CERVICAL EPIDURAL ELECTRICAL STIMULATION RESTORES VOLUNTARY ARM 1 CONTROL IN PARALYZED MONKEYS


 Regaining arm motor control is critical for people with paralysis. Despite promising results on grasping, no technology could restore effective arm control. Here, we show that electrical stimulation of the cervical spinal cord enabled three monkeys with cervical spinal injury to execute functional arm movements. We designed an epidural interface that engaged surviving spinal circuits via the recruitment of large sensory afferents to produce movement. Simple stimulation bursts produced sustained joint movements which, triggered by movement-related intracortical signals, enabled monkeys with arm paralysis to perform an unconstrained, three-dimensional reach and grasp task. This restoration of voluntary motor control was enabled by the synergistic integration of spared descending commands and electrical stimulation within the spinal cord. The simplicity of this technology promises realistic clinical translation.


Figure 1. Experimental framework. (A)
Monkeys were trained to reach for, grasp, and pull a target object placed at the end effector of a robotic arm. We measured 3D forces applied to the robot joints, full-limb kinematics, electromyographic (EMG) activity from eight muscles of the arm and hand, and intracortical signals from primary sensorimotor areas. (B) Conceptual scheme of the experimental protocol: (1) A decoder running on a control computer identified movement attempts and (2) delivered electrical spinal cord stimulation to the appropriate spinal roots. (3) Stimulation produced arm and hand movement that we recorded and analyzed off-line. (C) Stick diagram decomposition of arm movement during a reach, grasp and pull movement in intact monkeys (S = shoulder, E = elbow, W = wrist). We considered a movement complete when a target spatial threshold was crossed during pull. Copyright Jemère Ruby. segment 28,29 . Therefore, we designed a spinal interface that could target each of the roots independently by placing contacts on the lateral aspect of the cord to target the entry zone of 110 each individual root 28 . Since each monkey possessed a unique anatomy, we tailored the design 111 of our interface to each specific subject. For this, we measured white matter diameter and  Examples of muscle selectivity (polar plot) and muscle recruitment obtained by stimulating (1 Hz) at C5, C6/C7, and T1 spinal segments (Mk-Yg). Below, average muscle activations elicited from C7 and T1 contacts in n=3 monkeys (Grey bullets: for each animal, average recruitment across all stimulation currents. Big bullets: mean of average recruitments across animals). (B) Muscle recruitment obtained during delivery of pulse trains in anesthetized monkeys. Recruitment was estimated by computing the energy of EMG signals for each muscle and each stimulation contact. Stimulation frequencies ranged from 20 to 120 Hz (n = 2). For each muscle, energy values were normalized to the maximum value obtained across all frequencies and contacts. (C) Single joint angles excursions induced by stimulation at C7 (blue) and T1 (yellow) roots. Stimulation frequencies ranged from 20 to 100Hz (n = 2). Black bullets: mean. Line: interpolation of the mean values.
caudal contacts elicited spinal reflexes mostly in the hand and forearm muscles, while rostral 138 contacts recruited biceps and deltoids.

140
To ensure that this segmental selectivity translated into functional arm and hand movements, we targeting the C7 root (innervating triceps) produced clear elbow extension; instead, caudal 145 contacts (C8/T1) elicited grasping and wrist movements ( Figure 3C, Extended Data Figure 4).

146
All single joint angles excursions were gradually modulated by varying the stimulation frequency 147 ( Figure 3C). In most of the upper arm muscles we found a monotonic relationship between 148 muscle activation and stimulation frequency. However, for some muscles (e.g. abductor pollicis), 149 responses were lower at higher frequencies (Extended Data Figure 3B). We identified the 150 optimal stimulation range to be around 50-60 Hz (Figure 4). Movements elicited at frequencies 151 lower than 40 Hz were often too weak to complete a joint movement; bursts at frequencies 152 between 50 and 60 Hz produced smooth 31 and full-range movements and maximal forces, while 153 frequencies higher than 60 Hz produced either abrupt movements or incomplete movements 154 ( Figure 4A) due to attenuation of muscle responses during stimulation of sensory afferents 28,32,33 155 (Extended Data Figure 4B). We identified three stimulation contacts that could consistently elicit 156 arm extension (reach), hand flexion (grasp) and arm flexion (pull) (Figure 4B). We then verified 157 that this selection of few contacts could be used to sustain reaching, grasping and pulling 158 movements. By sequentially executing bursts on these three contacts, we could trigger whole arm 159 movements that mimicked smooth 31 and natural multi-joints movements ( Figure 4C, Video 1).

160
Extension, grasping and pulling movements produced clear EMG bursts as well as robust and 161 smooth kinematics. These data demonstrate that with only three contacts, stimulation bursts can 162 engage functionally relevant muscles that produce whole arm movements and sustained muscle 163 activation and forces. Therefore, we planned to link the delivery of these bursts to movement 164 onsets that we derived from intra-cortical signals. Indeed, since our lesions were not complete, 165 movement onsets could be reliably detected even after SCI from intra-cortical signals ( Figure   166 4D). Similarly to other spinal cord stimulation studies we could not identify contacts that selectively 167 produced finger extension 34 . This is likely caused by the overlap of extensor motor-pools in the 168 forearm 27 (Figure 2A), but possibly also because stronger flexors may dominate kinematics in 169 the case of co-contraction at rest.

173
We next tested whether our stimulation protocol could improve functional outcomes of upper limb 174 movements. Specifically, we tested the efficacy of EES to improve muscle activation, pulling forces, functional task performance, and kinematic quality of three-dimensional movements after 176 SCI. In all monkeys, the unilateral lesion led to motor deficits of the left arm and hand. Each  Figure 2D). Generally, animals showed severe paralysis immediately after lesion, and then monkeys and triggering stimulation manually to encourage the animal to perform the task. After the first week, all monkeys spontaneously attempted to perform the task, making it possible to of stimulation promoting reach or grasp/pull respectively. Outcomes were computed for each 189 animal independently and compared between EES on, and EES off conditions. EES significantly 190 enhanced muscles activity and forces ( Figure 5B,D) compared to no stimulation. In terms of 191 functional task performances, without stimulation, the monkeys were rarely capable of completing 192 any part of the task (defined as reach, grasp and pull). Instead, with the support of EES, the rate 193 of successes was significantly and robustly improved ( Figure 5C, Video 2,3,4). EES did not only 194 improve task performance and strength but also overall quality of movement ( Figure 5D). Indeed,  (2) smooth trajectory, extended movement and medium force (40 and 50Hz), (3) abrupt and very extended movement and low force (80 and 100Hz). The range 40-50 Hz was selected as the best optimization of sufficient movement, smoothness and force production. (B) Schematic representation of arm and hand kinematics during stimulation delivered from the selection of three contacts to produce elbow extension (blue), hand and wrist flexion (yellow and black), and elbow flexion (yellow). (C) Example of comparison between EMG activity during intact movement (left) and movement elicited by chaining stimulation from the three selected contacts (right). (D) Scheme illustrating how stimulation is triggered from movement-related intra-cortical signals. On the right, online performances of movement attempt decoder in two animals with SCI. Pie charts represent percentage of predicted (blue) and unpredicted (black) reach events by our decoder.

Figure 5. EES improves task performance, muscle strength and movement quality. (A) Snapshots of
Mk-Yg performing the task before SCI, after SCI without EES, and after SCI with EES. A full successful trial is composed of a reach, a grasp, and a pull. After SCI, Mk-Yg could only perform reaching movements without EES, while when EES was delivered the full task could be performed. (B) Violin plots of signal energy of triceps and FDS EMG profiles during reach (Mk-Br and Mk-Sa) and pull (Mk-Br and Mk-Yg). All individual data points are represented by bullets. Black lines correspond to means of the distribution. Statistical analysis with Wilcoxon Ranksum test. On the right, example raw EMG data after SCI with and without EES. (C) Bar plots report the rate of successful movements after SCI, without and with stimulation. Data are presented as mean ± STD and normalized on the mean value in stimulation condition. Statistics was performed with Bootstrap. (D) Example PC analysis of kinematic features (See methods). Top-left, first and second PC space. Bottom left, stick diagram representation of arm kinematics during pull in intact conditions, after SCI without and with EES. At the immediate right (both bottom and top), euclidean distance in the feature space of trials without stimulation (black) and with stimulation (blue) from the centroid of the trials in intact condition. At the extreme right, example violin plots of movement quality features in the three conditions: intact, after SCI, and after SCI with stimulation. Statistics with Wilcoxon Ranksum test. Asterisks: *p<0.05, **p<0.01, ***p<0.001.
performed wider movements, and generated stronger forces ( Figure 5D), getting closer to normal kinematic trajectory patterns without any long-term training.

202
Sensory feedback and cortical inputs shape EES efficacy.

203
We then investigated the role of spinal circuits and residual cortical inputs in the regaining of 204 voluntary movements that we observed. Indeed, since activation of motoneurons was pre-205 synaptic, both spinal reflexes and residual cortical inputs could shape motor output during 206 EES 19,35 . First, we assessed the influence of sensory inputs on EES-generated motor output.

207
Under propofol anesthesia, we delivered bursts of EES targeting the elbow flexion in isometric 208 conditions ( Figure 6A). We found that induced EMG activity was highly correlated with measured For an example M1 channel, the stimulation that evoked movement (blue, right) corresponded to more spiking activity than the same stimulation evoking no movement (yellow, left). (Bottom) Distribution of average firing rates across all M1 channels during stimulation trains that evoked no movement (yellow) and movement (blue). (C) (Top) State space view of M1 activity for all time points during rest (gray) and preceding attempted movement (orange). The brain states during successful stimulation (blue) were similar to those preceding attempted movements, while the unsuccessful stimulation (yellow) overlapped with the rest states. (Bottom) We computed a relative Mahalanobis distance between the two stimulation conditions and the cluster of neural states at rest. For both monkeys, neural states during stimulation periods with no movement were close to rest. did not generated any muscle activity in relation to M1 activity at rest or during movements without 224 stimulation. We found that motor cortex was significantly more active when EES produced 225 movement ( Figure 6B) than when it did not. We then applied PC analysis to reduce the M1

235
We showed that EES of cervical spinal cord immediately improved muscle activation and strength, 236 task performances and movement quality during a natural-like reach and grasp task in monkeys 237 with unilateral cervical SCI. Moreover, these results were obtained with simple stimulation 238 protocols engaging up to three contacts (one for reach, one for grasp and one for pull) that 239 enabled multi-joint movements. We believe that the design of our interface was key to achieve 240 this result. The dorsal roots are a robust anatomical target that we could easily identify through 241 standard imaging to personalize surgical planning and interface design. Our simple protocol only 242 required the detection of movement onset signals to trigger pre-determined stimulation bursts.

243
Therefore, stimulation control could be simplified and brain recordings may not be required in 244 clinical applications that might exploit more practical residual movements in patients with 245 incomplete paralysis 14,36 .

247
By engaging spinal circuits, EES generated smooth and functional muscle activations that 248 enabled the production of forces sustaining the weight of the arm. Moreover, EES was sensitive 249 to the action of residual descending cortical inputs allowing the cortex to shape voluntary muscle 250 activation and inhibition to produce a desired kinematic output 37,38 . Indeed, the analysis of brain 251 data during voluntary execution of moments with EES suggested that the cortex must be in a 252 movement-permissive state to enable movement with EES. Indeed, in order to produce a 253 functionally relevant motor output, stimulation bursts had to be coherent to motor intention. These 254 features might be regarded as limitations: activating muscles with segmental specificity implies 255 the impossibility to achieve single-muscle recruitment, and the sensitivity and dependence on 256 residual cortical inputs implies a potential failure of EES in motor complete injuries. However, 257 previous studies showed that even completely paralyzed subjects retain residual but functionally silent descending inputs 12,14,19 . Therefore, residual cortical activity may help shaping EES efficacy even in severe patients. In summary, we believe that by exploiting the functionality of residual

298
Data and materials availability 299 All software and data will be available upon reasonable request to the corresponding author.

484
We implemented a custom C++ software application running a control suite that used the

501
In order to obtain a comprehensive measure of muscle recruitment for each contact that would signals, for each implanted muscle. Energy of EMG signals during stimulation were computed on 512 each segment in which stimulation was delivered after the animal started a movement attempt.

513
Energy of EMG signals without stimulation were computed on each segment in which stimulation 514 was not delivered and the animal started a movement attempt. A movement attempt was defined 515 as an increased EMG activity of the Biceps and Deltoid muscles.