Neuromotor prosthetic to treat stroke-related paresis: N-of-1 trial.

Background: Functional recovery of arm movement typically plateaus following a stroke, leaving chronic motor deficits. Brain-computer interfaces (BCI) may be a potential treatment for post-stroke deficits.
Methods: In this n-of-1 trial (NCT03913286), a person with chronic subcortical stroke with upper-limb motor impairment used a powered elbow-wrist-hand orthosis that opened and closed the affected hand using cortical activity, recorded from a percutaneous BCI comprised of four microelectrode arrays implanted in the ipsilesional precentral gyrus, based on decoding of spiking patterns and high frequency field potentials generated by imagined hand movements. The system was evaluated in a home setting for 12 weeks.
Results: Robust single unit activity, modulating with attempted or imagined movement, was present throughout the precentral gyrus. The participant acquired voluntary control over a hand-orthosis, achieving 10 points on the Action Research Arm Test using the BCI, compared to 0 without any device, and 5 using myoelectric control. Strength, spasticity, the Fugl-Meyer scores improved. Conclusions: We demonstrate in a human being that ensembles of individual neurons in the cortex overlying a chronic supratentorial, subcortical stroke remain active and engaged in motor representation and planning and can be used to electrically bypass the stroke and promote limb function. The participant's ability to rapidly acquire control over otherwise paralyzed hand opening, more than 18 months after a stroke, may justify development of a fully implanted movement restoration system to expand the utility of fully implantable BCI to a clinical population that numbers in the tens of millions worldwide.

Introduction

Stroke is a leading cause of disability[1] with a global prevalence of over 42 million people in 2015[2], affecting over four million adults in the United States alone with 800,000 new cases per year[3]. Stroke leads to permanent motor disabilities in 80% of cases[4], and half of stroke survivors require long term care. Brain computer interface (BCI) technologies offer a potential solution to restore functional independence and improve health in people living with its effects. In the past decade, intracortical BCI technology has continued to advance, with multiple groups demonstrating the safety and efficacy of this approach to derive control signals[5-7] to restore communication and control. In parallel, wearable robotic orthosis technology can benefit patients with weakened limbs[8,9]. This single-patient pilot clinical trial sought to prove that a commercially available powered arm orthosis could be linked to the cerebral cortex in an adult with the most common form of chronic stroke. A direct path from the brain’s motor centers to the orthotic could reanimate a paralyzed limb to enable useful hand and arm function.

Several signal sources have been coopted to provide commands to move paralyzed limbs. Electromyographic (EMG) control of a powered orthosis or functional electrical stimulation (FES) of muscles, has proven problematic either because users could not generate sufficient or reliable activity to provide a good control signal, or because voluntary activation of those recorded muscles (that were intended to generate the command) was opposed by the stimulator’s effects[10]. Contralaterally controlled electrical stimulation- where activity from the unaffected arm triggers stimulation on the paretic arm- is a useful therapeutic intervention to improve function in the weaker limb[11] but it is not clear how this unnatural command source could be generalized to continuously-worn devices that enable independent arm movements. Several groups have explored scalp EEG, which is closer to the command’s origins, to derive control signals to drive robotic braces, and in one case, FES[8,10,12,13]. While using EEG-derived signals may be promising for rehabilitation therapy, it would not be feasible for daily independent function because skin sweat and hair can cause impedances to fluctuate, compromising signal quality. Daily application of even a subset of contacts to the same skin sites can lead to skin breakdown and cellulitis. Further, EEG signals are limited in the commands that can be easily and reliably derived from the available signal. By contrast, intracortical interfaces offer a rich source of high resolution, multidimensional control signals, since it is the origin of such signals in healthy adults, in non-human primates and in people with spinal or brainstem disorders[7,14,15].

While most strokes involve cerebral white matter and direct parenchymal damage, intracortical neuromotor prosthetics have not been tested in people with strokes above the mesencephalon. It is not known whether motor cortex remains a reliable signal source in this large population. A proof-of-concept that a brain-computer interface, based on micro-electrode arrays implanted in intact cortex above a subcortical stroke, could restore behaviorally useful independent, voluntary movement, could lead to the development of a fully implantable medical device that, in principle, could reverse the motor deficits caused by stroke. The purpose of this study was to show whether an assistive brain-computer interface, when in use, could provide a behaviorally useful benefit in motor function.

Here we found that neural activity recorded from the cerebral cortex overlying a chronic, subcortical stroke, generated activity patterns related to both actual and imagined movements in the contralateral, weakened limb of a person. The recorded neural activity was decoded in real-time to control a motor on a brace worn on the weakened arm, and the participant was able to use this decoded signal to voluntarily open and close the hand to perform tasks in a home setting.

Methods

Approval for this study was granted by the US Food and Drug Administration (Investigational Device Exemption) and the Thomas Jefferson University Institutional Review Board (protocol number 17D.459). The participant described in this report has provided permission for photographs, videos and portions of his protected health information to be published for scientific and educational purposes. After completion of informed consent on March 4, 2020, medical and surgical screening procedures, two MultiPorts (Blackrock Microsystems, UT), each comprising two 8 × 8 platinum tipped microelectrode arrays tethered to a titanium pedestal connector, were implanted into the cortex of the precentral gyrus using a pneumatic insertion technique[16,17]. Details of the human surgical procedure are in preparation for publication and followed other similar studies. Trial selection criteria are available online (see Clinicaltrials.gov, NCT03913286 and Supplementary Table 1). The trial was designed with the implantation phase to last a maximum of three months (Fig. 1).

Fig. 1

The clinical trial timeline.

This figure shows the sequence of the Cortimo clinical trial.

Participant

The participant was a right-handed male who experienced right hemispheric stroke, manifest as acute onset dense left hemiparesis and expressive aphasia, at which time he was age between ages 35 and 40. Due to unknown time of onset and hypertension at presentation, the participant was not a candidate for thrombolysis. CT angiogram showed occlusion of the right posterior cerebral artery and high-grade stenosis of the left posterior cerebral artery in the proximal P2 segment. MRI of the brain showed acute infarcts in the right basal ganglia/corona radiata and right occipital lobe. He was started on dual antiplatelet therapy for 3 weeks and then was transitioned to aspirin 81 mg once daily, along with atorvastatin and anti-hypertensives. He had left-sided hemiparesis, dysphagia, left homonymous hemianopsia and dense left visual neglect and was transferred for inpatient rehabilitation. Over a period of three months, aphasia and dysphagia resolved and he learned to ambulate independently, albeit with a persistent left foot drop. Neuroimaging showed evidence of multi-focal strokes, and prior silent strokes. The participant had previously been in good health and did not have any known stroke risk factors such as diabetes or smoking. There was a history of loud snoring and the participant had not been evaluated for obstructive sleep apnea. Transthoracic and transesophageal echocardiography were normal as were serial hypercoagulability panels; the participant was adopted and the biological family history unknown. The participant was deemed to have had embolic strokes of unknown source. Although serial electrocardiography since the stroke was normal, the participant is being scheduled for a loop recorder to survey for possible paroxysmal atrial fibrillation. The participant had learning disabilities and was presumed to have had mild cognitive impairment prior to the stroke. Screening formal neuropsychological testing identified neurocognitive problems (full scale IQ 59) and concluded that the participant remained fully capable to provide proper informed consent and to participate in this trial, meeting its demands and requirements. The participant provided both verbal and written informed consent, both to participate in the trial and to share his identifying information with the public. He had been working full time at the time of the stroke and has been unable to return to work since the stroke. Clinical assessment of the participant included neurological exams one year, six months and one month prior to enrollment in the setting of routine outpatient care; the exam was serially repeated once enrolled. Clinical assessment included detailed history review and confirmation of meeting all selection criteria. The trial was constructed to include a six-week screening phase (Fig. 1), during which the participant underwent occupational therapy (1 h per session, three times per week) to assess how well the participant could understand and master use of the MyoPro device. The neuropsychological testing also took place during the screening phase. Predefined outcome measures (described subsequently) were also recorded during the screening phase. During this phase, therapy consisted of evaluation of active and passive range of motion, strength, hypertonicity, and goal setting with ADLs and IADLs. Treatment of left upper extremity spasticity was completed utilizing stretching, functional electrical stimulation, and creating a splinting schedule. This pre-implant occupational therapy was considered screening in that it was decided prior to enrolling the participant that proceeding to implant would only occur if the occupational therapist felt that the participant could adequately comply with the therapy. Therapeutic exercise and activity were incorporated to improve postural control and non-volitional movements with left upper extremity. Introductory use of MyoPro device was incorporated within treatment. Pre-implant physical therapy included baseline functional balance measures with interventions focused on open/closed chain strengthening, static/anticipatory/dynamic postural control and gait training.

Pre-operative fMRI

The participant underwent MRI on a 3 T Philips Ingenia MRI scanner. A 1 mm isotropic 3-D T2 FLAIR was obtained for structural localization. A single shot echoplanar gradient echo imaging sequence with 80 volumes, repetition time (TR) = 2 s, echo time (TE) = 25 msec, voxel size = 3 × 3 mm2, slice thickness = 3 mm, axial slices = 37. The participant was asked to visualize movements of his paretic left hand during the MRI. Each motor trial consisted of a block design featuring a 20 s rest block and a 20 s active block repeated. This block design was repeated between 4 times for a total of 240 s scans. Visual stimuli comprised a 20 s video depicting a 3D modeled limb at rest, followed by a 20 s video of the limb performing the desired task. Motor tasks included repeated hand open/clench or arm extension elbow and were either active (participant performed or attempted to perform motion) or passive (physician manually moved participant’s arm). In active tasks, the participant was instructed to follow the movements in the video or concentrate on following for the paretic limb. Task prioritization was based on pre-exam training of the participant’s capabilities and examination of BOLD activation observed during the scan. Post processing including motion correction, smoothing, and general linear model estimation performed using SPM software (www.fil.ion.ucl.ac.uk/spm) and Nordic brain EX software (NordicNeuroLab, Bergen, Norway). Statistical maps were overlaid on the 3D T2 FLAIR image for visualization of activation.

Cortimo system

‘Cortimo’ is the designation provided to the FDA to represent the overall system (Fig. 2) that comprised two percutaneous Multiports (Blackrock Microsystems), each in turn having two multi-electrode array sensors, the cabling, amplifiers, software and the powered MyoPro orthosis. Each sensor is an 8 × 8 array of silicon microelectrodes that protrude 1.5 mm from a 3.3 × 3.3-mm platform. At manufacture, electrodes had an impedance ranging between 70 KOhm and 340 KOhm. The arrays were implanted onto the surface of the MI arm/hand region guided by the pre-operative fMRI; with electrodes penetrating the cortex to attempt to record neurons in layer V. Recorded electrical signals pass externally through a Ti percutaneous connector secured to the skull. Cabling attached to the connector during recording sessions routes signals to external amplifiers and a computer that process the signals and convert them into different outputs, such as servo motor position of the MyoPro brace or screen position of a neural cursor. Currently, this system must be set up and managed by an experienced technician.

Fig. 2

The cortimo system.

The overall device comprises two Multiports, each with two 8 × 8 microelectrode arrays, patient cables linked to external amplifiers, a decoding computer and a wearable, powered arm orthosis.

MyoPro brace

The MyoPro (Myomo, Inc, Cambridge, MA) is an FDA-cleared myoelectric powered arm orthosis designed to support a paretic arm[9]. The rigid brace incorporates metal contacts attached to soft straps that can be adjusted such that contacts rest on the biceps and triceps proximally, and on wrist flexors and extensors distally, on the paretic upper extremity. The sensors continuously record the root mean square of underlying muscle activity. Thresholds are manually set such that signals exceeding them will trigger one of the MyoPro motors. Because the participant retained residual elbow flexion and extension strength, the motor at the elbow was set up such that biceps activation triggered elbow flexion and triceps activation triggered elbow extension. For hand opening, the MyoPro was set up to either use myoelectric control, or to use BCI-based control. Since the participant was unable to voluntarily extend the wrist or open the fingers, the myoelectric mode was set up such that the default state was with the hand open, and it would only be closed by activating enough wrist flexor activity.

MyoPro EMG control

The MyoPro device is designed to use four EMG channels to achieve control user’s control over two electric motors that control hand and elbow motion. The four EMG channels are fixed and they are: (1) forearm extensors; (2) forearm flexors; (3) triceps; and, (4) biceps. It is important noticing that the MyoPro does not use raw EMG signals for hand/elbow control mechanisms, but it calculates in real-time the rectified RMS of the raw EMG signals and the RMS signals are eventually compared against the user selected thresholds for motion classification.

Depending on the control mode selected, a combination of the above channels can be used to control either a single joint at a time or both. Namely, the available modes are dual mode, open mode or close mode. In dual mode, hand motion controlled by two manually set thresholds based on forearm EMGs. Elbow motion controlled by two different manually set thresholds based on biceps’ and triceps’ EMGs. The dual mode control principle is as described in the eq. below:MyoPro DUAL Mode Motion Control Strategy: In DUAL mode the MyoPro uses flexors and extensors for each of the two joints (hand and elbow) and two thresholds for motion control decision making.

In open mode, hand motion controlled by one manually set threshold based on forearm extensors. Elbow motion controlled by one different manually set threshold based on triceps EMGs. The open mode control principle is as described in the eq. below:MyoPro OPEN Mode Motion Control Strategy: In OPEN mode the MyoPro moves the joint into full flexion if the extensors are below the extension threshold while the joint is moved into full extension if extensors are above the extensor threshold.

In close mode, hand motion controlled by one manually set threshold based on forearm flexors. Elbow motion controlled by one different manually set threshold based on biceps. The close mode control principle is as described in the eq. below:MyoPro CLOSE Mode Motion Control Strategy: In CLOSE mode the MyoPro moves the joint into full extension if the flexors are below the flexion threshold while the joint is moved into full flexion if extensors are above the flexion threshold. where Flexors and Extensors are the corresponding EMG signals and ExtTh and FlexTh are respectively manually set thresholds for extensors and flexors.

Pre-specified outcome measures

More than a year prior to enrolling the participant, the trial was designed to assess specific outcome measures one month prior to device implantation, and again three months post-implantation. These metrics included the Fugl-Meyer Motor Impairment Score[18], the Action Research Arm Test (ARAT)[19], the Motricity Index[20], the Hand and Recovery Scales within the Stroke Impact Scale[21]. Although the Giving Them A Hand scale[22] was initially included as a metric on the original IDE and IRB filings, the occupational therapist co-investigators who subsequently joined the team felt that this was not a well-validated measure and it was decided, prior to enrolling the participant, that it would not be used. During the implant phase, a component of the Jebsen-Taylor measure (picking up, moving and putting down five objects, one at a time, one after the other)[23] was added because it was found that the participant was able to perform this task more consistently and easily with the orthosis than the ARAT, inspiring greater participant motivation and engagement and hence facilitating comparison of myoelectric vs BCI control modes. The outcome measures were performed by clinicians who were trained and standardized in administration, and each measure was assigned to specific co-investigators to perform serially to minimize inter-rater variability across time.

Recording sessions

Research sessions were scheduled five days per week at a temporary residence, adjacent to the hospital, provided to the participant. Sessions could be canceled or ended early at the participant’s request. Sessions would commence with neural recording and spike discrimination. While initial sessions included filter building and structured clinical endpoint (cursor control) trials, in the final month of the trial, training-less algorithms were used with the participant proceeding directly to BCI-controlled hand action once patient cables were connected. Performance of computer tasks, orthosis control and occupational therapy exercises followed. The electrodes and neural signals selected immediately before filter building remained constant for any given session’s orthosis control trials.

Decoder filter building

Units were extracted using an automatic thresholding approach based for each electrode channel, based on Root Mean Squared multipliers[15]. For each session, single and multiunit data or high frequency (100–1000 Hz) local field potentials derived from multiple channels (20–30) were used to create a linear filter to convert these real-time multidimensional neural features into either a one or a two-dimensional (position or velocity) output signal. Motor activity and motor imagery approaches were tested for filter building, including imagining opening and closing the paretic hand, passively flexing and extending the elbow, passively opening and closing the hand, and observing a computer cursor displayed on a monitor moving up and down without any specific instruction. Training data for building the linear filter were collected with the participant gazing at a screen where a target cursor was moved slowly up and down for one minute (5 s to go from the top to the bottom of the screen or vice versa, at 20° visual angle). After this preliminary filter was built (see next section for more details), a new 1-minute re-training session was performed, this time the manually controlled target cursor was accompanied by a prediction cursor that was neurally controlled by the participant. Using this additional training set, a second filter was built and then tested on a simple target acquisition game in which the y-position of the predicted output was discretized into zones such that positions on the upper part of the screen would cause an animation sprite to move up by a fixed distance (1 cm), and positions on the lower part of the screen would cause the sprite to move down by the same fixed distance.

Decoder design

The Cortimo system has been designed to provide a series of real-time decoding methods. Namely, two types of decoders have been implemented both discrete and continuous (i.e., filters). The available discrete classifiers were Linear Discriminant Analysis (LDA), Decision Trees, Support Vector Machine (SVM), k-nearest neighbors (KNN); while the available continuous decoders were: a position-velocity Kalman filter[24] and a linear filter[25]. All decoders could be quickly trained using data and labels recorded during training sessions (filter building sessions). The Cortimo participant and the BCI system achieved the best brace control performance when the linear filter was used and the training sessions were guided by a one or a two-dimensional cursor control task. Briefly, a cursor was displayed on the computer screen and the cursor position was controlled using a weighted linear combination of (ad-hoc) selected features derived from real-time participant’s brain activity. The neural features mostly used in this trial were either the cumulative (across all selected channels) spike count binned in 200 ms windows or the cumulative (across all selected channels) spectral density of Local Field Potentials (LFPs) calculated in the frequency range 100–500 Hz with 50 Hz frequency steps.

To achieve better and more reliable brace control in closed-loop tasks, the continuous (linear) filter output, originally corresponding to a specific screen location, was fed into a discretization block where the decision-making rules were either:The Cortimo Discretization Block: Discretization decision rule number 1. where PBM represents the hand motion at time step t, x is the linear filter output at time step t and AT is a position threshold chosen to maximize user’s control.The Cortimo Discretization Block: Discretization decision rule number 2. where PBM represents the hand motion at time step t, x is the linear filter output at time step t and AT and AT are respectively a lower and higher position thresholds chosen to maximize user’s control.

BCI orthosis use

The discrete output was then used to control the aperture of the hand via the MyoPro’s hand brace motor. The up-down mapping on the screen was translated into closed-open positions of the hand. The participant then performed a series of functional tasks including grasping and then dropping an object, the Action Research Arm Test[19], and a variation on Jebsen Taylor item moving test[23]. These were tested with both the participant seated and standing.

Training-less mapping

When the participant would attempt to overpower the orthosis motors with residual finger flexion strength, a ‘training-less’ approach was invented and deployed in which a rolling 1-second baseline of the LFPs signals was used to calculate spectral power in the high gamma band (100–500 Hz). Namely, 1-second long LFP continuous voltages were used for computing the average spectral density estimation in the frequency band 100–500 Hz, using non-overlapping frequency bins with a 50 Hz width. Spectral density was computed using the Matlab periodogram method. Values were updated every 500 ms, using 1-second-long rolling windows with 50% overlap. Real-time spectral features derived from the 20 most neuromodulated channels were averaged across channels to produce a single high gamma band value for each 500 ms software update. Orthosis hand-closure would be triggered by an increase in this mean spectral power from the resting baseline ranging between 0.5 and 3 V2/Hz to values greater than 10 V2/Hz, where real-time values above this threshold would make the hand motor close.

Concomitant occupational and physical therapy

Since being discharged from acute rehabilitation 60 days after the initial stroke, the participant enrolled in outpatient physical and occupational therapy. Prior to the device implantation, the participant completed a six-week course of occupational therapy screening phase. Following device implantation, the participant continued occupational therapy, twice per week, and physical therapy, once per week, each session lasting approximately one hour. In the three-month implantation phase, the participant hence received 24 one-hour sessions of occupational therapy and 12 one-hour sessions of physical therapy. In addition, clinical trial assistants practiced therapy exercises with the participant and accompanied him to a gym for aerobic conditioning (either stationary bicycling or NuStep combined arm and foot cycle, for 20 min to a target heart rate of 120 beats per minutes): these sessions were approximately one hour and were practiced daily, including weekends for a total of 91 days. Occupational therapy focused on postural training while seated and walking, donning, and doffing the MyoPro, repetitive trials of hand open/close elbow flex/extend wth the MyoPro, and using the MyoPro for functional activities. Timed functional electrical stimulation[26] (e.g., pincer grasp programs using the XCite brand FES unit from Restorative Therapies) and vibration therapy (5 to 10 min of focal muscle vibration) were used for spasticity management[27]. Physical therapy exercises included scapular mobilization, progressive range of motion, weight bearing, forced use with game-related activities to encourage left UE volitional control, and aerobic endurance exercise. The exact exercises performed (passive and active range of motion stretching, neuromuscular education, electrical stimulation, orthosis use), blood pressure, and subjective pain reports, were logged for every rehabilitation session; in addition, clips of several sessions were recorded by video (see Supplementary Occupational Therapy Log and Supplementary Videos 1–12 in particular Supplementary Video 12).

Table 1

Arm Research Action Test.

	12 days (1.5 weeks) post-implant	10 weeks post-implant)	12 weeks post implant
CONDITION	NO BRACE	EMG-MYOPRO	BCI-MYOPRO
Block 10 cm3	0	0	0
Block 2.5 cm3	0	1	1
Block 5 cm3	0	0	1
Block 7.5 cm3	0	0	0
Cricket ball	0	0	0
Sharpening stone	0	1	1
Pour water from one glass to another	0	0	0
Displace 2.25-cm alloy tube from one side of table to the other	0	1	1
Displace 1-cm allow tube from one side of table to other	0	1	1
Put washer over bolt	0	1	1
Ball bearing held between ring finger and thumb	0	0	0
Marble held between index finger and thumb	0	0	1
Ball bearing held between middle finger and thumb	0	0	0
Ball bearing held between index finger and thumb	0	0	0
Marble held between ring finger and thumb	0	0	0
Marble held between middle finger and thumb	0	0	0
Hand to behind the head	0	0	0
Hand to top of head	0	0	0
Hand to mouth	0	0	3
SUM	0	5	10

The participant was unable to perform any aspect of the test with the unassisted, paralyzed arm. Conditions performed with the assistance of the MyoPro powered orthosis, whether under myoelectric or brain-computer interface control, are shown in non-italicized boldface. Conditions performed with the assistance of the MyoPro powered orthosis, only under brain-computer interface control, are shown in italicized boldface.

Table 2

Object Release Times.

BCI Control	3 4 1 1 2 6 1 8 5 1 7 18 13 7 26 24 2 3 2 3 1 2 5 2 9 4 2 5 9 14 7 4 12
EMG Control	45 8 13 24 5 1 7 19 3 4

Hand release times measured in seconds over multiple trials spanning multiple days of this identical task were faster under BCI control than myoelectric control (p = 0.04, two-sample t test).

Table 3

Motricity Index, Fugl-Meyer, Stroke Impact Scale.

Motricity Index (without orthosis assistance)	One month pre-implant	Two months post-implant	Three months post-implant
	48	61.5	75.5
Fugl-Meyer Upper Extremity Score (without orthosis assistance)	One month pre-implant	One month post-implant	7 weeks post-implant
	30	36	38
Stroke Impact Scale (without orthosis assistance)	One month pre-implant		Six months post-implant (three months post-explant)
	232		269

Table 4

Manual Muscle Testing.

Manual Muscle Testing	15 months pre-implant	8 months pre-implant	Day of implant (pre-op)	7 weeks post-implant	10 weeks post-implant	11 weeks post-implant	11 weeks, 3 days post-implant	Day of explant (pre-op) 3 months post-implant	3 days post-explant
Finger flexion			0	0		3	3	4	3
Finger extension			0	0		0		0	0
Wrist extension	0	0	0	0	3	3	3	1	2
Wrist flexion	3	1	0	0	5	3	3	3	2
Elbow extension triceps	2	0	4	4	5	5	4	2	4
Elbow flexion biceps	1	1	5	5	5	5	3	4	3
Shoulder abduction			5	4	4		4
Shoulder adduction			5		5		5