Human Cervical Epidural Spinal Electrogram Topographically Maps Distinct Volitional Movements

Abstract

Little is known about the electrophysiologic activity of the intact human spinal cord during volitional movement. We analyzed epidural spinal recordings from a total of five human subjects of both sexes during a variety of upper extremity movements and found that these spinal epidural electrograms contain spectral information distinguishing periods of movement, rest, and sensation. Cervical epidural electrograms also contained spectral changes time-locked with movement. We found that these changes were primarily associated with increased power in the theta (4–8 Hz) band and feature increased theta phase to gamma amplitude coupling, and this increase in theta power can be used to topographically map distinct upper extremity movements onto the cervical spinal cord in accordance with established myotome maps of the upper extremity. Our findings have implications for the development of neurostimulation protocols and devices focused on motor rehabilitation for the upper extremity, and the approach presented here may facilitate spatiotemporal mapping of naturalistic movements.

Significance Statement

The electrophysiology of the human spinal cord remains incompletely characterized. We build on our previous work in describing a novel method of recording spinal epidural electrograms from awake human participants by showing that spinal electrograms recorded from the cervical spinal cord during volitional upper extremity movements demonstrate spectral changes time-locked to movement that feature prominent increase in theta band power and theta phase to gamma amplitude coupling. These spectral changes can also be topographically mapped to the cervical spine in a myotome distribution broadly consistent with maps generated from intraoperative stimulation studies in humans and direct stimulation experiments in monkeys. Our methodology may aid in the developing spatiotemporal maps for neurostimulation protocols to recapitulate naturalistic movements.

Introduction

The spinal cord is traditionally conceptualized as a simple channel for transmitting top–down motor programs and receiving bottom–up sensory information. However, mounting evidence suggest the existence of intrinsic spinal cord circuits capable of producing movements in the absence of supraspinal inputs (Dimitrijevic et al., 1998). For instance, epidural electrical stimulation (EES) of the spinal cord in specific spatiotemporal patterns can induce lower extremity movements and restore walking in patients with incomplete and complete spinal cord injury (SCI; C. A. Angeli et al., 2018; Wagner et al., 2018; Rowald et al., 2022; Lorach et al., 2023) with at least one patient experiencing long-term improvement in walking and lower extremity strength even when stimulation was off (Lorach et al., 2023). This suggests that the spinal cord can undergo enduring plasticity in response to electrical stimulation to enable motor rehabilitation. EES is thought to facilitate volitional movement and promote long-term neurorehabilitation by recruiting motoneuron pools via large diameter sensory afferents (Kathe et al., 2022; Dorrian et al., 2023). Specifically, recent work has identified a population of excitatory interneurons in the lumbar spinal cord that is necessary and sufficient to enable recovery of locomotion following SCI (Kathe et al., 2022).

While the restoration of volitional upper extremity movement following SCI has remained challenging due to the inherent complexity of upper limb movements, several strides have been made in this direction. EES of the cervical spinal cord has been shown to recruit upper extremity motoneurons in addition to improving strength and functional abilities in macaque monkeys with cervical SCI (Greiner et al., 2021; Barra et al., 2022). Improvements in strength, control, and several upper extremity tasks have also been seen in human tetraplegic patients as well those with poststroke upper limb paresis (Lu et al., 2016; Powell et al., 2023). In practice, many EES applications depend on a combination of prior anatomic knowledge of motor neurons pools and their respective innervation, computational modeling, and stimulation-based mapping to deliver functional improvements for patients (C. Angeli et al., 2024). Developing more precise topographic maps of upper extremity movements could enable development of EES applications that recapitulate naturalistic volitional movements.

We have previously described a novel method to study neurophysiological activity of the spinal cord during volitional motor tasks in humans by recording from externalized spinal epidural electrodes in patients undergoing spinal cord stimulation (SCS) trial to treat their chronic pain (Burke et al., 2021). In the present study, we expand on our previous work by exploring whether volitional upper extremity movements can drive spinal electrogram (SEG) activity and whether these movement-related changes can be topographically mapped in the intact human spinal cord. By recording SEGs from five patients undergoing SCS evaluation with epidural electrodes covering Cervical Spine Levels 2–7, we found spectral changes in the theta (4–8 Hz) frequency band that are time-locked with upper extremity movements. Additionally, different regions of the cervical spinal cord are tuned to distinct upper extremity movement-driven theta power changes in accordance with established myotome maps of the upper extremity. This knowledge could have direct implications for EES optimization for neurorehabilitation.

Materials and Methods

Patient selection

Adult patients with chronic pain undergoing a SCS trial from the Pain Management Center and the Department of Neurological Surgery at the University of California San Francisco were recruited for the study. Patients without previously existing neurological injury and the ability to perform motor and sensory tasks according to task instructions are included in the study. Patients who had known motor deficits were excluded from the study. All study protocols were approved by the institutional review board (IRB), and all patients provided informed consent for participation (IRB number 9-29038). A total of five human subjects of both sexes were selected for this study.

Epidural electrode placement and operative procedure

Electrodes were selected and placed as previously described (Burke et al., 2021). In brief, FDA-approved epidural electrodes with either 8 or 16 leads arranged in either cylindrical or paddle lead configurations were selected at the discretion of the subjects’ treating physician with cervical spinal level targets chosen to optimize pain control. For the SCS trial, distal epidural electrodes are placed in the cervical spine and the proximal end either exits the skin directly in the patient's back, or it is connected to another lead extender which then exits at the skin. The trial period typically lasts 7–10 d during which a temporary stimulator is connected to the externalized leads to determine the efficacy of SCS.

Behavioral task

During each recording session, the patient was shown a behavioral task that gave the patient visual commands to move. The behavioral task was written using a custom Python (Python Software Foundation) software using the PyEPL toolbox (Geller et al., 2007) that was implemented on a separate laptop. This laptop was placed in front of the patient and displayed instructions to move the limb. The task began with a 30 s rest period, during which the patient was instructed to sit still in a relaxed position. Following rest, subjects were prompted to perform both active movements allowing full range of motion and isometric contractions of upper extremity muscles (biceps, triceps, wrist extensors, and flexors) where researcher opposed the movement. Subjects performed alternating left elbow flexion and extension, left wrist flexion and extension, right elbow flexion and extension, and right wrist flexion and extension, followed by isometric variants of each of these eight movements. Three sets of 10 repetitions were cued with 30 s pauses for rest between movement types and brief rests between sets. Subjects were given 2 s between cues to complete movements. Subjects then proceeded to the sensation task in which the examiner brushed the anterior forearm using fingertips for 1 s, repeated 10 times on each arm.

SEG recording protocol

Data from epidural electrodes, surface electromyography (EMG), and other kinematic sensors (accelerometer) were obtained using the recording setup and hardware as previously described (Burke et al., 2021). Briefly, proximal ends of the epidural electrodes were connected to a customized connector cable and plugged into an external amplifier (Tucker-Davis Technologies, or Neuro-Omega, Alpha-Omega Engineering). One of the distal contacts in the epidural electrode was used for reference. EMG and kinematic recordings were obtained from a wireless system (Delsys, 4.7.9) using a separate, dedicated task computer. Recordings from subjects in our study were sampled at either 2,750 or 3,052 Hz.

To synchronize multiple data streams, we sent the transistor–transistor logic (TTL) pulses from a LabJack U3-LV (LabJack) during the recording session, and the pulses were recorded by the external amplifier and the Delsys computer. The TTL pulses were created by a custom code implementing the behavioral task, and pulse times were based on the CPU time of the laptop executing the behavioral task. TTL pulses times from the behavioral laptop were then synchronized with the sample times on both the amplifier and the EMG system. A custom code was used to align these TTL pulses off-line to synchronize the data.

Data analysis

Identification of movement events

Movement onset times were determined from manual review and annotation of EMG as well as kinematic data. EMG and kinematic data were standardized using a z-transformation, and for most movements, motion onset was selected as the start time of a nonzero signal burst in the temporal vicinity of a cue for movement presented to the subject. In cases of unclear EMG change from the resting baseline, changes in acceleration from a zero baseline were taken as movement onset times. Data from relevant sensors were considered for each movement, for instance, left triceps EMG and kinematic data were used to assess left arm extension.

Preprocessing

SEG and EMG data were processed using off-line signal processing tools that were custom written using MATLAB (MathWorks, version 2021b). SEG data were bandpass-filtered from 4 to 150 Hz using a fourth-order Butterworth filter and passed through Butterworth notch filters centered around 60, 120, and 180 Hz to reduce 60 Hz electrical noise and its harmonics.

Channel selection

The power spectral density for each contact and its corresponding channel was plotted using Welch's method for (550 or 610.4 Hz sampling rate from downsampling signal, 10 s windows, 8,192 sample FFT length, Hamming windows) the duration of the behavioral task recording for the purpose of selecting channels for bipolar montage rereferencing. Channels which were used as references, had poor signal-to-noise ratio, contained clear frequency domain artifacts, or lacked the characteristic 1/f decay of neural signals were excluded. All 8 channels for Subject 1, 10/16 channels for Subject 2, 7/8 channels for Subject 3, 4/8 channels for Subject 4, and 6/8 channels for Subject 5 were included. Included channels were then rereferenced through subtraction of adjacent contacts to further reduce noise.

Artifact rejection

To remove movement-related and other artifacts in SEG data that persisted after filtering, we used a custom algorithm which used normalized power in the high gamma band (75–150 Hz) to label regions of data as potentially artifactual (Extended Data Fig. 2-1). We utilized the multitaper transform as implemented in the FieldTrip toolbox (“mtmconvol”) to first generate time–frequency spectrograms. For the parameter specifications of this transformation, center frequencies were determined using the “cwtfilterbank” function of MATLAB using the entire length of the signal, the sampling rate of the SEG recording, 16 voices per octave, a time bandwidth of 60, and filter limits of 4–150 Hz. To focus on high gamma band, we retained only center frequencies >75 Hz for artifact estimation. Additional configuration parameters for the multitaper transform included “dpss” for taper type and time window of 0.1 s. The amount of spectral smoothing (parameter “tapsmofrq”) was determined through adjustment of desired frequency resolution (20–40 Hz) about our center frequencies of interest. The output of the transform was then standardized using median absolute deviation. We set a threshold of an absolute value of eight times the median absolute deviation, above which regions were considered potentially artifactual. Any trials containing artifact between −500 and +1,500 ms of the movement onset were removed from further analysis. Additionally, any trials with signal >7 standard deviations above the signal mean in the time domain within this specified timeframe were also removed. The total number of samples and proportion rejected for each movement for each bipolar contact are shown in Extended Data Figure 2-2.

Time–frequency analysis

After filtering, time domain SEG signals were transformed into a time–frequency representation using the wavelet transform. The “cwtfilterbank” function was first used to identify the center frequencies for the transform with filter limits of 2–250 Hz, 20 voices per octave, and Morlet wavelets (“amor” for wavelet type). Time–frequency spectrograms were produced for regions of signal containing all repetitions of a particular movement, and trial-averaged spectrograms were generated by extracting −500 to +1,000 ms of signal around movement onset times before averaging these trials together.

Power spectral density

Welch's method was used to compute estimates of power spectral density. First, individual movements were identified based on movement onset times. The 500 ms of signal before the movement onset was considered the baseline, and the 1,000 ms following onset was considered the postmovement. For each of these individual movement samples, the power spectral density was estimated for the baseline and the postmovement signal using the “pwelch” function with 0.5 s windows (1,375 or 1,526 samples), 50% overlap, and Hamming windows which produced a frequency resolution of 1.34–1.49 Hz. Fold change in postmovement power divided by baseline power was assessed in several canonical bands, which were defined as 2–4 Hz for delta, 4–8 Hz for theta, 8–12.5 Hz for alpha, 13–20.5 Hz for low beta, 21–30 Hz for high beta, 30.5–50 Hz for low gamma, and 50.5–150 Hz for high gamma. Aggregate power spectral density curves for rest, movement, and sensation were produced from computing the mean and 95% confidence interval of all individual movement samples across trials for each bipolar rereferenced contact.

Normalized power analysis

Analysis of percent normalized power was done by first obtaining power spectral densities (PSDs) from Welch's method as described above and dividing by the sum of power in all bins from 4 to 150 Hz. The mean percent normalized power was used to compare trials and compute aggregate statistics.

Tuning curve analysis

The location for each contact was overlayed onto a 45 × 8 grid spanning the C2–C7 epidural space of the cervical vertebrae. For each contact, mean fold change in theta power for the four movements, left elbow flexion/extension and left wrist flexion/extension, was plotted against relative location of the contact on the spinal cord. A smoothing spline function with smoothing parameter of 0.9 was then fitted to this data to determine the locations of peak responses and their magnitudes using the “findpeaks” function.

Movement-related phase analysis

To interrogate the relationship between movement onset and theta phase, one representative contact was chosen for each subject by selecting the contact with the greatest mean fold change in theta power compared with the baseline and most consistent fold changes for the movements of left elbow flexion, left elbow extension, left wrist extension, and left wrist flexion. All trials were bandpass-filtered from 4 to 8 Hz using a third-order Butterworth filter, and phase angle was extracted using a Hilbert transform. Phase angles at the movement onset were plotted onto polar histograms to visualize their distribution. Summary statistics and assessment of significance were computed using the Circular Statistics Toolbox in MATLAB (Berens, 2009).

Phase–amplitude coupling

Phase–amplitude coupling (PAC; De Hemptinne et al., 2013, 2015) between theta phase and gamma amplitude was computed for the representative contacts selected by criteria above during baseline (−500 to 0 ms relative to the movement onset) and movement (0–1,000 ms relative to movement onset) periods. SEGs were bandpass-filtered for theta (4–8 Hz in increments of 1 Hz) and gamma (50–150 Hz in increments of 4 Hz) frequencies. The Hilbert transform was applied, and the instantaneous phase and the instantaneous amplitude were extracted from the low- and high-frequency–filtered signal, respectively. Modulation indices were computed for each phase–amplitude pair as previously described (Tort et al., 2010). To better contextualize our values for modulation index, for each contact and each movement, we generated a distribution of 200 surrogate modulation indices by computing the modulation index between the instantaneous amplitude and phase with random time lags for all concatenated baseline and movement trials, respectively. These surrogate distributions were used to compute Z-scores for baseline and movement modulation index values. Modulation indices were considered significant if their Z-scores remained significant after Bonferroni’s correction for the number of phase–amplitude bins. Finally, all significant modulation indices were summed and normalized by dividing by the number of bins to compare change in modulation index for baseline versus movement.

Statistical analysis

Assessment of statistical significance for fold change in power from the baseline to postmovement was performed using two-sided, Wilcoxon rank-sum tests. Each movement was assessed in each bipolar rereferenced contact independently such that band power determined from Welch's method in the baseline and postmovement signals from the same n samples were compared. Overall power and power in the delta, theta, alpha, low beta, high beta, low gamma, and high gamma bands were evaluated, and FDR adjustment using the Benjamini and Hochberg procedure was applied to adjust for these multiple comparisons. Wilcoxon rank-sum tests and FDR adjustment were also used in the assessment of statistical significance for normalized power analysis. The Rao test was used to test the assumption that phase data were uniformly distributed. Wilcoxon signed-rank tests were used to compare change in modulation index for PAC analysis. All p values reported have been adjusted with FDR correction for multiple comparisons.

Results

We recorded SEGs from cervical epidural electrodes in five patients during repeated trials of volitional movement and sensory stimuli. After performing bipolar rereferencing of signals between adjacent contacts, we obtained high-quality recordings from 32 contact pairs spanning the dorsal column from C2 to C7. Individual epidural electrode placement and locations of recorded bipolar contacts are shown in Figure 1. Subjects performed both active upper extremity movements which included left and right elbow flexion/extension and wrist flexion/extension, as well as isometric muscle contractions of biceps, triceps, and wrist extensors/flexors. Sensation trials involved brushing the anterior aspect of the forearm on both left and right. As reported previously, the power spectral densities of these SEGs displayed a 1/f characteristic, a common feature of neural signals (Extended Data Fig. 4-1A).

Figure 1.

A, Epidural electrode leads from all five individual subjects overlaid on a common cervical spinal cord model with relative positions determined from imaging. B, Map of all contacts considered for analysis after bipolar referencing of channels from electrode leads.

SEGs demonstrate movement-related spectral changes

While we have previously shown that SEGs contain movement-related information (Burke et al., 2021), the temporal relationship of this signal with volitional movement remains unknown. To further explore this relationship, we used surface EMG and kinematic measurements during volitional movement trials aligned with SEGs. To minimize the impact of artifacts from movement and other sources such as cardiac and breathing from our SEG analysis, we bandpass-filtered our signal from 4 to 150 Hz and implemented a custom artifact rejection algorithm as described in methods (Extended Data Fig. 2-1). While epidural sliding of electrodes during active movement tasks could produce artifactual signals, such artifacts would be diminished during isometric movement tasks. Furthermore, the waveform and spectral similarities between active and isometric movements shown in Figure 2 do not support motion of leads as a significant source of artifactual signal.

Figure 2.

A, PSDs before movement, after movement, and during sensory stimulus aggregated for all subjects and normalized to total power in the 4–150 Hz range. Distribution of normalized power before and after the movement onset (B) shows greatest increases in theta and alpha bands. Similarly, percentage of normalized power distribution following sensory stimulus (C) shows greatest increase in theta power. A visualization of the artifact rejection pipeline required to produce these PSDs is shown in Extended Data Figure 2-1, and the number of trials used for downstream analysis is shown in Extended Data Figure 2-2.

Figure 2-1

Visualization of artifact rejection pipeline applied to section of data. The four regions bounded by red lines indicate potentially artifactual signals because of their high amplitude and broadband signal characteristics that persist after bandpass filtering from 4-150  Hz. Our artifact rejection pipeline identifies signal regions with absolute value of Z-score of filtered signal greater than 7 and detects broadband signals using the magnitude of the median absolute deviation of the high gamma (75-150  Hz) band determined using a multitaper transform. Here a wavelet transform is shown over 2-150  Hz for additional context. Download Figure 2-1, TIF file ^{(11.4MB, tif)} .

Figure 2-2

(A) Total number of trials used for downstream analysis after artifact rejection for each movement and bipolar referenced contact pair. (B) Proportion of trials removed from further analysis by artifact rejection pipeline. Download Figure 2-2, TIF file ^{(4.2MB, tif)} .

Our analysis revealed that SEGs not only capture low-frequency oscillations time-locked with movement but also reliably record similar waveforms across subjects for a diversity of volitional movements (Fig. 2). During all upper extremity movements, SEGs showed increased amplitudes in lower-frequency (4–12 Hz) activity compared with those in the premovement baseline (Fig. 2). While these oscillations are more prominent in the active flexion and extension time domain SEG signal, inspection of trial-averaged time–frequency spectrograms reveals that the magnitude of low-frequency waveforms is consistent across movement types and subjects (Fig. 3). Furthermore, trial-averaged spectrograms and plots of all trials reveal that the majority of the movement-related SEG changes occur within 500 ms of the movement onset for multiple movement types and subjects.

Figure 3.

Aligned EMG, filtered SEG, and wavelet transformed SEG data for active and isometric movements of wrist and elbow for various subjects. Movement onset times are indicated by green lines. For all depicted movements, low-frequency spectral changes (4–12 Hz) time-locked with the movement onset are seen in SEG data. Mean fold change in theta band power is computed for each movement type and contact and shown in Extended Data Figure 3-1. Distributions of theta phase angle at the time of the movement onset are displayed in Extended Data Figure 3-2.

Figure 3-1

(A) Mean fold change in theta (4-8  Hz) band power across all trials for each pair of movements and bipolar referenced contacts. (B) -Log10 of p-values obtained using Wilcoxon rank rum test comparing within trial theta band power before and after movement onset. Values represent FDR adjusted p-values. Download Figure 3-1, TIF file ^{(6.8MB, tif)} .

Figure 3-2

Polar histograms showing the distribution of phase angles in the theta band (4-8  Hz) at movement onset for pairs of representative contacts and movements. Mean phase angles for each of the distributions are listed. Download Figure 3-2, TIF file ^{(8.1MB, tif)} .

Movement-related spectral changes show greatest change in the theta band

We next sought to investigate the spectral characteristics of signals recorded during volitional movement. Trial-averaged spectrograms were computed from all trials which passed our artifact screening pipeline (Extended Data Fig. 2-2). These spectrograms demonstrated the greatest increase in signal amplitude at low frequencies, particularly in the theta band (4–8 Hz; Fig. 3). Increase in theta–alpha frequency amplitude was found across all subjects, for all movement types (elbow and wrist flexion/extension), and for active and isometric movements.

These movement-related amplitude changes were also reflected in power spectral densities of subjects during postmovement onset, premovement baselines, and sensing (Fig. 4; Extended Data Fig. 4-1A–C). While postmovement SEGs featured statistically significant changes in mean normalized power in all frequency bands except low beta compared with premovement baseline, we found the greatest increases in theta and alpha frequency bands (movement vs premovement baseline, percentage of normalized power, theta 4.61 vs 2.92, p = 1.81 × 10⁻¹⁴; alpha 2.51 vs 2.06, p = 6.34 × 10⁻⁸; low beta 1.44 vs 1.40, p = 0.20; high beta 1.06 vs 1.13, p = 0.0029; low gamma 0.94 vs 1.02, p = 1.07 × 10⁻⁶; high gamma 0.64 vs 0.73, p = 2.03 × 10⁻⁹; post hoc FDR-adjusted Wilcoxon rank sum).

Figure 4.

Box plots showing increase in theta (4–8 Hz) band power after normalization to total power in the 4–150 Hz range for depicted movements. Time–frequency spectrograms averaged over all trials show low-frequency signal time-locked to movement. Aligned SEG tracings of all trials over 4–150 Hz and separately filtered to include only the theta band redemonstrate low-frequency oscillations occurring with the movement onset. Change in normalized power for all bands for all subjects individually is shown in Extended Data Figure 4-1 and included are PSDs for after movement, before movement, and during sensory stimulus. Examples of trial-averaged sensory stimulus trials are shown in Extended Data Figure 4-2.

Figure 4-1

(A) Power spectral densities (PSDs) for each subject across pre-movement, post-movement, and response to sensory stimulus generated using Welch’s method. (B) PSDs showing power normalized to total power in the 4-150  Hz range. (C) Bar graphs showing distribution of normalized power in each canonical band before and after movement onset or in response to sensory stimulus. Download Figure 4-1, TIF file ^{(6MB, tif)} .

Figure 4-2

Trial averaged spectrograms for sensation trials (left) and movement trials (right) for representative contacts. First row shows contact 5A during left arm sensing and left elbow extension. Second row shows contact 4D during right arm sensing and right wrist extension. Third row shows contact 3F during right arm sensing and right elbow extension. Download Figure 4-2, TIF file ^{(2MB, tif)} .

A similar trend was seen in SEG recordings during sensation compared with those during rest with theta band showing statistically significant increases in normalized power (sensing vs rest, percentage of normalized theta power, 4.03 vs 3.39, p = 0.0033), whereas other frequency bands trended toward decreases in normalized power (sensing vs rest percentage of normalized power, alpha 2.16 vs 2.18, p = 0.68; low beta 1.25 vs 1.53, p = 6.13 × 10⁻⁸; high beta 1.04 vs 1.1, p = 0.084 low gamma 0.94 vs 0.96, p = 0.68; high gamma 0.71 vs 0.72, p = 0.52; post hoc FDR-adjusted Wilcoxon rank sum) of which only the chance in low beta was significant. A comparison of sensation trials and movements in representative contacts is shown in Extended Data Figure 4-2, highlighting the greater magnitude of theta and alpha changes during movement compared with sensation.

We also computed fold change in power from the 500 ms before the movement onset to the 1 s following the movement onset. Across all movements, theta again demonstrated the greatest mean fold increase in power [5.95 (95% CI, 5.57–6.33)] with moderate changes in the other bands [alpha 3.6 (95% CI, 3.40–3.78), low beta 2.1 (95% CI, 2.03–2.15), high beta 1.7 (95% CI, 1.66–1.74), low gamma 1.32 (95% CI, 1.3–1.34), high gamma 1.08 (95% CI, 1.07–1.09)]. Of the 336 pairs of movements and bipolar referenced contacts in our dataset, 169/336 produced statistically significant increases in theta power (Extended Data Fig. 3-1). Similarly, sensation-related trials also produced the greatest fold change in power in the theta bands when compared with a resting baseline. Mean fold increase in theta power for sensation trials was 2.79 (95% CI, 2.34–3.24), alpha 2.36 (95% CI, 1.9–2.8), low beta 1.7 (95% CI, 1.51–1.89), high beta 1.86 (95% CI, 1.73–1.98), low gamma 1.97 (95% CI, 1.84–2.09), and high gamma 1.89 (95% CI, 1.79–1.99). The 32/64 sensation–contact pairs showed statistically significant increases in power.

We next explored whether movement-related changes produced differences in PAC between theta phase and gamma amplitude during movement compared with the baseline. Representative movements and contacts were selected, and modulation indices were computed for each phase–amplitude bin. Surrogate distributions of comodulation were generated to determine thresholds for significant modulation index values (Extended Data Fig. 5-1). Our analysis showed that the normalized sum of significant modulation index values was greater during movement for all contacts and movements analyzed compared with the baseline (Fig. 5). For example, the mean value of the normalized modulation indices during left elbow flexion for contact 2F was 0.0060 (95% CI, 0.0058–0.0061) which was greater compared with the value for the baseline [0.0018 (95% CI, 0.0016–0.0020), p = 5.12 × 10⁻⁶, post hoc FDR-adjusted Wilcoxon signed-rank].

Figure 5.

PAC between theta (phase, 4–8 Hz) and gamma (50–150 Hz, amplitude) during premovement baseline and movement. Comodulograms shown are grand averaged across all trials for representative contacts for each subject for each movement, and color intensity represents modulation index. Whisker plots compare the sum of modulation index values in significant bins during baseline versus movement for all trials, normalized by number of bins. An example of Z-scoring and significance masking for comodulograms is shown in Extended Data Figure 5-1.

Figure 5-1

Sample phase amplitude coupling comodulograms from contact 4B left elbow flexion trial 8. Values represented are z-scored modulation indices computed from a surrogate distribution. Significance masks show remaining z-scored modulation indices after thresholding for significance. Download Figure 5-1, TIF file ^{(2.1MB, tif)} .

We also investigated the relationship between theta phase angle at the movement onset and movement type using the same set of representative movements and contacts; however, we did not observe consistent phase locking of movement onset (Extended Data Fig. 3-2). In addition, we assessed theta phase coherence using phase locking values but did not observe changes in coherence or connectivity for the representative contact from Subject 5 during left elbow extension.

SEGs topographically map distinct volitional movements by the spinal level

Given the broad distribution of the bipolar contacts throughout the cervical spinal cord in our analysis, we examined whether SEG recordings could encode the topographic specificity of upper extremity movements. Despite the overlapping innervation of muscle groups in the upper extremities, it has been established that some movements derive their majority contribution from anterior horn cells of distinct spinal levels based on anatomical and stimulation studies. We found that SEG recordings captured similar anatomical relationships as contacts near C3–C5 vertebrae showed greater fold increase in theta power for elbow flexion, while contacts at C7/T1 vertebrae showed the greatest fold increase in power for elbow extension (Fig. 6). A similar pattern was observed for wrist movements as more proximal contacts at C4/C5 showed higher theta power increase during wrist extension, while distal contacts at C5–C7 showed higher theta power increase during wrist flexion. To further characterize whether SEG activities are spatially tuned to different upper extremity movements, we performed a smoothing spline fit to identify electrode contact locations that showed maximal theta power increase over the baseline for each of the movements. We found that C3 and C5–C6 SEGs are tuned to elbow flexion, C6–C7 to elbow extension, C4–C6 and C7 to wrist flexion, and C5–C6 and C7 to wrist extension (Fig. 7). Of note, wrist flexion had a greater peak fold change in theta at C7 compared with wrist extension.

Figure 6.

Fold change in theta (4–8 Hz) band power during movement compared with the premovement baseline for left elbow extension, left elbow flexion, left wrist isometric extension, and left wrist isometric flexion for all subjects. Individual bipolar referenced contacts are displayed in the heatmap grids overlying a cervical spinal cord model. Black squares represent contacts without any movement trials.

Figure 7.

Tuning curve analysis of fold change in theta power following the movement onset by vertebral level of recording contact. Curves for individual movements are smoothing spline fits over mean theta fold change over all active and isometric trials for a given movement and recording contact.

Discussion

In this study, we demonstrate the ability of SEGs recorded from the epidural surface of the spinal cord in awake human subjects to encode task-specific information with frequency-specific oscillations. Cervical epidural electrodes could detect oscillatory changes time-locked with the movement onset and distinguish periods of movement from rest and sensory stimuli. The resulting spectral changes were seen across volitional upper extremity movement tasks in all five subjects and were characterized by increased theta power as well as theta–gamma PAC during movement. This movement-related theta signal could also be mapped topographically to the cervical spinal cord illustrating tuning of contacts at the C3–C5 level for elbow flexion, C7–C8 for elbow extension, and proximal versus distal differentiation among wrist extension and flexion with C4–C5 being tuned for extension and C5–C7 tuned for flexion, consistent with current myotome maps of the upper extremity compiled from intraoperative stimulation and lesion studies (Schirmer et al., 2011; McIntosh et al., 2023b; Sonoo, 2023).

In addition to our prior work, there have been many other reports of electrophysiologic signals recorded from the spinal cord, typically in the setting of SCS for chronic pain (Russman et al., 2022). Motor mapping of the spinal cord has also been attempted through intraoperative stimulation experiments (Schirmer et al., 2011; McIntosh et al., 2023b) and computational modeling (Greiner et al., 2021). Recent work in intraoperative EES of the cervical spinal cord has shown that elbow flexion can be elicited by stimulation at the C5/C6 spinal nerve level, while elbow extension is evoked more inferiorly around C7/C8 nerves (McIntosh et al., 2023b; Sonoo, 2023). In addition, while Levels C5–T1 provide innervation for wrist flexors and extensors, wrist extensors have major contributions from C6 to C7 nerve roots, whereas wrist flexors have main contributions from C7 to T1 spinal nerves (Schirmer et al., 2011; Sonoo, 2023). While stimulation mapping of the spinal cord can activate motor neuron pools at the level to produce muscle responses, they are not representative of naturalistic movement and may have less spatial specificity given the current spread to adjacent levels. Recording SEG during volitional movements can potentially provide a topographic map of naturalistic movements which involves integration of motor and sensory information. We find that the topographic cervical spinal cord map produced in our study (Figs. 6, 7) is generally consistent with stimulation maps, with greatest theta power increase around C3 and C5–C6 for elbow flexion and C6–C7 for elbow extension. Furthermore, we see wrist flexion distributed over C4–C6 and C7 and wrist extension over C5–C7. In addition to these movement-specific foci, our recordings also support activation of overlapping circuitry or potential current spread given the partially overlapping distributions of our theta power fold change maps. Overlapping activation could also occur as a function of opponent process like activation of antagonistic muscles to ensure smooth movement.

The movement-related spectral changes reported in this analysis showed the greatest fold change in power primarily in the theta and, to a lesser extent, alpha frequency oscillations compared with pre-movement baselines. We also observed spectral changes in response to sensory stimulation in the theta band, although of lesser magnitude than movement-related theta changes. While recorded signals may represent the integration of motor, sensory, and proprioceptive potentials within the spinal cord, the spectral changes shown in our analysis could be driven by several different processes. These changes in the theta and alpha frequency bands could represent the transmission of motor cortex commands. In the human neocortex, theta and alpha oscillations are traveling waves that can coordinate activity across multiple regions and can propagate in task-specific fashion (Zhang et al., 2018). Cortical theta activity has been implicated in error-monitoring, motor reprogramming, and sensorimotor integration during motor tasks in human subjects (van Driel et al., 2012; Hori et al., 2013; Pellegrino et al., 2018; Carlsen et al., 2023). During a finger isometric force task, an early increase in theta band coupling between EEG electrodes overlying motor sensory areas and force modulation shown by EMG was seen (Novembre et al., 2019), suggesting an effect of theta corticomuscular coherence on tonic corticospinal drive. Elevated theta activities within the basal ganglia–thalamocortical network have been found in patients with dystonia (Hendrix and Vitek, 2012; Neumann et al., 2015, 2017), implicating its role in producing pathological muscle contractions with excessive synchronization. Therefore, the movement-driven increases in theta band SEG that we describe may reflect top–down corticospinal drive to generate muscle contractions.

Another possibility is that these oscillatory changes reflect neural activities from the local spinal circuits. The increase in theta oscillations in the spinal cord after the movement onset could represent recruitment of motor neuron pools with volitional motor control. Low-frequency oscillations starting from 2–3 to 40–70 Hz have been noted in a subprimary firing range of mouse spinal motoneurons and are thought to regulate the force output of motor units (Iglesias et al., 2011). In rat spinal cord organoid cultures, spontaneous low-frequency bursts from 3.5 to 14.3 Hz have been observed (Czarnecki et al., 2008). Inference of spinal motoneuron inputs from EMG recordings in human neonates showed that delta band oscillations were associated with force generation (Del Vecchio et al., 2020), and similar analysis underscored the role of low-frequency (0–35 Hz) inputs to motor neuron pools for modulating force control in digit tasks (Del Vecchio et al., 2019). Further studies would be needed to elucidate the source (cortical vs local spinal) of these observed oscillations.

Another interesting finding revealed by our study are the increases in theta phase to gamma amplitude coupling in SEGs during movement. There is ample evidence for the role of theta–gamma PAC volitional motor control in humans. In the human motor cortical areas, theta coupling to gamma frequencies has been implicated in governing adaptive motor control (Spooner and Wilson, 2022) and is associated with motor recovery after stroke (Rustamov et al., 2022). Local field potentials recorded from the subthalamic nucleus of patients with Parkinson's disease showed increased theta–gamma PAC at the onset of voluntary muscle contraction of the contralateral arm (Kato et al., 2016). Together, these findings suggest that coupling between fast gamma frequency amplitude during specific phases of theta oscillation across the motor system may be a general mechanism for regulation of motor behavior.

Our findings have direct implications for the development of novel neurostimulation approaches directed at restoring upper extremity movements. Mapping the spatiotemporal encoding of movements in the spinal cord may enable stimulation protocols to better recapitulate naturalistic movements. A recent report suggests that a combined cortical and spinal stimulation approach produces a synergistic effect in generating motor-evoked potentials, highlighting the importance of spinal cord electrophysiologic characterization in the development of therapeutic neuromodulation modalities (McIntosh et al., 2023a). In addition, epidural spinal recordings may aid detection of residual descending motor pathways in incomplete SCI as one cohort of individuals with motor complete SCI still demonstrated electrophysiologic evidence of voluntary control over clinically paralyzed muscles shown through EMG (Sharma et al., 2023).

The results of our study should be interpreted with the consideration of several limitations. First, our sample size was limited to five subjects with a combined 32 contact pairs distributed over the cervical spinal cord. This number of contacts also limits the spatial resolution of our topographic analysis. Next, our behavioral task of volitional upper extremity movements did not encompass the full range of upper extremity movement and did not feature movements of individual digits. In addition, although our data were passed through a rigorous artifact rejection pipeline and filtered to exclude broadband and low-frequency signals, we cannot completely rule out the presence of all sources of noise and artifactual signals. Finally, the intrinsic ability of CSF in the spinal cord to act as a volume conductor may limit speculation of the anterior or posterior origin of recorded signals.

Conclusion

Spinal epidural electrograms recorded from the cervical spinal cord in human subjects performing volitional movements show spectral changes time-locked to movement that feature prominent increase in theta band power and increased theta–gamma PAC. These spectral changes can also be topographically mapped to the cervical spine in a myotome distribution broadly consistent with maps generated from intraoperative stimulation studies in humans and direct stimulation experiments in monkeys. Our methodology may aid in the developing spatiotemporal maps for neurostimulation protocols to recapitulate naturalistic movements.

Data Availability

The MATLAB scripts used for the generation of results described in this study are available upon reasonable request made to the corresponding author.