Restoring Movement in Paralysis witha Bioelectronic Neural Bypass Approach:Current State and Future Directions

Chad E. Bouton

Feinstein Institute for Medical Research, Northwell Health, Manhasset, New York 11030

Correspondence: cbouton@northwell.edu

Bioelectronic medicine is a rapidly growing field that explores targeted neuromodulation innew treatment options addressing both disease and injury. New bioelectronic methods arebeing developed to monitor and modulate neural activity directly. The therapeutic benefit ofthese approaches has been validated in recent clinical studies in various conditions, includ-ing paralysis. By using decoding and modulation strategies together, it is possible to restorelost function to those living with paralysis and other debilitating conditions by interpretingand rerouting signals around the affected portion of the nervous system. This, in effect, createsa bioelectronic “neural bypass” to serve the function of the damaged/degenerated network.By learning the language of neurons and using neural interface technology to tap into criticalnetworks, new approaches to repairing or restoring function in areas impacted by disease orinjury may become a reality.

Traumatic injuries, especially those involvingthe brain and spinal cord, can be particularly

difficult to treat. Researchers have attempted todevelop pharmaceutical and stem cell approach-es to promote healing in spinal cord injuries, buthave been met with limited success. For injuriesthat have led to motor impairment, assistive andrehabilitative technologies using neuromuscularstimulation and robotics have been developed.Robot-assisted therapy has been shown to beeffective in stroke patient recovery, particularlyfor upper-limb rehabilitation (Volpe et al. 2008;Lo et al. 2010). In the case of traumatic braininjury, closely monitoring the patient multipleparameters in the brain is crucial. To allow this,multimodal sensing technologies have been re-

cently developed for sensing key parameters inthe brain simultaneously, providing valuable in-formation to the clinical team during recovery(Li et al. 2009). Also, stimulation strategies suchas trigeminal nerve stimulation have been stud-ied with the goal of improving traumatic braininjury outcomes (Chiluwal et al. 2017).

In cases in which disease or injury has led tocomplete paralysis, such as in spinal cord injury(SCI), stroke, and motor neuron disease, brain–computer interface (BCI) technology has beenof particular interest. BCI technology allows sig-nals in the brain to be recorded and decoded todetermine the user’s thoughts, which can allow aparalyzed or locked-in patient to communicate,control devices, and, when combined with neu-

Editors: Valentin A. Pavlov and Kevin J. TraceyAdditional Perspectives on Bioelectronic Medicine available at www.perspectivesinmedicine.org

Copyright © 2019 Cold Spring Harbor Laboratory Press; all rights reservedAdvanced Online Article. Cite this article as Cold Spring Harb Perspect Med doi: 10.1101/cshperspect.a034306

1

ww

w.p

ersp

ecti

vesi

nm

edic

ine.

org

on April 1, 2020 - Published by Cold Spring Harbor Laboratory Press http://perspectivesinmedicine.cshlp.org/Downloaded from

romuscular stimulation technology, regain voli-tional movement. To create a BCI system, sev-eral technologies are required, including brainelectrodes, machine learning algorithms to de-code intentions, and a device to be controlled(e.g., a computer, robotic arm, neuromuscularstimulators). To decode brain activity, it is crit-ical to understand how information is encodedin the brain so that electrical activity can bemonitored and decoded to determine a patient’sintent (e.g., to communicate, move, etc.). Earlystudies were conducted in the 1960s to deter-mine how sensorimotor information is encoded,which paved the way for BCIs (Evarts 1968; Per-kel and Bullock 1968). Later, neural modulationpatterns were studied during active and passivemovements (Fetz et al. 1980) and force-relatedneurons were identified (Cheney and Fetz 1980).As electrode technology developed, populationsof neurons could be studied, and it was dis-covered that neurons in the motor cortex havea “preferred” direction, known as “directionaltuning” (Georgopoulos et al. 1986). As comput-er technology further developed, our under-standing of modulation patterns developed(Warland et al. 1997; Donoghue et al. 1998).All of these important studies laid the ground-work for subsequent advancements in develop-ing BCI technology for debilitating conditions.

One of the critical components of bioelec-tronic medicine that uses BCI devices is the neu-ral interface. This interface provides a way tosense, record, and modulate neural signals innerves and in the brain. Recent advances havebeen made in flexible nerve cuff electrodes al-lowing acute and chronic recordings (Caravacaet al. 2017). Multielectrode arrays for the brainhave been under development for many yearswith researchers facing many challenges. Initial-ly, electrodes were made by hand in the labora-tory, which limited the number of recordingsites each had. Microfabrication techniqueswere later developed and progress accelerated.Researchers at University at Utah began to useetching techniques to form electrode arraysfrom a solid piece of silicon. Precise electrode“spikes” could be formed with uniform spac-ing. Encapsulation methods for this style ofelectrode arrays were also refined for chronic

applications in neural disorders (Hsu et al.2009). Researchers at the University of Michi-gan also developed brain and nerve electrodesthrough a thin-film process to make flexibleelectrodes, and chronic brain recordings havebeen achieved (Vetter et al. 2004). Electrocorti-cography (ECoG) electrodes can also be used forrecordings on the surface of the brain. Theseelectrodes typically are not intended for captur-ing single-unit action potentials and the spatialresolution is lower than the Utah and Michiganstyle electrodes.

Once neural signals are accessed through aneural interface, they must be processed anddeciphered to extract useful information. Inthe late 20th century, a branch of artificial intel-ligence, known as machine learning, began toaccelerate (Shavlik and Dietterich 1990; Michieet al. 1994). Computers were becoming morepowerful and the interest in studying signals re-corded in the brain was growing. The rate atwhich neurons “fire,” conveys useful informa-tion that can be used to determine the user’sintentions such as moving a cursor on a com-puter screen in a particular direction. This newarea of neural decoding was born and manyresearchers were trying to achieve what initiallyseemed impossible—read someone’s mind.

Researchers studied brain activity in pri-mates and linked their thoughts to roboticarms (Chapin et al. 1999; Wessberg et al. 2000;Serruya et al. 2002; Taylor et al. 2002; Carmenaet al. 2003; Lebedev et al. 2005). It was shownthat these activity patterns were related tomove-ment in three-dimensional space (Taylor et al.2002; Velliste et al. 2008). Taking this a stepfurther, primates were able to regain volitionalcontrol over their own (temporarily paralyzed)limbs (Moritz et al. 2008; Ethier et al. 2012).Clinical studies also were launched to determinewhether these technologies and methods couldbe used to restore speech (Kennedy and Bakay1998) or be used to assist patients living withparalysis caused by SCI, stroke, and amyotro-phic lateral sclerosis (ALS) interface with theircomputers, directly with their minds (Hochberget al. 2006; Bouton 2009). Later, control of ro-botic arms was then shown in a clinical study(Hochberg et al. 2012; Collinger et al. 2013).

C.E. Bouton

2 Advanced Online Article. Cite this article as Cold Spring Harb Perspect Med doi: 10.1101/cshperspect.a034306

ww

w.p

ersp

ecti

vesi

nm

edic

ine.

org

on April 1, 2020 - Published by Cold Spring Harbor Laboratory Press http://perspectivesinmedicine.cshlp.org/Downloaded from

Finally, functional movement in a paralyzed hu-man was recently shown through the use of aneural interface implanted in the brain and neu-romuscular stimulation (Bouton et al. 2016).

As the field of bioelectronic medicine con-tinues to grow rapidly, multidisciplinary teamswork closely to generate preclinical mechanisticinsight, define new options in disease diagnosisand treatment, and validate their clinical rele-vance in the context of various disorders includ-ing paralysis. Modulating neural pathways, suchas the vagus nerve–based inflammatory reflex,has a positive effect in chronic inflammatoryconditions including rheumatoid arthritis andCrohn’s disease (see Tracey 2002, 2015; Pavlovand Tracey 2015). Furthermore, listening in onneural pathways and decoding the acquired sig-nals can give us insight into changes in inflam-matory proteins and other important substancesin our bodies, potentially leading to earlier andbetter diagnostics one day (see Bouton 2017;Olofsson and Tracey 2017; Zanos et al. 2018).We need to deepen our understanding of under-lying molecular mechanisms, develop robustneural interfaces, and elucidate complex neuralcircuitry so we can better diagnose and treatconditions under our sites. As we continue tounderstand the human nervous system and thelanguage it uses, we will be better able to detectearly warnings of disease and to address a widearray of conditions through neural decoding andmodulation. Many open questions and chal-lenges remain in the field of bioelectronic med-icine. However, recent advances have shownwhat is possible for patients of the future.

RESTORING MOVEMENT IN A HUMAN

Decades of research in understanding howmovement information is encoded in the brain,development of chronic neural interfaces, andthe development of neural decoding methodshas laid the foundation for new advances andthe development of bioelectronic neural bypasstechnology. Until 2014, restoration of corticalcontrol of movement in a paralyzed personhad not been achieved. In that year, a clinicalstudy was launched with that goal in mind. A24-year-old young man, whowas left quadriple-

gic after a diving accident, was enrolled in thestudy.Within the first year, he was able to regainfunctional use of his right hand through the useof an artificial bypass, which combined a corticalneural interface, neural decoding algorithms,and transcutaneous neuromuscular stimulation(Bouton et al. 2016). The neural implant (4 × 4mm across with 96 electrodes, each 1.5 mmlong) was placed in his left primary motor cor-tex and was linked through a computer to aneuromuscular stimulator with >100 electrodesplaced on his forearm (see Fig. 1). This approachallowed the study participant to regain volitionalcontrol of hand grasping and individual fingermovements.

Another group of researchers showed laterthat a BCI, with assistance from a motorizedarm, could allow volitional movement in a C4participant (Ajiboye et al. 2017).

As shown in Figure 2, the participant wasable to achieve functional movements such aspicking up a cup using a cylindrical grasp andvolitionally switching to a pinch grasp. Thisallowed him to pick up smaller objects, such asthe stir stick shown in Figure 2, with finer dex-terity.

DECODING RHYTHMIC MOVEMENTS

Rhythmic movements, such as scratching,brushing your teeth, or playing amusical instru-ment, are all activities related to a person’s in-dependence and quality of life. Neural circuits inthe spinal cord and brain stem facilitate theserhythmic movements (Brown 1914; Marderand Bucher 2001). However, when they aredamaged through injury or stroke, restoring vo-litional rhythmic movement can be particularlychallenging or impossible. With a neural bypassapproach, signals are rerouted around these im-portant circuits, and, therefore, lost functionmust be replaced. To do this, neural circuitswithin the spinal cord must be emulated. Thesecircuits are called central pattern generators(CPGs), which producemotor patterns for rhyth-mic movement (Marder and Bucher 2001). Re-cently, artificial CPGs have been created in soft-ware and linked to decoded brain activity sensedin primary motor cortex of a paralyzed human

Restoring Movement with Bioelectronic Neural Bypass

Advanced Online Article. Cite this article as Cold Spring Harb Perspect Med doi: 10.1101/cshperspect.a034306 3

ww

w.p

ersp

ecti

vesi

nm

edic

ine.

org

on April 1, 2020 - Published by Cold Spring Harbor Laboratory Press http://perspectivesinmedicine.cshlp.org/Downloaded from

(Sharma et al. 2016). In this work, CPGbehaviorwas emulated by using a numerical model of anoscillatory neural circuit. This circuit is shown inFigure 3.

In primates, corticomotoneurons in theprimary motor cortex monosynaptically project

to spinal motor neurons allowing more directmuscle control. This type of connection is read-ily observed in the hand and arm areas of themotor cortex, and is believed to allow highlyskilled movements to be learned (Rathelot andStrick 2009). In the study involving restoration

Figure 2. Functional movements achieved by a paralyzed study participant using an electronic neural bypasslinking decoded brain activity to muscle activation in real time. (From Bouton et al. 2016; reprinted, withpermission.)

Time (sec)

Tria

l

0

5

10

15

20 −2

0

2

Mea

n w

avel

et p

ower

−1

0

1

E

0 1 2 3 4

Firi

ng r

ate

(spi

kes/

sec)

0

80604020

100

Tria

l

WF

5 101520

WE WR WUD

Time (msec)

µV

Channel 37, unit 1300200100

0100200300400

0 0.5 1 1.5

CB

Overlap

Array

A

Arm movement

Figure 1. Brain–computer interface (BCI) system for movement restoration in a paralyzed human study partic-ipant. (A) Cortical implant location, (B) muscle stimulation sleeve, (C) experimental setup, and (D,E) neuralactivity for imagined/attempted wrist movements (extension, flexion, and radial/ulnar deviations). (From Bou-ton et al. 2016; reprinted, with permission.)

C.E. Bouton

4 Advanced Online Article. Cite this article as Cold Spring Harb Perspect Med doi: 10.1101/cshperspect.a034306

ww

w.p

ersp

ecti

vesi

nm

edic

ine.

org

on April 1, 2020 - Published by Cold Spring Harbor Laboratory Press http://perspectivesinmedicine.cshlp.org/Downloaded from

of rhythmic movement in a human, it was hy-pothesized that two uniquemodulation patternswould be observed when the participant imag-ined a static versus rhythmic joint movement. Itwas found that when the participant visualizedthese different types of movements of the samejoint, differing neuronal patterns were indeedobserved. These patterns could also be accurate-ly decoded, which allowed the participant to ini-tiate and switch volitionally between static andrhythmic movements.

CORTICAL CONTROLOF MUSCLECONTRACTION LEVEL

Achieving variable force for various movementsis also important in our daily lives to do taskssuch as holding a cup of coffee or picking up anegg. Until recently, restoring cortical control ofsuch movements in a paralyzed person had not

been achieved. In a clinical study with a quadri-plegic participant previously described in Bou-ton et al. (2016), a new experimental paradigmwas created to explore the possibility of restoringgraded muscle contractions (Friedenberg et al.2017). In this study, the participant was initiallyasked to imagine flexing their wrist (against aload) to different angles. The participant was notable to flex their wrist without use of the artificialneural bypass under study because of their in-jury. Neural activity in primary motor cortexwas first recorded and decoded to allow the par-ticipant to control a virtual pointer on the com-puter screen. Beta regression methods were firstused to decode the neural activity and identifymodulation patterns of interest on specific elec-trode channels (1–96). As shown in Figure 4,some channels show a linear increase in activi-ty as measured through mean wavelet power(MWP) when target angle increases. Further-

Extensorneuron

1

0.5

0

Osc

illat

or o

utpu

t

–0.5

–10 0.5 1 1.5

Time (sec)

2 2.5 3

Input from Output to+μ12y2 μ21y1

βv2

βv1

– control NMES, youtneural decoder, c

Flexorneuron

Figure 3. Two neuron oscillator models and resulting output. (From Sharma et al. 2016; reprinted, with per-mission, from Scientific Reports in conjunctionwith a Creative CommonsAttribution 4.0 International License.)

Restoring Movement with Bioelectronic Neural Bypass

Advanced Online Article. Cite this article as Cold Spring Harb Perspect Med doi: 10.1101/cshperspect.a034306 5

ww

w.p

ersp

ecti

vesi

nm

edic

ine.

org

on April 1, 2020 - Published by Cold Spring Harbor Laboratory Press http://perspectivesinmedicine.cshlp.org/Downloaded from

more, some channels show nonlinear behavioras the angle changes.

Because of the nonlinear relationship be-tween neural activity and wrist angle, a supportvector regression (SVR) method was chosen foruse during reanimation of the wrist (throughneuromuscular stimulation) in the physical task.

In this task, a rubber band was used to provideresistance as the participant attempted to flex hiswrist to various target angles. The SVR algo-rithm also used a regularization parameter to“weed out” channels that were not contributingeffectively to decoding. In a series of real-timetasks using the SVR algorithm, and after recal-

45

0

90

Arm-supportingframe

Force

Lb1.32

135

Time

F

Force decoding

Angle measurement

Channel 96

Channel 1Channel 1

100 msec raw voltage Mean wavelet power Decoding algorithms

Restangle

30

6090

120

150

180

Maxangle

Decodedangle

A

C

B

Targetangle

0

100 msectime window

Dec

oded

forc

e

I ∝ f(F)

180

Stimulation-activated

movement0

Arm-supportingframe

TimeI ∝ f(F)

180

Stimulation-activated

movementStimulator

Figure 4. Decoding graded muscle contraction from intracortical activity. (A) Example screen used duringexperiment showing a target angle and acceptable window (green) and decoded angle (thick black line) updatedin real time. (B) Photograph of the setup. (C) Flow chart where intracortically recorded voltage data are convertedtomeanwavelet power (MWP) features, then decoded to produce force, F, and translated into a set of stimulationparameters, I, which results in wrist flexion movement recorded by the load cell and an overhead camera. (FromFriedenberg et al. 2017; reprinted, with permission, from Scientific Reports in conjunction with a CreativeCommons Attribution 4.0 International License.)

C.E. Bouton

6 Advanced Online Article. Cite this article as Cold Spring Harb Perspect Med doi: 10.1101/cshperspect.a034306

ww

w.p

ersp

ecti

vesi

nm

edic

ine.

org

on April 1, 2020 - Published by Cold Spring Harbor Laboratory Press http://perspectivesinmedicine.cshlp.org/Downloaded from

ibrating the system to account for muscle fa-tigue, the participant was able to achieve anoverall accuracy of 87.5%± 6.8%.

Finally, in the study involving restoration ofgraded muscle contraction, it was investigatedwhether the participant could flex his wrist toangles that were not previously encounteredduring training. The rest (0°) andmaximum an-gle were not changed, but the three target anglesbetween were changed to new angles. The par-ticipant was successful in 14 of the 18 angle cuesgiven and an overall accuracy of 89.6%± 4.4%was achieved.

CHRONIC NEURAL RECORDINGS

One of the outstanding challenges in the fieldof bioelectronic medicine is how to achievechronic (long-term) neural recordings. Specifi-cally, brain interfaces can be particularly chal-lenging owing to the glial response of the brain(Williams et al. 1999). Single-unit activity (SUA)results from axonal output close to a recordingelectrode (Rousche andNormann 1998) and candegrade or disappear only a few months after

implanting an electrode in the brain (Sharmaet al. 2015) as shown in Figure 5. Multiunit ac-tivity stems from combined neuronal outputfrom multiple neurons within several hundredmicrons of an electrode and has been shown tobe more robust over time (Stark and Abeles2007; Scheid et al. 2013; Sharma et al. 2015).Finally, local field potential (LFP) represents alarge number of neurons firing together and typ-ically is considered to be in the range of <100 Hzand also robust over time (Scherberger et al.2005; Sharma et al. 2015).

To combat degrading signals in chronicbrain recordings, two strategies have been used.On the hardware side, thinner, more flexibleelectrodes made of carbon have been exploredand showed a reduced inflammatory response(Guitchounts et al. 2013). A second strategy in-volves using signal processing methods thatextract MUA features for decoding. These fea-tures have been shown to be more robust overtime and still contain useful information fordecoding purposes (Bouton et al. 2016). It is,however, extremely important to select the ap-propriate frequency range so that neighboring

Channel: 65

1.5

MW

P

Angle

1.00.50.0

–0.5

1.51.00.50.0

–0.5

1.51.00.50.0

–0.5

0 50 100 150 0 50 100 150 0 50 100 150 0 50 100 150

Channel: 24 Channel: 3 Channel: 62

Channel: 67 Channel: 41 Channel: 34 Channel: 31

Channel: 15 Channel: 45 Channel: 82 Channel: 10

P-value65

A B

343

10

15

62

31

24

67

41 45 82

Pattern

<=0.01

0.01 to 0.05>0.05

IncreasingDecreasing

Nonlinear

Not significant

Figure 5.Neural modulation patterns with imagined gradedmuscle contractions. (A) The P-values from the betaregression are overlaid over the physical layout of the microelectrode array (MEA). Selected individual channels(highlighted in A) shows a range of neural modulation trends in response to the cued angles. (B) Each columnshows selected individual channels in which the average mean wavelet power (MWP) (column 1) increased withincreasing cue angles (green), (2) decreased with increasing nonzero cue angles (purple), (3) showed a nonlinearrelationship with cue angle (blue), and (4) did not show significant response to the cue angles (pink). (FromFriedenberg et al. 2017; reprinted, with permission, from Scientific Reports in conjunction with a CreativeCommons Attribution 4.0 International License.)

Restoring Movement with Bioelectronic Neural Bypass

Advanced Online Article. Cite this article as Cold Spring Harb Perspect Med doi: 10.1101/cshperspect.a034306 7

ww

w.p

ersp

ecti

vesi

nm

edic

ine.

org

on April 1, 2020 - Published by Cold Spring Harbor Laboratory Press http://perspectivesinmedicine.cshlp.org/Downloaded from

electrodes do not contain redundant informa-tion. Because LFP, for example, results from alarge number of neurons (within a larger radiusaround the electrode), it could be expected thatthe similarity of LFP signals for electrodes inclose proximity to each other would be highbecause their large “listening radii” would over-lap considerably. In fact, one might expect ageneral relationship (with exceptions owing toneural connections between more distant neu-rons) between correlation, signal frequency, anddistance. This relationship has been studied(Sharma et al. 2015) through coherence analy-sis. It is clear that, on average, LFP signals (lowfrequencies) are highly correlated for neighbor-ing electrodes, and multiunit activity (mediumfrequencies) and SUA (high frequencies) aremuch less correlated and can provide better(more independent) signals for decoding pur-poses for closely spaced electrodes. However,because the SUA signals have been observed todegrade over time,multiunit signals have provento be effective and more robust for chronic de-coding.

Another feature, called midfrequency MWP(mf-MWP) spanning 234 Hz to 3750 Hz, hasbeen studied in a human participant over ex-tended time periods (Zhang et al. 2018). Thisfeature was shown to be stable over extendedperiods of time and to correlate with MUA (de-fined as 300–6000 Hz in this study). Further-more, the overall decoding accuracy for mf-MWP features in two different tasks (involvingimagined hand grasping and wrist rotationmovements) was found to be 91.05%± 2.79%and 83.97%± 3.99%, respectively.

FEATURE SELECTION

A critical step in the process of decoding neuralsignals, and in any machine learning endeavor,is feature selection and extraction. Useful fea-tures in the input data to a decoding algorithmthat contain useful information must be identi-fied. An example of a feature in neural data isfiring/spiking rate (if SUA is being used). Withany neural signal, the power of that signal invarious frequency subbands would be another

example of useful features. Various signal pro-cessing methods can be used to extract thesefeatures from data sets.

One signal processing method that can beused is bandpass filtering. This type of filter canbe used to suppress unwanted components ofthe signal outside the frequency band of interest.After filtering is applied, a root-mean-squareoperation can be used to estimate the powercontained in the filtered signal. Variation inthis power may be useful as a feature because itmay change during various decoding tasks (e.g.,imagined hand movements in a motor decod-ing task). Furthermore, a sliding window can beused to further filter the power estimate to re-duce variability. This type of filter is formed bymultiplying the N most recent data points intime by either a uniform set of coefficients (e.g.,boxcar/moving average) orotherwindow shapes(e.g., exponential filter).

Fast Fourier transform (FFT) filters can alsobe used to estimate power levels in frequencybands of interest. This approach can have a re-duced temporal resolution, which may not beappropriate for transient events. In this case,wavelet techniques can be used to achieve suffi-cient resolution in both time and frequency do-mains (Graps 1995). When performing wavelettransformations, a “scale” represents a frequen-cy range. Various scales can be used to representSUA, MUA, and LFP activity. Wavelet coeffi-cients can then be used as features for real-time decoding purposes (Bouton et al. 2016).

NEURAL DECODING ALGORITHMS

Once a set of features have been created, variousmachine learning methods can be used to con-struct a neural decoding algorithm. They cantake on the form of a classifier that recognizes(classifies) neural data into known, discreteclasses, which could be physiological statessuch as hypoglycemic/hyperglycemic or imag-ined movements someone believes about open-ing or closing their hand. A decoding algorithmcan also perform regression in which a contin-uous output (prediction) is produced instead ofa discrete class.

C.E. Bouton

8 Advanced Online Article. Cite this article as Cold Spring Harb Perspect Med doi: 10.1101/cshperspect.a034306

ww

w.p

ersp

ecti

vesi

nm

edic

ine.

org

on April 1, 2020 - Published by Cold Spring Harbor Laboratory Press http://perspectivesinmedicine.cshlp.org/Downloaded from

Before a decoding algorithm can makeany predictions, it must be trained. One canuse supervised or unsupervised training meth-ods. Supervised training involves the use of as-sociation to train the algorithm, the same way apersonwould learn a new language—we listen tothe sounds for a word/phrase as someone iden-tifies the associated object/idea. In supervisedneural decoder training, neural data is collectedsimultaneously with label information. For ex-ample, in decoding imagined movements, thelabel information could be series of codes repre-senting what the person is supposed to be imag-ining at each point in time. One challenge inusing supervised training is that the label in-formation/data must be synchronized with theneural data, accounting for reaction time, and itcan be hard to know whether a person actuallyimagined what they were supposed to (they mayhave been distracted at various points duringthe training). To address this, unsupervisedtraining can also be used (Jain et al. 1999).With this approach, algorithms automaticallyidentify and group/cluster different neural pat-terns; however, each group/clustermust be iden-tified at some point through an operator or sen-sor data.

Different types of machine learning meth-ods can be used to build a classifier to regressionmachine. Support vectormachines (SVMs) havebeen used for many years, and have becomefaster and more powerful as the computationalpower of devices have increased. Specific meth-ods such as L1-SVM can be used to effectively“prune” a feature set so that only the most usefulfeatures are passed on to the decoding algorithm(Humber et al. 2010). Nonlinear/kernel meth-ods can also be used to map neural data into ahigh dimensional space (Scholkopf et al. 1997).Thesemethods can often improve decoding per-formance (Muller et al. 2003).

REMAINING CHALLENGES AND FUTUREDIRECTIONS

There has been tremendous progress in the fieldof bioelectronic medicine in recent years. How-ever, there are critical open questions and tech-nical challenges that must be addressed before

the next generation of clinical devices can bedeveloped. How information is encoded in thebrain and the peripheral nervous system is stilllargely unknown. We have learned how certaintypes of arm and hand movements are encodedin the brain and have even learned how to de-code certain movements with a respectable levelof accuracy. However, there are still many ques-tions surrounding how information is encodedand transmitted on afferent fibers in the periph-eral nervous system. It has been shown that spe-cific neural signaling occurs after exposure tointerleukin (IL)-1β and tumor necrosis factor(TNF) (Zanos et al. 2018), but additional cyto-kines and resulting neural signals remain tobe studied. This vast amount of informationcould be used for closed-loop devices, allow-ing automatic adjustments to stimulation andadaptation to changing conditions over time.These responsive, smart devices of the futurecould also lead to superior diagnostic and treat-ment options for patients, and, in some cases,completely replacing the need for drugs.

At the heart of a bioelectronic medicinedevice is the neural interface. This interfaceprovides the critical functions of stimulatingand/or recording neural signals. As previouslydiscussed, the spatial density/spacing requiredfor next-generation electrodes depends on theinformation being recorded/sensed. Currently,electrode spacing is actually quite suitable forcertain decoding applications in the brain; how-ever, the total number of electrodes and theamount of surface area being covered is lacking.For example, in motor neuroprosthetic appli-cations for quadriplegic users, penetrating neu-ral interfaces only cover a small portion of thehand/arm area, providing only a fraction ofthe information available. Increasing the totalnumber of electrodes would allow improveddecoding accuracy and/or an increased numberof movements. Surface electrodes are being im-proved and are increasing in size; however, thereis a limit to how small a single recording area canbe while providing quality signals. Special coat-ings can combat this and further development ofnew materials and electrode designs will helpadvance efforts toward next generation neuralinterfaces.

Restoring Movement with Bioelectronic Neural Bypass

Advanced Online Article. Cite this article as Cold Spring Harb Perspect Med doi: 10.1101/cshperspect.a034306 9

ww

w.p

ersp

ecti

vesi

nm

edic

ine.

org

on April 1, 2020 - Published by Cold Spring Harbor Laboratory Press http://perspectivesinmedicine.cshlp.org/Downloaded from

As our knowledge in the field of bioelec-tronic medicine increases and new technologiesare conceived and developed, a significant ex-pansion of addressable conditions will occur.New bioelectronic devices will become morereadily available and easier to apply. Noninva-sive and smaller devices that are easier to im-plant, will also become available. These newversions will become more efficacious, adaptive,and robust over chronic period. The combina-tion of molecular biology, neuroscience, and en-gineering—as a core foundation of bioelectronicmedicine will continue to pave the way to a newfuture with many new treatment options forpatients. This new approach will allow doctorsto target a long list of diseases and conditionsthat currently have no cures, helping the mil-lions battling paralysis, motor neuron disease,Alzheimer’s, Parkinson’s, traumatic brain inju-ry, and many other debilitating conditionsworldwide.

REFERENCES

Ajiboye AB, Willett FR, Young DR, Memberg WD, MurphyBA, Miller JP, Walter BL, Sweet JA, Hoyen HA, KeithMW, et al. 2017. Restoration of reaching and graspingmovements through brain-controlled muscle stimulationin a person with tetraplegia: A proof-of-concept demon-stration. Lancet 389: 1821–1830. doi:10.1016/S0140-6736(17)30601-3

Bouton C. 2009. Decoding neural activity from an intracor-tical implant in humans with tetraplegia. In BiomedicalScience & Engineering Conference, 2009. First AnnualORNL. Oak Ridge, TN.

Bouton C. 2017. Cracking the neural code, treating paralysisand the future of bioelectronic medicine. J Int Med 282:37–45. doi:10.1111/joim.12610

Bouton CE, Shaikhouni A, Annetta NV, Bockbrader MA,Friedenberg DA, Nielson DM, Sharma G, Sederberg PB,Glenn BC, Mysiw WJ, et al. 2016. Restoring cortical con-trol of functional movement in a human with quadriple-gia. Nature 533: 247–250. doi:10.1038/nature17435

Brown TG. 1914. On the nature of the fundamental activityof the nervous centres; Together with an analysis of theconditioning of rhythmic activity in progression, and atheory of the evolution of function in the nervous system.J Physiol 48: 18–46. doi:10.1113/jphysiol.1914.sp001646

Caravaca AS, Tsaava T, Goldman L, SilvermanH, Riggott G,Chavan SS, BoutonC, TraceyKJ, DesimoneR, BoydenES,et al. 2017. A novel flexible cuff-like microelectrode fordual purpose, acute and chronic electrical interfacingwiththe mouse cervical vagus nerve. J Neural Eng 14: 066005.doi:10.1088/1741-2552/aa7a42

Carmena JM, LebedevMA, Crist RE,O’Doherty JE, SantucciDM, Dimitrov DF, Patil PG, Henriquez CS, NicolelisMA.2003. Learning to control a brain–machine interface forreaching and grasping by primates. PLoS Biol 1: e42.doi:10.1371/journal.pbio.0000042

Chapin JK, Moxon KA, Markowitz RS, Nicolelis MA. 1999.Real-time control of a robot arm using simultaneouslyrecorded neurons in the motor cortex. Nat Neurosci 2:664–670. doi:10.1038/10223

Cheney PD, Fetz EE. 1980. Functional classes of primatecorticomotoneuronal cells and their relation to activeforce. J Neurophysiol 44: 773–791. doi:10.1152/jn.1980.44.4.773

Chiluwal A, Narayan RK, Chaung W, Mehan N, Wang P,Bouton CE, Golanov EV, Li C. 2017. Neuroprotectiveeffects of trigeminal nerve stimulation in severe traumaticbrain injury. Sci Rep 7: 6792. doi:10.1038/s41598-017-07219-3

Collinger JL, Wodlinger B, Downey JE, Wang W, Tyler-Kabara EC, Weber DJ, McMorland AJ, Velliste M, Bo-ninger ML, Schwartz AB. 2013. High-performance neu-roprosthetic control by an individual with tetraplegia.Lancet 381: 557–564. doi:10.1016/S0140-6736(12)61816-9

Donoghue JP, Sanes JN, Hatsopoulos NG, Gaál G. 1998.Neural discharge and local field potential oscillations inprimate motor cortex during voluntary movements. JNeurophysiol 79: 159–173. doi:10.1152/jn.1998.79.1.159

Ethier C, Oby ER, Bauman MJ, Miller LE. 2012. Restora-tion of grasp following paralysis through brain-con-trolled stimulation of muscles. Nature 485: 368–371.doi:10.1038/nature10987

Evarts EV. 1968. Relation of pyramidal tract activity to forceexerted during voluntary movement. J Neurophysiol 31:14–27. doi:10.1152/jn.1968.31.1.14

Fetz EE, Finocchio DV, Baker MA, Soso MJ. 1980. Sensoryand motor responses of precentral cortex cells duringcomparable passive and active joint movements. JNeurophysiol 43: 1070–1089. doi:10.1152/jn.1980.43.4.1070

Friedenberg DA, Schwemmer MA, Landgraf AJ, AnnettaNV, Bockbrader MA, Bouton CE, Zhang M, Rezai AR,Mysiw WJ, Bresler HS, et al. 2017. Neuroprosthetic-en-abled control of graded arm muscle contraction in a par-alyzed human. Sci Rep 7: 8386. doi:10.1038/s41598-017-08120-9

GeorgopoulosAP, Schwartz AB, Kettner RE. 1986.Neuronalpopulation coding of movement direction. Science 233:1416–1419. doi:10.1126/science.3749885

Graps A. 1995. An introduction to wavelets. IEEE ComputSci Eng 2: 50–61. doi:10.1109/99.388960

Guitchounts G, Markowitz JE, Liberti WA, Gardner TJ.2013. A carbon-fiber electrode array for long-term neuralrecording. J Neural Eng 10: 046016. doi:10.1088/1741-2560/10/4/046016

Hochberg LR, Serruya MD, Friehs GM, Mukand JA, SalehM, Caplan AH, Branner A, Chen D, Penn RD, DonoghueJP. 2006. Neuronal ensemble control of prosthetic devicesby a human with tetraplegia. Nature 442: 164–171.doi:10.1038/nature04970

Hochberg LR, Bacher D, Jarosiewicz B, Masse NY, SimeralJD, Vogel J, Haddadin S, Liu J, Cash SS, van der Smagt P,

C.E. Bouton

10 Advanced Online Article. Cite this article as Cold Spring Harb Perspect Med doi: 10.1101/cshperspect.a034306

ww

w.p

ersp

ecti

vesi

nm

edic

ine.

org

on April 1, 2020 - Published by Cold Spring Harbor Laboratory Press http://perspectivesinmedicine.cshlp.org/Downloaded from

et al. 2012. Reach and grasp by people with tetraplegiausing a neurally controlled robotic arm.Nature 485: 372–375. doi:10.1038/nature11076

Hsu JM, Rieth L, Normann RA, Tathireddy P, Solzbacher F.2009. Encapsulation of an integrated neural interface de-vice with Parylene C. IEEE Trans Biomed Eng 56: 23–29.doi:10.1109/TBME.2008.2002155

Humber C, Ito K, Bouton C. 2010. Nonsmooth formulationof the support vectormachine for a neural decoding prob-lem. arXiv. http://arxiv.org/abs/1012.0958v1

Jain AK,Murty N, Flynn PJ. 1999. Data clustering: A review.ACM Comput Surveys 31: 264–323. doi:10.1145/331499.331504

Kennedy PR, Bakay RA. 1998. Restoration of neural outputfrom a paralyzed patient by a direct brain connection.Neuroreport 9: 1707–1711. doi:10.1097/00001756-199806010-00007

LebedevMA, Carmena JM, O’Doherty JE, Zacksenhouse M,Henriquez CS, Principe JC, Nicolelis MA. 2005. Corticalensemble adaptation to represent velocity of an artificialactuator controlled by a brain–machine interface. J Neu-rosci 25: 4681–4693. doi:10.1523/jneurosci.4088-04.2005

LoAC, Guarino PD, Richards LG,Haselkorn JK,WittenbergGF, Federman DG, Ringer RJ, Wagner TH, Krebs HI,Volpe BT, et al. 2010. Robot-assisted therapy for long-term upper-limb impairment after stroke. N Engl J Med362: 1772–1783. doi:10.1056/NEJMoa0911341

Marder E, Bucher D. 2001. Central pattern generators andthe control of rhythmic movements. Curr Biol 11: R986–R996. doi:10.1016/S0960-9822(01)00581-4

Michie DM, Spiegelhalter DJ, Taylor CC, ed. 1994.Machinelearning, neural and statistical classification. Ellis Hor-wood, Chichester, England.

Moritz CT, Perlmutter SI, Fetz EE. 2008. Direct control ofparalysed muscles by cortical neurons. Nature 456: 639–642. doi:10.1038/nature07418

Muller KR, Anderson CW, Birch GE. 2003. Linear andnonlinear methods for brain–computer interfaces. IEEETrans Neural Syst Rehabil Eng 11: 165–169. doi:10.1109/TNSRE.2003.814484

Olofsson PS, Tracey KJ. 2017. Bioelectronic medicine: Tech-nology targeting molecular mechanisms for therapy. JIntern Med 282: 3–4. doi:10.1111/joim.12624

Pavlov VA, Tracey KJ. 2015. Neural circuitry and immunity.Immunol Res 63: 38–57. doi:10.1007/s12026-015-8718-1

Perkel DH, Bullock TH. 1968. Neural coding. Neurosci ResProgram Bull 6: 221.

Rathelot JA, Strick PL. 2009. Subdivisions of primary motorcortex based on cortico-motoneuronal cells. Proc NatlAcad Sci 106: 918–923. doi:10.1073/pnas.0808362106

Rousche PJ, Normann RA. 1998. Chronic recording capa-bility of the Utah Intracortical Electrode Array in cat sen-sory cortex. J Neurosci Methods 82: 1–15. doi:10.1016/S0165-0270(98)00031-4

Scheid MR, Flint RD, Wright ZA, Slutzky MW. 2013. Long-term, stable behavior of local field potentials during brainmachine interface use. Conf Proc IEEE Eng Med Biol Soc2013: 307–310.

Scherberger H, JarvisMR, Andersen RA. 2005. Cortical localfield potential encodes movement intentions in the pos-

terior parietal cortex. Neuron 46: 347–354. doi:10.1016/j.neuron.2005.03.004

Scholkopf B, Sung KK, Burges CJC, Girosi F, Niyogi P,Poggio T, Vapnik V. 1997. Comparing support vectormachines with Gaussian kernels to radial basis functionclassifiers. IEEE Trans Signal Process 45: 2758–2765.doi:10.1109/78.650102

Serruya MD, Hatsopoulos NG, Paninski L, Fellows MR,Donoghue JP. 2002. Instant neural control of amovementsignal. Nature 416: 141–142. doi:10.1038/416141a

Sharma G, Annetta N, Friedenberg D, Blanco T, Vasconce-los D, Shaikhouni A, Rezai AR, Bouton C. 2015. Timestability and coherence analysis of multiunit, single-unitand local field potential neuronal signals in chronicallyimplanted brain electrodes. Bioelectron Med 2: 63–71.

Sharma G, Friedenberg DA, Annetta N, Glenn B, Bock-brader M, Majstorovic C, Domas S, Mysiw WJ, RezaiA, Bouton C. 2016. Using an artificial neural bypass torestore cortical control of rhythmic movements in a hu-man with quadriplegia. Sci Rep 6: 33807. doi:10.1038/srep33807

Shavlik JW, Dietterich TG. 1990. Readings in machine learn-ing. Morgan Kaufmann, San Mateo, CA.

Stark E, Abeles M. 2007. Predicting movement from multi-unit activity. J Neurosci 27: 8387–8394. doi:10.1523/jneurosci.1321-07.2007

Taylor DM, Tillery SI, Schwartz AB. 2002. Direct corticalcontrol of 3D neuroprosthetic devices. Science 296: 1829–1832. doi:10.1126/science.1070291

Tracey KJ. 2002. The inflammatory reflex. Nature 420: 853–859. doi:10.1038/nature01321

Tracey KJ. 2015. Electronic medicine fights disease. Sci Am312: 28–35. doi:10.1038/scientificamerican0315-28

VellisteM, Perel S, SpaldingMC,WhitfordAS, SchwartzAB.2008. Cortical control of a prosthetic arm for self-feeding.Nature 453: 1098–1101. doi:10.1038/nature06996

Vetter RJ,Williams JC, Hetke JF, Nunamaker EA, Kipke DR.2004. Chronic neural recording using silicon-substratemicroelectrode arrays implanted in cerebral cortex.IEEE Trans Biomed Eng 51: 896–904. doi:10.1109/TBME.2004.826680

Volpe BT, Lynch D, Rykman-Berland A, Ferraro M, Gal-gano M, Hogan N, Krebs HI. 2008. Intensive sensori-motor arm training mediated by therapist or robot im-proves hemiparesis in patients with chronic stroke.Neurorehabil Neural Repair 22: 305–310. doi:10.1177/1545968307311102

Warland DK, Reinagel P, Meister M. 1997. Decoding visualinformation from a population of retinal ganglion cells. JNeurophysiol 78: 2336–2350. doi:10.1152/jn.1997.78.5.2336

Wessberg J, Stambaugh CR, Kralik JD, Beck PD, LaubachM,Chapin JK, Kim J, Biggs SJ, SrinivasanMA, NicolelisMA.2000. Real-time prediction of hand trajectory by ensem-bles of cortical neurons in primates.Nature 408: 361–365.doi:10.1038/35042582

Williams JC, Rennaker RL, Kipke DR. 1999. Long-term neu-ral recording characteristics of wire microelectrode arraysimplanted in cerebral cortex.Brain Res Brain Res Protoc 4:303–313. doi:10.1016/S1385-299X(99)00034-3

Restoring Movement with Bioelectronic Neural Bypass

Advanced Online Article. Cite this article as Cold Spring Harb Perspect Med doi: 10.1101/cshperspect.a034306 11

ww

w.p

ersp

ecti

vesi

nm

edic

ine.

org

on April 1, 2020 - Published by Cold Spring Harbor Laboratory Press http://perspectivesinmedicine.cshlp.org/Downloaded from

Zanos TP, Silverman HA, Levy T, Tsaava T, Battinelli E,Lorraine PW, Ashe JM, Chavan SS, Tracey KJ, BoutonCE. 2018. Identification of cytokine-specific sensory neu-ral signals by decoding murine vagus nerve activity. ProcNatl Acad Sci 115: E4843–E4852. doi:10.1073/pnas.1719083115

Zhang M, Schwemmer MA, Ting JE, Majstorovic CE, Frie-denberg DA, Bockbrader MA, Mysiw WJ, Rezai AR, An-nettaNV, BoutonCE, et al. 2018. Extractingwavelet basedneural features from human intracortical recordings forneuroprosthetics applications. Bioelectron Med 4: 11.doi:10.1186/s42234-018-0011-x

C.E. Bouton

12 Advanced Online Article. Cite this article as Cold Spring Harb Perspect Med doi: 10.1101/cshperspect.a034306

ww

w.p

ersp

ecti

vesi

nm

edic

ine.

org

on April 1, 2020 - Published by Cold Spring Harbor Laboratory Press http://perspectivesinmedicine.cshlp.org/Downloaded from

published online February 11, 2019Cold Spring Harb Perspect Med  Chad E. Bouton Approach: Current State and Future DirectionsRestoring Movement in Paralysis with a Bioelectronic Neural Bypass

Subject Collection Bioelectronic Medicine

DiseaseMedicine in Treatment of Chronic Inflammatory Neural Control of Inflammation: Bioelectronic

Centa, et al.Michael Eberhardson, Laura Tarnawski, Monica

Disease Diagnosis and Treatmenton the Inflammatory Reflex to New Approaches in Bioelectronic Medicine: From Preclinical Studies

J. TraceyValentin A. Pavlov, Sangeeta S. Chavan and Kevin

Therapeutic ImplicationsPathways Using Ultrasound and Its Current Noninvasive Neuromodulation of Peripheral Nerve

Christopher Puleo and Victoria Cotero

SystemVagus Nerve Stimulation and the Cardiovascular

Lance B. BeckerMichael J. Capilupi, Samantha M. Kerath and

Enteric Neuromodulation for the Gut and BeyondYogi A. Patel and Pankaj J. Pasricha Treatment of Inflammation-Mediated Diseases

Harnessing the Inflammatory Reflex for the

ChernoffYaakov A. Levine, Michael Faltys and David

SystemOptogenetic Control of the Peripheral Nervous

Rui B. Chang and Biomarkers of DiseaseRelated to Changes in Physiological Parameters Recording and Decoding of Vagal Neural Signals

Theodoros P. Zanos

and Translational ResearchClosed-Loop Neuromodulation in Physiological

Stavros Zanos State and Future DirectionsBioelectronic Neural Bypass Approach: Current Restoring Movement in Paralysis with a

Chad E. Bouton

Assessment: An OverviewElectrical Impedance Methods in Neuromuscular

Seward B. Rutkove and Benjamin Sanchez

Ethical Concerns−−Bioelectronic Medicine

HaridatSamuel Packer, Nicholas Mercado and Anita

Solutions Precision-Guided by LightOptogenetic Medicine: Synthetic Therapeutic

Haifeng Ye and Martin Fussenegger

Use of Bioelectronics in the Gastrointestinal TractLarry Miller, Aydin Farajidavar and Anil Vegesna

NanoparticleTechnobiology's Enabler: The Magnetoelectric

Sakhrat KhizroevGut Interactions

−Vagus Nerve Stimulation at the Interface of Brain

Bruno Bonaz, Valérie Sinniger and Sonia Pellissier

http://perspectivesinmedicine.cshlp.org/cgi/collection/ For additional articles in this collection, see

Copyright © 2019 Cold Spring Harbor Laboratory Press; all rights reserved

on April 1, 2020 - Published by Cold Spring Harbor Laboratory Press http://perspectivesinmedicine.cshlp.org/Downloaded from