Epidural stimulation of the cervical spinal cord for post-stroke upper-limb paresis

Cerebral strokes can disrupt descending commands from motor cortical areas to the spinal cord, which can result in permanent motor deficits of the arm and hand. However, below the lesion, the spinal circuits that control movement remain intact and could be targeted by neurotechnologies to restore movement. Here we report results from two participants in a first-in-human study using electrical stimulation of cervical spinal circuits to facilitate arm and hand motor control in chronic post-stroke hemiparesis (NCT04512690). Participants were implanted for 29 d with two linear leads in the dorsolateral epidural space targeting spinal roots C3 to T1 to increase excitation of arm and hand motoneurons. We found that continuous stimulation through selected contacts improved strength (for example, grip force +40% SCS01; +108% SCS02), kinematics (for example, +30% to +40% speed) and functional movements, thereby enabling participants to perform movements that they could not perform without spinal cord stimulation. Both participants retained some of these improvements even without stimulation and no serious adverse events were reported. While we cannot conclusively evaluate safety and efficacy from two participants, our data provide promising, albeit preliminary, evidence that spinal cord stimulation could be an assistive as well as a restorative approach for upper-limb recovery after stroke. Electrical stimulation of cervical spinal circuits facilitates arm and hand movements in two participants with moderate and severe chronic post-stroke hemiparesis.

Cerebral strokes can disrupt descending commands from motor cortical areas to the spinal cord, which can result in permanent motor deficits of the arm and hand. However, below the lesion, the spinal circuits that control movement remain intact and could be targeted by neurotechnologies to restore movement. Here we report results from two participants in a first-in-human study using electrical stimulation of cervical spinal circuits to facilitate arm and hand motor control in chronic post-stroke hemiparesis (NCT04512690). Participants were implanted for 29 d with two linear leads in the dorsolateral epidural space targeting spinal roots C3 to T1 to increase excitation of arm and hand motoneurons. We found that continuous stimulation through selected contacts improved strength (for example, grip force +40% SCS01; +108% SCS02), kinematics (for example, +30% to +40% speed) and functional movements, thereby enabling participants to perform movements that they could not perform without spinal cord stimulation. Both participants retained some of these improvements even without stimulation and no serious adverse events were reported. While we cannot conclusively evaluate safety and efficacy from two participants, our data provide promising, albeit preliminary, evidence that spinal cord stimulation could be an assistive as well as a restorative approach for upper-limb recovery after stroke.
Globally, 1 in 4 people will suffer from a stroke 1 . Of these people, nearly three-quarters will exhibit lasting deficits in motor control of their arm and hand 2 that cause enormous personal and societal impact 3 . These motor deficits persist partly due to the failure of current neurorehabilitation approaches to substantially reduce upper-limb impairement 4 .
Patients with chronic stroke exhibit a stereotypical motor syndrome of the upper limb that can be decomposed into independently quantifiable deficits 5 : loss of strength; reduced dexterity; intrusion of aberrant synergies; and disorders of muscle tone. This 'paresis' phenotype emerges from damage to the corticospinal tract (CST), which disrupts connections between the cortex and the cervical spinal circuits controlling arm and hand movements [5][6][7] .
Given that in most cases damage to the CST is incomplete, we posited that voluntary motor control could be restored by amplifying the capacity of the residual CST. Specifically, we hypothesized that modulating the excitability of intact sublesional spinal circuits would increase their responsiveness to remaining CST neurons, thereby restoring the ability of these supraspinal inputs to drive movement. A century of research has shown that primary sensory afferent neurons in the dorsal roots provide a pathway for influencing spinal circuit We performed a tractography analysis using high-definition fiber tracking to compare the integrity of the CST axons between the lesioned and healthy hemispheres (Extended Data Fig. 1). Relative white matter integrity was then determined by comparing the fractional anisotropy (FA) of the lesioned hemisphere to the nonlesioned hemisphere by calculating a FA symmetry score (FAS; see Methods for definition). We obtained FAS = 0.17 for SCS01 and FAS = 0.35 for SCS02 (FAS = 0 no impairment; Methods), which indicated important unilateral damage to the CST in both participants. This was reflected in prestudy Fugl-Meyer motor assessments of 35 out of 66 (SCS01) and 15 out of 66 (SCS02); indicative of moderate and severe impairment, respectively.
This pilot study was designed to quantify the immediate assistive effects of continuous SCS on post-stroke motor deficits, including muscle weakness, impaired dexterity of arm and finger movements, intrusion of aberrant flexor synergies and abnormal tone. Based on previous work in SCI, we expected any immediate benefit to reverse once SCS was turned off 15,[20][21][22]24 . In this pilot trial, we did not incorporate activity-based training exercises into the protocol and instead focused on measuring immediate improvements attributable to the direct effects of SCS in facilitating motor function in the arm and hand. Testing began 4 d after implantation of the SCS leads and continued for four weeks, during which the individuals participated in assessment sessions five times per week, approximately 4 h per day. After 29 d, the percutaneous leads were removed. In each session, we evaluated function with and without stimulation, which we delivered continuously through a custom-built microcontroller-based system connected via percutaneous access to the SCS leads (Extended Data Fig. 2).

Cervical SCS achieves segment-level muscle recruitment
For pain, clinical SCS leads are placed along the midline to broadly stimulate dorsal columns 39 . In contrast, we showed previously in monkeys and humans that more selective recruitment of primary afferent fibers in the cervical dorsal rootlets can be achieved by positioning the clinical SCS leads laterally, near the dorsal root entry zone 15,16,38 . These primary afferents innervate motoneuron pools according to a well-defined rostro-caudal somatotopy 40 , and we predicted that stimulating specific nerve roots would lead to excitation of the corresponding motoneurons 40 (Fig. 1c). Consequently, we hypothesized that selective stimulation of rostral roots (for example, C4) would facilitate muscle activation in the upper arm, while stimulation of the caudal roots (for example, C8-T1) would target distal muscles including the forearm and hand 16 . Therefore, we designed a surgical approach to implant two linear electrodes mediolaterally spanning the dorsal roots C4-T1 (Fig. 1b). During implantation, we guided surgical placement with neurophysiological intraoperative monitoring 22 and verified that reflex-mediated muscle responses could be obtained reliably across all muscles of the arm and hand. Intraoperative data showed that SCS followed a clear rostro-caudal segmental specificity in both participants ( Fig. 1d and Extended Data Fig. 3). Monopolar stimulation of rostral contacts induced activity in the deltoids and trapezius while caudal contacts recruited intrinsic hand muscles (Fig. 1f,h and Extended Data Fig. 3). To verify that stimulation responses resulted from afferent-mediated recruitment of the motoneurons, and not by directly recruiting ventral roots, we stimulated through the same contact at different frequencies (1.1, 2, 5, 10 and 20 Hz). Reflex-mediated responses are well known to show frequency-dependent suppression phenomena 10,16 . The peak-to-peak amplitude of evoked muscle activity was reduced in a frequency-dependent manner confirming that motor neuron activation was occurring trans-synaptically (Extended Data Fig. 4). Repeated X-rays showed minimal rostro-caudal displacement of the leads from the implant (Extended Data Fig. 1), which did not affect functional performance, and we used the same contact configurations from week 2 (when they were finalized) to week 4 (Fig. 2). Stimulation intensity was adjusted daily to levels that enabled volitional movements but did not produce any passive joint movement or torques at rest 19-21 . excitability [8][9][10][11][12][13][14][15] . Furthermore, we and others have demonstrated that epidural spinal cord stimulation (SCS), a clinically approved technology, can be used to directly recruit these afferents [16][17][18] , thus providing a means to test our CST augmentation hypothesis in humans with chronic stroke.
Clinical support for SCS comes from studies exploring its use in the recovery of locomotion after spinal cord injury 19-25 (SCI). While most of these studies focused on quantifying the cyclic patterns of movement involved in walking 21,23,26 , several groups reported that SCS enabled people with complete leg paralysis to produce voluntary single-joint movements 19,20,22 . This movement facilitation was immediate and required SCS to remain on: an assistive effect 19,20,24 . Importantly, this SCS-related assistive effect is distinct from the involuntary movements elicited by technologies such as functional electrical stimulation 27,28 . Rather than directly producing movement, SCS facilitates the ability of residual propriospinal and supraspinal inputs to activate spinal motoneurons, thereby enabling volitional movement 20,29-31 . Moreover, when the assistive effect of SCS is combined with prolonged behavioral training, it promotes long-lasting recovery of voluntary leg motor function even in the absence of stimulation, thus demonstrating a therapeutic effect 22,25 .
Despite these encouraging findings for leg motor recovery, epidural stimulation of the cervical spinal cord to target upper-limb recovery has been largely unexplored (see these animal studies 15,16,32,33 and a single pilot in humans 34 ). In addition to paucity of evidence, application of SCS to the post-stroke upper-limb motor syndrome is hindered by disease-specific scientific and technical challenges. Dexterous control of the arm and hand relies more heavily on corticospinal input 35 than does control of locomotion 11,36 , so for the same degree of residual CST, the degree of spinal circuit potentiation required may be greater. Moreover, the physiology of post-stroke motor deficits is different from SCI; therefore, the response to SCS may be qualitatively different. For example, pharmacological treatments are more effective in treating spasticity after SCI than spasticity after stroke 37 . From a technical perspective, the cervical enlargement is longer and larger than the lumbosacral spinal cord, which makes current clinical paddle leads insufficient to cover all the segments innervating upper-limb muscles 16 . Here, building on our work in monkeys 15,16 we sought to overcome these challenges and tested the efficacy of cervical SCS to restore upper-limb motor deficits in humans with post-stroke arm and hand paresis. Specifically, we designed a neurosurgical approach that implants two staggered linear leads on the lateral aspect of the cervical cord to target each dorsal root innervating the arm and hand muscles at their entry into the spinal cord 15,16,38 . To determine whether continuous SCS could improve corticospinal control, we devised a battery of scientific and clinical assessments to quantify the assistive effects of SCS on strength, dexterity, synergies and spasticity.

Study design and experimental setup
In this study we report results from the first two participants of an ongoing study (NCT04512690) that seeks to collect feasibility and preliminary efficacy data on the effects of cervical SCS on motor control in people with chronic post-stroke upper-limb paresis using continuous trains of stimulation pulses delivered at a fixed amplitude and rate through selected contacts (Fig. 1). SCS01, female, 31 years old, suffered a right thalamic hemorrhagic stroke secondary to a cavernous malformation nine years before participation in the study. SCS02, female, 47 years old, suffered a right ischemic middle cerebral artery (MCA) stroke secondary to a right carotid dissection resulting in a large MCA territory infarct three years before participation in the study. Written informed consent was given by both study participants before enrollment. SCS01's lesion was localized to the internal capsule, midbrain and pons, while SCS02's lesion was larger, affecting the corona radiata of the right hemisphere ( Fig. 1d and Extended Data Fig. 1 Fig. 1 | Experimental setup and stimulation arrangement. a, Schematic of the experimental apparatus and paradigm. We measured wireless EMG activity from the muscles of the arm and hand while participants performed upper-limb motor tasks. We delivered electrical stimulation to the cervical spinal cord via two 8-contact leads (rostral, R; caudal, C) implanted in the cervical spinal cord. Simultaneous stimulation through selected contacts was controlled via percutaneous connections using an external stimulator. b, X-rays of both participants showing the location of the contacts of the rostral (blue) and caudal (dark gray) leads with respect to the midline (in red). c, Location of the motoneurons of the arm and hand muscles in the human spinal cord in relation to spinal segments (light yellow) and vertebrae (gray). We estimated the rostrocaudal position of motoneuron pools (blue) from Schirmer et al. 40 . Apb, abductor pollicis; Adm, abductor digiti minimi; Bic, bicep; Del, deltoid; Ext, wrist extensor; Flx, wrist flexor; Pro, pronator teres; Trap, trapezius; Tri, tricep. d, Graphical representation of muscle activation obtained by stimulating through selected contacts (labeled in red on the left of each human figurine). Each human figurine represents the front (left half) and back views (right half) of the arm muscles (Extended Data Fig. 3). Each muscle is colored with a color scale (on the left) representing the normalized peak-to-peak amplitude of EMG reflex responses obtained during 1-Hz stimulation at the stimulation amplitude indicated on the left. Peak-to-peak values for each muscle are normalized to the maximum value obtained for that muscle across all contacts and all current amplitudes. On the left, the MRI scan of each participant is shown with the segmented lesion in red.

Article
https://doi.org/10.1038/s41591-022-02202-6 In summary, we showed that accurate placement of clinical leads over the dorsolateral cervical spinal cord produces selective muscle activation according to preexisting myotomal maps and that stimulation activates motor activity through sensory afferents in the dorsal roots.

Arm and hand strength immediately improved with SCS on
To determine whether SCS would lead to an increase in strength, we asked participants to apply their maximum force during isometric flexion and extension at single arm joints. Forces were applied to a robotic platform that measured joint torque (HUMAC NORM) (Fig. 3g). We compared the torques produced with and without continuous SCS targeting muscles of the tested joint (Fig. 3). We found that SCS01 consistently increased flexion and extension strength for the shoulder and elbow; mean torques at the elbow more than doubled when SCS was provided (day 9: 9.8 versus 22.0 Nm; day 23: 11.6 versus 24.6 Nm) (Fig. 3a,c,e). SCS01 was unable to generate wrist extension torques detectable by our system even with SCS. However, we could measure consistent improvement in wrist flexion torques (Fig. 3d).
As an example of the functional relevance of these improvements, we show a video in which SCS01 can raise her arm above her head during SCS (Supplementary Video 1). In SCS01, we tested multiple stimulation frequencies (20, 40 and 60 Hz) during elbow flexion and extension isometric tests and found that 60 Hz yielded maximal torques (Extended Data Fig. 5a). The severity of SCS02's impairment hindered consistent assessment of some joints. Specifically, she could produce detectable torques during shoulder flexion and extension and demonstrated significant improvements in elbow flexion torque (Fig. 3a,c,e) similar to those observed in SCS01 (40% average increase) but we were not able to detect elbow extension or wrist torques either with or without SCS.
We also tested isometric grip strength using a handheld dynamometer (Fig. 3f). SCS led to a 40% increase in SCS01 and a 108% increase in SCS02, suggesting that SCS can potentiate both arm and hand function. This result was particularly striking for SCS02 who had near-complete left-hand paralysis and was unable to consistently produce detectable hand grip forces (as measured with a hand dynamometer) without SCS. Additionally, on the first day of testing, SCS01, for the first time in the nine years since her stroke, immediately reacquired the capacity to fully and volitionally open her hand (Supplementary Video 1). We also compared the root mean square values of electromyography (EMG) signals measured from the anterior deltoid, biceps and triceps during elbow extension (SCS01) and elbow flexion (SCS02). EMG was substantially higher with stimulation than without for these muscles in both participants (>100% increase; Fig. 3a,b) indicating that SCS potentiated the participant's ability to recruit muscles.
Because participants were always aware that stimulation was on due to SCS-induced paresthesias 38 , we performed sham trials in which nonoptimal stimulation was delivered without participant knowledge to control for motivational bias. In these sham trials, we selected electrodes that preferentially activated muscle groups that were antagonistic to the movement performed (for example, electrode 8 R facilitated extension and 2 R facilitated flexion). SCS01 still experienced paresthesia over the shoulder and arm during stimulation and was unable to distinguish optimal from suboptimal configurations. As expected, while even antagonist stimulation led to some increase in strength (+19% extension using 2 R, 2.2 mA, 60 Hz; 7 C, 3.6 mA, 60 Hz and +16% flexion using 8 R, 2.4 mA, 60 Hz; 7 C, 3.6 mA, 60 Hz), the largest improvements in strength occurred only when SCS was optimized for the intended movement (agonistic stimulation; +82% extension using 8 R, 2.4 mA, 60 Hz; 7 C, 3.6 mA, 60 Hz and +25% flexion using 2 R, 2.2 mA, 60 Hz; 7 C, 3.6 mA, 60 Hz) (Extended Data Fig. 6a). In summary, we showed that SCS led to immediate and substantial improvements in strength and muscle activity of the arm and hand when optimal contacts were used.

SCS improved arm motor control during planar reaching
In addition to strength, we evaluated the benefit of SCS on arm dexterity and muscle synergies. For this, both participants performed planar reach and pull tasks using a robotic platform (KINARM) that supported the weight of their arm (Fig. 4a). Importantly, these reaching tasks were performed in the horizontal plane to dissociate the effects of shoulder weakness and compensatory movements from the capacity of participants to extend their arm toward a target 41 .
SCS01 was asked to reach toward different targets positioned to maximize active elbow extension since this was particularly difficult for the participant due to the intrusion of flexor synergies. During continuous stimulation, SCS01 was able to successfully reach all targets, whereas without stimulation, she was never able to reach the central target (Fig. 4b). The movements to targets that she could consistently reach with and without stimulation were significantly smoother with stimulation active (Fig. 4c; 34% (left target) and 47% (right target) fewer velocity peaks). Similarly, speed ( Fig. 4c; 41% (left target), 37% (right target) faster), trajectory, variability and maximum distance reached, were all improved with stimulation compared to controls (Extended Data Fig. 5b).
Due to the severity of SCS02's impairment and to reduce frustration, she performed a simpler task in which she was instructed to reach the furthest of three horizontal lines spaced at 10-cm intervals (Fig. 4d). Despite the simpler concept, the task assessed the same reaching and pulling arm kinematics as SCS01. Without stimulation, SCS02 was never able to reach the farthest line but with stimulation on she was able to reach it on every trial due to the facilitation of elbow extension. This was reflected in the elbow excursion angle, which increased 23 degrees with stimulation (Fig. 4e). The maximum distance reached was 7.8 cm greater and total movement time was 37% faster with stimulation ( Fig. 4e). Like SCS01, her movements was also smoother during stimulation (20% fewer velocity peaks; Extended Data Fig. 5c) and her trajectory variance and total path length also significantly improved (Extended Data Fig. 5c). Arm extension kinematics and elbow angle were strongly modulated by stimulation frequency in SCS02 showing peak performances at 100 Hz (Extended Data Fig. 5a).
We hypothesized that the improvements in reaching were attributable to facilitation of elbow muscle activity and changes in flexor and extensor synergies. To test this, we inspected EMG activity and extracted muscle flexor and extensor synergies associated with the extension and flexion movement phases using dimensionality reduction ( Fig. 5; Supplementary Information). Without stimulation, muscle activity was very low at the elbow muscles and very high at the shoulder muscles, probably indicating a compensatory strategy dominated by shoulder muscles and allowing the elbow to extend passively during the reach. This was reflected in the strength of the components of each synergy that showed a greater contribution of shoulder muscles in both participants (Fig. 5d,f,i). Instead, with stimulation, the contribution of elbow muscles increased and was dominant in both synergies, which suggests a reduction of compensatory shoulder movements (Fig. 5d,f,i).
To test if stimulation specificity was necessary for optimal motor control, we performed a sham-controlled task in which frequency and amplitude-matched nonoptimal stimulation was delivered without participant knowledge in the center-out task (sham: 4 R, 4.4 mA, 50 Hz; 7 R, 4.8 mA, 100 Hz; 8 C, 4.6 mA, 50 Hz versus optimized: 1 C, 4.4 mA, 50 Hz; 1 R, 4.6 mA, 50 Hz; 5 C, 4.8 mA, 100 Hz). Extended Data Fig. 6b-e shows the dramatic impact of incorrect stimulation configuration on SCS02's task performance. Specifically, during sham-stimulation, arm kinematics suffered dramatically and her performance worsened, even compared to her ability with stimulation off, affecting her ability to reach designated targets. Instead, with optimal stimulation she reached all targets 100% of the time.
In summary, despite differences in deficits and task difficulty, we showed that SCS targeting dorsal roots at specific cervical segments immediately improved dexterity and enabled both participants to perform smooth and effective arm movements enabling full elbow extension, improving elbow extension and flexion synergies and reducing compensatory shoulder activity.

Functional benefits of SCS
Finally, we sought to determine whether these improvements in strength and control translated to functional movements and improved performance during activities of daily living (ADLs) (Fig. 6). For this, we personalized tasks according to impairment level and chose ADLs based on observations of early-study improvements and individuals' rehabilitation goals. We first evaluated the ability of SCS01 to perform three-dimensional (3D) reaching movements. We asked SCS01 to reach as fast as she could toward six targets placed on two different horizontal planes that required both planar and upward reaching movements against gravity. Continuous SCS enabled her to reach targets faster, approximately reducing by half the time required to complete the six-target circuit (Fig. 6f). We also asked SCS01 to perform a classic manipulation task: the box and blocks task, in which she was instructed to move small cubic objects from one side of a box to the other by grasping and lifting them over a barrier. With stimulation on, she consistently performed better; on day 17 post-implant, she more than doubled the number of blocks transferred when stimulation was off. Her score increased from six blocks without stimulation to 14 blocks during stimulation ( Fig. 6e

Fig. 2 | Optimized continuous stimulation protocols.
Stimulation protocol used to achieve maximum assistive benefit for SCS01 (top) and SCS02 (bottom). Top, for SCS01, contacts 1R and 8R on the rostral lead and 7C on the caudal lead were simultaneously and continuously activated at a fixed 60-Hz frequency and 200-µs pulse width. These electrodes corresponded to the shoulders and biceps (1R); triceps, extensors and hand opening (8R); and hand grasp (7C). Amplitudes were changed daily based on participant preference and were set to 2.4-2.6 mA (1R), 2.1-2.7 mA (8R) and 3.3-6.2 mA (7C). Bottom, for SCS02, contacts 1R on the rostral lead and 1C, 5C and 8C on the caudal lead were simultaneously and continuously stimulated. These electrodes corresponded to muscles related to shoulder support (1R), elbow flexion (1C), elbow extension and wrist flexion (5C) and hand grasp (8C). Contacts 1R and 1C were stimulated at 50 Hz while 5C and 8C were stimulated at 100 Hz all at a fixed pulse width of 400 µs. A reduced frequency was used on contacts corresponding to elbow flexion to bias the assistive benefit of stimulation toward elbow extension. Multifrequency stimulation was achieved by skipping every other period of a 100-Hz stimulation protocol on channels stimulating at 50 Hz. Location of the motoneurons of the arm and hand muscles in the human spinal cord in relation to spinal segments (light yellow) and vertebrae (gray) is shown on the left for SCS01 and SCS02. We estimated the rostro-caudal position of motoneuron pools (green) from Schirmer et al. 40 .
Article https://doi.org/10.1038/s41591-022-02202-6 score was 31 out of 57. At the end of the study, we administered the test both with and without stimulation, with resulting scores of 45 out of 57 and 36 out of 57, respectively, representing a 14-point improvement while SCS was active. Finally, we increased the complexity of the tasks by presenting ADLs that required high skill and dexterity, such as drawing a spiral, reaching for and lifting a soup can, eating with a fork and opening a lock. SCS increased her overall dexterity, allowing her to produce smoother and more consistent drawings (Fig. 6a). Stimulation also enabled simultaneous reaching, forearm supination and grasp allowing SCS01 to reach, grasp and lift a soup can. Without stimulation, forearm pronation and supination were not possible ( Fig. 6b and Supplementary Video 3). SCS also enabled fine motor skills, such as opening a lock and manipulating utensils to eat independently ( Fig. 6c and Supplementary Video 4), tasks that she had been unable to perform for nine years. Since SCS02 was unable to sustain the weight of her arm against gravity, we adapted the 3D reaching task using a clinically approved actuated exoskeleton robot (Hocoma Armeo Power) to provide titrated assistance during the task (50% arm weight support). We endeavored to keep the movements as similar as possible to those performed by SCS01 to allow for comparison by having the participant collect virtual objects from a room and place them on a target (Fig. 6g). With stimulation, SCS02 was more efficient at the task and managed to consistently reach toward more targets than without stimulation across three sessions (Fig. 6g). We then asked her to perform a skilled motor task where she had to remove a hollow cylinder from a wooden dowel and slip it over another. With SCS she was not only able to grasp and lift the metal cylinder but also to place it on the adjacent dowel without any weight support ( Fig. 6d and Supplementary Video 5). Without SCS she could not complete any of the steps required for this task. These results from the first two participants in our pilot study show that the assistive effects of SCS may lead to important improvements in functional performances and ADLs.

Tone, spasticity and lasting effects on motor control
To ensure that increased excitation to the spinal cord via SCS did not exacerbate spasticity or muscle tone, we measured the Modified Ashworth Scale (MAS) on each day of testing. To minimize daily assessment duration, we limited the joints tested for each participant to those with MAS > 1 before the study. However, elbow and digit flexion, shoulder external and internal rotation, and shoulder abduction were tested in both participants, for consistency, regardless of previous history. Over the course of four weeks, we found that SCS did not lead to any worsening nor amelioration in MAS scores ( Fig. 6i and Extended Data Table 1). As this pilot was designed to study the assistive rather than the therapeutic effects of SCS, participants did not receive concomitant physical or occupational therapy over the four weeks. Thus, we did not expect to observe sustained improvements when SCS was turned off. Nevertheless, when we compared the participants' pre-and poststudy Fugl-Meyer scores, SCS01 improved from 35 points at enrollment to 47 points and SCS02 from 15 points to 18 points. These scores decreased by one point at the four-week follow-up visit ( Fig. 6h and Extended Data Table 2).

Safety and tolerability of the stimulation
Concerning preliminary safety data, no serious adverse events were reported. SCS01 experienced phlebitis several days after the explantation procedure at the end of the study that was resolved with oral antibiotics and was not related to the use of stimulation to the spinal cord. Moreover, the two participants did not report increased rigidity nor painful sensations during SCS. In fact, both participants described the stimulation as a 'feeling of power in the arms' or a feeling of 'being able to control my arm as if I know what I should do to move it'.

Discussion
In this study we report preliminary evidence from two participants showing that continuous SCS targeting cervical dorsal roots could immediately improve upper-limb strength, motor control and function in two individuals with moderate-to-severe post-stroke hemiparesis. This assistive effect was lost when SCS was turned off. However, both participants showed some lasting improvements in motor function by the fourth and final week of the study that were retained even without stimulation. While we cannot conclude on safety and efficacy from two participants, we discuss the first results, to the best of our knowledge, in humans reporting the effects of epidural SCS on post-stroke upper-limb hemiparesis.
Although SCS for SCI has recently generated considerable excitement 21-23 , anecdotal reports of the beneficial motor effects of SCS in people with SCI, multiple sclerosis, cerebral palsy and even stroke date back more than 40 years 43,44 . Unfortunately, a lack of understanding of the mechanisms of SCS led to considerable variability in implant location, which affected the size and consistency of the observed effects. We now know that SCS engages spinal motoneurons via recruitment of primary afferents, providing excitatory input to motoneurons and interneurons directly connected to these afferents [16][17][18]29,30 . Thus, it could be hypothesized that by raising the membrane potential of spinal neurons, SCS increases responsiveness to residual cortical inputs and immediately improves voluntary motor control 16,19,31,45 . We define this as the 'assistive effect' of SCS. By implanting epidural electrodes over the lateral aspect of the cervical spinal cord, we focused this

Synergy vector
Left target

SCS01
Center target

Reach phase
Pull phase  assistive effect on the arm and hand motor pools most needed for each participant 15,16 . Our data show that SCS had broad assistive effects on motor control but we did not assess individuated finger control, which is highly CST dependent. Nevertheless, we observed improved dexterity in functional hand tasks, particularly in participant SCS01.
Notably, Supplementary Video 4 shows SCS01 grabbing a fork using a pinch grasp during SCS. This suggests that SCS is tapping into residual CST function. We argue that the large effect sizes we measured were possible because, unlike in SCI, the stroke lesion spares cervical spinal circuits and usually some supraspinal pathways. Indeed, studies of cervical SCS in SCI have not shown the magnitude of immediate assistive effects at the arm and hand as those that we report in this study post-stroke 34,46 . Despite large differences between the two participants in terms of severity, age and time since the stroke, during SCS, both participants showed a significant improvement in strength (Fig. 3), arm kinematics (Fig. 4) and functional task performance (Fig. 6) compared to their baselines without SCS. However, clinical adoption of SCS for stroke will require the application of simple and standardized parameter optimization protocols. Additionally, in this study we focused on the immediate assistive effects of SCS with temporary implants; future studies should focus on demonstrating the long-term safety and efficacy of a fully implanted SCS system combined with protocol-based upper-limb rehabilitation in larger randomized controlled studies.
The lateral epidural placement of the SCS leads enabled selective targeting of specific dorsal roots, which allowed us to use simple yet personalized continuous stimulation protocols. Our surgical approach simplified the personalized parameter tuning by leveraging clear somatotopic organization of the cervical spinal cord. Final configurations were selected during the first two weeks after the procedures similar to works in SCI 47 . While noninvasive alternatives to SCS are being investigated in SCI 46 , our results depended on fine-tuning of stimulation parameters at particular contacts, which would not be possible with the limited specificity of transcutaneous SCS 48 . We believe that the simplicity and robustness of our protocol could facilitate translation to the clinic. Indeed, implantation and programming procedures are similar to routine applications of SCS for refractory pain 49 .
The most important limitation of our study is that results are presented in only two participants, which hinders definitive conclusions on safety or efficacy. Other limitations include the absence of fast-reaching (f) tasks performed on multiple days. Individual data points are also shown because some datasets contained fewer than five data points. g, Picture of the 3D reaching task using the Armeo Power for SCS02 and relative task performance on multiple days. Individual data points are also shown because some datasets contain fewer than five data points. h, Fugl-Meyer assessment at different time points for SCS01 and SCS02 including four weeks poststudy. a protocol-based upper-limb behavioral intervention and the short duration of the study (four weeks) that may have reduced the amount of recovery that participants could obtain. In this study, we placed an emphasis on quantifying the immediate assistive effects of SCS. In contrast, almost all rehabilitative stroke studies concern recovery after the intervention is over 5,50,51 , which makes comparison with the literature difficult. That said, although our two participants received no protocol-based training, we observed some motor recovery that, for SCS01, was above minimally clinical important differences for the Fugl-Meyer score (+12 Fugl-Meyer points; minimally clinical important difference: +7.25). Given the short duration of our study compared to similar works in SCI 22,52 , we did not expect to observe restorative effects. Therefore, we believe that our results are a preliminary but promising indication that, by combining SCS with protocol-based upper-limb rehabilitation, future studies may lead to higher improvements and promote a true post-stroke restorative effect 45 . This restorative avenue is especially exciting considering the advent of new and more effective impairment-focused behavioral interventions for stroke 5,53,54 that could be combined with SCS into an effective therapy for post-stroke hemiparesis.

Online content
Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41591-022-02202-6.

Methods
In addition to those reported in this article, detailed methods on surgical procedures, EMG analysis and the stimulation control software can be found in Supplementary Information.

Trial and participant information
All experimental protocols were approved by the University of Pittsburgh Institutional Review Board (IRB) (protocol STUDY19090210) under an abbreviated investigational device exemption. The study protocol is published on ClinicalTrials.gov (NCT04512690). Two females (aged 31 and 47) participated in the study. Both participated in every experiment. Some procedures were modified depending on specific individual deficits and such cases are indicated where appropriate. Both participants provided informed consent according to the procedure approved by the IRB of the University of Pittsburgh and participants were compensated for each day of the trial and for travel and lodging during the study period.
Inclusion criteria. Individuals between 21 and 70 years of age who had suffered from an ischemic or hemorrhagic stroke more than six months before the start of the study were eligible for participation. All individuals had hemiparesis affecting their upper limb and had a prestudy Fugl-Meyer upper-extremity score between 7 and 45. Before the study, participants were screened via a medical evaluation. Candidates with severe comorbidities, previously implanted medical devices, claustrophobia or who were pregnant or breastfeeding were excluded from the study. Individuals were not receiving any anti-spasticity, antiepileptic or anticoagulation medications for the duration of the study period.
Study design and data reported. The goal of this exploratory clinical trial is to obtain preliminary evidence of safety and efficacy of SCS to improve motor control in people with chronic post-stroke upper-limb hemiparesis. The study is designed as a single-group, open-label, prospective study in which we expect to enroll and retain up to ten participants with chronic stroke. Given the pilot nature of the study, to minimize safety risks, SCS leads are implanted for a maximum period of 29 d, after which the electrodes are explanted. We designed our primary and secondary outcomes to primarily assess safety and obtain preliminary clinical and scientific evidence of both the assistive and long-term effects of SCS. Briefly, after screening and prestudy baselines, participants are implanted with clinical SCS leads. Starting from day 4 post-implant, participants undergo scientific sessions five times per week, 4 h per day, for a total of 19 sessions until explantation day. Tasks and measurements during the first 5-7 sessions are focused on identifying optimal stimulation configurations that are then maintained for the remaining sessions.
A detailed description of primary and secondary outcomes can be found in Supplementary Table 1 and on ClinicalTrials.gov NCT04512690.
Briefly, primary outcomes are focused on reporting serious adverse events and assessing pain and discomfort. Specifically, we consider the trial to be successful if there are no serious adverse events related to the use of the stimulation. We then ask participants to report and rate, if present, any pain or discomfort that arises from the stimulation with the goal of understanding if intensities required for motor effects are pain-and discomfort-free. Secondary outcomes are geared toward scientific and preliminary efficacy goals. In terms of clinical efficacy, we quantify immediate improvements in strength by measuring isometric forces with and without SCS once per week. We rate motor deficits by assessing the Fugl-Meyer evaluation, ARAT assessment prestudy and on the last day of implantation, and spasticity via the MAS daily. We then evaluated function by measuring the kinematics of the arm and hand during two-dimensional and 3D reaching and grasping tasks. Finally, we perform a battery of imaging and electrophysiology testing to assess the excitability and plasticity of spinal circuits pre-, during and poststudy. Below we detail the methods for each of the measurements reported in this trial that are part of the secondary outcome assessments. In this manuscript, we report the preliminary results from the first two participants of all the primary and secondary outcomes except those that require additional data on more participants to obtain meaningful analyses, such as imaging and electrophysiology outcomes.

Participant information.
In this work, we report the results from the first two individuals participating in our trial, both of whom were white females. SCS01 (31 years old) had a right thalamic hemorrhagic stroke secondary to a cavernous malformation nine years before participating in the study. Her interim history involved several bleeding events with eventual ablation of the malformation with gamma knife radiosurgery. At the time of her participation in our trial, her post-stroke residual was a left-sided spastic hemiparesis for which she was receiving botulinum neurotoxin injections in her biceps, brachioradialis and pronator teres. Botulinum neurotoxin treatments were suspended starting six months before the study period and continuing through the end of the study. In the years between her initial infarct and participation, she also underwent a C5-C6 anterior cervical discectomy and fusion to treat cervical stenosis and flexor tendon lengthening surgery due to spasticity and suffered arm and wrist fractures in her affected arm. For SCS01, we included in this work analysis of 138 isometric force test repetitions at multiple joints (54 stimulation off and 84 stimulation on) and 36 planar reaches (18 with SCS and 18 without SCS). We also report the results of simulated ADLs and other motor tasks that were performed at least one session per week (Fig. 6). Because of technical and participant availability reasons, we could not perform transcranial magnetic stimulation tests before the study to obtain motor evoked potential (MEP) maps on SCS01.
SCS02 (47 years old) had a right ischemic MCA stroke secondary to a right carotid dissection resulting in a large MCA territory infarct three years before participating in the study. Her post-stroke residual at the time of participation was a left-sided spastic hemiparesis complicated by a left wrist flexion contracture despite treatment with splinting. For SCS02, we included in this work the analysis of 42 isometric force tests repetitions at multiple joints (21 stimulation off and 21 stimulation on) and 57 planar reaches (38 with SCS and 19 without SCS) that were obtained across multiple days during the study. We also report the results of simulated ADLs and other motor tasks that were performed at least one session per week (Fig. 6). Transcranial magnetic stimulation measurements obtained over nine locations and 11 muscles confirmed that SCS02 was MEP-negative (for example, no MEP present in any of the muscles of the paretic arm).

Safety
We recorded all adverse events and reported them to the data safety and monitoring board (DSMB) and to the IRB to determine whether these would be related to the delivery of electrical stimulation to the spinal cord. Both participants successfully completed the protocol with no serious adverse events. SCS01 experienced phlebitis several days after the explantation procedure at the end of the study that was resolved with oral antibiotics. The DSMB and IRB determined that this adverse event was not serious and not related to the delivery of SCS to the cervical spinal cord.

Medical imaging
X-ray imaging. X-ray images were acquired at weekly time points in both axial and sagittal views to ensure the stability of lead position.
Lesion segmentation. Magnetic resonance imaging (MRI) was acquired using a 3T Prisma system (Siemens) using a 64-channel head and neck coil. A T1-weighted structural image was captured using a Nature Medicine Article https://doi.org/10.1038/s41591-022-02202-6 magnetization-prepared rapid gradient echo sequence (repetition time = 2,300 ms; echo time = 2.9 ms; field of view = 256 × 256 mm 2 ; 192 slices, slice thickness = 1.0 mm, in-plane resolution = 1.0 × 1.0 mm). Lesion segmentation was performed manually for each slice of the sequence using the MRIcron image viewer (NITRC); the resulting region of interest was smoothed on all planes using a Gaussian smoothing kernel with full-width at a half-maximum of 2 mm. MRIcroGL (NITRC) was used to visualize and export the resulting segmented overlays.
High-definition fiber tracking. The same 3T MRI scanner was configured to use a diffusion spectrum imaging scheme to capture a total of 257 diffusion samples. The maximum b-value used was 4,000 s mm −2 and the in-plane resolution and slice thickness were 2 mm. The accuracy of b-table orientation was examined by comparing fiber orientations with those of a population-averaged template 55 .
The diffusion data were reconstructed in the Montreal Neurological Institute space using q-space diffeomorphic reconstruction 56 to obtain the spin distribution function 57 . A diffusion sampling length ratio of 1.25 was used. The output resolution in diffeomorphic reconstruction was 2 mm isotropic. The restricted diffusion was quantified using restricted diffusion imaging 58 . The tensor metrics were calculated and a deterministic fiber tracking algorithm 59 was used to reconstruct the CST fibers. A tractography atlas 55 was used to map the left and right corticospinal tracts with a distance tolerance of 16 mm. For fiber tracking, we used an anisotropy threshold of 0.035, an angular threshold of 50 degrees and a step size of 1 mm. Tracks with lengths shorter than 10 mm or longer than 200 mm were discarded. A total of 1,000,000 seeds were placed. Topology-informed pruning 60 was applied to the study tractography with 16 iterations to remove false connections. We then calculated the mean FA values for the left and right CST and the percentage of asymmetry was computed using Stinear's formula: where FA L is the mean FA value of the CST in the lesioned hemisphere and FA H is the mean FA value of CST in the intact hemisphere.

Efficacy assessments: single-joint isometric torque
Maximum isometric strength was measured for the shoulder, elbow and wrist joints (when possible) using a robotic torque dynamometer (HUMAC NORM, CSMi). To measure torque, the robot's manipulandum was positioned and held at a fixed angle and the participant was asked to apply their maximum force while flexing or extending the joint under test for a sustained period of 5 s followed by a 10-15-s break. This procedure was repeated five times to complete a set. For each joint, the system was configured such that the joint was at a nominal and comfortable angle and so that it was aligned with the manipulandum's center of rotation. To isolate single-joint function, participants were constrained with tight straps at the shoulders and additional straps and bracing specific to each joint configuration of the HUMAC NORM. For example, while testing elbow strength, the upper arm and elbow were stabilized against the back of the chair while holding the manipulandum at a 90-degree angle. This ensured minimal shifting. Additionally, while testing wrist strength, the forearm was strapped to the robot's joint stabilization attachment. The HUMAC NORM's suggested configurations were used, when possible, but SCS02 was unable to support the weight of her arm and so was placed in a seated position to measure the elbow and shoulder torques instead of the suggested supine position. In addition, a splint was used to secure SCS02's hand to the manipulandum to assist her in holding the handle securely and a counterweight was used where appropriate to offset the mass of the manipulandum and allow for more sensitive measurements. The maximum torque value within each repetition was considered for analysis.
Grip force was measured using a hand dynamometer. Participants were asked to hold the dynamometer and apply their maximum grasping force for 5 s. Before every repetition and after the participant's impaired hand was around the dynamometer, the device was zeroed out to ensure that the baseline grip force at rest was zero. Each measurement comprised the highest force produced on each of three attempts and data were combined across days to assemble enough data for statistical comparison.

Efficacy assessments: clinical impairment scales
Fugl-Meyer upper-extremity assessment. The Fugl-Meyer upper-extremity assessment is a standardized evaluation of upper-limb motor control and sensory function 61 . It includes seven categories of assessments including passive and active range of motion, joint pain, proprioception and tactile sensation. In total, there are 126 possible points. However, all scores reported in this manuscript correspond to the 'motor function' subscore, which has a maximum value of 66. A trained medical professional conducted and scored the exam at four different time points: prestudy; mid-study (approximately two weeks after implant); end of study (four weeks); and poststudy (one month after explantation).
ARAT. ARAT is another assessment of upper-limb motor function that focuses on object interaction and manipulation. It consists of four categories including grasp, pinch, grip and gross movement 42 . Scores can range from 0 to 57 with 57 representing the highest functional performance 62 . A trained medical professional evaluated SCS01's ARAT performance both before the study and at the end of the study. While SCS02's score was recorded at the prestudy time point, she did not consent to perform the test again at the end of the study because of fatigue; hence, these data points are not available for SCS02.

MAS.
To ensure that SCS was not exacerbating joint spasticity, we performed the MAS on each session day at the beginning of the session. To minimize daily assessment duration, we largely limited the joints tested for each participant to those with spasticity before the study. However, elbow and digit flexion, shoulder external and internal rotation and shoulder abduction were tested in both participants, for consistency, regardless of previous history. This assessment involves passive manipulation of each joint and ranking spasticity levels from 0 to 4 (0 being no spasticity). A trained medical professional performed and scored the assessment each day. In this article, we report both a full breakdown of all joint scores measured on each day for both participants as well as a 'summary score'. The summary score was taken to be the average score across all joints for each day.

Functional assessments: planar reach and pull kinematics
To evaluate upper-limb motor control during directed reach and pull movements, we used a robotic augmented reality exoskeleton system (KINARM, BKIN Technologies). Participants were secured in a modified wheelchair and their arms were suspended in the exoskeleton to remove the effects of gravity. The platform displayed virtual targets onto a dichroic augmented reality display in front of the individual that allowed them to visualize their hand position relative to the virtual graphics. The robot's motorized joints permitted the application of a mechanical load to the participant's movements.
Center-out task. For this task, participants were asked to reach from a central starting position to one of three targets displayed using the augmented reality display, then return to the starting position. On each trial, the starting position was displayed and the robot moved the participant's arm into position, locking it in place. Next the target was presented and the exoskeleton was unlocked. An audio cue was played after a randomized 100-700-ms delay indicating that the participant could begin their movement. The participant was given 10 s (SCS01)

Nature Medicine
Article https://doi.org/10.1038/s41591-022-02202-6 or 15 s (SCS02) to complete each trial. A target was considered acquired when the participant's index finger was within a 0.5-cm radius of the target center for 500 ms. An audio cue indicated the end of the reach phase. If the participant was unable to reach the target, the robot returned the arm to the starting position and the next target was presented. If the trial was successful, the participant's finger was positioned in the center of the target in preparation for the pull phase and locked in place. After a 500-ms delay, the arm was unlocked followed by a final audio cue after another 100-700-ms delay indicating the start of the pull phase and the participant was required to return their hand to the starting position. In some trials, a load of −30 Ns m 2 was applied isotropically to the movement using the exoskeleton to increase the task difficulty. Each target was presented six times in random order (unless otherwise noted). For each participant, appropriate targets were selected based on their individual range of motion.
The following metrics were calculated for each trajectory to compare kinematic quality. Trajectory smoothness was calculated as the number of peaks in the velocity profile for both the reach and pull phases. We also measured the total time of the combined reach and pull phases. Total path length was calculated and normalized to the Euclidean distance between the starting position and the target; more efficient movements had a lower value. Finally, the variance of each trajectory was calculated as the mean deviation of the actual trajectory from the mean trajectory calculated across all five repetitions of the movement. Open-ended reaching task. The participant was presented with three equally spaced horizontal lines (approximately 15, 25 and 30 cm away from them) and was asked to reach from a starting position to the furthest line they could. In this way we assessed how far they could reach in an open-ended manner.
During each task, the participant started with their hand as close to their body as they could (maximum elbow flexion). After a verbal cue, they began their movement with the goal of passing the farthest line possible. Once the participant indicated that they had reached their maximum distance, another verbal cue indicated that they should return to their initial position. Task events were manually labeled during the trial by the experimenter. Each set consisted of five repetitions.
As in the center-out task, a set of metrics was calculated for each trajectory; the reach and pull phases were considered separately. Movement duration was calculated as the time it took from the beginning of each phase for the participant to cross the second horizontal line (25 cm) during reach and the first horizontal line during pull (15 cm). The maximum distance was measured as the axial distance between the point closest to the participant and the point furthest from the participant in each phase. Range of motion of the elbow during the task was considered as the angle difference between the most acute and most obtuse elbow angles achieved during each phase. As a metric of smoothness, the number of peaks in the elbow angle velocity profile was counted. Total path length measured the total length of the trajectory from the starting point to the second line (25 cm; reach phase) or from the end position to the first line (15 cm; pull phase) and was normalized by the phase duration. Finally, as a measure of variance, we calculated the distribution of each trajectory time point from the mean trajectory. A distribution skewed toward the left indicated that more samples were close to the mean trajectory, whereas a distribution with values toward the right indicated large deviations from the mean trajectory and therefore more variance.

Functional assessments: 3D reaching
Fast-reaching task. The participant was presented with six targets, all axially equidistant from them but at varying heights and lateral positions. The three 'lower' targets were at table surface height and the three 'upper' targets were raised to require shoulder flexion beyond 90 degrees. There was a left, center and right target at each height. A seventh position was placed directly in front of the participant and was used as a 'home' position. Starting with their arm outside the working area, the participant was asked to first touch the home position then touch each of the six targets, returning to the home position after each target. For this task, we asked the participant to perform the sequence as fast as possible. The total time it took to reach all six targets was recorded.
Robotic 3D reaching task. As an alternative to the fast-reaching task, we used an exoskeleton robot (ARMEO POWER, Hocoma) to assist 3D movements when the participant was unable to lift their arm against the force of gravity (SCS02). This robotic system provides motorized support at each joint of the arm and measures kinematic variables in real time allowing for a participant's real-world movements to be displayed in a virtual video game environment. For this task, objects were presented within a virtual room and the participant was asked to reach toward each object and move it to a different position within the room (ARMEO POWER cleanup game). The robot was configured to provide 50% weight support and assist movements at the 'low support' setting. Game difficulty was set to 'easy'. Each game lasted 3 min and the goal was to move as many objects as possible within the time limit. The score was then recorded based on the number of objects successfully moved.

Box and blocks.
When possible, we also evaluated the participant's performance in the 'box and blocks' task. This is a standardized assessment in which a participant must grasp one small block at a time from one side of a box, lift it over a divider and drop the block in the other half of the box. The total number of blocks moved from one side to the other within 1 min was the participant's score.

Functional assessments: ADLs
We chose ADLs after an initial assessment phase based on participant ability and preferences. In some instances, we chose tasks that emulated the everyday activities participants had identified as being difficult to perform before the study but that they wished to try after having experienced the stimulation. As ADLs were customized for each participant, we did not evaluate prestudy performance for these tasks.
Drawing a spiral. We asked the participant to draw a spiral shape using a marker on a plain piece of white printer paper taped down to a table. The goal of the task was to make the curves as smooth as possible and attempt not to overlap each of the concentric rings. The participant was allowed to comfortably position the pen in their hand using their unaffected hand before starting to draw.
Object manipulation. We placed a full, sealed can of soup on a table in front of the participant. They were asked to grasp the object from the side, requiring them to supinate their forearm, lift the can and place it at an adjacent target. This task evaluated their ability to reach, grasp, lift and release a moderately heavy object. They were not allowed to use their unaffected arm to assist in grasping the object.
In an alternative object manipulation task, we asked them to hold a wooden plank with vertical dowels (similar to a tower of Hanoi toy) on their lap using their unaffected hand. We then placed a metal cylinder over one of the dowels. The participant was required to grasp the cylinder, lift it off of the first dowel, align it and place it onto a second dowel, and release the cylinder. An experimenter helped to position the hand on the cylinder before the start of the trial. All other movements were performed by the participant entirely on their own. Lock opening. As a measure of hand dexterity, we positioned a wooden panel with a shackle-style key-actuated lock on a table in front of the participant, who was asked simply to open the lock using their affected limb. To do this task, they were required to grasp and stabilize the lock with one hand (for example, the unaffected hand), use a pinch grip to Nature Medicine Article https://doi.org/10.1038/s41591-022-02202-6 grasp the key with the other hand (for example, the affected hand) and supinate the forearm to twist the key and unlock the lock. The participant then removed the lock from its latch on the wooden panel, replaced it by realigning the shank with the latch and relocked the lock by aligning and pressing the shank back into the body.
Self-feeding. The subject was presented with small bite sized portions of food on a plate and a plastic fork. They were tasked with first picking up the fork from a table, using it to secure a piece of food, and perform the complex movement of orienting the food toward their mouth in preparation to eat it. Here, the subject was required to initiate picking up the fork with their affected hand but was allowed to reposition it using their unaffected hand before attempting to pick up the food.

EMG analysis
Isometric contraction (root mean square analysis). During isometric contractions, an EMG was acquired from appropriate muscles using wireless sensors as described above. Empirically, we observed that deltoid EMG signals contained stimulation artifacts during trials where stimulation was active due to the proximity of deltoid muscles to the stimulating electrodes. We removed these artifacts by blanking the signal coinciding with stimulation pulses. All signals were bandpass-filtered (25-300 Hz, 5th order Butterworth digital filter) and the root mean square value was calculated from the filtered data over the full duration of each trial for statistical analysis.
Planar reaching (muscle synergy analysis). Coordinated movements such as reaching and pulling require the timed coactivation of appropriate muscles to produce accurate and controlled limb motion. We measured which muscles were simultaneously active during planar reaching movements by calculating muscle synergies using nonnegative matrix factorization (NNMF), a dimensionality reduction technique 63 .
EMG preprocessing was different for SCS01 and SCS02 due to large amplitude stimulation artifacts present in SCS02's EMG data that were not present for SCS01. For SCS01, the stimulation artifact was removed by blanking and the resulting data were bandpass-filtered (20-500 Hz, 5th order Butterworth digital filter). For SCS02, EMG data were first bandpass-filtered using a narrower pass band (10-200 Hz, 5th order Butterworth digital filter) to remove high-frequency components of the stimulation artifact. Notch filters (5th order Butterworth) at 50, 100 and 150 Hz were then used to remove low-frequency harmonics of the stimulation artifact. The resulting signals from both participants were rectified, low pass-filtered (5 Hz, 5th order Butterworth digital filter) and normalized to the maximum EMG value recorded from that muscle over the whole day. Processed EMG data were extracted from the reach and pull phases of each movement. Muscle synergies were identified using NNMF.
NNMF decomposes the EMG signals into a synergy activation matrix using the temporal correlation between the activity of individual muscles 63 . The result is a set of one-dimensional time series signals for each muscle synergy identified. Each synergy in turn comprises contributions from multiple muscles as described by a synergy vector. We implemented NNMF with two factors, which were selected by observing the point of inflection in the residuals versus the number of the synergies curve 64 . For each phase of the movement (reach and pull), the primary synergy for that movement was identified as the one that most positively correlated (increased) with the movement. All repetitions of the movement were used to perform the dimensionality reduction. Finally, the contributions of the deltoid and elbow muscles were quantified and compared using the primary synergy's synergy vector.

Statistics
Bootstrapping. All statistical comparisons of the means presented in this manuscript were performed using the bootstrap method, a nonparametric approach that makes no distributional assumptions on the observed data. Instead, bootstrapping uses resampling to construct empirical confidence intervals (CIs) for quantities of interest. For each comparison (for example, comparing stimulation on versus stimulation off for shoulder torque in SCS01; shown in Fig. 3c), we constructed bootstrap samples by drawing a sample with replacement from the observed measurements, while preserving the number of measurements in each condition. We constructed 10,000 bootstrap samples and, for each, calculated the difference in means of the resampled data. A 95% CI for the difference in means was obtained by identifying the 2.5th and 97.5th quantiles for the resulting values. The null hypothesis of no difference in the mean was rejected if 0 was not included in the 95% CI. If more than one comparison was being performed at once, we used a Bonferroni correction by dividing this alpha value by the number of pairwise comparisons being performed. Bootstrap statistical analysis was only performed when at least five data points were obtained.
Comparison of distributions. Statistical comparison of distributions was done using a two-sample Kolmogorov-Smirnov nonparametric test using MATLAB R2021 a/b. Again, an alpha value of 0.05 was used. We used this test to compare the variability of kinematic trajectories during two-dimensional planar reaching (the open-ended reaching task). The deviations of each trajectory from the mean trajectory were used to build a distribution of deviations. The resulting distributions for two conditions (stimulation off and stimulation on) could then be compared using the Kolmogorov-Smirnov test.

Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability
Anonymized data will be uploaded to the Database Archive of the BRAIN Initiative (DABI, https://dabi.loni.usc.edu/home). Source data are provided with this paper. Article https://doi.org/10.1038/s41591-022-02202-6 Extended Data Fig. 3 | Muscle recruitment curves. In each panel we show the recruitment curves obtained with stimulation at 1 Hz at increasing current amplitude for 11 arm and hand muscles: TRAP: trapezius, A, P, M DEL: anterior, posterior and medial deltoid respectively, BIC: biceps, TRI: triceps, EXT: Extensor carpi, FLX: flexor carpi, PRO: pronator teres, ABP: abductor pollicis and ADM: abductor digiti minimi. Below each set of recruitment curves we report the graphical representation of the muscle activation obtained at the amplitude indicated on the left of each human figurine. Interpretation of human figurines is reported in the bottom right. Each muscle is colored with a color scale (on the left) representing the normalized peak-to-peak amplitude of EMG reflex responses obtained at the stimulation amplitude indicated on the left. Peak-topeak values for each muscle are normalized to the maximum value obtained for that muscle across all contacts and all current amplitudes.

Nature Medicine
Article https://doi.org/10.1038/s41591-022-02202-6 Extended Data Fig. 4 | Frequency dependent suppression. To demonstrate that SCS recruits arm and hand muscles via direct activation of the primary afferents we performed stimulation at multiple frequencies. The figure reports the spinal reflexes obtained when stimulating at 1, 5, 10 and 20 Hz from multiple contacts and multiple muscles. Each plot on the top shows the normalized peakto-peak reflex amplitude as a function of frequency showing in the muscles that respond to the specific contact substantial frequency dependent suppression at stimulation frequencies greater than 10 Hz. On the bottom, we report raw EMG traces that show the classic phenomenon. At 5 Hz each pulse of stimulation corresponds to a clear evoked potential in the EMG albeit amplitude slightly diminishes at each pulse. At 10 Hz, modulation of peak-to-peak amplitudes becomes more evident, at 20 Hz almost complete suppression of EMG evoked responses subsequent to the first is shown. Example is taken from Pronator muscles, contact 1 C, (highlighted in darker grey in the top panel). (a) Effect of stimulation frequency shown for SCS01 and SCS02. In SCS01, quantification of isometric torques during single joint flexion and extension is shown for the elbow during no stim (dark grey), 20 Hz (blue), 40 Hz (blue), and 60 Hz (blue). In SCS02, maximum reached distance and elbow angle excursion (max-min) are reported during reach and pull of the reach-out task for no stim (dark grey), 20 Hz (blue), 40 Hz (blue), and 60 Hz (blue). Raw endpoint trajectories for SCS02 are shown in the reach out task during no stim (dark grey), 20 Hz (blue), 40 Hz (blue), and 60 Hz (blue) where SCS02 was tasked to reach beyond the third horizontal line to complete the task. Reach (solid line) and pull (dashed line) trajectories are represented in separate plots. Darker lines represent average trajectories, shaded lines represent single trajectories. (b) Quantification of kinematic features for SCS01, path length for completed reach and pull of three targets in cm and variance of the path between trials are reported for no-stim (dark grey) and stim condition (blue). Center target could not be calculated for no-stim condition because SCS01 did not complete the task. (c) Quantification of kinematic features for SCS02, movement smoothness (velocity peaks) and path length in cm for reach and pull separately are reported for no-stim (dark grey) and stim condition (blue). The distribution of deviations from the mean path trajectory is shown in cm (equivalent to variance in SCS01). Statistics Distributions of deviations were compared using a two-sample Kolmogorov-Smirnov non-parametric test with alpha=0.05 where p~=0 (where the value was smaller than able to be stored in a double precision variable). All other quantifications are reported using box-plots. For each box, the central circle indicates the median while the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. The whiskers extend to the minima and maxima data points, not considering outliers. Any outliers are plotted individually with additional circles. Inference on mean differences is performed by bootstrapping the n = 5 repetitions obtained for each measurement, with n = 10,000 bootstrap samples, and by using a Bonferroni correction when performing multiple comparisons; * indicates statistical significance and rejection of the null hypothesis of no difference with a 95% confidence interval.

Nature Medicine
Article https://doi.org/10.1038/s41591-022-02202-6 Extended Data Fig. 6 | Optimized SCS leads to best improvement. (a) Quantification of isometric torques during single joint flexion and extension of the elbow during no stim (dark grey), non-optimal stim (light blue), and optimal stim (blue) for SCS01. (b) Quantification of performance for three targets of the center-out task during no stim (dark grey), non-optimal stim (light blue), and optimal stim (blue) normalized from 0 (SCS02 never reached target) and 1 (SCS02 reached target in all trials). n = 3 (c-e) Raw endpoint trajectories by SCS02 for three targets of the center-out task during no stim (dark grey), non-optimal stim (light blue), and optimal stim (blue). Darker lines represent average trajectories, shaded lines represent single trajectories. Statistics For quantifications reported using box-plots, the central circle indicates the median while the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. The whiskers extend to the minima and maxima data points, not considering outliers. Any outliers are plotted individually with additional circles. Inference on mean differences for (a) were performed by bootstrapping the n = 5 repetitions obtained for each measurement, with n = 10,000 bootstrap samples, and by using a Bonferroni correction when performing multiple comparisons;