al. 2005) and PAC in sensorimotor cortex (Swann et al. 2015). Of note, cortico-cortical coherence has been assessed using pairwise coherence across all channels of an EEG montage, which we refer to as “distributed coherence” after the terminology of Silberstein et al. (2005). In this study we recorded EEG from a new cohort of PD patients On and Off levodopa and from age-matched healthy controls. Our goal was twofold: to utilize a new data set to confirm previous EEG studies showing that PAC and coherence are modulated by disease state and medication, and to evaluate the hypothesis that these metrics are related to specific motor signs.

METHODS

Patient selection.

Patients with PD were recruited from the University of California, San Francisco (UCSF) Movement Disorder and Neuromodulation Center and from the Parkinson’s Disease Research, Education and Clinical Center at the San Francisco Veteran’s Affairs Medical Center. All patients were undergoing clinical evaluation for possible DBS therapy. PD patients with tremor [Unified Parkinson’s Disease Rating Scale Part III (UPDRS III) resting tremor score >1 on either side in the Off medication state] were excluded, because tremor may introduce movement artifacts in the EEG signal and because the physiology underlying tremor is distinct from other PD symptoms (Qasim et al. 2016). Age-matched healthy control subjects were recruited via informational flyers. Other exclusion criteria for patients were presence of a neurological disorder other than PD or use of related medications, or a total UPDRS III score <15 Off medication. Exclusion criteria for controls were presence of any neuropsychiatric disorder or use of psychiatric medication. Informed written consent was obtained before initiation of study procedures under a protocol approved by the UCSF Institutional Review Board.

EEG recording procedure.

Scalp EEG was recorded using a 64-channel BioSemi ActiveTwo system (Amsterdam, The Netherlands) at a sampling rate of 2,048 Hz. With the same system, surface electromyography (EMG) was recorded from biceps brachii, extensor carpi radialis, and tibialis anterior. EMG was recorded from patients’ more symptomatic sides and controls’ hand dominant sides.

For patients with PD, EEG was recorded Off and then On antiparkinsonian medications. In the Off medication condition, patients were instructed to hold their medication for at least 12 h before the recordings began. After the Off medication recordings, patients took their regular dose of levodopa and waited 1 h, without interval removal of the scalp cap, before On medication recordings commenced. Patients’ motor symptoms were quantified in both conditions by a movement disorders neurologist using the UPDRS III within 15 min of EEG recordings. For patients and controls, and for both medication conditions, EEG was recorded in two behavioral conditions: “rest” and “movement.” For rest recordings, subjects were asked to remain still and keep their eyes open for 5 min. For movement recordings, patients performed tasks using the hand on their more clinically affected side, and controls used their dominant hands. Subjects performed either verbally cued intermittent hand opening/closing (10 s of movement followed by 10 s of no movement, repeated 5 times) or a simple intermittent reaching task requiring tapping multiple targets on a touchscreen (3 s of movement followed by 7 s of no movement, repeated 20 times). Thus the movement condition included epochs of movement alternating with no movement, and due to the brevity of the epochs, they were not analyzed separately.

Signal processing.

EEG data were analyzed using a combination of custom MATLAB scripts (version R2014A; The MathWorks, Natick, MA) and the EEGLAB toolbox (Delorme and Makeig 2004). The mean of each channel was subtracted to remove direct current offset, and then data were referenced to the common average of all channels. Channels with obvious noise (electrical or muscle related), identified either visually in the raw signal or where power spectra failed to display the canonical 1/f pattern (Pritchard 1992), were excluded from the common average calculation. A two-way, 1-Hz high-pass filter (Butterworth, 3rd order) was applied to the data to remove low-frequency activity. Data were then visually inspected for periods of movement artifact and eyeblinks, which were marked for exclusion. Periods of artifact were not excluded until after all filtering was performed on data to reduce the impact of edge artifacts. On average, 30.6% of each recording was rejected due to artifact, and two patients who had >50% of their recorded data removed due to artifact were excluded from analysis. Further analysis of resting-state recordings was done using the first minute of data without artifacts. Movement start and stop indexes were extracted using the EMG recordings with a semiautomated threshold detection method and confirmed by visual inspection.

To reduce the contribution of possible contamination from muscle activity and volume conduction, EEG data were computationally re-referenced using a current source density (CSD) procedure (CSD toolbox, using a spherical spline with medium flexibility, m = 3; Kayser 2009), which isolates spatially specific features of the electrophysiological signal.

PAC was calculated using the Kullback-Leibler-based modulation index method (de Hemptinne et al. 2013; Tort et al. 2008). The EEG signal from each electrode was filtered using a two-way finite impulse response (FIR) filter (modified EEGLab toolbox function eegfilt.m with FIR1 parameters). Phase was extracted using a Hilbert transform on signals filtered from 4- to 50-Hz frequencies with a 2-Hz bandwidth. Amplitude was extracted using a Hilbert transform on signals filtered from 4 to 200 Hz with a 4-Hz bandwidth. For each phase-amplitude frequency pair, the distribution of the instantaneous amplitude envelope was computed for every 20° interval of the instantaneous phase, giving 18 phase binds in total. From this distribution, a Kullback-Leibler coupling measure (modulation index, MI) is derived, based on its deviation from a uniform distribution. This process is illustrated in Fig. 1. The MI values for each frequency pair can be visualized using a comodulogram (as shown in Fig. 2). For each patient in each condition, we averaged all MI values in the phase frequency range 13–30 Hz and amplitude frequency range 50–150 Hz. To confirm the significance of our sensorimotor PAC values, we also analyzed PAC after converting each patients MI values recorded from C3 and C4 to z scores, following the method of Canolty et al. (2006). In brief, this involves creating a surrogate distribution of PAC values by calculating the MI value after shifting the amplitude signal relative to the phase signal by a random value, and repeating this process with 200 different random shift values. The z score is then calculated by comparing the raw MI value to this surrogate distribution. Our results using z scores did not differ significantly from those obtained using raw MI values, and therefore our reported results use the raw MI values.

Fig. 2.

Phase-amplitude coupling (PAC) observed in sensorimotor cortex of patients with Parkinson’s disease (PD), On and Off medications, and healthy controls, at rest and during movement. A: median comodulograms showing PAC differences across all subjects. Each color pixel represents the strength of coupling (modulation index, MI) between the phase of the low-frequency band defined on the x-axis (Hz) and the amplitude of the high-frequency band defined on the y-axis (Hz), both extracted from the same sensorimotor contact (C3 or C4). Mean beta-broadband gamma PAC was calculated by averaging the indexes within the red box (13–30 Hz for beta phase and 50–150 Hz for broadband gamma amplitude). B: boxplots showing mean beta-broadband gamma PAC. Significant differences were observed between PD patients Off vs. On medication (two-tailed, paired Wilcoxon signed rank test, P = 0.025) and between patients Off medication vs. healthy controls (two-tailed, unpaired Wilcoxon rank-sum test, P = 0.016). C: boxplots showing mean sharpness ratios. There was no difference in waveform sharpness ratio within PD patients Off medication compared with On medication (P = 0.227, two-tailed, paired Wilcoxon signed-rank test) or between controls and patients Off medication (P = 0.767, two-tailed, unpaired Wilcoxon rank-sum test). D: median comodulograms showing reduction in PAC with voluntary movement (in Off medication state; P = 0.011, two-tailed, paired Wilcoxon signed-rank test).

Because PAC may be related to the “sharpness” of the peaks and troughs in the EEG time series, we also calculated waveform sharpness using the method of Cole et al. (2017), as adapted for scalp EEG data (Jackson et al. 2019). In brief, the common average re-referenced EEG time domain signal was bandpass filtered from 13 to 30 Hz. Zero-crossings were used to segment the waveform into beta-band cycles, and local maxima and minima were identified. A 12-ms window around each local maximum or minimum was then used to calculate the sharpness of each cycle’s peak and trough, respectively. The maximum of the peak-to-trough (or trough-to-peak) ratio was taken for each cycle, log₁₀ transformed, averaged for all cycles recorded from the sensorimotor electrodes used for PAC analysis, and compared between controls and patients and within patients by medication state.

For PAC analysis, data from the sensorimotor electrode contralateral to the patient’s more symptomatic side were analyzed (C3 for patients with primarily right-sided symptoms, C4 for patients with primarily left-sided symptoms). Each patient’s more symptomatic side was determined using hemibody rigidity and bradykinesia subscores of UPDRS III data. In healthy controls, PAC was analyzed from the sensorimotor electrodes contralateral to the subject’s dominant hand (C3 for right-handed, C4 for left-handed). For analysis of PAC in response to movement, we utilized the rest and movement conditions defined in EEG recording procedure. To perform group statistical analysis, a PAC value was derived for each subject in each condition by averaging the MI value over the beta range (13–30 Hz) for phase frequency and over the broadband gamma range (50–150 Hz) for amplitude frequency. Frequency ranges were selected on the basis of a previous publication (Swann et al. 2015). Mean log power spectral density (PSD) was also analyzed from the electrode contralateral to each patient’s more symptomatic side. PSD was calculated using the Welch method (MATLAB function pwelch; 512-ms window, 256-ms overlap).

Coherence was computed using the mscohere function in MATLAB (512-ms window, 256-ms overlap). Using previously reported methods (George et al. 2013; Silberstein et al. 2005), we analyzed, for every possible pair of electrodes (64 electrodes, 2,016 unique pairs), the correlation of mean coherence (from 2 to 50 Hz in 2-Hz bins) with the subjects’ UPDRS III scores across all 14 subjects. For each subject, coherence values were averaged over the entire recording length (excluding periods of artifact), giving one coherence value for each channel pair in each 2-Hz frequency bin. We then checked for a relationship between cortical coherence and severity of parkinsonian motor signs in two ways: first, by calculating the relative prevalence of positive vs. negative correlations of coherence vs. UPDRS scores in the set of all 2,016 unique coherence pairs (separately for the Off and On medication states), and second, by calculating the relative prevalence of positive vs. negative correlations between the levodopa-induced change in coherence and the levodopa-induced change in UPDRS scores for all coherence pairs. These relative prevalences were calculated separately for each of the 25 frequency bins and plotted to show differences in positive vs. negative correlations at each frequency following previously published methods (George et al. 2013; Silberstein et al. 2005).

Significance for these numbers of correlations was evaluated by using permutation testing to determine whether there were more or fewer significant correlations than would be expected by chance (Pitman 1937; Puth et al. 2014). For each coherence vs. UPDRS comparison, the coherence values were randomly permuted across subjects while the UPDRS scores were kept the same. A new correlation coefficient was computed on these shuffled data, and this procedure was repeated for n = 200 permutations, for all 2,016 electrode pairs, giving a null distribution of numbers of significant correlations in each frequency bin. Significant numbers of correlations for a given frequency bin were then defined as those that differed significantly from the null distribution [P < 0.025, false discovery rate (FDR) corrected].

Statistical analysis.

A paired, two-tailed Wilcoxon signed-rank test was used to compare mean MI values Off and On levodopa and during rest vs. movement. An unpaired, two-tailed Wilcoxon rank sum test was used to compare median MI values in patients vs. controls. Spearman’s ρ was used to analyze correlations between change in either PAC or coherence and change in UPDRS III scores and subscores. Nonparametric statistical tests were used because MI values are generally not normally distributed across patient populations, and nonnormality of our data set was confirmed using a Lilliefors test (P < 0.001 for patients Off and On levodopa and for controls).

RESULTS

Patient characteristics.

We recorded scalp EEG activity from 14 patients with rigid-akinetic PD and from 14 age-matched healthy control subjects. Patient clinical characteristics are summarized in Table 1. The mean age was 61.2 ± 8.8 yr for patients and 66.1 ± 6.3 yr for controls (P = 0.10 for the difference in age between patients and controls, two-sided t-test). Patients had a mean UPDRS III score of 37.0 ± 11.34 Off medication (maximum score = 56) and 22.8 ± 8.9 On medication, with all patients’ scores improving after levodopa administration. Of note, 19 patients with PD enrolled in the study, but 5 patients were excluded for the following reasons: 2 patients had limb tremor UPDRS scores >1 (out of a maximum of 8 per side) during Off medication recording, 2 had recordings contaminated by strong environmental noise, and 1 had too low a UPDRS III score (7) during Off medication recording to be included in this cohort.

Table 1.

Patient characteristics

Patient	Age, yr/Sex	Handedness/More Symptomatic Side	Disease Duration, yr	Medications, Daily Total, mg	UPDRS Off Medication (Lateralized Scores Left/Right)	UPDRS On Medication (Lateralized Scores Left/Right)
1	46/M	R/L	5	400	39 (15/14)	21 (10/9)
2	60/M	R/L	15	1,000	44 (16/13)	31 (12/8)
3	73/M	R/L	9	800	46 (20/11)	34 (13/9)
4	49/M	R/L	5	800	36 (12/8)	22 (9/6)
5	54/M	R/L	8	750	47 (11/9)	28 (8/8)
6	58/M	R/R	8	600	37 (9/12)	30 (7/10)
7	73/M	R/R	15	1,000	41 (11/12)	21 (3/9)
8	53/M	R/R	11	1,000	60 (15/16)	30 (2/7)
9	59/M	R/R	9	980	41 (10/13)	34 (8/9)
10	68/M	R/L	11	1,250	31 (14/14)	23 (11/10)
11	67/M	R/L	11	1,000	28 (14/11)	8 (6/1)
12	63/M	R/L	6	800	31 (19/9)	16 (9/5)
13	66/F	R/L	10	300	23 (8/7)	12 (3/3)
14	51/M	R/L	6	1,250	14 (7/6)	9 (4/4)

Lateralized Unified Parkinson’s Disease Rating Scale (UPDRS) scores include only scores for items measuring hemibody rigidity, bradykinesia, and tremor. M/F, male/female; R/L, right/left.

PAC is elevated Off levodopa and greater over the more affected hemisphere.

At rest, PAC analyzed from sensorimotor electrodes (C3 or C4) was elevated in patients with PD Off levodopa compared with On (P = 0.025, two-tailed, paired Wilcoxon signed-rank test; Fig. 2, A and B), with 13 of 14 patients showing this effect. PAC was also higher in patients Off levodopa compared with healthy controls (P = 0.016, two-tailed, unpaired Wilcoxon rank sum test). No significant difference was observed between patients On medication and healthy controls (P = 0.31). In patients Off medication, PAC was stronger over sensorimotor electrodes contralateral to the more symptomatic side than over ipsilateral sensorimotor electrodes (P = 0.031, two-tailed, paired Wilcoxon signed-rank test), but this difference was not present in the On medication state (P = 0.27, two-tailed, paired Wilcoxon signed-rank test). Of note, the PAC difference between the Off and On states was only significant when PAC was analyzed from sensorimotor electrodes contralateral to patients’ more symptomatic side, and no difference in PAC was observed between Off vs. On medication using either data from less symptomatic sides (P = 0.27) or the average of PAC recorded over both hemispheres (P = 0.22).

Previous studies have identified an effect of medication and DBS state on waveform sharpness ratio using EEG (Jackson et al. 2019) and electrocorticographic (ECoG) data (Cole et al. 2017), respectively. There was no difference in waveform sharpness ratio between controls and patients Off medication (P = 0.767) or between controls and patients On medication (P = 0.744) using the two-tailed, unpaired Wilcoxon rank sum test, or between sharpness ratio in PD off medication compared with On medication (P = 0.227, two-tailed, paired Wilcoxon signed-rank test).

Elevated PAC values were also observed in nonsensorimotor electrodes, especially those near frontalis and temporalis muscles. Prior work showed that muscle artifact can also result in high PAC values (Swann et al. 2015). However, when PAC was averaged over all frontal or all temporal contacts, there was no difference between PAC in the Off vs. On medication states (P = 0.72 and P = 0.51, respectively, two-tailed, paired Wilcoxon signed-rank test). Thus the effect of medication on PAC appears to be specific for sensorimotor contacts and is unlikely to be driven by a contribution from muscle activity. To underscore this point, we plotted whole scalp topographical maps to characterize the spatial distribution of PAC differences Off vs. On medications (Fig. 3). The differences were strongest over sensorimotor areas, suggesting they are less likely to be driven by muscle activity. PAC was also significantly elevated over sensorimotor electrodes compared with nonmotor areas that are distal to muscle areas (posterior parietal electrodes P1 and P2; P = 0.003, two-tailed Wilcoxon signed-rank test).

Fig. 3.

Whole scalp topography of phase-amplitude coupling (PAC) comparisons. A and B: for all 64 scalp electrodes, statistical differences in PAC values were computed between all Parkinson’s disease (PD) patients Off and On levodopa (A) and all patients Off levodopa and healthy controls (Ctrl; B). The normalized test statistic for each comparison is plotted for each electrode that showed a significant difference in PAC (P < 0.05; Wilcoxon signed-rank test for A, Wilcoxon rank sum test for B). For each comparison, the test statistic was normalized to the percentage of the maximum test statistic for all channels to account for the different ranges of the paired and unpaired test statistics.

PAC values may be affected by spectral power of the underlying signal at phase frequencies. Beta-band spectral power calculated from C3 and C4 electrodes did not differ between PD Off and PD On (P = 0.68, two-tailed, paired Wilcoxon signed-rank test) or between PD Off and healthy controls (P = 0.55, two-tailed, unpaired Wilcoxon rank-sum test.) Thus the observed differences in PAC are unlikely to be driven by differences in beta power between groups.

PAC is attenuated during voluntary movement.

PAC was originally defined using invasive ECoG, a technique with much higher spatial resolution and more favorable signal-to-noise characteristics than EEG. ECoG studies in both patients with PD (de Hemptinne et al. 2015) and in humans without movement disorders (Miller et al. 2012; Yanagisawa et al. 2012) showed a movement-related PAC reduction in sensorimotor areas. To confirm that PAC derived from EEG in the present study shows this canonical movement-related reduction, we examined recordings during movement vs. rest. In the Off levodopa state, PAC was decreased during recordings with movement compared with the resting recordings (P = 0.011, two-tailed, paired Wilcoxon signed-rank test; Fig. 2D). No difference in PAC was observed between rest and movement recordings in the On levodopa state or between Off levodopa patients and healthy controls.

Levodopa-associated improvements in bradykinesia correlate with PAC change.

To examine the relationship between PAC and motor symptoms, we correlated levodopa-induced changes in PAC with the levodopa-induced change in UPDRS III scores collected during the same visit. Statistics were corrected for multiple comparisons using a nonparametric permutation-based method (Groppe et al. 2011). The percent decrease in PAC was correlated with percent decrease in the bradykinesia subscale (Spearman’s ρ = 0.60, P = 0.030; Fig. 4A). No significant correlation was observed between decrease in PAC and decrease in rigidity (ρ = −0.0176, P = 0.73; Fig. 4B) or with percent decrease in total UPDRS III (ρ = 0.33, P = 0.25; Fig. 4C). No significant correlations were observed between baseline (Off levodopa) PAC and baseline severity of bradykinesia (ρ = 0.18, P = 0.12), rigidity (ρ = 0.10, P = 0.26), or total UPDRS (ρ = 0.01, P = 0.64).

Fig. 4.

Correlations between phase-amplitude coupling (PAC) and patient symptoms. A: correlation between percent decrease in modulation index (MI) and percent decrease in contralateral bradykinesia subscale of Unified Parkinson’s Disease Rating Scale (UPDRS; ρ = 0.060, P = 0.030). B: nonsignificant correlation between percent decrease in MI and percent decrease in contralateral rigidity subscale of UPDRS (ρ = −0.018, P = 0.73). C: nonsignificant correlation between percent decrease in MI and percent decrease in total UPDRS Part III scores (ρ = 0.33, P = 0.25). Statistics shown are corrected for multiple comparisons.

Distributed cortico-cortical coherence.

Several prior studies have examined levodopa-induced changes in cortical synchronization with EEG, by computing the pairwise coherence across the entire recording montage. We replicated those results in our data set using the same methodology as prior studies (George et al. 2013; Silberstein et al. 2005), analyzing the correlation of coherence with UPDRS scores across all 14 subjects, separately for each contact pair and each frequency band (example in Fig. 5A). Consistent with previous findings (George et al. 2013; Silberstein et al. 2005), we found a prevalence of positive correlations between Off medication coherence and Off medication UPDRS III scores and between levodopa-induced change in coherence and levodopa-induced change in UPDRS III scores (Fig. 5B). We also conducted this analysis using the bradykinesia and rigidity subscales of the UPDRS III and observed significant numbers of positive correlations between levodopa-induced change in coherence and levodopa-induced change in rigidity. Significant relationships between distributed coherence and bradykinesia subscores were not observed.

Fig. 5.

Effect of levodopa on pairwise EEG coherence and its relationship to levodopa-induced changes in motor function, as previously described (George et al. 2013; Silberstein et al. 2005). A: example matrices of correlations between coherence and Unified Parkinson’s Disease Rating Scale (UPDRS) scores at one frequency bin (30 Hz) for all EEG contacts, averaged across all 14 subjects, arranged anatomically. Left: correlation between coherence at 30 Hz and UPDRS score Off levodopa. Right, correlation between medication-induced change in coherence and medication-induced change in UPDRS score. Both plots are thresholded for significant (P < 0.05) positive (light) and significant negative (dark) correlations. Each pixel represents one pair of scalp electrodes. Inset at far right illustrates the correlation plot for a single pixel of the right matrix (FC4-C3 coherence Off-On). B: the correlation analysis in A was performed for each frequency 0–50 Hz, in 2-Hz bins. Plots show percentage of electrode pairs (out of 2,016 possible pairs) in each frequency bin showing significant positive (top) and negative (bottom) correlations between coherence and UPDRS scores in the Off medication state (left) or between levodopa-induced changes in coherence and changes in UPDRS scores (right). Positive correlations outnumber negative correlations. Asterisks indicate significant numbers of correlations relative to a null distribution (correlations that would be expected to occur by chance alone).

The above-described analysis of cortico-cortical coherence utilizes a full EEG recording montage, a disadvantage compared with methods requiring only a single electrode. To determine if fewer contacts could yield a coherence measure that differed by medication state, or between PD patients Off medication and control subjects, we also examined, across all patients, interhemispheric beta-frequency (13–30 Hz) coherence between sensorimotor electrodes (C3 and C4) and intrahemispheric coherence between motor and sensory electrodes (C3 and CP3; C4 and CP4) but found no differences between patients Off and On medication (P = 0.10, P = 0.12, P = 0.28; two-tailed, paired Wilcoxon signed-rank test) or between patients Off medication and healthy controls (P = 0.32, P = 0.28, P = 0.55; two-tailed, unpaired Wilcoxon rank-sum test). We also tested a simpler method of comparing distributed coherence across subjects: averaging pairwise coherence, in the beta-frequency range (13–30 Hz), across all contact pairs within a single subject and then determining if the grand average of coherence across all subjects within a group distinguished disease state or medication state. However, this metric did not differ between patients Off and On medication (P = 0.95, two-tailed, paired Wilcoxon signed-rank test).

DISCUSSION

We evaluated two previously described EEG measures of elevated neuronal synchronization in PD. Utilizing a new, well-characterized patient cohort, we confirm prior findings that PAC recorded from sensorimotor contacts is elevated in PD patients Off levodopa compared with both On levodopa and age-matched healthy control subjects (Swann et al. 2015). We add several novel findings regarding EEG measurement of PAC in PD: sensorimotor PAC is higher over the more affected hemisphere compared with the less affected hemisphere, and across subjects, the levodopa-induced PAC decrease correlates most specifically with reduction in bradykinesia. In the same data set, we also confirm previous work related to the effect of levodopa on distributed coherence, the ensemble of pairwise coherence measurements across all 64 EEG contacts. In the Off medication state, distributed coherence across a group of subjects correlates with the severity of parkinsonian motor signs, and levodopa-induced reduction in distributed coherence correlates with levodopa-induced reduction in the severity of motor signs. Compared with PAC, however, changes in distributed coherence were more related to changes in rigidity than bradykinesia.

Electrophysiological signatures of the parkinsonian state.

There is considerable interest in utilizing electrophysiological methods to identify circuit-level abnormalities related to specific parkinsonian motor signs. In addition to providing insight into the origin of motor dysfunction, such physiological “biomarkers” could prove useful in diagnosis, as objective monitors of the effects of therapy or as control signals for feedback-controlled deep brain stimulation (Swann et al. 2018). Whereas the earliest work in this area was based on invasive microelectrode recording of the rates and patterns of single-unit discharge in basal ganglia nuclei (Lin et al. 2008; Remple et al. 2011), subsequent work has emphasized the importance of synchronized neural activity across an ensemble of neural elements, especially oscillatory synchronization in the beta band. Local field potential (LFP) recording from DBS electrodes implanted in the subthalamic nucleus shows that oscillations in the beta band are reduced by effective therapy (both levodopa and DBS; Jenkinson and Brown 2011) and may correlate with the severity of motor signs across patients (Neumann et al. 2016). Another manifestation of elevated beta-band synchronization is PAC, the tendency of high-frequency activity to occur at a preferred phase of the dominant motor beta rhythm. This has been studied in basal ganglia using LFP recordings (López-Azcárate et al. 2010; van Wijk et al. 2016; Yang et al. 2014), as well as in primary motor cortex using ECoG (de Hemptinne et al. 2013; Kondylis et al. 2016; Malekmohammadi et al. 2018). In the cortex, elevated beta phase–high-gamma amplitude could be a surrogate measure of the entrainment of population spiking to the motor beta rhythm (de Hemptinne et al. 2013).

However, all of these measures are invasive and thus restricted to patients with PD undergoing surgical procedures. There are obvious advantages to detection of biomarkers noninvasively, including applicability to a much larger group of patients, comparison with healthy controls, and use in longitudinal study designs. Although magnetoencephalography (MEG) has been tested as a tool for studying PD-related circuit alterations (Stam 2010; Stoffers et al. 2008), EEG is technically simpler, less expensive, and more widely available.

EEG coherence vs. PAC as noninvasive biomarkers.

Cortico-cortical coherence, measured by EEG, is a conceptually appealing and potentially straightforward metric of elevated synchronization in PD. However, the simplest analytic techniques, such as calculation of pairwise coherence between contacts near sensorimotor cortex, do not appear sensitive enough to assay effects of therapy or show differences across subjects of varying severity. Although the “distributed coherence” method previously described, and replicated here, is sensitive to effects of therapy, it requires a multichannel recording montage, and its sensitivity to effects of therapy has only been statistically demonstrated by assessing correlations of distributed coherence to severity of motor signs across a group of subjects with varying degrees of motor impairment and varying responses to levodopa. The finding that medication-induced changes in distributed coherence correlate specifically with changes in rigidity suggest that rigidity arises from a spatially distributed cortical network.

Other previously described measures of distributed coherence have proven useful in studying movement-related changes in neuronal synchronization, but not as a resting-state biomarker (Weiss et al. 2015). In contrast, the effect of levodopa on PAC is spatially specific to sensorimotor cortex and yields a metric that can characterize a single subject, for comparison across subjects or with nonparkinsonian conditions. This may be more practical for noninvasive assessment of effects of therapy or as a noninvasive signal for feedback-controlled neurostimulation. With advancements in wearable technologies, use of EEG for long-term monitoring is becoming more and more achievable (Mullen et al. 2015), suggesting that simple approaches for acquiring EEG at home are possible. Cortical PAC has also been shown to be related to isolated dystonia and the therapeutic effects of DBS in dystonia (Miocinovic et al. 2015), which suggests there may be value in exploring these EEG biomarkers in other movement disorders.

Limitations.

Neurologists were not blinded to medication state during clinical ratings, because the recording sessions were combined with patients’ ongoing clinical evaluations associated with DBS candidacy. However, UPDRS III scores were permanently logged into clinical and research records before EEG analysis, and the clinicians who rated symptoms did not analyze the EEG data. UPDRS ratings were not done during the recordings themselves, but several minutes before Off medication state recordings and just after On medication state recordings, respectively, because the wires from the EEG setup hindered patient performance of UPDRS tasks, and the voluntary movements during the UPDRS examination would themselves tend to reduce measures of synchronization (Pfurtscheller and Aranibar 1979). The stationarity of PAC over time is not well understood, so the sampling time required for a representative estimate is unknown. Calculation of PAC involves measurement of high-frequency signals (50–150 Hz), which are low in amplitude, and the ability to detect these frequencies by scalp EEG is controversial. However, our observation of the canonical movement-related reduction in EEG PAC, expected to occur based on prior invasive studies that used the higher signal amplitude technique of ECoG (de Hemptinne et al. 2015; Miller et al. 2012; Yanagisawa et al. 2012), provides reassurance that the PAC studied here from EEG is physiological rather than artifactual. We acknowledge that the physiological interpretation of PAC is controversial (Aru et al. 2015; Cole et al. 2017; Kramer et al. 2008). PAC may reflect the asymmetry between “sharpness” of peaks and troughs in cortical recordings, and the biophysical origin of these waveform changes are not clear. Although Jackson et al. (2019) demonstrated that sharpness ratio decreased in response to levodopa, we did not observe this difference in our data set. Finally, in our data set, one patient had PAC values that were very low (median value On and Off <1 × 10⁻⁶). This suggests that meaningful PAC may not be detectable in all subjects, and at this time this metric cannot be used to reliably distinguish PD from the normal state on a single-subject basis. Longer recording times could potentially improve single-subject reliability in future studies.

Conclusions.

Our results support the hypothesis that cortical oscillatory synchronization in patients with PD is elevated in the Off medication state, compared with both the On medication state and healthy controls, and that this difference can be detected noninvasively. Medication-related changes in PAC are spatially specific to the sensorimotor cortex, are closely related to parkinsonian bradykinesia, and could be useful as an objective index of the effects of therapy.

GRANTS

This work was supported by National Institutes of Health Grant R01-NS090913-01 (to P. A. Starr), the Bachman-Strauss Dystonia and Parkinson Foundation, the American Brain Foundation Clinical Research Training Fellowship (to S. Miocinovic), the UC President’s Postdoctoral Fellowship (to N. C. Swann), and patient gifts.

DISCLOSURES

Several authors (N. C. Swann, C. de Hemptinne, J. L. Ostrem, P. A. Starr) have intellectual property related to the work presented (US patent no. 9,295,838).

AUTHOR CONTRIBUTIONS

A.M.M., S.M., N.C.S., C.d.H., and P.A.S. conceived and designed research; A.M.M., S.M., N.C.S., and S.R. performed experiments; A.M.M., S.M., D.D., and R.G. analyzed data; A.M.M., S.M., N.C.S., D.D., R.G., C.d.H., J.L.O., and P.A.S. interpreted results of experiments; A.M.M. prepared figures; A.M.M. drafted manuscript; A.M.M., S.M., N.C.S., S.R., D.D., C.d.H., J.L.O., and P.A.S. edited and revised manuscript; A.M.M., N.C.S., and P.A.S. approved final version of manuscript.