Data from 12 participants with PD (4 females; mean age = 64.6 ± 1.5 years) undergoing bilateral STN-DBS implantation surgery were included (see Table 1). All participants completed an extensive pre-surgical assessment including detailed neurological examination, structural MRI, levodopa challenge and formal neuropsychological evaluations. They participated under procedures approved by the University of Iowa’s Institutional Review Board, and all provided written informed consent prior to study participation.

STN-DBS implantation was performed such that bilateral DBS electrodes were implanted sequentially during a single stereotactic procedure. Indirect framed stereotactic targeting of the STN was refined by multielectrode recordings from 0.4 to 0.8 MΩ tungsten microelectrodes (Alpha Omega Co., Inc., Alpharetta, GA, USA). Three to five MER tracks were used for each hemisphere, typically consisting of three simultaneous (anterior, middle, and posterior) trajectories each separated by 2 mm center-to-center. These simultaneous trajectories allowed detailed (i.e., sub-millimetric) sampling of the anterior/posterior and dorsal/ventral axes of STN. STN margins were defined by the functional and electrical properties from STN clinical MER, including presence of large and medium amplitude single units demonstrating burst, pause, or tremor related activity, consistent with standard reported techniques with current commercial software (e.g., Alpha Omega Co., Inc.; for similar studies see: Benazzouz et al., 2002; Cuny et al., 2002; Gross et al., 2006; Thompson et al., 2018). Behavioral and neurophysiological data were acquired simultaneously using a Tucker-Davis Technologies (TDT) multi-channel recording system (System 3, Tucker-Davis Technologies, Alachua, FL, USA). MERs were sampled at 24 kHz, amplified and passed from the Alpha Omega system into the TDT system so that spike, voice, and limb movement data had a common timescale.

Because MERs are clinically informative, participants were not exposed to extra electrode penetrations to participate in this research. Surgery was typically prolonged by < 15 min per hemisphere to conduct speech and limb testing. Participants did receive analgesic and sedative medications in the operating room; these were short-acting agents and were stopped > 1 h prior to MER, allowing participants to be maximally awake for necessary clinical testing and research. All participants were off PD medications for ∼12 h prior to participation.

Electrode locations within the physiologically defined boundaries of STN were included in the study. Since multiple simultaneous microelectrodes were utilized, recording locations for each participant were determined relative to the dorsal border of STN. All recording locations for this study were localized to the dorsal half of STN as this is the accepted motor subregion of STN and target for DBS leads. Intraoperative time constraints precluded receptive field mapping of STN neurons (e.g., face versus limb vs. potential speech responsive intra-STN differences). Each participant had at least two microelectrodes in dorsal STN during experiments. Per clinical protocol, we did not routinely probe the medial/lateral STN axis, nor did our clinical protocol include routine post-operative brain imaging. Therefore, we are unable to comment on the relative medial/lateral distribution of recording locations within STN.

Once microelectrodes were positioned within the physiologically defined STN, participants performed interleaved trials of speech and limb motor tasks (Figure 1). Participants received a verbal cue from the experimenter instructing them to either speak or tap their finger. Speech consisted of repetitions (3.4 ± 0.3 s mean duration) of one of two tokens often used in PD clinical speech evaluation: diadochokinesis (/tɑtɑtɑ/) and sustained phonation (/ɑ/). Speech tokens were interleaved throughout ∼10 min recording sessions to produce roughly equal amounts of/tɑtɑtɑ/and/ɑ/trials. Participants were told to speak at a rate and loudness that they would use in everyday conversation. Speech signals were captured via a condenser microphone (Beta 87A, Shure, Inc, Niles, IL, USA) positioned close to the mouth, amplified (Audio Buddy, M-Audio LLC, Cumberland, RI) and recorded with MERs using a common time scale. Speech onsets and offsets were detected visually in Matlab from the recorded speech signal.

The limb motor task consisted of blocks of repetitive finger tapping (9.6 ± 0.8 s mean duration) and was captured with a goniometer (SG65, Biometrics Ltd, Newport, UK) fixed to the thumb and first finger. Participants performed finger tapping using the hand contralateral to the STN recorded. Onset and offset of finger tapping were derived from the goniometer signal reflecting when recorded amplitudes deviated from background levels visually in Matlab. The speech and finger tasks were interleaved and chosen to provide short duration, high-yield tasks to assess speech and limb motor function yet provide enough repetitions to allow averaging and correlation with neuronal firing rates.

Raw voltage time series data were sampled from STN MERs at 24 kHz and high pass filtered at 2 kHz to lessen contamination by speech or movement artifacts. The choice of this high pass filter was due to occasional presence of microphonic artifacts below 2 kHz. In a recent study, we showed that microelectrodes that are used to collect spike/LFP data are susceptible to microphonic artifacts (Berger et al., 2022). Although a high pass of 2 kHz can distort the normal spike waveforms and potentially impact signal-to-noise ratio (SNR) and by discarding spikes with dominant frequencies below 2 kHz, it ensures that the data were not substantially contaminated by microphonic artifacts and in our previous study the majority of genuine physiological spikes remained identifiable. Prior to filtering, we used the demodulated band transform (DBT) method to detect and remove transient noise and artifacts from the filtered signal (Kovach and Gander, 2016). In addition to filtering, two additional techniques were utilized to ensure adequate MER quality and mitigate potential contamination by electromechanical artifacts. First, visual inspection was performed and trials containing any remaining microphonic artifacts arising from speech trials were excluded from analysis. Second, trials were evaluated for waveform instability, or non-stationarity. Non-stationarity can result from decreased spike amplitude at high firing rates (Fee et al., 1996; Stratton et al., 2012) or sub-millimetric displacement of the MER during heartbeat and respiration (Joshua et al., 2007). Units affected by non-stationarity appear as multiple smeared clusters in principle component space yet display characteristic inter-spike interval distributions consistent with single units (Fee et al., 1996; Snider and Bonds, 1998). We hypothesized that non-stationarity trials might occur more frequently during tasks due to head movement and vibration, and therefore tested whether louder speech could increase waveform non-stationarity due to speech-induced vibrations causing the STN microelectrode to be momentarily displaced relative to its neuronal target.

To assess the degree of speech-locked non-stationarity in our presumptive single units with respect to voice loudness, root mean square values (RMS) were calculated. RMS is defined as [∑1nxi2n]in which x_i represents a voltage sample. For RMS_signal, x_i was comprised of the samples recorded in a time window (−0.5 ms to +1.2 ms] surrounding the threshold crossing of n spikes of a presumptive single unit. For RMS_sound, x_i was comprised of the speech audio samples recorded in the same time window (−0.5 ms to +1.2 ms from spike detection) used in RMS_signal. RMS_sound was calculated to determine the correlation of the amplitude of audio waveforms captured by the condenser microphone (i.e., voice loudness) with the peak-to-trough amplitude of spike waveforms. RMS_signal was regressed over RMS_sound to determine if increased voice loudness coincided with a nonstationary spike waveform. Following the method of Stratton et al. (2012), SNRs were calculated as a ratio of RMS values (RMS_signal/RMS_noise; RMS_noise was calculated from the entire recording after extracted spikes removed) for each presumptive single unit and found to be satisfactory given the neuronal density of STN: (mean 2.7 +/−0.8, range 1.8–5.1; Zwirner et al., 2017).

After rejection of contaminated MERs based on post-filtering visual inspection and assessment of non-stationarity, offline spike detection was performed by thresholding at ± ∼3 std. Spikes were sorted using principal component analyses (PCA) in Plexon Offline Sorter (Plexon, Inc., Dallas, TX). Single units were identified as having (i) consistent waveform shape; (ii) separable clusters in principal component space; and (iii) a consistent refractory period of at least 1 ms in interspike interval (ISI) histograms. Figure 2 shows an example of putative single neuron waveforms and ISI distributions. We further quantified the quality of our single units clusters using isolation distances (Joshua et al., 2007) and refractory period violation (Hill et al., 2011). Isolation distance (IsoD), an index of clustering quality for single units, was calculated for single units clusters in Plexon (Joshua et al., 2007). The mean IsoD index wherein more than one unit was recorded on a single channel (n = 8) was 24.97 in PCA space. Moreover, refractory period violation was calculated for all 28 single neuron clusters (for more details, see: Hill et al., 2011). We found that these neurons had a mean of 5.06% of spikes occurring within 3 ms of each other. Additionally, a more stringent cutoff (ISIs < 1 ms) for refractory period violation showed that 0.007 percent spikes occurred within 1 ms each other.

Spike times from individual neurons were examined relative to the onset and offset of speech and limb movement using the co-recorded speech and goniometer recordings, respectively. Spike-time activity was aligned according to trial onset or offset, followed by binning to create peri-event time histograms (PETH). PETH data were constructed for each neuron and task.

As others have reported, STN neurons have demonstrated task-related modulations at both single and multi-unit levels (Tankus et al., 2017; Lipski et al., 2018; London et al., 2021). Hence, multi-unit analyses were performed here to investigate the modulations of neuronal clusters at a larger population level for speech and limb movement. All parameters for multi-unit sorting were similar to the single unit analysis. MountainSort toolbox was used to perform multi-unit analysis (Chung et al., 2017). Briefly, data were first bandpass filtered between 2,000 and 6,000 Hz and whitened for spike waveform detection, and then run through the MountainSort algorithm. This offers a fully automated, fast approach to spike detection, involving PCA following amplitude thresholding and a proprietary algorithm for clustering [(ISO-SPLIT) Magland and Barnett, 2015]. Sorted spikes were imported into Matlab for visualization and manual curation. Clusters were labeled as multi-unit if more than one percent of spikes had an interspike interval < 1 ms, consistent with our single unit analyses and others (e.g., Roy and Wang, 2012).

Absolute firing rate in spikes per second (Hz) and response modulation index (RMI) were used to quantify neuronal patterns in STN and determine speech and limb movement-related changes. Firing rate was calculated on a single trial basis and was calculated by dividing total spikes during a given trial by trial length in seconds (Hz = spikes/s). RMI, as a measure of percent change in firing between two conditions, was used to assess firing rate modulation occurring around behavioral events (Eliades and Wang, 2003, 2008; Eliades and Tsunada, 2018). Note absolute firing rate is a complementary and more general measure of firing rate changes following the onset of the speech and limb task. RMI was used as a normalized index which can measure more transient modulations around the onset or offset of the events. RMI was calculated using the formula RMI = (R_interval –R_baseline)/(R_interval+ R_baseline) where R_baseline is spikes per second pre-event (−0.5 to 0 s) and R_interval is spikes per second post-event-i.e., after task started (onset) or stopped (offset) (0 to +0.5 s). Baseline for speech and limb offset was 0.5 before the offset of the task. RMI was used to analyze STN neuronal modulation occurring around task onset and task offset, i.e., the time at which subjects started or stopped a given trial. Trials were excluded from RMI analysis if R_baseline contained any carryover task performance (e.g., speech occurring during the R_baseline period of a limb movement trial, or vice versa). Additionally, we visually inspected trials for both tasks and excluded any trial that had overlap with the baseline of the next trial. RMI > 0 indicates increased firing rate during a behavioral event and RMI < 0 indicates decreased firing rate. RMI values were calculated around two behavioral events (task onset: RMI_onset; task offset: RMI_offset) for each task (speech and limb movement trials).

RMIs and absolute firing rates were separately submitted to a linear mixed effects model with task (speech vs. limb movement), epoch (trial onset vs. trial offset for RMIs), and interaction of task × epoch as fixed effects, and subjects and units as random effects. We used Satterthwaite approximation as the estimation method for degrees of freedom and variances. Note, we did not see significant differences between single and multi-unit firing rates [t (125.95) = 0.69, p = 0.49] and RMIs [t (64.33) = −0.39, p = 0.70] during our tasks; therefore we combined our single and multi-units and performed statistical analysis on all 69 STN units except for correlations with clinical PD measures where single and multi-units were analyzed separately as detailed below.

To delineate the characteristics of neuronal units in the current study, we classified identified units into 4 categories based on their RMIs: units that were modulated by (1) speech onset or offset only (2) limb onset or offset only, (3) both speech and limb movement onset or offset, and (4) neither speech nor limb movement. To label units as modulated or not modulated by tasks, first RMIs were calculated at the single trial level, then the 95% confidence interval of the mean RMI was obtained for each of 69 units. We tested mean RMI of each unit against the null value of 0. When the null value 0 was not within the 95% confidence interval limits of RMI distribution, units were labeled as “modulated.” Next, units classified as modulated were submitted to a logistic mixed effects model with task (speech vs. limb movement), epoch (onset vs. offset), and their interaction as fixed effects. Additionally, we adopted subjects and units as random effects since measures were repeated across subjects and units.

Linear mixed effects modeling was performed to determine if four selected clinical variables interacted with speech and limb movement-related firing rate or RMI differences: Unified Parkinson’s Disease Rating Scale (UPDRS) and disease duration. These pre-operative clinical variables were 1. Δ UPDRS III _total; 2. Δ UPDRS III _speech; 3. Δ UPDRS III _{fingerbradykinesia}; 4. PD disease duration (i.e., time since PD diagnosis). Δ UPDRS III refers to the difference in OFF vs. ON levodopa scores in preoperative UPRDS Part III motor testing; higher UPDRS scores indicate greater impairment. Participants with greater Δ UPDRS III values showed greater overall improvement of PD-related motor impairments when administered levodopa medications during preoperative testing; these delta metrics are measures of levodopa sensitivity. We also performed Pearson correlation analyses between clinical measures of PD and RMIs and absolute firing rates separately for single and multi-units. The rationale for this correlation was based a previous study that showed longer PD duration was associated with increases in the firing rates of single neurons in STN (Remple et al., 2011). We are not aware of any study reporting correlation between multi-units and PD duration. All hypothesis testing was two-sided at α = 0.05 and statistical analyses were conducted in R and Matlab.

Our observation of increased STN firing in single units of PD participants with longer disease duration is consistent with the findings of Remple et al. (2011). They report a median firing rate of 29 Hz in an “early” group (mean = 3.1 years of PD medication use) versus 36 Hz in advanced group (mean = 10.3 years of PD medication use). They report differences in OFF vs. ON UPDRS III scores in their early and advanced groups (15.1 vs. 24.8 respectively). Remple et al. only explored this correlation in single units similar to what we observed, namely a significant correlation between PD duration and absolute firing rates of single, but not multi, units. This finding implies that STN firing rates at the single unit level may be a more reliable indicator of PD severity. Future study is needed to further expand upon the relationships between STN firing rates, disease duration and sensitivity to PD medication.

Our study has several limitations but does provide directions for future studies. First, we have recorded from only a fraction of the STN, i.e., the dorsal portions, given that the dorsal region is accepted to represent the motor subregion and effectuates the greatest clinical improvement during chronic DBS Additional studies are also needed to interpret how our evidence of differentiated speech- and limb movement-related firing in individual neurons aligns with the current understanding of STN functional organization. It is also unclear how greatly the density of speech-modulated neurons at the STN recording site might vary across subjects. Our recording locations within STN were limited to the dorsal half of the nucleus with a combination of locations in the anterior, center, and posterior thirds of the nucleus. The number of clusters we obtained does not yet allow rigorous differentiation of these subregions. Of note, one recent study did not observe differences in speech-related STN neuronal modulation as a function of recording location (Lipski et al., 2018) while another reported improved voice function with STN-DBS application at the dorsal anterior portion of the nucleus (Jorge et al., 2020); this may be consistent with more robust speech motor representation in dorsal STN. Moreover, we and previous studies did not have a control task to confirm the speech-specificity of “speech-related” neuronal modulation we report. Speech production requires varying degrees of movement of orofacial structures, and future work should incorporate non-speech orofacial movements to differentiate between speech-specific modulation and orofacial movement-related modulation. There is evidence from previous studies demonstrating that STN neurons are modulated by non-speech facial/jaw movement (Delong et al., 1984; Abosch et al., 2002; Golshan et al., 2018). For example, Abosch et al. (2002) isolated 9 of 64 neurons within STN that were modulated by orofacial non-speech movement. Hence, we acknowledge that modulation of firing rates in STN neurons during our speech task, like all other reports to date and including both speech onset and speech offset, may not be solely induced by speech. It is reasonable to assume that orofacial movement during speech tasks may have contributed to the modulation of firing rates we report as well as previous studies reporting STN neuronal activity during speech (e.g., Lipski et al., 2018; Tankus and Fried, 2019; Tankus et al., 2021). However, we think that future studies will benefit from our findings to further elucidate the role of STN neurons in speech and limb movement.