SCHOOLANDSYMPOSIUMONADVANCEDNEUROREHABILITATION(SSNR2016)

Proceedings

June6-10,2016Baiona(Spain)

TableofContents

Page

UnsupervisedMappingofMotorUnitActionPotentialTrainstoHandKinematicsUsingComplexPrincipalComponentAnalysisArashAndalib,FrancescoNegro,DarioFarina,andJoséPrincipe…………………………………………………4Real-TimeEMGDrivenModelingforNeurorehabilitationGuillaumeDurandau,MassimoSartori,andDarioFarina……………………………………………………………..6Inter-LimbTransitioninGaitCoordinationTasksPost-StrokeisAffectedbyAgeJessicaL.Fujan-Hansen,TroyJ.Rand,PierreFayadandMukulMukherjee……………………………………8Self-BalancingWheeledMobilityDeviceAmitGoffer………………………………………………………………………………………………………………………………...10TheNon-linearRelationshipBetweenSensoryandMotorPrimitivesDuringReachingMovementsRussellL.Hardesty,MatthewT.Boots,SergiyYakovenko,andValeriyaGritsenko…………………….…12AutomatedMovementTherapyforNeurologicalRehabilitationM.KlöcknerandB.Kuhlenkötter………………………………………………………………………………………………...14Howdoweplanmovements?:AgeometricanswerRakeshKrishnan,NiclasBjörsell,andChristianSmith………………………………………………………………...16HumanoidNeurorobotics-Posture,BalanceandMovementControlVittorioLippiandThomasMergner……………………………………………………………………………………………18EffectofPulseShapeontheRecruitmentSelectivityofaCombinedInterfascicularandCuffElectrode(CICE)inaninvitroPigNerveModelL.E.Lykholt,W.Jensen,andK.R.Harreby…………………………………………………………………………………..20ImprovingTargetLocalisationwithCollaborativeBCIsAnaMatran-Fernandez,andRiccardoPoli…………………………………………………………………………………22Cyclingwithplantarstimulationincreasescutaneomuscular-conditionedspinalexcitabilityinsubjectswithincompletespinalcordinjuryStefanoPiazza,DiegoSerranoMuñoz,JulioGómez-Soriano,DiegoTorricelli,Gerardo

Avila-Martin,IrianaGalan-Arriero,JoseLuisPons;andJulianTaylor………………………………………....24Neuralcorrelatesofadaptationtonovelforce-field:anexploratoryERPstudySaraPizzamiglio,DuncanL.Turner……………………………………………………………………………………………26Analysisoffootpathspatio-temporalvariabilityduringgait.Acasestudyofalow-lumbarlevelMyelomeningocelepatientClaudiaN.Lescano,SilviaE.Rodrigo…………………………………………………………………………………………..28TranscranialDirectCurrentStimulation(tDCS)protocolsforimprovingresultsofdetectionintentionofpedalinginitiationthroughEEGsignalsMarisolRodríguez-Ugarte,EduardoIáñez,ÁlvaroCosta,JoséM.Azorín……………………………………..30DetectionofmuscleactivityduringmotorimaginationandattemptedmovementsusingultrasoundimagingAnnaJ.Sosnowska,HenrikGollee,IanD.Loram,andAleksandraVuckovic………………………………….32

Unsupervised Mapping of Motor Unit Action Potential Trains to HandKinematics Using Complex Principal Component Analysis

Arash Andalib, Francesco Negro, Dario Farina, and Jose C. Prıncipe

Abstract— We propose an unsupervised approach to map mo-tor unit action potential (MUAP) trains to hand kinematics. Toquantify the spatio temporal information in MUAP trains, wedevelop a novel complex principal component analysis (CPCA)representation that includes muscle synergies for antagonisticmovements in a single dimension, so it makes the direct controlof each motor of a bionic hand straightforward by a simpledimensionality reduction. We further explore the importanceof a linear transformation (rotation) in the projection spacethat best aligns the PCA subspace with the canonical projectionspace of the bionic hand degrees of freedom. We perform someexperiments to estimate hand flexion-extension from MUAPtrains identified during voluntary movements, with promisingpreliminary results.

I. INTRODUCTION

RECENT advances in non-invasive EMG decompositionare the basis for the development of novel strategies for thecontrol of hand prostheses, ultimately aimed to allow the userto perform composite movements in a natural and intuitiveway. Current control strategies use either classification basedapproaches, which only provide sequential control of a fewdegrees of freedom (DoFs) [1], or regression based methodsthat produce continuous outputs intended for independentproportional and simultaneous control [2]. The taxonomyalso considers if the methods are supervised/unsupervised,linear/nonlinear, or if they rely on sensory feedback or open-loop control. Here, we present a rather different approach tobionic hand control based on the idea of empowering theamputee as the natural feedback controller, which exploresthe fundamental inner works of the nervous system usingthe perception-action cycle [3]. The pre requisites for theapproach are the availability of high density recordings ofmyoelectric signals to separate them in MUAP trains, andproper training with the availability of sensory feedbackincluding visual and touch feedback. Amputees learn throughplasticity how to operate a bionic hand in a similar wayhumans learn how to use their bodies. We introduce a novel,unsupervised framework to extract muscle synergies inherent

This work is sponsored in part by the Defense Advanced ResearchProjects Agency (DARPA) BTO under the auspices of Dr. Doug Weberthrough the Space and Naval Warfare Systems Center, Pacific Grant/ContractNo. N66001-15-C-4018 to the University of Florida. This text is approvedfor Public Release by DARPA, Distribution Unlimited.

This work is supported in part by the European Research CouncilAdvanced Grant DEMOVE Contract No. 267888.

A. Andalib and J. C. Prıncipe are with the Department of Electrical andComputer Engineering, University of Florida, Gainesville, FL 32611 USA(e-mail: andalib, principe@cnel.ufl.edu).

F. Negro and D. Farina are with the Institute of Neurorehabilitation Sys-tems, Bernstein Focus Neurotechnology Gottingen, University Medical Cen-ter Gottingen, Georg-August University, Gottingen, GERMANY (e-mail:fblack, dario.farina@bccn.uni-goettingen.de).

in the neuromuscular data, and we believe this leads toan easier learning process because our signal processingapproach is exploiting the structure of the myoelectric signalproduced by the nervous system in the control of movement.

II. METHODOLOGY

Hundreds of motor unit spike trains, extracted from highdensity EMG signals, constitute the high-dimensional featurespace (HDS). We exploit their spatial and temporal structureprojecting the data in a metric preserving sense to the low di-mension subspace (LDS) of the actuators for each DoF of thebionic hand, which create an external reference basis (ERB).Here, we use spike rates of MUAPs in HDS to design and testthe simplest of unsupervised linear metric projections knownas principal component analysis (PCA). This approach hasthe following properties: 1) it preserves the metric in thefeature space because it uses the Mahalanobis distance [4],defined by the MUAP trains multidimensional covariance,2) it produces hand trajectories without using trajectories aslabels, and 3) it projects the data to an orthogonal LDScontrolled by the data. The first property guarantees that(dis)similar neural responses in HDS are projected close to(away from) each other, and the second is important becausein amputees only the neural outputs are available for training.The third property is beneficial, because the desired space ofthe bionic hand (ERB) is also orthogonal, so if the axes ofthese two subspaces are aligned, then the problem reduces toan appropriate assignment (connection) of PCA componentsto motors. However, the two coordinate systems are still notthe same (PCA scale uses variance, and eigenvectors are notsparse) so we attempt to compensate this difference withan orthogonal rotation to align the principal eigen-directionswith the ERB. Fig. 1 illustrates the general procedure, wherethe d-dimensional space of MUAP trains is projected into anorthogonal p-dimensional PCA space (p < d), and then thePCA sub-space is orthogonally rotated for the assignment tothe motor control space.

PCA projection

(dxp)

PCA Space (Nxp)MUAP data in Nxd HDS

The top three eigenvectors

[e1 e2 e3]Selected PCs

(Nx3)

Rotation

Estimated Lables in ERB (Nx3)

3D PCA Space before rotation

mc2

mc3

mc1 e1

e2e3

e1

e2

e3e4

. . .ep

Fig. 1. PCA projection and rotation of PCA subspace (red axes) toalign with the orthogonal motor control space ERB (blue axes). For bettervisualization, a three-dimensional subspace is considered here.

Among an infinite number of rotation matrices, the optimalone linearly transforms the principal components to empha-size the strong spatial correlation between the channels, soeach component after rotation is described by a sparse basisfunction. We use the Householder transformation [5] (imple-mented by QR decomposition) to decompose the projecteddata matrix as the product of a unitary matrix (rotated com-ponents) and an upper triangular matrix (rotation matrix),that guarantees sparseness on the largest eigendirection.

The last step in our methodology is how to best exploit thetemporal structure in HDS, which conventional PCA cannotdo (i.e., if we randomly shuffle the MUAP trains per channel,the PCA projection remains the same). In an experimentalstudy of the multidimensional MUAP trains during antago-nistic movements (e.g. wrist flexion-extension), we noticedthey are organized in an on-off arrangement in time, i.e.,there is a set of motor neurons that actuates the flexion, whileanother set actuates the extension at different times. This isconsistent with the fact that muscles can only pull and notpush, however, it means each individual principal componentof MUAP trains in the subspace will represent just onedirection of movement and we need two eigendirectionsto map to a single ERB axis (positive/negative excursion)of the external motor space. This destroys the conceptualprojection operator framework of Figure 1. Our goal is tomap a single eigendirection to a motor coordinate, hencethe information for flexion and extension must appear in thesame principal component. To accomplish this, we proposeto perform the PCA decomposition in the complex domain(CPCA) to capture the motor synergism. We construct thecomplex numbers by assigning the real part to flexion,and the imaginary part to extension (or vice versa). Inorder to decompose the channels into these two subsets weselect MUAP channels based on their low temporal cross-correlation (using a user-defined threshold). The complexcross-covariance matrices determined from these complexinputs are decomposed into complex eigenvectors, but theprojected CPCA eigenvalues are real. The CPCA subspaceis rotated using the procedure explained earlier, and noweach single dimension represents a full agonist-antagonisticmovement.

III. RESULTS AND CONCLUSIONS

The motor unit spike trains were decomposed from seg-ments of 192 channels of high density EMG signals of 30s duration, recorded from an able volunteer. The decom-position was performed using a recently proposed methodfor convolutive blind source separation of HD EMG signals[6]. The subject performs flexion-extension with residualadduction-abduction and the target labels are sampled at2048 Hz. More information on the experimental set up, theelectrodes, and the motion tracking system is available in [7].The input to the CPCA is built from spike counts (200msbin size) on every channel, and the projection is done to a12-dimensional data manifold. The recorded labels are onlyused for performance analysis. The original and projectedkinematics after the Householder rotation are shown in Fig.

2. In simple PCA, the first and second principal componentsare combined to estimate the flexion-extension movement.However, for complex PCA, we only use one (the first)complex component, which includes real and imaginaryparts, corresponding to flexion and extension, respectively.The eigen spectrum of CPCA vs. PCA, shown in Fig. 3,verifies that most of power is preserved in a single (first)principal component, when using CPCA, while this powerappears in first two principal components of PCA. A bitof power in the second (third) component in CPCA (PCA)corresponds to the secondary adduction-abduction.

Time (sec)

0 2 4 6 8 10

Fle

xion

-Ext

ensi

on

-1

0

1OriginalPCA (1st and 2nd PCs)CPCA (1st complex PC)

Fig. 2. Original vs. projected kinematics when using simple PCA (the redtrace) and complex PCA (green trace).

dimensions 0 2 4 6 8 10 12

pow

er

0

0.2

0.4

0.6

0.8

CPCAPCA

Fig. 3. The eigen spectrum of CPCA (solid line) vs PCA (dashed line).

To conclude, we introduced a novel unsupervised frame-work to extract muscle synergies inherent in the neuromus-cular data of simple movements. This method is based onthe orthogonal rotation of CPCA space and should imply aneasier learning process for the user in the loop because oursignal processing is exploiting the way the nervous systemcontrols movement.

REFERENCES

[1] E. Scheme, and K. Englehart, “Electromyogram pattern recognitionfor control of powered upper-limb prostheses: State of the art andchallenges for clinical use,” J. Rehabil. Res. Develop., vol. 48, no. 6,pp. 643-659, 2011.

[2] N. Jiang, J. L. Vest-Nielsen, S. Muceli, and D. Farina, “EMG-basedsimultaneous and proportional estimation of wrist/hand dynamics inuni-lateral trans-radial amputees,” J. Neuroeng. Rehabil., vol. 9, no. 1,pp.42, 2012.

[3] R. W. Sperry, “Neurology and the mind-brain problem,” AmericanScientist, vol. 40, pp. 291-312, 1952.

[4] H. H. Harman, Modern factor analysis, University of Chicago Press,1993, p. 474, 1967.

[5] A. S. Householder, “Unitary Triangularization of a NonsymmetricMatrix,” J. ACM, vol. 5, no. 4, pp. 339-342, 1958.

[6] F. Negro, S. Muceli, A. M. Castronovo, A. Holobar, and D. Farina,“Multi-channel intramuscular and surface EMG decomposition byconvolutive blind source separation,” J. Neural Eng., vol. 13, no. 2,2016.

[7] J. M. Hahne, F. Biessmann, N. Jiang, H. Rehbaum, D. Farina,F. C. Meinecke, K-R Muller, and L. C. Parra, “Linear and non-linearregression techniques for simultaneous and proportional myoelectriccontrol,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 22, pp. 269-79,2014.

Real-Time EMG Driven Modeling for Neurorehabilitation

Guillaume Durandau, Massimo Sartori, and Dario Farina

Abstract— Neuromusculoskeletal (NMS) modeling is a com-puter based representation of the human musculo-skeletalsystem as controlled by the nervous system. With the knowledgeof the kinematics and dynamics parameter of the musculo-skeletal system, a NMS model can compute the output forcedeveloped by the muscle and thus the force developed aroundarticular degrees of freedom. This gives access to internalbody paramaters that can be used to inform personalizedrehabilitation interventions and technologies. In this studywe describe the development of a real-time NMS modelingsystem that can be used in the context of neurorehabilitationtechnologies.

I. INTRODUCTION

Neuromusculoskeletal (NMS) modeling has gained a lotof tractions lately thanks to the release of modeling soft-ware such as OpenSim [1], Anybody1, Biomechanics ofBodies2. Current modeling software are still expensive com-putationally and are not really optimized for EMG drivenapproaches. For EMG driven modeling, the offline CalibratedEMG-informed Neuromusculoskeletal Modeling (CEINMS)toolbox[2] is a software created for the offline computationand calibration of EMG subject-specific NMS models. Ourobjective was to create a real-time algorithm that can beused on wearable devices. We also aim for a generic solutionsupporting any model including upper and lower limb model.Another goal is the possibility to easily interface with exter-nal device like EMG amplifiers, motion capture systems orwearable robotics systems. Giving access to body parametersin real-time can help create personalized controllers forrehabilitation systems like prosthesis or exoskeleton. Real-time modeling can also increase the quality of rehabilitationtherapy and help to assess the progress made by the patientduring the whole rehabilitation process. Despite the progressmade on NMS modeling [3], there still are limited progressesin the effective translation to real-time neurorehabilitationscenarios.

II. METHOD

We extended CEINMS to work in real-time by creatinga system of plug-in for connecting to external device usingTCP/IP protocol. We also extended it to work with external

G. Durandau, M. Sartori and D. Farina are with theInstitute of Neurorehabilitation Systems, UniversitatmedizinGottingen, Gottingen, Germany. Corresponding author e-mail:guillaume.durandau@bccn.uni-goettingen.de.

This work was supported by the ERC Advanced Grant DEMOVE[267888]

1http://www.anybodytech.com/2http://www.prosim.co.uk/BoB/

NMS model described via XML format. For having real-time capacity, we integrated the Multidimensional Cubic B-Spline (MCBS) software [4] which computes the musclelength (LMT) and muscle moment arm (MA) from the jointangle. The joint angle are computed from marker positionrecorded by a motion capture system. We use the inversekinematic library from OpenSim to compute in real-timethe value of the joint angle. With access to the EMG,LMT and MA in real-time, CEINMS can compute the jointtorque. First, it compute the neural activation from the EMG.Then, CEINMS compute the muscle force using an Hillmuscle model. Finally, With the MA of the muscle thejoint torque is computed. For better result, we created acalibration software extended from the calibration procedureof CEINMS. We extended it by making it compute the B-spline coefficients, then we compute a pre-scaling of theoptimal fiber length and tendon slack length we then calibratethe model using experimental data. The computation of theb-spline coefficient and the pre-scaling are extended from theOpenSim library. In the calibration procedure we minimizethe error between the joint torque computed by CEINMS andthe joint torque computed by the OpenSim inverse dynamicfrom experimental data. The parameters computed by thecalibration are the tendons slack length, optimal fibers lengthand muscles force. Furthermore, we added a GUI which plotsthe different joint torque and filtered EMG, we also plot anhistogram of the muscles force and have a 3D representationof the skeletal model. CEINMS gives us now access in real-time to parameters like pennation angle, muscles tendonlength, fiber length, fiber velocity, moments arm and tendonlength.

III. EXPERIMENT

we used a lower limb model composed of 23 degrees offreedom and 92 muscles. For our experiments all the jointsangles were computed but only the right knee and right anklemuscles were used in the joint torque computation.

EMGs are recorded (2048Hz) using a 256-channel EMGamplifier (OTBioelettronica, Italy) from 10 muscle groups:rectus femoris, biceps femoris, vastus medialis and lateralis,lateral and medial hamstrings, gastrocnemious medialis andlateralis, soleus and tibialis anterior. For the motion capturesystem we use a Qualisys motion capture system at 256Hzand a Bertec force plate connected to the motion capturesystem and captured at 2048Hz. A plug-in was created forhaving access in real-time to the data from the motion capturesystem and EMG amplifier using TCP/IP. The EMG werelow-pass filtered (4Hz, fourth order), then high pass filtered(100 Hz, fourth order) and normalised against the maximum

Fig. 1. Real-time torque from inverse dynamic (ID) (Black is the mean andGrey is the standard deviation) and from CEINMS-RT (Blue is the meanand light blue is the standard deviation) of the ankle joint from 3 fast gaittrials normalized by the body weight.

of the EMG in real-time. the Force plate and marker werealso filtered using a 6Hz low pass filter.

We first record maximum voluntary contraction (MVC)for having the maximum EMG then we do gait trial, squatfinishing by a calf rise and static trials for the scaling and thecalibration of the model. For the scaling of the model we usethe scaling tool from OpenSim. After calibration we realisedreal-time computation on forward gait, backward gait, squatfinishing by a calf rise, jump, one leg squat and fast gait. Allthese trial were repeated at least three time.

IV. RESULT

A pilot of the experiment was realised on one subjectshowing encouraging results (see Fig. 1 ).

The average Processing time was 2.654e−4 ± 95%SDseconds for the NMS model, 3.5538e−4± 93%SD secondsfor the cubic B-spline computation, 8.899e−3±79%SD forthe inverse kinematic computation and 3.3658e−2±29%SDfor the total delay (most of the delay is due to EMG packetsreceived from the amplifier that add a fix delay).

In table I, we can see the error in Newton/Meter of theforce between the inverse dynamics and CEINMS forwarddynamic of the different trials. For the gait and fast gaittrials knee joint the result for the knee inverse dynamic werenot correct due to a problem caused by the motion capturesystem and so are not reported here.

V. DISCUSSION

The advantage of NMS modelling over machine learningmethod is the possibility to have access to a lot of differentinternal parameter and the extrapolation out of the initialdataset used for calibration. This, add the possibility torealise short calibration on a reduced data trials (gait, squatfinishing by a calf rise and static) and used the calibratedmodel on other kind of trials (jump, fast gait ect..).

TrialKnee Joint Error Ankle Joint Error

Mean (N/m) SD (%) Mean (N/m) SD (%)

Squats/Calf Rise 7.05 ±61% 15.90 ±35%

Gait 17.28 ±47%

Jump 22.69 ±71% 8.53 ±61%

One Leg Squat 30.51 ±50% 17.80 ±51%

fast Gait 18.89 ±51%

TABLE I

Real-time modeling can be also applied on robotic systemfor rehabilitation like exoskeleton or prosthesis, so we alsowant to test the software on wearable robotics device in aclosed loop fashion, where the torque computed by the modelis used to control the device.

Our algorithm was also successfully compiled and usedon a wearable computer (Raspberry Pi 2, Raspberry PiFoundation, UK) with real-time capabilities.

Finally, the model can easily be improved for com-puting new body parameters in real-time like joint load[5], stiffness[6], muscle energy consumption and muscleprimitives[7].

VI. CONCLUSION

In conclusion we have created a real-time NMS model thatused the EMG signals and joints angles to compute in real-time joints torques and internal body parameters. We intendto use our software for controlling wearable robotics devicein real-time for personalized rehabilitation.

REFERENCES

[1] S. L. Delp, F. C. Anderson, A. S. Arnold, P. Loan, A. Habib, C. T. John,E. Guendelman, and D. G. Thelen, “Opensim: open-source softwareto create and analyze dynamic simulations of movement,” BiomedicalEngineering, IEEE Transactions on, vol. 54, no. 11, pp. 1940–1950,2007.

[2] C. Pizzolato, D. G. Lloyd, M. Sartori, E. Ceseracciu, T. F. Besier,B. J. Fregly, and M. Reggiani, “Ceinms: A toolbox to investigatethe influence of different neural control solutions on the predictionof muscle excitation and joint moments during dynamic motor tasks,”Journal of biomechanics, vol. 48, no. 14, pp. 3929–3936, 2015.

[3] M. Sartori and D. Farina, “Neural data-driven musculoskeletal modelingfor personalized neurorehabilitation technologies,” 2016.

[4] M. Sartori, M. Reggiani, A. J. van den Bogert, and D. G. Lloyd,“Estimation of musculotendon kinematics in large musculoskeletalmodels using multidimensional b-splines,” Journal of biomechanics,vol. 45, no. 3, pp. 595–601, 2012.

[5] P. Gerus, M. Sartori, T. F. Besier, B. J. Fregly, S. L. Delp, S. A.Banks, M. G. Pandy, D. D. D’Lima, and D. G. Lloyd, “Subject-specificknee joint geometry improves predictions of medial tibiofemoral contactforces,” Journal of biomechanics, vol. 46, no. 16, pp. 2778–2786, 2013.

[6] M. Sartori, M. Maculan, C. Pizzolato, M. Reggiani, and D. Farina,“Modeling and simulating the neuromuscular mechanisms regulatingankle and knee joint stiffness during human locomotion,” Journal ofNeurophysiology, vol. 114, no. 4, pp. 2509–2527, 2015.

[7] M. Sartori, L. Gizzi, D. G. Lloyd, and D. Farina, “A musculoskeletalmodel of human locomotion driven by a low dimensional set of im-pulsive excitation primitives,” Frontiers in computational neuroscience,vol. 7, 2013.

Abstract— Stroke has been viewed as a phenomenon which predominately affects the aged, however, these occurrences are now inclusive of younger populations. To assess the current manner to which we rehabilitate stroke survivors regardless of age, we investigated the effects of age on a common gait paradigm used in stroke rehabilitation. Two groups of young and older stroke survivors performed a gait protocol designed to induce motor learning. Significance between the groups was shown to exist in both step time and double support time. This difference in control of complexity during inter-limb transition while learning a gait coordination task may play an important role during the gait rehabilitative process of stroke survivors.

I. INTRODUCTION ISTORICALLY stroke has been viewed as a health concern which predominately affects the aged,

however, there has been a recent increase in the prevalence of stroke in younger and middle age groups [1]. Twenty percent of strokes occur in people younger than 65 years of age [2]. This age range represents working-age adults in our society and poses a significant economic impact to the global bottom line. This spectrum comprises those at the brink of their societal contributions, as well as those in the heart of their careers. With the escalation of stroke in younger populations, this continuing study will allow us to take steps to determine whether applying the current rehabilitative techniques to all stroke survivors regardless of age is adequate in order to improve lifestyles and reduce these societal burdens.

In addition to the economic influence these individuals hold, motor related disorders such as stroke have a substantial impact on research strategies employed to improve variability. Optimal levels of variability in our movement patterns are a signal of healthy motor behavior [3], whereas the changes in movement variability often observed in motor related disorders represent a reduction in adaptive capabilities of the neuromuscular system [4], [5].

This work was supported by the National Institutes of Health Centers of

Biomedical Research Excellence (COBRE) grant. J. L. Fujan-Hansen is with the Department of Biomechanics of the

University of Nebraska at Omaha, USA (JFujan@unomaha.edu). T. J. Rand is with the Department of Biomechanics of the University of

Nebraska at Omaha, USA. P. Fayad is with the Department of Neurological Sciences of the

University of Nebraska Medical Center, Omaha, NE, USA. M. Mukherjee is with the Department of Biomechanics of the University

of Nebraska at Omaha, USA.

Nonlinear analysis of variability in movement patterns generated as a result of motor related disability, such as stroke, provide us a manner to analyze the neuromuscular status and allow the opportunity to improve functioning of the complex strategies used to control movement in this population [6], [7], [8].

Preliminary results from an ongoing study investigated the variability present in gait as an indicator of the adaptive capabilities of varying age ranges of stroke survivors. This is important in order to gauge where different age groups lie on the variability continuum in order to employ meaningful rehabilitative techniques aimed to achieve restoration of healthy variability in each age population. Given stroke survivors are amongst those who suffer from major functional mobility impairments post-stroke [9], [10], [11], it is pertinent to seek manners to restore healthy adaptability in each of these age groups in order to achieve a higher quality of life.

II. MATERIAL AND METHODS

A. Participants This research involved six subjects inclusive of the

general stroke survivor population within the two age groups of younger (20 to 60 years of age) and older (61 years of age and older) individuals. All stroke subjects were at least three months removed from their date of stroke. Inclusion criteria consisted of: 1) supratentorial ischemic or hemorrhagic stroke (based on imaging data), 2) ability to stand unsupported without dizziness and walk 10m without assistance, 3) ability to follow instructions (Folstein Mini-Mental exam ≥24), 4) free of major post-stroke complications (e.g. recurrent stroke, hip fracture, myocardial infarction), and 5) free of wheelchair. There was no restriction based upon gender, race or socioeconomic status.

B. Data Collection All participants provided University of Nebraska Medical

Center Internal Review Board approved written consent and standard demographic information. The protocol consisted of participants walking on an instrumented dual-belt treadmill (Bertec Corp., Columbus, OH, USA) using a paradigm intended to induce locomotor adaptation. The paradigm consisted of familiarization and baseline walking conditions, split-belt adaptation periods, removal of stimulus, and readaptation (Fig. 1). The lab is fitted with an 8-camera motion capture system (Vicon, Centennial, CO,

Inter-Limb Transition in Gait Coordination Tasks Post-Stroke is Affected by Age.

Jessica L. Fujan-Hansen, Troy J. Rand, Pierre Fayad and Mukul Mukherjee

H

USA). Each participant wore a tight fitting outfit to allow for the placement of motion analysis markers according to the lower body Plug-In-Gait marker set allowing kinematic data to be captured as the participants walked on the treadmill. All subjects were fitted with an overhead harness to minimize risk of falling during the protocol.

Fig. 1. Split-belt walking paradigm.

C. Data Processing and Statistics Three dimensional marker position time series data was

sampled at 100 Hz using Nexus software (Vicon, Centennial, CO, USA), exported in ASCII format and processed using MatLab software (Mathworks Inc., Natick, MA, USA) to calculate spatio-temporal parameters. Mean and variability analysis (sample entropy) was performed on the processed data. These values were used to identify differences between the age groups of stroke survivors.

The independent variable was age (two levels-young and old) and time (three adaptation levels-early, late, readaptation). The dependent variable was sample entropy as a measure of the complexity of gait variability of spatio-temporal variables (step length, step time, double support time, etc.). Two-way repeated measure ANOVAs were performed. Significance was set at 0.05.

III. RESULTS The spatio-temporal variables of step time and double

support time both revealed significance. Age was found to significantly impact participants’ ability to transition between slow and fast belt speeds during the split-belt trials. The split-belt learning process was significant across time, but not different between the two age groups.

IV. DISCUSSION AND CONCLUSION In a preliminary study we investigated the differences in

the spatio-temporal structure of variability in younger and older stroke survivors. Sample entropy distinguishes not only the regularity of stride time series between the stroke participants, but also between the affected and less affected legs. Such discrepancies between the two limbs impact coordination needed for accomplishing gait tasks. Given that an optimal level of variability in our movement patterns signifies healthy motor performance [12], such discrepancies in variability demonstrate abnormal movement patterns.

Age significantly impacts a stroke survivor’s ability to control the complexity of inter-limb transition during learning a gait coordination task. Complexity of gait plays an important role during the learning process, specifically during the transitional states. Complexity allows richness of behavior that is conducive to enhanced exploration during

learning or rehabilitation. This may be an important consideration when gait rehabilitation of stroke survivors is being designed and performed.

Future studies should expand upon these results with larger number of participants in order to deem whether the significant results found here continue to persist. The application of such results may have an important impact on rehabilitative strategies targeting younger stroke survivors.

REFERENCES [1] B. M. Kissela, et al., “Age at stroke: Temporal trends in stroke

incidence in a large, biracial population”, Neurology, vol. 79, 2012, pp. 1781-1787.

[2] F. Orsucci, “The paradigm of complexity in clinical neurocognitive science”, The Neuroscientist, vol. 12, 2006, pp. 390-397.

[3] B. Goldstein, D. Toweill, S. Lai, K. Sonnenthal, and B. Kimberly, “Uncoupling of the autonomic and cardiovascular systems in acute brain injury”, American Journal of Physiology, vol. 275, 1998, pp. R1287-R1292.

[4] J. C. Chien, D. J. Eikema, M. Mukherjee, and N. Stergiou, “Locomotor sensory organizational test: A novel paradigm for the assessment of sensory contributions in gait”, Annals of Biomedical Engineering, vol. 42, 2014, pp. 2512-2523.

[5] M. C. Cirstea and M. F. Levin, “Compensatory strategies for reaching in stroke”, Brain, vol. 123, 2000, pp. 940-953.

[6] S. J. Harrison and N. Stergiou, “Complex adaptive behavior and dexterous action”, Nonlinear Dynamics, Psychology, and Life Sciences, vol. 19, 2015, pp. 345-394.

[7] K. Z. Li and U. Lindenberger, “Relations between aging/ sensorimotor and cognitive functions”, Neuroscience and Behavior, vol. 26, 2002, pp. 777-783.

[8] R. D. Seidler, et al., “Motor control and aging: Links to age-related brain structural, functional, and biochemical effects”, Neuroscience and Biobehavioral Reviews, vol. 34, 2010, pp. 721-733.

[9] J. M. Finley, A. J. Bastian, and J. S. Gottschall, “Learning to be economical: The energy cost of walking tracks motor adaptation”, The Journal of Physiology, vol. 591, 2013, pp. 1081-1095.

[10] H. S. Jorgensen, et al., « Outcome and time course of recovery in stroke. Part ii : Time course of recovery The Copenhagen stroke study », Archives of Physical Medicine and Rehabilitation, vol. 76, 1995, pp. 406-412.

[11] T. E. Lockhart, J. L. Smith, and J. C. Woldstad, « Effects of aging of the biomechanics of slips and falls », Human Factors, vol. 47, 2005, pp. 708-729.

[12] P. M. Rothwell, et al., « Change in stroke incidence, mortality, case-fatality, severity, and risk factors in Oxfordshire, UK from 1987 to 2004 (Oxford Vascular Study) », Lancet, vol. 363, 2004, pp. 1925-1933.

Abstract—Few conditions are as disabling as paralysis that

results in confinement to a wheelchair. Loss of self-esteem,

daily struggle, poor health, poor quality of life and high

expenses are among the consequences of this condition.

This article presents a wheeled mobility device that enables

upright mobility as well as seated mobility. The device,

UPnRIDE , is not merely a stand-able wheelchair; its design is

equally optimized to both standing and seated mobility.

UPnRIDE automatically balances the system, so the user

remains vertical on forward, backward and lateral slopes. In

addition, the user’s center of gravity (COG) is substantially

independent of the mode of mobility, and therefore enabling

safe and stable indoors and outdoors mobility and over

different surfaces in both standing and sittings.

I. INTRODUCTION

Advances in healthcare have aided individuals with

serious injuries and severe disabilities in living longer.

Current estimates in the U.S.A. of persons with spinal cord

injury (SCI) range between 240,000 and 337,000 [1]. About

3.6 million people use wheelchairs and 11.6 million people

use assistive technology devices for mobility [2]. In the last decade, powered exoskeletons (e.g., ReWalk,

Indego, and Ekso) have emerged, holding the promise of

restoring walking for individuals suffering from severe

walking impairment [3]-[5].

However, powered exoskeletons can serve only a segment

of the wheelchair user community; so far, individuals

suffering from conditions like quadriplegia, MS (multiple

sclerosis), CP (cerebral palsy), and stroke are not suitable for

powered exoskeletons and wheelchairs are still their only

solution.

The rate of wheelchair usage has increased; as such, the

demands for better wheelchairs and seating systems have led

to an expanded market. Similar to individuals without

disabilities, maintaining an active lifestyle is essential

among people with disabilities. Innovative wheeled mobility

technology is essential for users to maintain an active

lifestyle for enhanced function, increased independence, and

greater accessibility in the home, work and community

environments. As a result, the wheelchair is the primary

mobility component for this segment of society and as the

individual begins to adapt to their disability, the wheelchair

is soon considered as an extension of their bodies. Although

wheelchairs provide mobility, people who are non-

ambulatory are at risk for many secondary medical

A. Goffer is with UPnRIDE Robotics Ltd., Yokneam, Israel,

(amit@upnride.com).

consequences due to the extreme amount of time spent

sitting.

In the non-paralyzed community, the modern day office

chairs are well-designed with ergonomic advantages for

sitting, but long sitting in an office chair can cause the

development of joint and back pain and many other

problems.

Frequent repositioning and standing is recommended in

the able-bodied population while sitting. Similar to sitting in

an office chair, frequent changes of position in a wheelchair

are highly recommended [6], [7] and may help to prevent, or

mitigate some of the secondary medical consequences of

extreme inactivity. Although the seating function systems

provide changes of position in a wheelchair, one drawback

that remains for wheelchair users is that they lose the ability

to be at eye level with others who are standing.

II. DESCRIPTION OF THE DEVICE

A. Specifications

The wheeled mobility device has two modes of mobility:

wheelchair (seated mobility) and a standing wheeled

mobility. The device possesses the following qualities:

Serves individuals suffering from paraplegia, quadriplegia, MS, CP and similar conditions.

COG of the system (device and user) remains substantially independent of mobility modes (Fig. 1).

A dynamic autonomous balancing mechanism keeps the

user vertical on forward, backward and lateral slopes. Training mode: activation of lower limbs using the lifting

motors. Indoors and outdoors mobility and over different

surfaces.

B. Mechanical Structure

The device consists of four main systems:

Wheels system that comprises: a pair of driving wheels and a pair of swivel wheels, two DC motors and their drivers, wheel-driving computer, and a battery pack.

Self-balancing system that comprises: a balancing plate (upon which the user is situated), two balancing motors, and an IMU (Inertial Measurement Unit).

A seat and raising system that holds the user and contains the lifting mechanism.

Computerized electronics that hosts the system software and control algorithms.

Unlike traditional stand-able wheelchairs, the user is

secured to the system by an exoskeletal mechanism,

attached to the sides of the user, thus the lifting process from

sitting to standing, is not carried out by pushing the seat and

Self-Balancing Wheeled Mobility Device

Amit Goffer

back supports. Such an arrangement ensures a better fitting

of the user to the system; oftentimes, in traditional stand-

able wheelchairs, the sit-stand-sit cycle disarrays the posture

of the user, while the exoskeletal mechanism doesn’t.

C. On the Balancing Algorithm

The IMU produces pitch, roll and yaw angles and the

corresponding angular velocities. The pitch and roll angles

are fed into a closed-loop control algorithm that keeps the

surface, upon which the exoskeletal system stands,

horizontal.

The yaw angular velocity is an indicator to the rotational

speed of the device; high rotational speed on steep slopes

may endanger the user, hence the yaw is used as an input to

the warning and safety system.

Fig. 1. The UPnRIDE wheeled mobility device; in a standing mode

(left) and in a wheelchair mode (right). The red lines denote the COG

which isn’t sensitive to mode of mobility.

III. RESULTS

The efficacy of the device and user satisfactory has been

proven in in-house tries (e.g., Fig. 2); official safety,

efficacy, and clinical studies are yet to come.

Fig. 2. The UPnRIDE on a slope: the user (A. Goffer, SCI-C5

complete) is vertical on a sloped surface of 10.

IV. CONCLUSION

The innovation of the UPnRIDE is threefold: it

improves user’s health and dignity, returns the investment

(its cost) within couple of years, and benefits society by

reducing medical expenses. The stability and self-balancing

qualities make the UPnRIDE suitable for both indoors and

outdoors mobility.

REFERENCES

[1] National Spinal Cord Injury Statistical Center, “Spinal Cord Injury

(SCI) Facts and Figures at a Glance 2015,”

www.nscisc.uab.edu/PublicDocuments/fact_figures_docs/Facts%2020

15.pdf.

[2] M. W. Brault, “Americans with Disabilities: 2010,” US Census

bureau, Health & Disability Statistics Branch, July 27, 2012.

(www.census.gov.edgekey-

staging.net/newsroom/cspan/disability/20120726_cspan_disability_sli

des.pdf).

[3] A. Esquenazi, “New Bipedal Locomotion Option for Individuals with

Thoracic Level Motor Complete Spinal Cord Injury,” Journal of The

Spinal Research Foundation, VOL. 8 No. 1, Spring 2013.

[4] A. Goffer, “ReWalk™: An Exoskeleton for Overcoming Vertical

Mobility Impairment; a superior ambulation alternative to wheelchair

users, ”Int. Trade Show and World Congress for Prosthetics, Orthotics

and Rehabilitation Technology, Leipzig, Germany, May 2012.

[5] A. Spungen et al, “Exoskeletal-Assisted Walking for Persons with

Motor-Complete Paraplegia,” STO Human Factors and Medicine

Panel (HFM) Symposium, Milan, Italy, April 2013.

[6] H. Miller, ”Supporting the Spine When Seated ,”

www.hermanmiller.com/content/dam/hermanmiller/documents/solutio

n_essays/se_Supporting_the_Spine_When_Seated.pdf.

[7] Brubaker, C., S. Ross, and C. McLaurin, “Effect of seat position on

handrim force,” 5th Annual Conference on Rehabilitation

Engineering, 1982.

Abstract—The objective of this study is to quantify the role that Ia sensory feedback may play in the assembly of motor primitives for motor control of the human arm. We integrated a musculoskeletal model of the human arm with a model of Ia primary afferent discharge to analyze motion and muscle activation patterns during reaching movements in virtual reality. Clustering analysis of electromyographic and predicted Ia primary afferent signals showed a divergent relationship between muscle activation and sensory feedback patterns. The muscle activations were more task-specific than sensory feedback patterns for all subjects (N=9). Altogether, these results suggest that sensory feedback is non-linearly related to muscle activation profiles.

I. INTRODUCTION OVEMENTS emerge as a result of interaction between the musculoskeletal dynamics of the limb and

neural control signals that control it with predictive and sensory feedback pathways. Redundant musculoskeletal anatomy and dynamical interactions of multi-segmented limbs contribute to control complexity. The frontier in understanding how the nervous system overcomes muscle redundancy and controls limb dynamics is the concept of motor primitives. These are defined as groups of muscles that are activated by a subset of sequential or partially overlapping neural signals. While sensory feedback from proprioceptors is known to contribute to the motor control at multiple levels of the neural hierarchy, its contribution to the formation of motor primitives is less well understood.

II. METHODS

A. Experimental design and human subjects We have recruited 9 healthy adults (5 males, 4 females,

24.3 ± 1.8 years old, 76.3 ± 14.5 kg) to perform upper extremity reaching movements in a virtual reality (VR) environment. All procedures were approved by the West Virginia University Institutional Review Board (Protocol #1309092800). The movement task was created with a

This study is supported by a scholarship from NIH/NIGMS

U54GM104942 (RLH), Ruby Fellowship (MTB), and NIH CoBRE P20GM109098 (SY, VG). The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.

R. L. Hardesty is with the Centers for Neuroscience, West Virginia University (WVU), Morgantown, WV, USA (rlhardesty@mix.wvu.edu).

M. T. Boots is with the Department of Mechanical and Aerospace Engineering, WVU, Morgantown, WV, USA (mboots@mix.wvu.edu).

S. Yakovenko is with the Division of Exercise Physiology, WVU, Morgantown, WV, USA (seyakovenko@hsc.wvu.edu).

V. Gritsenko is with the Division of Physical Therapy, WVU, Morgantown, WV, USA (corresponding author vgritsenko@hsc.wvu.edu).

wearable VR helmet (Oculus Rift) integrated with motion capture. Subjects moved to virtual targets on cue with the visual feedback of their arm position. To minimize the inter-subject differences in motion, the locations of all targets were calculated mathematically based on subject’s segment lengths and shoulder and elbow joint angles, so that in movement #1 subjects’ shoulder extended while elbow flexed, in Movement #2 both shoulder and elbow flexed, and in Movement #3 shoulder extended while elbow flexed from a different starting position. This defined diverse dynamical contexts where the movement was largely passive (#1), interaction torques were resistive with increasing gravitational load (#2), and interaction torques were assistive with decreasing gravitational load (#3). During movement, we recorded kinematics of shoulder, elbow, and wrist joints and electromyography (EMG) of 12 muscles that span those joints. Motion capture data were recorded at 480 Hz using the Impulse system (PhaseSpace) and EMG was recorded at 2000 Hz with MA400-28 (MotionLab Systems).

B. Model To estimate the sensory contribution from muscle spindles

during movement, we have used a published model of Ia primary afferent discharge [1]. The spindle model relates afferent firing rate (Ia) to the rate of change of muscle length (v), muscle length (l), and normalized EMG (a) as follows:

𝐼𝑎 = 65 ∙ 𝑣!.! + 200 ∙ 𝑙 + 30 ∙ 𝑎 + 80 (1)

Muscle length was normalized to vary between 0 and 1,

which corresponded to minimal and maximal possible muscle lengths respectively, calculated as described below. Velocity was in units of rest length per second (RL/s); RL was defined as described below. EMG was high-pass filtered at 5 Hz, rectified, low-pass filtered at 20 Hz, and normalized to the maximum across all movements.

Muscle lengths were calculated using a modified musculoskeletal model of the human arm in OpenSim [2]. The model was scaled for each participant’s segment lengths adjusting the muscle origin and insertion points and muscle paths to individual dimensions. We used the model to estimate muscle length changes during movement from recorded motion capture data. Muscle RL was generalized as the average muscle length within the physiological range of motion. These data and EMG were used in equation (1).

C. Analysis We used regression analysis to explore the relationships

The Non-linear Relationship Between Sensory and Motor Primitives During Reaching Movements

Russell L. Hardesty, Matthew T. Boots, Sergiy Yakovenko, and Valeriya Gritsenko

M

between muscle activations and Ia sensory feedback. We created normalized EMG and Ia profiles for each muscle and each movement per subject by averaging signals across 20 repetitions. Then, we calculated correlation coefficients (r) between all pairs of EMG and Ia profiles. These r values were then converted into the heterogeneous variance explained (HVE) as follows:

𝐻𝑉𝐸 =1 − 𝑟!, 𝑟 > 0, 𝑝 < 0.051 + 𝑟!, 𝑟 < 0, 𝑝 < 0.05

1, 𝑝 ≥ 0.05 (2)

This transformed the large positive r values that were

characteristic of agonistic relationships into short distances close to 0 and the large negative r values corresponding to antagonistic relationships into long distances close to 2. Lastly, we used hierarchical clustering across participants on unbiased HVE distance matrix using the linkage function with un-weighted average distance method.

III. RESULTS We found that EMG profiles were highly correlated with

each other, but not with Ia profiles in all three recorded movements (Fig. 1a). Also, Ia profiles were more closely correlated with each other, and less closely with EMG profiles in these movements (Fig. 1a). The clustering further supported this result by showing that EMG and Ia profiles formed two separate groups, with three exceptions (Fig. 1b, exceptions denoted with *). Furthermore, the clustering of Ia profiles tended to be more consistent across movements than the clustering of EMG profiles. The Ia clustering closely followed the functional distinction between synergistic and antagonistic muscle actions around the joints, e.g. Ia profiles of both wrist flexors clustered with each other and all three elbow flexors clustered with each other across all movements (Fig. 1b). In contrast, EMG clustering was varied across movements and less similar to the traditional flexor/extensor distinction. For example, EMG profiles of flexor carpi ulnaris and extensor carpi radialis clustered with each other, so did the long head of biceps and the short head of triceps, but only in movement #3 (Fig. 1b bottom).

IV. CONCLUSION Our results indicate that clustering of Ia profiles is distinct

from that of EMG profiles. Also the clustering of EMG profiles varies across movements more than that of Ia profiles. Together, this suggests that this proprioceptive feedback alone may not be sufficient to shape motor primitives observed in EMGs and that additional task-specific processing by the CNS is required [3-6].

The variability of EMG clustering across movements may be related to the variable dynamic conditions. For example, the clustering of antagonistic muscles in the movement #3 with assistive interaction torques may serve to increase joint stiffness. In contrast, the clustering of predominantly

antigravity flexors in movement #2 may serve to overcome resistive interaction torques and increasing gravity load. Thus, motor primitives may reflect the neural compensation for complex limb dynamics [7].

Fig. 1: The relationships between mean profiles of EMG and Ia across subjects. (a) Correlation matrixes. (b) Hierarchal clustering of HVE based on correlation matrixes in (a). Short linkages indicate close relationships between signals. Muscle abbreviations: ADelt and PDelt – anterior and posterior deltoids; Pec – pectoralis major; TMaj – teres major; BicL and BicS – biceps brachii long and short heads; TriL and TriS – triceps brachii lateral and short heads; BR – brachioradialis; ECR – extensor carpi radialis; FCR – flexor carpi radialis; FCU – flexor carpi ulnaris.

REFERENCES [1] A. Prochazka, “Quantifying proprioception”. Prog Brain Res, vol.

123, 1999, pp. 133–142. [2] K. R. Saul, et al. “Benchmarking of dynamic simulation predictions in

two software platforms using an upper limb musculoskeletal model.” Comp Meth Biomech Biomed Eng, vol. 18, 2015, pp. 1445–1458.

[3] I. Kurtzer, F. Crevecoeur, and S. H. Scott. “Fast feedback control involves two independent processes utilizing knowledge of limb dynamics.” J Neurophysiol, vol. 111, 2014, pp. 1631–1645.

[4] E. Chapman, M. Bushnell, D. Miron, G. H. Duncan, and J. Lund. “Sensory perception during movement in man.” Exp Brain Res, vol. 68, 1987, pp. 516–524.

[5] A. Prochazka, V. Gritsenko, and S. Yakovenko. “Sensory control of locomotion: reflexes versus higher-level control”. Adv Exp Med Biol, vol. 508, 2002, pp. 357–367.

[6] S. Yakovenko, V. Gritsenko, and A. Prochazka. “Contribution of stretch reflexes to locomotor control: a modeling study.” Biol Cybern, vol. 90, 2004, pp. 146–155.

[7] V. Gritsenko, J. F. Kalaska, and P. Cisek. “Descending corticospinal control of intersegmental dynamics.” J Neurosci, vol. 31, 2001, pp. 11968–11979.

EMG

Ia F

iring

Rat

e

EMG Ia Firing Rate

EMGIa Firing Rate

(a) (b)

Mov

emen

t 1M

ovem

ent 2

Mov

emen

t 3

EMG Ia Firing Rate

EMG

Ia F

iring

Rat

eEMG Ia Firing Rate

EMG

Ia F

iring

Rat

e

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

r

FCRFCU

BicLECRPecBRBicS

TMaj*TriS

ADeltPDelt

TMajPecBicL

PDeltBR

FCRADelt BicS TriL FCU

ECRTriS

TriL*

0.8 0.6 0.3 0.0

BicLBR

BicSFCR

FCUFCRPecTriS

ADeltPDelt

TMaj

TriLPec

BicLFCR

ECRADelt

TriS FCU TMajTriLPDelt

BicSBR

0.9 0.6 0.3 0.0

PDelTMaj

ADeltTriSPecTriLTMaj

TriSBicL

FCRFCU

ECR

BR

PDelt

ADeltBicS*

BicLBicS BR ECR

PecTriL

FCRFCU

0.9 0.6 0.3 0.0

Abstract— In this paper we describe a procedure which gives

the opportunity to evaluate manually performed therapy

methods with regard to their automation potential. This is very

important because automated movement therapy for

neurological rehabilitation includes a huge number of

advantages such as the relief of therapists and the purposive

enhancements of the patients’ training. The example of

hippotherapy serves to successfully validate the functionality of

the developed evaluation system.

I. INTRODUCTION

A. Motivation and aim of project

S part of the demographic change, the life

expectancy is increasing and the population is growing

older. This causes a higher need of medical care. Especially

the rise of neurological diseases and the lower stroke

mortality rate requires a higher capacity in the area of

rehabilitation. To generate this capacity, it is essential to

develop automated assistive devices for therapy. One

approach is the automation of existing and effective manual

therapy methods for patients with neurological diseases and

their application within the early phase of rehabilitation.

This serves to provide existing manual therapy methods to a

larger group of patients. It is very difficult to identify a

manual therapy method which can be automated wisely. One

reason is that therapists and engineers communicate in

different ways and speak different languages [1].

Furthermore, the idealism is not in agreement with the

occurring costs. For these reasons, some brilliant ideas of

therapists do not find their way into the stage of

development. This paper introduces an evaluation system as

a possible approach to tackle this problem.

B. State of the Art

In rehabilitation, very few therapists are responsible for a

high number of patients. This leads to the fact that patients

exercise too little and at rare intervals. In the early phase,

especially after an acute stroke, it is possible for patients to

quickly and considerably improve their physical conditions.

This can be reached by systematic and continuous exercising

and the use of the brains’ neuronal plasticity [2]. Current

inquiries have shown that in the early phase of rehabilitation

(phase A & B based on the phase model of neurological

rehabilitation [3]), there are hardly any or even no automated

device-supported therapy methods available. A possibility to

improve this situation is to develop evaluation systems to

M. Klöckner and B. Kuhlenkötter are employed at the Chair of

Production Systems, Ruhr-Universität Bochum, Germany. {kloeckner;

kuhlenkoetter}@lps.rub.de

identify the automation potential of existing manual therapy

methods.

Such systems are used in different application fields, e.g.

in software engineering [4] and in assembly processes [5]. In

the area of rehabilitation, specific safety and functionality

rules apply, which leads to the unwanted fact that

automation potential identification in this area is an

enormous challenge

II. MATERIAL & METHODS

A. Identification of automation relevant criteria

Specific criteria in evaluation systems are needed to

analyze the regarded processes/methods and make them

rateable. The identification of all relevant criteria for an

evaluation system is a complex iterative process. As a start,

we have built five relevant classes and pointed out their

relevant interconnections (cp. Fig. 1).

Fig. 1. Identified classes for criteria

The start event are the patient influences (1). They are

also the basis for the correct choice of therapy methods,

make requirements for the safety design and influence the

system design. Defined criteria are the human dimension,

phase of disease and patient weight. The therapy method (2)

and the engineered system (3) represent the complete

system. Moreover, the therapy method poses requirements

on the safety design and includes the criteria of the number

of repetition, accuracy and increase possibility of therapy.

The engineered system (3) consists of mechanic, electronic

and software components. Important criteria are the stability

of the setting, degrees of freedom, number of motors,

number of joints, parametrization of the therapy, feedback,

assistance as needed, individual settings of the driving path

and sensor systems. Standards and guidelines (4) define

Mechanics

Software/

Information

engineering

Elektronics/

Actuating

elements

Engineered system / final result

Therapy method

Patient

influencesSafety

Standards &

Guidelines

Complete system

Startevent

System boundary

Rehabili

tation s

pecific

1

2

3

45

Automated Movement Therapy for Neurological Rehabilitation

M. Klöckner and B. Kuhlenkötter

A

requirements for safety including the design and

construction of the systems’ man-machine interaction. The

regarded safety (5) criterion brings about therapy limitations

through safety arrangements.

B. Quantification of automation relevant criteria

Due to the patient-individual processing of the therapies,

a generalized quantification of the identified criteria is very

laborious. To perform a quantification of each criterion, we

have defined seven categories, which are grouped depending

on the affected body part. These categories include pelvis,

shoulder, knee & elbow, finger & wrist & anklejoint, spine,

hip, neck & cervical spine.

C. Evaluation method - theory

We have used a simple evaluation method, the output of

which is given in % (max 100 %), and divided it into two

parts. The focus of the first part rests on the existing manual

therapy method including criteria belonging to patient

influences, the chosen therapy method and safety classes.

The second part centers on various engineering possibilities

for a realization. Possible evaluation values for the criteria

are chosen from 1 to 5. Each value has an allocated meaning.

A higher value includes a superior automation potential. The

evaluation part results are calculated by the values of

quantifiers, the values of evaluation and the number of

regarded criteria. Evaluation part 1 supports the systematic

identification of manual therapy methods considering the

automation possibility with therapists’ and patients’ help.

Automation is possible and evaluation part 2 can be

executed if the result is more than 50 %. Evaluation part 2

evaluates the engineering possibilities of the chosen method.

Thereby a possible realization is checked by generating and

evaluating concepts.

III. RESULTS

Even therapeutic processes which are highly standardized

depend on individual patients, the diagnosis, the current

therapy procedure and the attending physician [6]. However,

it is shown on the hippotherapy example that an

improvement of manual therapy methods identification with

automation potential is possible. In hippotherapy typical

movement patterns, which are similar to the human walking

pattern, are transmitted to the human body while the patient

(neurological diseases) is riding a specially trained horse.

In evaluation part 1, hippotherapy gets a result of 85.24 %

(Table 1). This result includes the opportunity to execute

evaluation part two. To show that it is possible to use this

method, two concepts are regarded where producers have

adapted the trajectory of hippotherapy to automated systems:

JOBA EU 7805 (Panasonic) and HIROB (intelligent motion)

(Table 2). Joba is fixed on the ground and has three different

programs. Hirob is a combination of an industrial robot and

a replication of a horse back.

Joba gets a better result than Hirob because it is a very

simple realization of the hippotherapy and has only three

different programs which cannot be parameterized in a

patient-individual way. Hirob is a more laborious realization

of hippotherapy, which includes more features than Joba.

IV. CONCLUSION

Especially in the area of automation potential

identification in neurological rehabilitation, it is very

important to develop an evaluation system. It allows the

therapist to recognize early where an automated assistive

device would be possible. Although it is a huge challenge to

develop a generalized evaluation system for this application

area, the regarded hippotherapy example shows that it is

possible to use an evaluation system wisely for this task.

A further development of the evaluation system includes

the aim to strengthen the cooperation of therapists, patients,

medical doctors and developers. Furthermore, it can build a

communication basis which is acceptable and definite for

each of these groups. This approach is a perfect way to

enlarge and scoop out the potential of rehabilitation and to

handle the increasing number of patients.

REFERENCES

[1] J. Aston, E. Shi, H. Bullot, R. Galway, J. Crisp, Qualitative evaluation of regular morning meetings aimed at improving interdisciplinary

communication and patient outcomes, International Journal of Nursing

Practice 11 (2005), vol. 5, pp. 206–213. [2] J. A. Kleim, T. A. Jones, Principles of Experience-Dependent Neural

Plasticity, Journal of Speech Language and Hearing Research 51

(2008), vol. 1, p. 225. [3] P.W. Schönle, Frühe Phasen der Neurologischen Rehabilitation,

Allensbacher Forschungsinstitut für Rehabilitationsneurologie und

Neuropsychologie, 1996. [4] K. Hildebrand, Software Tools: Automatisierung im Software

Engineering, Berlin, New York, Springer-Verlag, 1990.

[5] M. Naumann, J. Wößner, Automatisierungspotenzialanalyse für die Montage, 2014.

[6] W. Korb, R. Riener, Cooperative Human-Machine Systems in Surgery

and Rehabilitation, In: Automatisierungstechnik 55 (2007), Nr. 10.

TABLE 2

EVALUATION PART II – JOBA 74,17 % HIROB 53,61 %

Criteria Quantifier Evaluation

Joba - Hirob

Stability of setting 0,01 4 3 Degrees of freedom 0,14 3 1

Number of motors 0,08 3 1

Number of joints 0,07 3 3 Parametrization of the therapy 0,15 3 2

Feedback 0,10 1 4

Assistance as needed 0,18 5 4 Individual settings of the driving path 0,14 5 3

Sensor systems 0,13 5 5

TABLE 1

EVALUATION PART I – HIPPOTHERAPY 85,24 %

Criteria Quantifier Evaluation

Human dimension 0,02 3 Phase of disease 0,21 4

Patient weight 0,10 5

Number of repetition 0,24 5 accuracy 0,19 3

Different therapy levels 0,19 5

Limitation of therapy (safety arrangements)

0,05 3

How do we plan movements?:A geometric answer

Rakesh Krishnan, Niclas Bjorsell, and Christian Smith

Abstract— Human movement is essentially a complex phe-nomenon. When humans work closely with robots, understand-ing human motion using robot’s sensors is a very challengingproblem. This is partially due to the lack of proper consensusamong researchers on which representation to use in suchsituations. This extended abstract presents a novel kinematicframework to study human intention using hybrid twists. Thisis important as the functional aspects of the human shoulder areevaluated using the information embedded in thoraco-humeralkinematics. We successfully demonstrate that our approachis singularity free. We also demonstrate that how the twistparameters vary according to the movement being performed.

I. INTRODUCTION

As we progress towards robot-assisted neuro-rehabilitation, robots are expected to have automaticunderstanding of human intention. In the case of humanupper limb, it is the fine motor skills that enable us toperform successful manipulation. The human shoulderacts as a base for the forearm and the hand in everydaymanipulation tasks. But, we are presented with importantchallenges in attempting to paramatrise human shouldermotion; like its complex anatomy, its large range of motioncapability, kinematic redundancy, and inherent motionvariability.

Current approaches in modelling human shoulder kine-matics is not suited for high-reliability applications likeneuro-rehabilitaion [1]. Because, numerical instabilities inwidely accepted shoulder kinematic representations intro-duces singularities, which in turn compromises the reliability.Therefore, we present a special class of spatial vectors thathas the potential to be useful in such applications. Also, wedemonstrate that our approach is singularity-free in SectionIII. The results discussed in this extended abstract is a subsetof our work under review [4]. We present a brief overviewof hybrid-twists in the next section.

II. HYBRID-TWISTS

For any generalized rigid body, the instantaneous velocityof a body-fixed point R is given in terms of the velocity vA

of the origin of body-fixed coordinate system A and angularvelocity vector ω ∈ R3 expressed as

vR = vA + ω × pAR. (1)

This work is supported by AXO-SUIT (AAL Call 6) project.R. Krishnan is with the CVAP lab at KTH (Royal Institute of Technol-

ogy), Stockholm and with the Department of Electronics at University ofGavle, Sweden (e-mail:rkth@kth.se, rahkrn@hig.se).

N. Bjorsell is with Department of Electronics at University of Gavle,Sweden (e-mail:nbl@hig.se).

C. Smith is with the CVAP lab at KTH (Royal Institute of Technology),Stockholm (e-mail:ccs@kth.se) .

Humeral Triad

Pedal point

Distance

ξ

Y

Z

XO

Global frame

T1

T2

T3

Thoracic Triad

H1

H2

H3

Fig. 1. Figure illustrates the concept of hybrid-twists in describing shouldermotion. The markers T1-T3 constitute the thoracic triad and H1-H3 formsthe humeral triad.

Any generalised motion of a rigid body as in (1) can beparametrised using spatial vectors or twists [2]. A generalisedhybrid-twist in terms of instantaneous velocities is defined as(2).

ξ =

[ωvA

]DA. (2)

The basis vector DA, that defines hybrid-twists belong to thePlucker basis. The axis defined by the hybrid-twists representa line in 3D-space along which the rigid body velocities areof minimum magnitude. Hybrid-twists have several advan-tages like they are singularity-free and decoupling in segmentkinematics. The concept of hybrid twists in parametrisinghuman shoulder kinematics, is shown in Fig.1. In the nextsection we will describe our experiments and computation.

III. EXPERIMENTS AND COMPUTATION

We mounted passive markers on brace supports on ahealthy male subject (aged 20 years, weight: 80 kg, dominantright hand). A total of six markers were mounted on thesubject: T1-T3 on the thoracic segment and H1-H3 on thehumeral segment. Using an Optotrak MOTIVE 17-camerasystem, the motion of the passive markers were capturedat a rate of 120 fps. The volunteer was asked to performthree basic shoulder motions at a self-selected pace, namely:1) Flexion-Extension in the sagittal plane, 2) Abduction-Adduction in the coronal plane and 3) Elevation-Depressionwhich is generalised raising and lowering motion of thearm. From the marker data the rigid segment velocities werecomputed following the procedure in [3]. Each movementwas repeated five times.

Then, the relative kinematics between the thoracic andhumeral segments were computed using steps in [2]. We thencalculate the pitch (h) of the screw given by the following

h =ω · vA

ω · ω= ±

‖vlin‖‖ω‖

. (3)

The value of h describes the instantaneous relationshipbetween the angular velocity and translational velocity: zerorepresents pure rotation, a positive value represents a right-handed screw, and a negative value represents a left-handedscrew. The results have been shown in the next section.

IV. RESULTS AND DISCUSSION

The figures Fig.2-Fig.4 presents the pitch values for thesubject during three basic shoulder movements. As can beseen in the results, the computed pitch curves have distinctfeatures for each of these activities. For instance, duringthe elevation-depression task there is a positive pitch duringelevation and corresponding negative pitch throughout thedepression phase as can be seen in Fig.4. Similarly, individ-ual pitch features can be seen in both flexion-extension (seeFig.2) and abduction-adduction (see Fig.3) movements.

As can be seen in our results, there are ringing features dueto the wobbling effects of the segment soft tissue. We wouldlike to propose pre-processing steps towards mitigating these;only then can we demonstrate the theoretically guaranteedhigh-reliability.

Despite our measurement limitations, it is evident that weare able to extract movement-related invariant features con-sistently. Parametrising complex shoulder movements usinghybrid-twists is an interesting problem we would like toexplore further. In the future, we look forward to answeringquestions related to movement planning using hybrid-twists.

0 2 4 6 8 10 12 14 16 18-10

-8

-6

-4

-2

0

2

4

6

8

10

12

Time, s

Pit

ch, h

Initial flexion

Later extension

Fig. 2. Computed pitch values for flexion-extension task. Early flexion ismarked by blue arrows and later extension is shown by brown arrows.

V. CONCLUSION

In short summary, we have proposed a well known kine-matic parametrisation in robotics onto parametrising humanmovement. Despite our measurement limitations, we haveshown that our approach can highlight movement-dependentkinematic features without being affected by numerical sin-gularities. Thus, hybrid-twists do hold a potential towardssuccessfully parametrising human shoulder kinematics.

0 2 4 6 8 10 12 14 16 18 20-10

-5

0

5

10

15

Time, s

Pit

ch, h

Initial adduction

Later abduction

Fig. 3. Computed pitch values for abduction-adduction task. Laterabduction phase is indicated by green arrows and early adduction by brownarrows.

0 2 4 6 8 10 12 14 16-6

-4

-2

0

2

4

6

8

10

Time, s

Pit

ch, h

Depression

Elevation

Fig. 4. Computed pitch values for elevation-depression task. Elevationphase is marked by green arrows and depression phase is marked by pinkarrows.

VI. ACKNOWLEDGMENT

We thankfully acknowledge the participant of this study.

REFERENCES

[1] J. Pons, Rehabilitation exoskeletal robotics, IEEE Eng. Med. Biol.Mag., vol. 29, no. 3, pp. 5763, 2010.

[2] R. Featherstone, Rigid Body Dynamics Algorithms, Springer US, 2008.[3] J. Angeles, Fundamentals of Robotic Mechanical Systems : Theory ,

Methods , and Algorithms., Second Edi. 2003.[4] R. Krishnan, N. Bjorsell, and C. Smith, Invariant Spatial Parametriza-

tion of Human Thoracohumeral Kinematics : A Feasibility Study, inIEEE/RSJ International Conference on Intelligent Robots and Systems(IROS 2016) (Under Review), 2016.

Abstract—Humanoid robotics finds its natural inspiration in

the study of humans themselves and its objectives in developing

robots with human-like skills. Neurorobotics in particular

exploits findings of neuroscience with the aim to implement bio-

inspired sensory input processing, movement control, decision

making, and more. Moreover neurorobotics provides a way to

test theories of human sensorimotor control and its

dysfunctions in patients by embodying them in real world

devices. The two aspects of designing robot control systems and

understanding the control of humans integrate and improve

each other. This will be shown here in the context of posture

control by presenting the human inspired Disturbance

Estimation and Compensation (DEC) concept in humanoids.

I. INTRODUCTION

A. Overview

OSTURE control is a fundamental task of humanoid

robots, since it is a prerequisite for other tasks such

standing, walking or performing a manipulation. This owes

to the intrinsically unstable nature of humanoid body stance.

In fact, posture control is an issue also for humans, where

postural adjustments provide the movement buttress that the

action-reaction law of physics prescribes and maintain body

equilibrium by balancing the body's center of mass (body

COM) over the base of support. Impairment of posture

control in humans causes severely disabling syndromes such

as ataxia, with jerky and dysmetric movements and postural

instability [1]. In controlling bipedal balance, humans are

still better than humanoids with respect to robustness and

versatility [2]. Human-likeness of bipedal control both in

walking and balancing is an open research topic in humanoid

robotics [3].

The embodiment of human-inspired neural control in

robots can be performed at several levels of abstractions:

From the reproduction of low level mechanisms such as

simulated spiking neural networks for the control of

voluntary movements [4] or central pattern generators for

gait [5] to the reconstruction of higher level phenomena such

as movement synergies [6]. The DEC concept [7] provides

an intermediate to high-level description of sensor fusion

and posture control. High-level signals are internally used in

the control loop. However, the concept leaves open where in

the brain or in the spinal cord the processing occurs. Also,

the concept takes in account low-level features such as

neural time delays, but does not specify behaviors (like

The work was supported by EU FP7 Grants 600698 and 610454. Authors are with the Uniklink Freiburg, Neurology. Freiburg, Germany

(e-mail: {Vittorio.lippi|Thomas.mergner}@uniklinik-freiburg.de).

synergies), which in the DEC are emerging from interactions

between control elements and body biomechanics. The DEC

concept has been embodied into humanoid robots in order to

test posture control in the same experimental context as

tested in humans (i.e. tracking body position while standing

on a moving platform or being pulled by an external force

acting on the body). Furthermore the concept has been

generalized to full body control of several humanoids

[8,9,10,11].

Fig. 1. Simplified scheme of DEC model.

B. The DEC concept

In the DEC concept, human posture control builds upon a

proprioceptive servo feedback loop that controls voluntary

movements (A in Fig. 1). The control also comprises an

intrinsic musculoskeletal stiffness and damping loop (B) and

DEC loops (C, four in a complete scheme). Estimates of the

external disturbances, achieved through sensor fusions,

command via negative feedback the servo to produce the

extra joint torque that is needed to compensate for the

disturbances. Assuming ideal compensation, the movement

control can function as if there were no disturbances. This

concept can be extended to multi-DOF, as shown in [9] for

two degrees of freedom and in [10] for the general case.

The effects of the external world on the body are described

in DEC by four external disturbances [7]: support surface

rotation (‘platform tilt’) and translation (‘platform

acceleration’) and field forces (such as gravity) and contact

forces (‘external torque’). For the estimates of the

disturbances, humans fuse signals of physical variables,

which they derive in turn from multisensory integration of

vestibular signals, vision, touch, force, and joint

proprioception [12]. Inspiration for this form of multisensory

integration was derived from studies on human self-motion

perception [13].

II. HUMANOID ROBOT CONTROL

A. Robots and tasks

The three humanoids developed at the University Clinics of

Freiburg are shown in Fig. 2. The first humanoid Posturob I

has been used to reproduce human upright stance in the

presence of moderate perturbations, i.e. balancing using the

Humanoid Neurorobotics - Posture, Balance and Movement Control

Vittorio Lippi and Thomas Mergner

P

ankle joints (single inverted pendulum, SIP, scenario). It is

able to compensate the four external disturbances while

performing voluntary movements, and it is able to control

balance on compliant support surfaces (or on a surface

rotating together with the body). Posturob II extends the

capabilities of Posturob I, adding the control of the hip joint.

This allowed the DEC control to reproduce human behavior

in terms of leg and trunk sway in the presence of support

surface tilt, to control the robot during superposition of

disturbances and voluntary movements, and to show the

emergence of inter-link coordination [9, 14]. This robot can

also be used to test body coordination patterns during leg

swing.

Posturob I Posturob II Posturob III, Lucy

Fig. 2. Humanoid robots used for neurorobotics experiments. These

platforms are equipped with a human inspired vestibular system and actuation is torque controlled.

Both Posturob I and II are actuated with pneumatic muscles,

optionally with springs in series (that can also be changed or

removed), demonstrating that the DEC system provides a

robust control in the presence of series elastic stiffness. The

Lucy robot (Posturob III, 14 DOF) was used to show how

the DEC can be applied to both frontal and sagittal planes

when maintaining balance in the three-dimensional space,

while perform voluntary movements in both the planes [11].

The DEC concept has been applied also to the TORO robot

platform from DLR. This robot has a very fast joint torque

control system that allowed simulating a parallel passive

stiffness, demonstrating that the DEC can integrate and

exploit it in the presence of time delays [8]. The DEC

concept is expected to be robust to time delays because they

are present in humans. On the other hand, tolerance to delays

may lower the performance requirements of the control

hardware.

B. Open issues and future work

The ongoing research is focused on the exploitation and

modulation of passive stiffness, the compensation of

coupling forces, prediction of self-produced disturbances

[15], the integration of learning processes, and the creation

of a model for human transient response to external stimuli

[16]. Posture control is expected by us to be a key for human

inspired algorithms for the control of humanoid gait.

Neurorobotics findings on posture control will find an

application in assistive devices for patient rehabilitation and

support.

III. CONCLUSION

Human-inspired sensorimotor control provides a robust

solution for humanoid posture control. This will help to

improve the versatility of humanoid robots and, because

these robots may serve as blueprints for exoskeletons, may

enhance acceptance of assistive devices.

REFERENCES

[1] A.J. Bastian. Mechanisms of ataxia. Physical Therapy, 77(6), 672-

675, 1997.

[2] S. Ivaldi, J. Peters, V. Padois, & F. Nori. Tools for simulating humanoid robot dynamics: a survey based on user feedback, in

Humanoid Robots (Humanoids), 14th IEEE-RAS International

Conference on (pp. 842-849), 2014. [3] D. Torricelli et al. Benchmarking human-like posture and locomotion

of humanoid robots: a preliminary scheme, in Biomimetic and

biohybrid systems. Springer International Publishing 320-331, 2014. [4] E. Rueckert, D. Kappel, D. Tanneberg, D. Pecevski, J. Peters.

Recurrent spiking networks solve planning tasks, in Scientific Reports.

2016; 6: 21142. Doi: 10.1038/srep21142. [5] A.J. Ijspeert, A. Crespi, D. Ryczko, & J.M. Cabelguen. From

swimming to walking with a salamander robot driven by a spinal cord

model, in Science, 315(5817), 1416-1420, 2007 [6] H. Hauser, G. Neumann, A.J. Ijspeert & W. Maass. Biologically

inspired kinematic synergies enable linear balance control of a

humanoid robot, in Biological cybernetics, 104(4-5), 235-249, 2006. [7] T. Mergner, A neurological view on reactive human stance control.

Annual Review Control, 34:77–198, 2010.

[8] C. Ott et al., Good Posture, Good Balance: Comparison of Bioinspired and Model-Based Approaches for Posture Control of

Humanoid Robots, in IEEE Robotics & Automation Magazine, vol.

23, no. 1, pp. 22-33, March 2016. Doi: 10.1109/MRA.2015.2507098, 2016

[9] G. Hettich, L. Assländer, A. Gollhofer & T. Mergner. Human hip–ankle coordination emerging from multisensory feedback control, in Human Movement Sci 37: 123–146. 2006

[10] V. Lippi, T. Mergner, and G. Hettich. A bio-inspired modular system for humanoid posture control. . In Proceedings of the International Conference on Intelligent Robots and Systems (IEEE IROS),Tokyo, Japan, 2013, Workshop on Neuroscience and Robotics, 2013.

[11] V. Lippi, T Mergner, M. Szumowski, M. Zurawska and T. Zielinska. Bioinspired humanoid posture control in frontal plane. Accepted at 21st CISM IFToMM Symposium on Robot Design, Dynamics and Control, Udine, Italy 2016

[12] F. Horak and L. M. Nashner. Central programing of postural movements: adaptation to altered support-surface configurations, Journal of Neurophysiology, vol. 55, pp. 1369–1381, 1986.

[13] T. Mergner, W. Huber, and W. Becker. Vestibular-neck interaction and transformations of sensory coordinates, Journal of Vestibular Research, vol. 7, pp. 119–135, 1997.

[14] M. Zebenay, V. Lippi and T. Mergener. Human-like humanoid robot posture control, in 12th international conference on informatics in control automation and robotics (ICINCO 2015) Colmar France 2015

[15] V. Lippi and T. Mergener. Coupling forces in human-like posture control, in 2015 IEEE/RSJ IROS Workshop on Dynamic Locomotion and Balancing of Humanoids: State of the Art and Challenges, Seattle, USA, 2015.

[16] V. Lippi, G.Hettich, and T. Mergner. Modeling postural control of support surface translations, in IEEE Humanoids, Workshop on cognition, perception and postural control for humanoids Madrid, Spain, 2014.

Abstract— The development of fully implantable functional

electrical stimulation devices is dependent on the availability of

highly selective, minimally invasive and biocompatible neural

interfaces. The objective was to examine the effect of

stimulation pulse shape on the recruitment selectivity of a novel

combined interfascicular and cuff electrode (CICE) in an in

vitro pig nerve model. The CICE consisted of a multi-contact

cuff electrode with 18 contact sites (6 contacts placed in three

rings, 60° spacing between contacts) and an interfascicular

electrode. The electrode was implanted on 10 median nerves

explanted from seven Danish Landrace pigs. Two different

pulse shapes and 12 different configurations were tested. The

results showed that by combining the use of the two pulses the

selectivity increased.

I. INTRODUCTION

Neural electrode constitutes the interface between the

hardware of the stimulator and the biological tissue of

the body. In some applications, it may be necessary to

activate specific nerve fibers to perform a desired task,

without activating other fibers in the same nerve

simultaneously (referred to as selectivity).

One example of a nerve-based electrode is the

Longitudinal Intrafascicular Electrode (LIFE) that consists

of a thin wire that is implanted inside the nerve fascicles by

penetrating the epineurium. The LIFE has proven to activate

fascicles with a high selectivity in comparison to other

electrodes like the cuff electrode, however the LIFE requires

surgical implantation of several electrodes to interface all or

a major subset of fascicles inside a nerve [1]. The cuff

electrode takes on different approaches to achieve selective

stimulation of nerves. The cuff electrode consists of an

isolating tube with contact sites mounted on the inside,

which encloses the nerve. The simplest cuff electrode has

only two contact sites that will allow whole nerve activation

[2]. A way to enable selectivity is to use a multi-contact cuff

electrode, which gives the opportunity to stimulate with

different contact sites and configurations. A disadvantage of

the cuff electrode is the risk of nerve compression and

potentially nerve damage following implant due to swelling

L. E. Lykholt is with the Center for Sensory-Motor Interaction at Aalborg University, Aalborg, Denmark (corresponding author to provide e-

mail: lel@hst.aau.dk).

W. Jensen is with the Center for Sensory-Motor Interaction, Aalborg University, Aalborg, Denmark (wj@hst.aau.dk).

K. R. Harreby is with the Center for Sensory-Motor Interaction, Aalborg

University, Aalborg, Denmark (krauhe@hst.aau.dk).

of the nerve. Cuff electrodes have proven to be a safe and

stable design in clinical applications [3]. In a study by Riso

et al. the selectivity of an interface that combined a multi-

contact cuff electrode and an intrafascicular electrode was

tested [4], and compared with results from stimulation with

the multi-contact cuff electrode alone. The combined

electrode was able to give modest improvement in achieving

selective activation in muscles. However the interface was

not able to reach absolute selectivity (where only one single

fascicle or muscle is activated without any activation of the

other fascicles or muscles) of the monitored muscle, and the

stimulation pulse waveform will also affect the selectivity

and the safety of the tissue during stimulation [5].

The typically applied pulses are mono- and biphasic

rectangular pulses [6]. Biphasic charge balanced pulses

prevent corrosion of the electrode and tissue damage.

However, charge balancing pulse may tend to block some of

the fibers just activated (if there is no intrapulse delay) thus a

higher amount of charge will be needed than when applying

a single pulse.

A combined interfascicular and cuff electrode (CICE)

design may contribute to a higher selectivity than obtained

with the LIFE or multi-contact cuff electrode alone.

However, there is limited knowledge in the literature that

directly compares the effect of different pulse shapes on the

recruitment selectivity. In addition, the majority of previous

work has been carried out in nerves or animal models with

few fascicles that cannot be directly compared with humans.

The objective of this study was to examine the effect of

stimulation pulse shape on the recruitment selectivity of a

novel CICE in and in vitro pig nerve model.

II. METHODS

In total 10 median nerves from seven Danish Landrace

pigs were used. The pigs were euthanized in accordance

with the Danish law. The median nerves were harvested

from the left and right forelimbs and immersed in a Na-

Krebs solution. In the laboratory the nerve was placed in a

bath that was circulated and oxygenated with Carbox (95 %

O2 and 5 % CO2) at 21°C. The epineurium was opened at the

distal end (i.e. corresponding to the pig anatomy) and up to

31 fascicles was isolated.

The multi-contact cuff electrode part of the CICE

consisted off a silicone tube ring with 18 platinum contact

sites placed in one of three parallel rings. The interfascicular

Effect of Pulse Shape on the Recruitment Selectivity of a Combined

Interfascicular and Cuff Electrode (CICE) in an in vitro Pig Nerve

Model

L. E. Lykholt, W. Jensen, and K. R. Harreby

A

part of the CICE was constructed using a thin Teflon coated

wire that was de-isolated for 2 mm. The cuff electrode part

was placed at the proximal end of the nerve, i.e. the

interfascicular part was implanted first and then the multi-

contact cuff electrode was placed to enclose the nerve. To

record the evoked responses a custom-made tungsten rod

electrode was placed in each fascicle and connected to an

amplifier and sampled. 12 different stimulation

configurations were applied, where either the contact sites at

the center ring or the interfascicular electrode were used as

cathodes. Two different charge-balanced stimulation pulses

were applied: Charge balanced biphasic (P1) and charge

balanced asymmetric biphasic (P2). In both cases the

frequency was 5 Hz. The amplitude range was 0-4000 µA

for nerves 1-6 and 0-2000µA nerves 7-10. There was an

increasing step of 250 µA for nerve 1-6 and of 50 µA for

nerve 7-10. Supramaximal stimulation was applied using P2

(contact sites of ring one and three were used as the cathode,

contacts in ring two as anode).

To assess the stimulation selectivity, the stimulation

artifacts were removed from the recorded signal and the

evoked response was gain corrected, filtered and averaged

across the stimulation intensities. Recruitment curves were

calculated based on the RMS method (period of 10 ms, were

the data was normalized according to the prior

supramaximal stimulation) and smoothed to estimate a 30 %

activation level [7]. To quantify and compare the recruitment

performance a selectivity index (SI) was defined, inspired

from [8] as the activity of a target fascicle (the only fascicle

desired to activate) with the mean activation of the activity

of all non-target fascicles (all the fascicles that was not to be

activated) subtracted. As such, a SI score of 100 %

corresponds to that one fascicle is recruited 100 % (no other

fascicles were recruited) and a score of 0 % corresponds to

that all fascicles are equally recruited (i.e. no selectivity). It

was estimated how many fascicles that were selectively

activated for each of the two pulses and how many times

each pulse was used to achieve the optimal stimulation for

each nerve.

III. RESULTS

During stimulation with increasing intensities an increase

in the response was observed. Some of the responses were

monophasic whereas others were biphasic, but they

maintained the same shape during the increase of the

intensity. We found a high variability in the recruitment

curves, dependent on the used stimulation configuration and

stimulation pulses. A clear relation between the selectivity

and anatomy was found, i.e. the area with the highest SI was

seen in the proximal branch for all nerves except for nerve

10. In some nerves a high SI was also seen for some

fascicles in the distal branches. It was observed that P1 and

P2 had a high degree of resemblance in the activation

patterns. It was observed that by carefully selecting and

combining the two different pulses the SI could be

increased. It was found that both pulses on average activated

between 70 % and 80 % of the fascicles. It was observed that

P1 was used more than P2 for optimal stimulation with the

CICE.

IV. DISCUSSION AND CONCLUSION

The comparison of the CICE and multi-contact cuff

electrode configurations alone is an important factor to

establish the effect of the CICE. It was clearly observed that

a higher overall selectivity was reached with the use of the

CICE throughout the experiments of the 10 nerves. In the

present study the effect of pulse shapes on the stimulation

selectivity of a CICE was investigated. It was observed that

by using a combination of the two different pulses increased

the selectivity compared to using only one of the pulses.

Further perspectives for the work are to include more types

of pulse shapes to investigate if this increases the selectivity

even more. Future research is needed to evaluate if the effect

of the CICE, stimulation configurations, and stimulation

pulses increases the selectivity during the use of FES in a

clinical application.

ACKNOWLEDGMENT

The authors wish to thank Sahana Ganeswarathas for help

with performing the experiment. The authors also wish to

thank the staff at the pathological institute, Aalborg

University Hospital for assistance during the nerve explant.

REFERENCES

[1] J. Badia, T. Boretius, D. Andreu, C. Azevedo-Coste, T. Stieglitz, and

X. Navarro, “Comparative analysis of transverse intrafascicular

multichannel, longitudinal intrafascicular and multipolar cuff electrodes for the selective stimulation of nerve fascicles.,” J. Neural

Eng., vol. 8, no. 3, p. 036023, Jun. 2011.

[2] M. Haugland, “A flexible method for fabrication of nerve cuff electrodes,” Proc. 18th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc.,

vol. 1, pp. 359–360, 1996.

[3] T. N. Nielsen, G. a M. Kurstjens, and J. J. Struijk, “Transverse versus longitudinal tripolar configuration for selective stimulation with

multipolar cuff electrodes,” IEEE Trans. Biomed. Eng., vol. 58, no. 4,

pp. 913–919, 2011. [4] R. Riso, a. Dalmose, D. Stefania, and M. Schuttler, “Addition of an

intrafascicular electrode at the site of application of a multipolar nerve

cuff enhances the opportunity for selective fascicular activation,” 2001 Conf. Proc. 23rd Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., vol. 1,

pp. 2–5, 2001. [5] W. M. Grill and J. T. Mortimer, “Stimulus waveforms for selective

neural stimulation,” IEEE Eng. Med. Biol. Mag., vol. 14, nr. 4, pp.

375-385, 1995. [6] D. R. Merrill, M. Bikson, and J. G. R. Jefferys, “Electrical stimulation

of excitable tissue: Design of efficacious and safe protocols,” J.

Neurosci. Methods, vol. 141, no. 2, pp. 171–198, Feb. 2005. [7] S. Bao and B. Silverstein, “Estimation of hand force in ergonomic job

evaluations.,” Ergonomics, vol. 48, no. 3, pp. 288–301, Feb. 2005

[8] W. Jensen and K. R. Harreby, “Selectivity of peripheral neural interfaces"

Improving Target Localisation with Collaborative BCIs

Ana Matran-Fernandez, and Riccardo Poli

Abstract— The N2pc event-related potential (ERP) appearson the opposite side of the scalp with respect to the visualhemisphere where an object of interest is located. We usedthis ERP to build brain-computer interfaces (BCIs) that ex-tract information on the spatial location of targets in aerialimages shown by means of a rapid serial visual presentationprotocol using single-trial classification. Images were shownto 10 participants at a presentation rate of 6 Hz whilerecording electroencephalographic signals. For each participantwe trained a neural network to predict the x-coordinate of thetargets. We then created collaborative BCIs by combining theoutputs of individual nets to explore the extent to which thelocalisation could be improved using groups. We show thatthe correlations between actual and predicted location improvewith group size (from 0.3 to 0.68 on average for individuals andgroups of size 9, respectively), reaching a maximum of 0.72.

I. INTRODUCTION

Brain-Computer Interfaces (BCIs) convert signals from thebrain into commands that allow users to control deviceswithout relying on the usual peripheral pathways [1]. Onerecent form of BCI explores the augmentation of humancapabilities, e.g., by means of collaborative BCIs (cBCIs) [2],rather than focusing on aiding people with severe motordisabilities, such as those who are locked-in. One of thesenew types of BCI focuses on the augmentation of the humanvisual perception capabilities to speed up the process offinding pictures of interest in large collections of images [3],[4], [5], [6]. This problem is particularly important in counterintelligence and policing [7].

Researchers typically use the Rapid Serial Visual Pre-sentation (RSVP) protocol, in which sequences of imagesare shown at high presentation rates over a fixed area onthe screen [8]. Observers are tasked with detecting targetconfigurations in the stream of images. Targets elicit distinctEvent-Related Potentials (ERPs, e.g., the P300) in elec-troencephalographic (EEG) signals which can be used todetermine which images contain targets [3].

In this work we will focus on the N2pc ERP [9], whichtypically appears 170–300 ms after stimulus onset in con-tralateral electrode sites with respect to the visual hemifieldwhere a target is located, presenting its maximum amplitudeat electrodes PO7/PO8 and P7/P8 [10].

In particular, we will use this ERP to extract informationon the spatial location of targets in complex real-life scenesfor a task of practical utility (aerial image sifting) presented

The early stages of this research were financially supported by the UK’sEngineering and Physical Sciences Research Council (EPSRC), throughgrant EP/K004638/1.

A. Matran-Fernandez (amatra@essex.ac.uk) and R. Poli arewith the Brain-Computer Interfaces and Neural Engineering Laboratory ofthe University of Essex, Colchester, United Kingdom.

(a) (b)

Fig. 1. Examples of (a) target and (b) non-target images.

using the RSVP protocol through cBCIs based on single-trialclassification. If the N2pc can be detected in such a scenario,it could be exploited, for example, to help circumscribethe area of the image where the target is located, therebyspeeding up the job of the person reviewing the potentialtargets detected by a BCI.

Whereas other similar RSVP-based BCI systems requireparticipants to press a key when they see a target (e.g., [3]),we decided against such an approach due to the severaldrawbacks of this method [5], including errors induced byvariations in reaction times and artefact contamination inthe EEG resulting from the keypress. In our system, nomusculoskeletal interactions are required.

II. MATERIAL AND METHODS

A. Data Collection

We gathered data from 10 volunteers with normal orcorrected-to-normal vision (24.5±3.83 years old, four fe-males, five left-handed).

Participants were seated at approx. 80 cm from the LCDscreen where the stimuli were presented. EEG data wereacquired at a sampling rate of 2048 Hz using a BioSemi Ac-tiveTwo system with 64 electrodes following the international10-20 system, plus one electrode on each earlobe, whosemean was used as a reference. Signals were band-pass-filtered between 0.15 and 28 Hz and downsampled to 64 Hzbefore correcting for eye blinks and ocular movements [11].

B. Experimental design

The stimuli for the experiment consisted of 2,400 aerialpictures of London. Images were converted to grayscaleand their histograms were equalised. Picture size was640×640 px2. Target (T) pictures contained a randomlyrotated and positioned airplane, the horizontal location ofwhich was recorded as the x-coordinate of its centroid. Non-target (NT) images were those that did not contain airplanes.Examples of T and NT images can be seen in Fig. 1.

TABLE IMEAN AND STANDARD DEVIATION OF THE CORRELATION COEFFICIENT

AND THE SLOPE OF THE REGRESSION LINE FITTED TO THE TEST SET

ACROSS ALL GROUPS FOR EACH GROUP SIZE.

Size Correlation (µ ± SD) Regression slope (µ ± SD)1 0.30±0.15 0.14±0.082 0.44±0.13 0.42±0.143 0.51±0.10 0.45±0.104 0.56±0.08 0.46±0.085 0.59±0.07 0.47±0.076 0.62±0.05 0.47±0.057 0.64±0.04 0.48±0.058 0.66±0.04 0.49±0.049 0.68±0.03 0.49±0.0310 0.67±0.00 0.47±0.00

Pictures were shown at a rate of 6 images/second insequences (or bursts) of 100. Each burst contained 10 targetsinserted at random positions in the sequence, with the onlyrestriction that there could not be two consecutive targets.

Participants were asked to mentally count the planes theysaw within each burst.

C. Feature extraction and classification

From the original EEG signals we computed four elec-trode differences: (PO7−PO8), (P7−P8), (PO3−PO4) and(O1−O2). We then extracted 200 ms epochs in the interval[200, 400] ms after stimulus onset. Data were referenced tothe mean value of the 200 ms interval before stimulus onset.

For each participant, we used 10-fold cross-validation totrain a neural network (NN) to predict the horizontal positionof targets within images. The training set of each fold wasused to find the optimal number of neurons of the hiddenlayer (10 or 20) and their activation function (hyperbolictangent or sigmoid). The output neuron was linear.

We then created a cBCI which optimally combines theoutputs of individual NN regressors. This was achieved usinga Linear Discriminant Analysis regressor, hence assigningdifferent weights to the different group members whenmaking the prediction of the location of the target.

III. RESULTS

Table I shows, for individuals and groups of sizes 2–10, the average correlation coefficient between the real andthe predicted x-coordinate of the targets on the test set.Fig. 2 shows the predicted vs. the real coordinate of alltarget images in the test set for one group of size 9 witha correlation coefficient of 0.67 and a slope of 0.47.

IV. DISCUSSION

There is information in the N2pc that can be used topredict the location of targets in images shown using theRSVP protocol, but, due to noise, it is difficult to extract itwith a single-user BCI. As we have shown here, however,cBCIs can markedly improve the localisation of targets.

Bigger groups have higher correlations between the pre-dicted and the real coordinates of the object of interest (an

150 200 250 300 350 400 450 500Real x-coordinate

50

100

150

200

250

300

350

400

450

500

Pred

icte

d x-

coor

dina

te

Fig. 2. Predicted vs. real x-coordinate of targets, for a group of size 9.

airplane) within an image, whilst increasing the slope of theregression line that can be fitted to the predicted locations.

The results in Table I are averages across groups of agiven size. However, the maximum correlation and slopewere as high as 0.72 (for a group of size 8) and 0.64 (fora group of size 3), respectively. These two measures givedifferent information regarding the capability of predictionof the system: correlations assess how much information iscaptured by the linear model, whereas the slope helps todetermine how a change in the input reflects on the output.Interestingly, the group sizes that maximise one measure donot necessarily maximise the other. In the future we willapply a participant selection method to carefully form thegroups to explore how it affects these two measures.

REFERENCES

[1] L. A. Farwell and E. Donchin, “Talking off the top of your head:toward a mental prosthesis utilizing event-related brain potentials,”Electroencephalography and Clinical Neurophysiology, vol. 70, no. 6,pp. 510–523, Dec 1988.

[2] Y. Wang and T.-P. Jung, “A Collaborative Brain-Computer Interfacefor Improving Human Performance,” PLoS ONE, vol. 6, no. 5, May2011.

[3] A. D. Gerson, L. C. Parra, and P. Sajda, “Cortically coupled computervision for rapid image search,” IEEE Transactions on Neural Systemsand Rehabilitation Engineering, vol. 14, no. 2, pp. 174–179, Jun.2006.

[4] L. Parra et al., “Spatiotemporal linear decoding of brain state,” IEEESignal Processing Magazine, no. January 2008, pp. 107–115, 2008.

[5] P. Sajda et al., “In a Blink of an Eye and a Switch of a Transistor:Cortically Coupled Computer Vision,” Proceedings of the IEEE,vol. 98, no. 3, pp. 462–478, Mar. 2010.

[6] H. Cecotti, A. Marathe, and A. Ries, “Optimization of single-trialdetection of event-related potentials through artificial trials,” IEEETransactions on Biomedical Engineering, vol. 62, no. 9, 2015.

[7] Y. Huang et al., “A framework for rapid visual image search usingsingle-trial brain evoked responses,” Neurocomputing, vol. 74, no. 12,pp. 2041–2051, 2011.

[8] K. I. Forster, “Visual perception of rapidly presented word sequencesof varying complexity,” Perception & Psychophysics, vol. 8, no. 4,pp. 215–221, 1970.

[9] S. J. Luck and S. A. Hillyard, “Spatial filtering during visual search:evidence from human electrophysiology,” Journal of ExperimentalPsychology: Human Perception and Performance, vol. 20, no. 5, pp.1000–1014, 1994.

[10] S. Luck, “Electrophysiological correlates of the focusing of attentionwithin complex visual scenes: N2pc and related ERP components,”Oxford Handbook of ERP components, 2012.

[11] P. Quilter, B. MacGillivray, and D. Wadbrook, “The removal of eyemovement artefact from EEG signals using correlation techniques,”in Random Signal Analysis, IEEE Conference Publication, vol. 159,1977, pp. 93–100.

Cycling with plantar stimulation increases cutaneomuscular-conditionedspinal excitability in subjects with incomplete spinal cord injury

Stefano Piazza1, Diego Serrano Munoz2, Julio Gomez-Soriano3, Diego Torricelli1,Gerardo Avila-Martin2, Iriana Galan-Arriero2, Jose Luis Pons1,4 and Julian Taylor2,5,6

Abstract— The aim of this study was to investigate the effectsof a rehabilitation exercise for people with incomplete SpinalCord Injury (iSCI), based on cycling and combined afferentelectrical stimulation (ES-cycling), to normalize spinal activityin response to a plantar cutaneous stimulation.

We studied Soleus H-reflex excitability following ipsilateralplantar electrical stimulation applied at 25-100ms inter-stimulusintervals (ISI´s), on 13 non-injured subjects and 10 subjects withiSCI. Reflexes were tested before and after a 10 minutes sessionof ES-cycling to evaluate the effects of the exercise. Plantar-conditioned H-reflex modulation increased in the iSCI groupafter ES-cycling, compared to the limited modulation observedbefore the exercise. Conversely, the non-injured group presentedpronounced modulation both before and after the exercise.

We conclude that ES-cycling improved plantar-conditionedspinal neuronal excitability in subjects with iSCI. Results couldbe used in the design of more effective leg-cycling therapies,to promote central neuroplasticity and rehabilitation in lowerlimb muscle activity following iSCI.

I. INTRODUCTION

After iSCI, the afferent feedback becomes essential for thereorganization of spinal circuits in the spinal levels belowthe lesion. To this account, afferent feedback potentiation bymeans of electrical stimulation have been previously appliedin association with functional movements, as walking [2] orcycling [6], with promising results.

Electrical stimulation has also been used to evaluate spinalprocessing of afferent information [3], [5]. The modulation ofSoleus H-reflex excitability following a plantar-conditioningstimulus, applied at a 3-90 ms inter-stimulus interval (ISI),was shown to reflect spinal modulatory mechanisms [3] thatare altered following iSCI [2].

Accordingly, in our previous study we studied cyclingeffects on plantar-conditioned H-reflex modulation in non-injured subjects and subjects with iSCI, showing minimal

This work is partially supported by: the Commission of the EuropeanUnion, FP7-ICT-2013.2.1-611695 (BioMot). This work is partially sup-ported by the Commission of the European Union, FP7-ICT-2013.2.1-611695 (BioMot), and the Ministerio de Ciencia e Innovacion - Spain,CSD2009-00067 (Hyper).

1 Neural Rehabiliation Group of the Spanish National ResearchCouncil, Madrid, Spain (corresponding author to provide e-mail:stefano.piazza@csic.es).

2 Sensorimotor Function Group, Hospital Nacional de Paraplejicos,Toledo, Spain.

3 Toledo Physiotherapy Research Group (GIFTO), Nursing and Physio-therapy School,Castilla La Mancha University, Toledo, Spain.

4 Tecnologico de Monterrey, Monterrey, Mexico.5 Stoke Mandeville Spinal Research, National Spinal Injuries Centre,

Aylesbury, UK.6 Harris Manchester College, University of Oxford, Oxford, UK.

but significant changes [4]. Here, we analyze the effects gen-erated by the combination of cycling with artificial afferentfeedback (ES-cycling) in non-injured subjects and subjectswith iSCI. Our hypothesis is that the exercise would increasecutaneomuscular-conditioned H-reflex modulation, altered insubjects with SCI.

II. MATERIALS AND METHODS

A. Participants

Thirteen healthy volunteers and ten subjects with iSCI(AIS C-D) participated to the study. The inclusion criteriafor subjects with iSCI included a neurological level of lesionbetween C5 and T10, less than ten months from the injury,and the ability to cycle autonomously. The exclusion criteriaincluded epilepsy, pregnancy, lower limb musculoskeletalinjury or peripheral nervous system disorders. An informedconsent was collected from each participant. The proto-col was approved by the Toledo Hospital clinical ethicalcommittee (CEIC 07/05/2013, CEIC 15/07/2013 and CEIC17/1/2012).

B. Procedures

Participants were asked to cycle for 10 minutes on a staticergometer, at a speed of 42 rpm and gear set by the subjectto comfortable. A non-noxious afferent electrical stimulationwas applied to the right foot of the subjects during the down-stroke cycling phase, mimicking the cutaneous feedback per-ceived during walking. Stimulation consisted in 1ms pulsesat 200Hz, at the maximal intensity just below the generationof visible muscle contraction.

Plantar-conditioned Soleus H-reflex modulation was as-sessed before and after the exercise, by delivering twostimuli (IntFES, Technalia, San Sebastian, Spain) at an Inter-Stimulus Interval (ISI) of 25, 50, 75 and 100m, the first toconditioning and the second to elicit the Soleus H reflex.The conditioning stimulus consisted in a train of five non-noxious 1-ms pulses at 200Hz, with an intensity of 80%the amplitude required to generate reflex activity in TAEMG. As anode, a 10mm cup electrode was placed betweenthe first and second metatarsal joints; as cathode, a gelledsurface electrode was placed in the dorsum of the foot, incorrespondence to the anode. The eliciting stimulus consistedin a 1-ms electrical pulse delivered to the right tibial nerve bya superficial bipolar electrode applied on the popliteal fossa,with an intensity set to generate a control H-reflex of 50% themaximum H-reflex amplitude. Reflex testing was performedwith the right pedal fixed at 90o crank position. H-reflexes

were recorded 15 times for each condition (without plantarconditioning and each ISI’s), with a minimum test intervalof 7s. The M-wave was also monitored to verify the constantactivation of the alpha motoneuron’s fibers.

EMG signals from the Soleus and TA muscles of the rightlower limb were recorded using bipolar surface electrodesand a conventional amplifier.. Signals were digitized at asampling rate of 2 KHz, and processed with a zero-phase-lag Butterworth second-order bandpass digital filter (20-400 Hz and 30, 50 and 100Hz notch filter). The H-reflexand M-wave amplitude were measured as the peak-to-peakvalues analyzed within a 37-69 and 10-25ms time window,respectively.

During the whole protocol, participants were seated on astationary ergometer, with both feet firmly tied to the pedals.An ankle-foot orthosis prevented ankle movements. A moviewas presented to entertain and distract the subjects.

C. Data Analysis and Statistics

R (R Core Team, 2015) and lme4 [1] were used to performlinear mixed effects analysis of H-reflex excitability. Thedependent variable of the model was the normalized con-ditioned H-reflex amplitude. As predictors, the ISI intervalwas entered into the model as a fixed effect, and by-subjectrandom intercepts were defined as random effects. Theprobability p-values were calculated using Satterthwaite’sapproximation, considering the significance level for theanalysis at 0.05, 0.01 and 0.001, using a factor of 4 as theBonferroni adjustment for multiple comparisons.

III. RESULTS

In the non-injured group, the analysis of cutaneomuscular-evoked Soleus H-reflex modulation before ES-cycling re-vealed inhibition at short ISI’s (-10.6±3%, p=0.001 at 25 msISI, -7.8±3%, p=0.04 at 50 ms ISI, 10±3%) and excitationat long ISI (10±3%, p=0.004 at 75 ms ISI and 9.4±3%,p=0.008 at 100 ms ISI). In iSCI, this modulation presentedan inhibition at 25ms ISI (-9,1±2.9%, p=0.008) and wasabsent for all other ISI’s.

After ES-cycling, reflex inhibition was found in the non-injured group at 25 ms ISI (-9.3±3%, p=0.009) and exci-tation at long ISI (16.1±3% at 75ms ISI, p<0.0001 and19.2±3% at 100 ms ISI, p<0.0001). In the iSCI group,the reflex inhibition previously observed at 25 ms ISI dis-appeared, but reflex excitation emerged at 50 ms (17±5%,p=0.006) and 75 ms ISI (20±5%, p=0.006) (Fig.1).

IV. DISCUSSION

Data analysis showed that cycling combined with affer-ent electrical stimulation promotes cutaneomuscular H-reflexmodulation in individuals with iSCI. After a conditioningstimulation, H-reflexes measured in non-injured subjectspresented inhibition at short-ISI (25-50ms) and excitationat long ISI (75-100ms), which is in line with [2] and [4].These results strongly support the notion that signals fromfoot mechanoreceptors participate in spinal reflex circuitsmodulation.

Fig. 1. Soleus H reflex amplitude versus conditioining stimulation, fornon-injured and iSCI groups, before and after the ES-cycling task.

The modulation was absent in subjects with iSCI. Dif-ferences in the modulation profile observed between non-injured subjects and subjects with SCI confirms our previousresults [4] and are in line with [3].

However, 10 minutes of ES-cycling strongly increased H-reflex excitability at 50 and 75ms ISI in the iSCI group,suggesting the start of a process of adaptation. Fung etBarbeau [2] showed that Soleus H-reflexes modulation insubjects with iSCI was increased during walking when a sim-ilar conditioning stimulation was used, which may explainthe results. On the contrary, non-injured subjects presentedlittle changes after ES-cycling, suggesting that the afferentfeedback may have effect only when the spinal network isnot fully expressed.

Our results suggest that the excitation of plantar cutaneousafferents during a cycling task might promote sensorimotorprocessing. These findings are relevant to the design of noveltherapies to reeducate spinal networks to improve rehabilita-tion.Full analysis of participants’ clinical and demographiccharacteristics, and comparison with normal cycling, will bepresented in a companion article, currently under revision.

REFERENCES

[1] D. Bates, M. Machler, B. M. Bolker, and S. C. Walker. Fitting linearmixed-effects models using lme4. Journal of Statistical Software,67(1):1–48, 2015.

[2] J. Fung and H. Barbeau. Effects of conditioning cutaneomuscularstimulation on the soleus H-reflex in normal and spastic paretic subjectsduring walking and standing. Journal of neurophysiology, 72(5):2090–104, nov 1994.

[3] M. Knikou. Plantar cutaneous input modulates differently spinal reflexesin subjects with intact and injured spinal cord. Spinal cord, 45(1):69–77, jan 2007.

[4] S. Piazza, J. Gomez-Soriano, E. Bravo-esteban, D. Torricelli, G. Avila-martin, I. Galan-arriero, J. L. Pons, and J. Taylor. Maintenance ofcutaneomuscular neuronal excitability after leg-cycling predicts lowerlimb muscle strength after incomplete spinal cord injury. ClinicalNeurophysiology, 2016.

[5] D. G. Sayenko, A. H. Vette, H. Obata, M. I. Alekhina, M. Akay,and K. Nakazawa. Differential effects of plantar cutaneous afferentexcitation on soleus stretch and H-reflex. Muscle & nerve, 39(6):761–9, jun 2009.

[6] T. Yamaguchi, T. Fujiwara, K. Saito, S. Tanabe, Y. Muraoka, Y. Otaka,R. Osu, T. Tsuji, K. Hase, and M. Liu. The effect of active pedalingcombined with electrical stimulation on spinal reciprocal inhibition.Journal of electromyography and kinesiology, 23(1):190–4, feb 2013.

Abstract— Adaptation to a force-field during arm reaching

results in reducing trajectory errors as result of muscle co-

activation and adaptive neural commands. The nature and role

of the neural changes responsible for adaptation to a novel force-

field are still unclear. Here we investigated the effects of motor

adaptation on the brain in terms of Event Related Potentials

(ERPs) as a first exploratory step. Subjects performed reaching

trials during a null force field, then during a velocity-dependent

force field and then again during a null force field. Reaching

trajectory error initially increased during adaptation but then

decreased gradually during motor adaptation. During the whole

adaptation period the activity of motor and premotor areas in

both hemispheres significantly decreased immediately before the

start of the movement. ERPs are an easy-to-measure evidence of

motor adaptation and encourage a deeper investigation of this

phenomenon.

I. INTRODUCTION

Conceptually, the underlying strategy operating during motor

adaptation to a physical disturbance has been viewed as the

combination of online error correction, as a result of sensory

(mainly proprioceptive) feedback and the development of a

new internal model of the new skill [1][6]. To date however,

very little is known about the changes in neural activity that

this process. Here we present a first investigation of neural

correlates during motor adaptation to novel force-field using

electroencephalography (EEG).

II. MATERIALS AND METHODS

A. Experimental Protocol

Twenty-three right-handed healthy adults (20-42 years old)

gave written informed consent to participate in the study.

Subjects sat in front of a shoulder/arm manipulandum

workstation (MIT-Manus, Interactive Motion Technologies,

Cambridge, US) and grasped the end-effector with their right

hand. A vertical screen situated at eye-level illustrated online

feedback on the position of the joystick. Subjects were

instructed to perform straight-line reaching trials from a

central starting point to a peripheral target (15 cm trajectory

length) within 1.0 – 1.2 seconds from the visual cue (i.e.

peripheral target changing color from black to red). The

experimental protocol was based on 3 conditions,

Familiarization (96 trials in null force-field), Motor

Adaptation (96 trials in a 25 Nsm-1 velocity-dependent force-

field counterclockwise direction perpendicular to the

trajectory of the joystick), and Wash Out (96 trials in null

Sara Pizzamiglio is with the Neurorehabilitation Unit of the University of

East London, London, UK (S.Pizzamiglio@uel.ac.uk).

force-field). The reaching task was always in north-west

direction on target board, 135°.

B. Recording Techniques

End-effector position and velocity (along the x and y axes),

and exerted forces (along x, y and z axes) for each reach trial

were automatically recorded at 200 Hz by the robotic device

with sensors incorporated in the joystick. Brain activity (EEG;

μV) was recorded with a high-density 64-channel Waveguard

cap (ANT Neuro, Enschede, Netherlands), amplified with

TMSi Refa_Ext, digitized at 1024 Hz, and band-pass filtered

from 0.1 to 500 Hz. Impedance was kept below 5 kΩ. During

the recording data were referenced to channel Fz.

C. Data analysis – Kinematics

Offline data analyses were run in MatLab 2013b (The

MathWorks, Inc.). All data were temporally aligned at a

frequency of 1 kHz. Reaching movements were described by

a starting time point (movement onset defined by a planar

compound velocity > 0.03 ms-1) and by an end time point

(movement offset defined by a planar compound velocity <

0.03 ms-1). Trial-by-trial trajectory error was quantified by

calculating the summed error (m), defined as the absolute

cumulative perpendicular distance between the actual

trajectory and the ideal straight line connecting central

starting point and peripheral target [5].

Measures were assessed trial-by-trial for each subject and

then averaged for each block trials across subject to obtain the

averaged population behavior per block (i.e. 48 trials).

Statistical analysis then focused on differences between 4

blocks of major interest: late Familiarization (second half,

Baseline), early Adaptation (first half), late Adaptation

(second half), and late Wash Out (second half).

Statistical analyses were run through SPSS 22 (IBM).

Averaged block data were first tested for normality with

Kolmogorov-Smirnoff test: the vast majority of data were

normally distributed so only parametric statistics are

presented. For each measure, 1 way repeated measures

analysis of variance with factor Block (RMANOVA; 4

Blocks) was performed in order to highlight the presence of

any variance across Blocks; paired-sample T-tests with

Bonferroni correction for multiple comparisons were then

employed to define differences between Blocks (p < 0.0125).

D. Data analysis - EEG

EEG data were pre-processed with EEGLab analysis

toolbox following a customized cleaning pipeline. Data were

Duncan L. Turner is with the Neurorehabilitation Unit of the University

of East London, London, UK (d.l.turner@uel.ac.uk).

Neural correlates of adaptation to novel force-field: an exploratory

ERP study.

Sara Pizzamiglio MBioEng, Duncan L. Turner PhD

first band-pass (0.5-100 Hz) and notch (50 Hz) filtered,

divided into epochs of 3 seconds each (-1 sec – 2 sec with

respect to visual cue) and visually inspected to remove bad

trials and any bad channels. Data were then re-referenced to

common average and inspected with Independent Component

Analysis (Infomax): artefactual components (e.g. eye blinks,

muscle artifacts) were rejected. Bad channels were

interpolated, data re-referenced to common average again and

eventually visually inspected to check the quality of the pre-

processing and remove additional bad trials. 3 subjects were

excluded from further analysis because of too many artefacts

within the data. Event-Related Potentials (ERPs) were then

calculated for each subject, condition and channel as simple

mathematical average with the support of Fieldtrip analysis

toolbox for MatLab [4]. After visual inspection of the ERPs

data we defined component-specific time windows of interest

as follow: N/P100 (90 – 110 ms), N/P130 (120 – 140 ms),

N/P170 (160 – 190 ms) and N300 (280 – 360 ms). ERPs

components during Baseline were then compared to early

Adaptation and late Adaptation responses. A third

comparison with late Wash Out was performed in order to

verify the complete disappearance of any adaptation effects.

A cluster-based non parametric permutation test with

Montecarlo technique as provided by FieldTrip toolbox [3]

was employed as statistical analysis on all channels ERPs.

III. RESULTS

Fig. 1. N300 ERPs component analysis. Average ERPs topoplots for the three

conditions (top); Baseline vs. Early Adaptation (bottom-left): statistics (p <

0.05), significant clusters and ERPs traces as averaged across statistically significant electrodes across all participants; Baseline vs. Late Adaptation

(bottom-right): statistics (p < 0.05), significant clusters and ERPs traces as

averaged across statistically significant electrodes across all participants.

A. Kinematics

RMANOVA value for summed error were significant (F =

71.113, p < 0.001). According to paired-samples T-tests, early

Adaptation error was higher than in the other blocks (all p <

0.001). Error returned to baseline after wash out according to

the comparison between Baseline and late Wash Out blocks.

B. EEG

Only the analysis of N300 reported significant differences

between late Familiarization (N300) and both early (N300e)

and late (N300l) Adaptation (Fig. 1). Specifically, both

comparisons reported a positive central cluster straddling the

brain midline (i.e. capturing premotor and motor cortex in

both hemispheres) with 14 electrodes for N300vs. N300e and

12 electrodes for N300 vs. N300l. Both N300e and N300l

were significantly more negative than N300 (p < 0.001).

IV. DISCUSSION

The motor adaptation process has been successfully

described by means of complementary measures, aligning our

work with previous findings. The typical disturbance of

reaching kinematics by a velocity-dependent force field was

accompanied by an increased movement error as shown

previously, which returned to its normal values once the

effects of the perturbation were washed away [8]. For the first

time event related responses associated to motor adaptation to

novel force-field are reported. A negative deflection over

fronto-central areas in both hemispheres is consistent with

findings claiming that motor-related cortical potentials extend

more than 100 ms after EMG onset [7], which are likely

associated to a somatosensory feedback from the movement

itself. The fact that the same negative component significantly

decreased its voltage value consistently during adaptation

could be an evidence of stronger somatosensory feedback

because of the novel force-field applied. Moreover, areas that

showed significant changes are in line with previous studies

of motor adaptation to novel force field using PET [2].

V. CONCLUSIONS

Our findings shed new light on the motor adaptation process

during protocols of force field learning and have some

implications on motor skill evaluation and development of

interventions in neurorehabilitation.

REFERENCES

[1] J. W. Krakauer and R. Shadmehr, "Consolidation of motor memory," Trends in neurosciences, vol. 29, pp. 58-64, 2006.

[2] H. I. Krebs, T. Brashers‐Krug, S. L. Rauch, C. R. Savage, N. Hogan, R.

H. Rubin, et al., "Robot‐aided functional imaging: Application to a motor learning study," Human brain mapping, vol. 6, pp. 59-72, 1998.

[3] E. Maris and R. Oostenveld, "Nonparametric statistical testing of EEG-

and MEG-data," Journal of neuroscience methods, vol. 164, pp. 177-190, 2007.

[4] R. Oostenveld, P. Fries, E. Maris, and J.-M. Schoffelen, "FieldTrip:

open source software for advanced analysis of MEG, EEG, and invasive electrophysiological data," Computational intelligence and

neuroscience, vol. 2011, 2010. [5] R. Osu, E. Burdet, D. W. Franklin, T. E. Milner, and M. Kawato,

"Different mechanisms involved in adaptation to stable and unstable

dynamics," Journal of Neurophysiology, vol. 90, pp. 3255-3269, 2003. [6] R. Shadmehr and T. Brashers-Krug, "Functional stages in the formation

of human long-term motor memory," The Journal of Neuroscience, vol.

17, pp. 409-419, 1997. [7] I. Tarkka and M. Hallett, "Topography of scalp-recorded motor

potentials in human finger movements," Journal of Clinical

Neurophysiology, vol. 8, pp. 331-341, 1991. [8] K. A. Thoroughman and R. Shadmehr, "Electromyographic correlates

of learning an internal model of reaching movements," The Journal of

Neuroscience, vol. 19, pp. 8573-8588, 1999.

Abstract— Considering the spatial and temporal variability of kinematic and kinetic gait patterns typically seen in Myelomeningocele (MMC) patients and based on the term motor synergy –defined as an organization of elemental variables from the central nervous system in order to stabilize a task-specific performance variable, such as footpath during walking– the behavior of this variable for a case study of a low-lumbar level MMC patient is here analyzed. The ultimate aim is to explore the applicability of this global performance variable as control variable of a robotic orthosis for gait rehabilitation in these patients. The results indicate a greater anteroposterior component intra-variability than the vertical one of the right malleolus gait trajectory on the sagittal plane, as well as a larger temporal than spatial intra-variability of both trajectory components. These results render clues to design a suitable control scheme for this type of robotic device.

I. INTRODUCTION YELOMENINGOCELE, a congenital birth defect affecting both sensory and motor functions, habitually

results in muscle paresis proportional to ascending spinal lesion level. Disorders in spatial and temporal organization of movements observed in these patients regarding healthy subjects are reflected in changes in kinematic and kinetic gait patterns, closely related to muscle paresis level [1]-[2].

In particular, it is known that the ankle plantar-flexor muscle group paresis typically seen in low-lumbar level Myelomeningocele (MMC) patients causes relevant changes of kinematic and kinetic gait patterns, mainly linked to the limitation of providing body centre-of-mass (COM) forward propulsion during late-stance phase of ambulation [1].

Such limitation is evident when we consider that during healthy gait the ankle joint alone produces more power than both the knee and hip together for forward propelling the body COM [3]. Bearing in mind the transition phase of these patients between being ambulatory and non-ambulatory, it is crucial therefore, to try to functionally recover their gait.

Among treatment options of gait neuromotor deficiencies

The authors thank the National University of San Juan and the Science, Technology and Innovation Secretariat belonging to the Government of San Juan province, Argentina (IDEA 1400-SECITI-0031-2014) for the financial support, as well the FLENI Institute for Neurological Research (Escobar, Argentina) for providing data from the studied MMC patient.

C. N. Lescano is with the National Council of Scientific and Technical Research Council (CONICET), and Medical Technology Cabinet, Engineering Faculty, National University of San Juan, Argentina (clescano@gateme.unsj.edu.ar).

S. E. Rodrigo is with the Medical Technology Cabinet, Engineering Faculty, National University of San Juan, Av. San Martin 1109 (o), San Juan, Argentina (srodrigo@gateme.unsj.edu.ar).

are included robotic orthoses or exoskeletons, based on the actuation of lower limb joints while practicing this activity in a treadmill combined with partial support of body weight. According to our knowledge, although such devices have not been used for gait rehabilitation of MMC patients, the experience achieved in spinal cord injury patients [4] shows promising results that could also benefit MMC patients.

A key aspect of designing an effective gait rehabilitation protocol is the type of variable used to control the active orthosis in order to promote gait functional recovery of these patients. In this sense, the footpath during gait –considered as a task-specific performance variable that is stabilized by the nervous system from the abundant degrees of freedom of muscles and joints [5]-[6] – could be applied for such aim.

As an example, it has been signaled on one hand, that the increased step-to-step variability in gait parameters shown by spinal injury patients may be linked to greater footpath variability while walking. On the other hand, preliminary results achieved from training of these patients to step on a treadmill with body weight support show that they learned to produce footpaths similar to those of healthy subjects [7].

From this background, in this work we propose analyzing the behavior of the footpath during gait for a case study of a low-lumbar level MMC patient. The procedure employed combines standard and novel methods for characterizing spatial and temporal intra-variability of the right footpath during gait trials of the studied patient. The final purpose is to explore the applicability of this global performance variable to control an exoskeleton for gait rehabilitation in these patients.

II. MATERIALS AND METHODS

A. Clinical evaluation A single case study of a patient with low-lumbar level

MMC (female, 22 years, 1.55 m in height and 83 kg in mass) participated in our research, previously reviewed and approved by the research Ethics Committee. After explaining the protocol to the patient, she signed the consent.

Besides, prior to gait analysis a clinical evaluation of the patient was made by the team of Gait Laboratory at FLENI. The severity of her spinal cord injury was rated from the standard Manual Muscle Test (MMT, 0_/5 scale), inaccordwith themuscle paresis classification previouslysetup[1]forthemajorlowerlimbmusclegroups,whereno grade higher than 4 was assigned for musclesinvolvedinparesis.

The studied MMC patient is within Group 2 signaled by

Analysis of footpath spatio-temporal variability during gait. A case study of a low-lumbar level Myelomeningocele patient

Claudia N. Lescano, Silvia E. Rodrigo

M

this research team [1], consistent with the MMT scale. In particular, the ankle plantarflexors showed 0 and 3-4 MMT values for the right and left limbs, respectively.

B. Data collection During the gait trials, the MMC patient was able to

walkataself-selectedspeed(0.73m/s)withoutwalkingaids.Kinematicandkineticdatawere recorded throughan Elite 2002 motion capture system (BTSBioengineering, Italy)with8 cameras (100Hz,22 retro-reflective skin markers over bony landmarks) and twoforceplates (Kistler 9281E,KistlerGroup, Switzerland).Electrical muscle activity data were recorded for thegastrocnemius and tibialis anterior muscle groups ofboth limbs using a Teleemg surface dynamicelectromyograph(2000Hz,BTSBioengineering,Italy).

C. Spatial and temporal variability measures From independent gait cycles (K=3) of the MMC patient,

the anteroposterior and vertical components of the right malleolus (footpath) trajectory on the sagittal plane were reconstructed and then smoothed with a 4th order Butter filter. Two methods were applied to describe the footpath variability. The standard method is based on linear rescaling of the time axis of footpaths. From the time-normalized footpaths, a reference trajectory is then computed by averaging for each time point. The intra-variability measure was defined from the absolute deviations between each trajectory and the reference one over the gait cycle time T.

In contrast to the above method and as an approach of the spatio-temporal behavior for footpath during MMC patient’s gait, the spatial and temporal shifts between individual paths and a reference path were calculated at N characteristic point, such as the main maxima and minima of the respective footpath curves, as well as the points for typical events of gait cycle (heel contact and toe-off). Measures for the spatial and temporal intra-variability were constructed by averaging separately the spatial and temporal deviations, defined by the space–time average of all gait cycles:

∑∑∑∑====

==N

1nk

K

1k

N

1nk

K

1k)n(

K1var)n(

K1var τξ τξ (1)

where ξk(n) and τk(n) represent the spatial and temporal shifts, respectively (see [8] for details).

III. RESULTS From the standard method, the footpath (average ± SD) of

MMC patient’s gait (Fig. 1) showed an increased variability for the swing phase of the gait cycle. Greater antero-posterior component variability in relation to the vertical one was also verified during late-stance phase (not shown here).

Besides, the results of spatial and temporal variability applying the second method for both components of the gait footpath are displayed in Table I, where a larger temporal than spatial intra-variability in both footpath components.

Finally, from both methods, a greater variability of the horizontal component as regards the vertical one was found.

x coordinate (m)

y co

ordi

nate

(m)

Fig. 1. Right limb footpath during gait trials of the MMC patients.

TABLE I Spatial and temporal variability of right malleolus path.

Spatial variability Temporal variability x trajectory 0.0502 0.053 y trajectory 0.01965 0.0367

IV. CONCLUSION The greater intra-variability found for the anteroposterior

component of the MMC patient's gait footpath could be linked to limitations of the ankle plantar-flexor muscles to provide body COM forward propulsion during late-stance phase [1]-[3]. As well, the larger temporal intra-variability shown could relate to a distinct combination of muscle activation patterns, as is seen in spinal injury patients [7]. In future works we propose to deeper explore these issues, in the search for clues to design a suitable control scheme of an exoskeleton for gait rehabilitation of these MMC patients.

V. REFERENCES [1] E.M. Gutierrez, A. Bartonek, Y. Haglund-Akerlind and H. Saraste,

“Characteristic gait kinematics in persons with lumbo-sacral myelomeningocele”, Gait Posture, vol. 18, Dec. 2003, pp.170–177.

[2] E.M. Gutierrez, A. Bartonek, Y. Haglund-Akerlind and H. Saraste, “Kinetics of compensatory gait in persons with myelomeningocele”, Gait Posture, vol. 21, Aug. 2004, pp.12-23.

[3] D.A. Winter, Biomechanics and Motor Control of Human Movement. John Wiley and Sons Inc., 2009.

[4] M. Wirz, D.H. Zemon, R. Rupp R, et al., “Effectiveness of automated locomotor training in patients with chronic incomplete spinal cord injury: A multicenter trial”, BMC Neurology, vol. 11, May 2011, pp. 1-5.

[5] M.L. Latash, J.G. Anson, “Synergies in health and disease: Relations to Adaptive Changes in Motor Coordination, Phys Ther, vol. 86, Aug. 2006, pp. 1151-1160.

[6] Y.P. Ivanenko, R. Grasso, V. Macellari, F. Lacquaniti, “Control of foot trajectory in human locomotion: role of ground contact forces in simulated reduced gravity”, J Neurophysiol, vol. 87, Jun. 2002, pp. 3070-89.

[7] Y.P. Ivanenko, R. Grasso, M. Zago, et al., “Temporal Components of the Motor Patterns Expressed by the Human Spinal Cord Reflect Foot Kinematics”, J Neurophysiol, vol. 90, Jul. 2003, pp. 3555–3565.

[8] W. Ilg, G.H. Bakır, M.O. Franz, M.A. Giese, “Hierarchical Spatio-Temporal Morphable Models for Representation of complex

Transcranial Direct Current Stimulation (tDCS) protocols for improvingresults of detection intention of pedaling initiation through EEG signals

Marisol Rodrıguez-Ugarte, Eduardo Ianez, Alvaro Costa, Jose M. Azorın

Abstract— Transcranial direct current stimulation (tDCS)has been a growing research topic in the last decade. Itis a noninvasive technique that stimulates specific parts ofthe brain producing cortical excitability changes. Our goalis to use tDCS to improve the rehabilitation and increaselong term potentiation plasticity in cerebrovascular accident(CVA) patients. This technique will be used to improve therehabilitation of the lower limbs during pedaling. In this paper,the procedure for detecting the intention of pedaling from EEGsignals without stimulation in healthy subjects and patients isexplained.

Two tDCS protocols are described: In the first one, stimula-tion is applied for 10 minutes before the session. In the secondone, the stimulation is applied while the subject is performingthe pedaling movement. The idea is to apply these two protocolsin future work.

I. INTRODUCTION

Transcranial current stimulation (tCS) is a noninvasivetechnique for brain stimulation which generates electric fieldsin the brain by delivering low currents. Its use for improvinghuman impaired condition has increased lately [1], [2]. Themost popular methods in this technique are: transcranialdirect current stimulation (tDCS), transcranial alternatingcurrent stimulation (tACS) and transcranial random noisestimulation (tRNS). In previous studies, the use of tDCSenhanced motor task, promoting motor recovery [3], [4], [5].

On the other hand, it is necessary to study the intentionof a particular motor task. Different motor task studies suchas hand movement or walking, have already been researched[6], [7].

This study is based on previous work [8], [9], [10], wherethe intention of pedaling initiation was studied. To improvethe results obtained before, the idea is to apply tDCS. In thiswork, the architecture of the system and the protocols thatare going to be followed are explained.

II. DETECTING PEDALING START INTENTION

A Brain-Machine Interface (BMI) is used to detect thepedaling start intention. This BMI is composed by a EEGrecorder and an algorithm to translate the signals intocomputer commands. Several electrode configurations were

This research has been carried out in the framework of the projectAssociate - Decoding and stimulation of motor and sensory brain activityto support long term potentiation through Hebbian and paired associativestimulation during rehabilitation of gait (DPI2014-58431-C4-2-R), fundedby the Spanish Ministry of Economy and Competitiveness and by theEuropean Union through the European Regional Development Fund (ERDF)A way to build Europe.

The authors are members of the Brain-Machine Interface Systems Lab,Miguel Hernandez University of Elche, Av. de la Universidad S/N, 03202Elche, Spain (email: maria.rodriguezu@umh.es).

studied in [8], [9], [10]. First of all, the algorithm pre-processed the full signals obtained by the EEG systemapplying a Notch filter to remove power line interferenceand then utilizing a Butterworth high-pass filter to eliminatethe low frequency. Secondly, signals were segmented. Therewere two types of classes: rest and start; and two types ofwindow selection: 4 and 2 seconds. Start class 4 secondswindows were centered at the “real start”, while the startclass 2 seconds windows were located just before the “realstart”. The rest class windows were chosen to be the sameduration as the corresponding start class windows. Thesewere placed before the start class windows with a gap of0.5 seconds between the rest and the start class windows toavoid overlap. Feature extraction was done based on a FastFourier Transform and the sum of three frequency ranges:mu, beta low and beta high. Thirdly, these characteristicswere classified by a Support Vector Machine classifier witha Radial Basis Function kernel.

The protocol followed by the subject consisted of 16 runs.Each run was based on 5 repetitions of the pedaling cycle.Each pedaling cycle consisted of remaining still around 10seconds and subsequently the subject started pedaling byabout 10 seconds.

The main goal is to improve the results obtained previouslyby using tDCS.

III. TDCS PROTOCOLS

Two protocols are going to be followed in order to improvethe results. In the first protocol, the stimulation is appliedbefore the pedaling activity. In the second protocol, thestimulation is applied while the subject is performing thetask. This is represented in Fig. 1.

Several factors are important for increasing motor corti-cal excitability according to the literature [11], [12], [13].Among them are: the time of stimulation, the location of thestimulation electrodes, the polarization of the electrodes andthe current density. In [1], [2], [14], the time of stimulationused was between 5 and 20 minutes. Taking this intoconsideration, time chosen for stimulation is 10 minuteswith the first protocol and 5 minutes for the second one.In [19], it is concluded that having the reference eitherin the extracephalic ipsilateral or contralateral, the currentdensities reach deeper brain regions. The polarization of theelectrodes has been widely studied. Indeed in, [15], [16],[17] it is concluded that anodal polarization improves theflexor muscles contraction, and the learning and workingmemory, as compared to cathodal polarization and shamstimulation. Therefore, as we want to increase the motor

cortical excitability of the brain, the polarization chosen isanodal. However, in addition to anodal stimulation, shamstimulation will be applied to verify if the placebo effectexists. An effective current density for stimulation is 0.06mA/cm2 which is inside the safety limits range [18]. As ourstimulation electrodes have an area of π cm2 the intensitycannot increase beyond 188.5 µA to fulfill the safety limits.

Fig. 1. tDCS protocols. In the first protocol (top) the stimulation willbe applied before the pedaling motor task. In the second protocol, thestimulation will be applied while the pedaling is taking place.

IV. DISCUSSION AND CONCLUSION

Offline and pseudo-online results show that 8 - 10 elec-trodes are the minimum required for having average resultsbetween 66 - 76 % of true positive rate, 65 - 72 % of accuracyand 4.3 - 6.6 of false positive per minute [8], [10].

Applying tDCS protocols is expected to improve theseresults. In future work, the protocols will be tested withpatients that have suffered a cerebrovascular accident.

V. ACKNOWLEDGMENT

The authors wish to thank Neuroelectrics for lending theequipment Enobio 32 used in the experiments.

REFERENCES

[1] D.T. Jeffery, J.A. Norton, F.D. Roy, M.A. Gorassini,Effects of tran-scranial direct current stimulation on the excitability of the leg motorcortex, Experimental brain research, vol. 182, 2007, pp. 281-287.

[2] P.S. Boggio, L.O. Castro, E.A. Savagim, R. Braite, V. Cruz,etal..,Enhancement of non-dominant hand motor function by anodaltranscranial direct current stimulation, Neuroscience letters, vol. 404,2006, pp. 232-236.

[3] S. Koyama, S. Tanaka, S. Tanabe, N. Sadato, Dual-hemisphere tran-scranial direct current stimulation over primary motor cortex enhancesconsolidation of a ballistic thumb movement, Neuroscience letters, vol.588, 2015, pp. 49-53.

[4] M.C. Chang, D.Y. Kim,D.H. Park, Enhancement of Cortical Ex-citability and Lower Limb Motor Function in Patients With Strokeby Transcranial Direct Current Stimulation,Brain stimulation, vol. 8,2015, pp. 561-566.

[5] M.A. Nitsche, W. Paulus, Excitability changes induced in the humanmotor cortex by weak transcranial direct current stimulation, TheJournal of physiology, vol. 527, Aug. 2000, pp. 633639.

[6] E. Hortal, A. Ubeda, E. Ianez, E. Fernandez, J.M. Azorın, Using EEGSignals to Detect the Intention of Walking Initiation and Stop,ArtificialComputation in Biology and Medicine, 2015, pp. 278-287.

[7] E. Lew, R. Chavarriaga,S. Silvoni, J. Millan, Detection of self-paced reaching movement intention from EEG signals, Frontiers inNeuroengineering, vol. 5, 2012.

[8] M. Rodrıguez-Ugarte, E. Hortal, A Costa, E. Ianez, A. Ubeda, et al.,Detection of intention of pedaling start cycle through EEG signals,38th Annual International Conference of the IEEE Engineering inMedicine and Biology Society, submitted for publication.

[9] M. Rodrıguez-Ugarte, A Costa, E. Ianez, A. Ubeda, J.M. Azorın,Pseudo-online detection of intention of pedaling start cycle throughEEG signals, International Conference on Neurorehabiliation, submit-ted for publication.

[10] M. Rodrıguez-Ugarte, A. Costa, E. Ianez and J. M. Azorın, Electrodeconfigurations comparison using offline and pseudo-online analysesto detect intention of pedaling initiation through EEG signals, IEEEInternational Conference on Systems, Man, and Cybernetics (SMC2016), submitted for publication.

[11] L.J. Lauro, M. Rosanova, G. Mattavelli, S. Convento, A. Pisoni, A.Opitz, N. Bolognini, G. Vallar, TDCS increases cortical excitability:Direct evidence from TMS–EEG, cortex, vol. 58, 2014, pp. 99111.

[12] T. Hunter, P. Sacco, M.A. Nitsche, D.L. Turner, Modulation of internalmodel formation during force field-induced motor learning by anodaltranscranial direct current stimulation of primary motor cortex , TheJournal of physiology vol. 587, 2009, pp. 29492961.

[13] A. Alonzo, J. Brassil, J.L. Taylor, D. Martin, C.K. Loo, Dailytranscranial direct current stimulation (tDCS) leads to greater increasesin cortical excitability than second daily transcranial direct currentstimulation, Brain stimulation vol. 5, 2012, pp. 208208–213213.

[14] J. Baudewig, M.A. Nitsche, W. Paulus, J. Frahm, Regional modulationof BOLD MRI responses to human sensorimotor activation by tran-scranial direct current stimulation,Magnetic Resonance in Medicine,vol. 45, 2001, pp. 196-201.

[15] F. Cogiamanian, S. Marceglia, G. Ardolino, S. Barbieri, A. Priori,Improved isometric force endurance after transcranial direct currentstimulation over the human motor cortical areas, European Journal ofNeuroscience, vol. 26, 2007, pp. 242249.

[16] M.A. Nitsche, A. Roth, M.F. Kuo, A.K. Fischer, D. Liebetanz, etal.., Timing-dependent modulation of associative plasticity by generalnetwork excitability in the human motor cortex, The Journal ofNeuroscience, vol. 27, 2007, pp. 38073812.

[17] S.H. Ohn, C.I. Park, W.K. Yoo, M.H. Ko, K-P- Choi,et al., Time-dependent effect of transcranial direct current stimulation on theenhancement of working memory, Neuroreport, vol. 19, 2008, pp.4347.

[18] M. Bikson, A. Datta, M. Elwassif, Establishing safety limits for tran-scranial direct current stimulation Clinical neurophysiology: officialjournal of the International Federation of Clinical Neurophysiologyvol. 120, 2009, pp. 1033.

[19] G.M. Noetscher, J. Yanamadala, S.N. Makarov, A. Pascual-Leone,Comparison of cephalic and extracephalic montages for transcra-nial direct current stimulationa numerical study, vol. 61, 2014, pp.24882498.

Detection of muscle activity during motor imagination and attemptedmovements using ultrasound imaging

Anna J. Sosnowska, Henrik Gollee, Ian D. Loram, and Aleksandra Vuckovic

Abstract— The aim of this study was to assess whether motorimagination in able-bodied people results in overt movementsand to compare the capability of electromyography (EMG) andultrasound imaging (USI) to detect muscle activity. Approach.Ten able-bodied participants performed a motor task of press-ing a force plate with a foot following cues in 4 different modali-ties: executed, attempted, visually imagined and kinaestheticallyimagined movements. Main Results. Motor imagery can resultin overt muscle activation of which the person is not aware.When performing visual and kinaesthetic imagination tasks,23.4% and 37.8% cues respectively led to real muscle movementdetected with USI. With EMG respectively, only 10.1% and18.7% activations were detected. The attempted movements ofsmall force outputs were recognized with both USI (96%) andEMG (89%). Significance. Presence of real movement duringimagination indicates that experimental paradigms which usemotor imagination in able bodied population might not replicatethis process in people with impaired sensory-motor processing.

I. INTRODUCTION

Movement imagination is the mental execution of an ac-tion without any overt movement or peripheral activation [1].Among other factors it relies on proprioception [2]. MI leadsto the activation of the same brain areas as actual movement,thus it is often used in brain computer interface (BCI)experimental paradigms for healthy population [3]. However,imagination could lead to unpredicted real, potentially smallmovements that could interfere with operation of BCI.

Surface electromyography (EMG) mostly reflects the ac-tivity of the superficial muscles and is prone to musclecross-talk [4], while ultrasound imaging (USI) enables tomonitor the muscle movement with better spatial resolution.The brightness-mode (B-mode) USI, where reflected soundechoes are displayed as bright dots in a straight line, resultsin 2D image of a scanned region. The contraction mechanismis expressed through pennation angle, muscle thickness andfascicle length parameters, however manually identifyingthem is not cost-effective. It is possible to automaticallyquantify movement by applying tracking algorithms to theultrasound videos [5].

The aim of this study is to compare the capability ofelectromyography (EMG) and ultrasound imaging (USI) todetect muscle activity during imagined and real motor tasks.

This work is supported by the Engineering and Physical Sciences Re-search Council (James Watt Research Scholarship).

A. J. Sosnowska, A. Vuckovic and H. Gollee are with theCentre for Rehabilitation Engineering, School of Engineering,University of Glasgow, Glasgow, UK (corresponding author:a.sosnowska.1@research.gla.ac.uk)

I. D. Loram is with Institute for Biomedical Research into HumanMovement and Health, Manchester Metropolitan University, Manchester,UK.

Both techniques are also used to evaluate whether motorimagination results in actual movement of the muscle.

II. MATERIALS AND METHODS

A. Experimental setup

Ten able-bodied participants performed an isometric con-traction when pressing on a force pedal with a foot (ankledorsiflexion) under 4 different conditions: visual imagina-tion, kinaesthetic imagination, attempted movement, andexecuted movement at 30% of maximum voluntary contrac-tion (MVC). During visual imagination, the participant wasinstructed to visualize themselves moving the foot, while thekinaesthetic task required imagining the sensations in themuscles and joint. Attempted movement was movement withminimum bodily awareness of performing a physical action.

The participants were seated in an upright position withfoot strapped to the platform of a dynamometer. They werefacing a computer screen, where motor initiation cues weredisplayed. Force output was digitized with a DAQcard-6024E (National Instruments, USA) at 1000 Hz. An ul-trasound probe (linear array, 7MHz, 40 frames/s, 65 mmdeep, connected to Echoblaster 128, Telemed, Lithuania) andsurface EMG electrodes were positioned over the belly of thegastrocnemius medialis muscle. EMG was recorded using theUSBamp (gTech, Austria) with bipolar Ag/AgCl electrodes,and sampled at 1200 Hz. For each of the 4 modalities, intotal 90 cues lasting 2.5 s each were shown in 2 minuteslong sessions.

B. Data processing and analysis

EMG signal was filtered to 10-300 Hz, rectified, andsmoothed over 120 samples, while force output wassmoothed over 100 samples. The ultrasound videos wereconverted to an avi-file format with 400x380 resolution,individual frames extracted and changed to grayscale. Thevalues of pixels were subtracted between adjacent framesand the absolute value of differences summed over the wholeframe. Plotting the sum of pixels difference allowed to detectchanges in muscle activation, as shown in Fig. 1 that displaysthe EMG, force and pixel difference for a representativeparticipant attempting movements.

For automatic movement detection, the thresholds foronsets were established by averaging resting periods duringexecuted tasks, and using mean+2SD for EMG and USI. ForUSI, this value was selected as suitable following empiricalanalysis of videos. The trials when muscle was not relaxed asverified with EMG were discarded. With the force platform,movement was defined when exceeding 5% of force of MVC.

Fig. 1. Recording of EMG, force output and difference in pixels betweenadjacent frames of ultrasound video for a representative attempted movementtask. Green and red vertical lines signify appearing and disappearing of cue.

The threshold value was applied to attempted and imagina-tion tasks to determine when the muscle was activated.

The activations observed with ultrasound were qualified asmovements if 2 consecutive peaks exceeding the thresholdwere registered following the cue (see Fig. 1). It was thenassessed whether these movements were detected with EMGand if 5% of MVC threshold of force was exceeded. Sin-gular peaks on USI, present only following onset or offsetof movement, corresponded to short lasting muscle fibresmovements and were defined as twitches.

III. RESULTS

Table I shows the percentage of all displayed cues thatwere followed by movements detected with USI and EMGaveraged across all participants. All executed movementswere registered, while attempted and imagined movementswere better detected with USI than with EMG. The attemptedmovements resulted in lower force outputs than real tasks(15 ± 10% of MVC). It was also observed, that some of thecues were followed by twitches (7.9% and 8.6% for visualand kineasthetic imagination, respectively). These muscleactivations were not shown by EMG or force output.

TABLE IMOVEMENTS DETECTED WITH USI AND EMG

Task, detection method Detected movements (%)USI EMG

Visual Imag. 23.4 ± 28.7 10.1 ± 17.5Kinaesthetic Imag. 37.8 ± 30.1 18.7 ± 28.6Attempted Move 96.2 ± 6.9 86.8 ± 33.7Executed Move 100.0 ± 0.0 100.0 ± 0.0

The variability of data between participants was substantialas can be seen in the boxplots shown in Fig. 2. These alsoindicate that the spread of the data is generally smaller forUSI than for EMG.

IV. DISCUSSION

The correlation between movement onset and offset de-tected with EMG and presence of peaks in ultrasound graphdemonstrates that comparison of pixels difference betweenframes could serve as a reliable, semi-automatic method to

Fig. 2. Variability in movements detected with USI and EMG for imaginedand attempted movements as percentage of displayed cues

detect muscle movements. Detected peaks have distinctivemorphology, and are distinguishable from artifacts intro-duced by for example activity of the blood vessels. Forfurther quantification of muscle activity, methods like featuretracking [5] could be used.

Able-bodied people can differentiate between movementof minimum detectable activity (attempted) and an executedmovement producing a relatively small force of 30% MVC.Imagined movements can produce muscle twitches and overtmovements that are not detectable by EMG or force plat-form. Kinaesthetic imagination of sensations results in moremuscle activations than visual, confirming its prominentrole [6]. This demonstrates importance of proprioception inmotor imagination and makes it uncertain whether resultsof imagery tasks performed by healthy volunteers can begeneralized to patient population. The variability betweenpeople when performing imagination needs to be addressedfurther. Future work will involve analysis of EEG duringimagery tasks and comparing cortical activity for trials withand without small movements/twitches detected by USI.

V. CONCLUSIONS

Ultrasound imaging can detect small muscle activationsmore precisely than EMG. Analysis of USI reveals substan-tial muscle activity during motor imagination, indicating thatable-bodied people activate proprioceptive pathways.

REFERENCES

[1] T. Mulder, ”Motor imagery and action observation: cognitive tools forrehabilitation,” J. Neural. Transm., vol. 114, no. 10, pp. 1265-1278,Oct. 2007

[2] C. Farrer, N. Franck, J. Paillard, and M. Jeannerod, ”The role ofproprioception in action recognition,” Consciousness and Cognition,vol. 12, no. 4, pp. 609-619, 2003

[3] G. Pfurtscheller and C. Neuper, ”Motor imagery and direct brain-computer communication,” Proc. IEEE vol. 89, pp. 1123-1134, 2001

[4] J.-Y. Guo, Y.-P. Zheng, Q.-H. Huang, and X. Chen, ”Dynamic mon-itoring of forearm muscles using one-dimensional sonomyographysystem,” J. Rehabil. Res. Dev., vol. 45, no. 1, pp. 15-64, 2008

[5] I. D. Loram, C. N. Maganaris, and M. Lakie, ”Use of ultrasoundto make noninvasive in vivo measurement of continuous changes inhuman muscle contractile length,” J. Appl. Phys., vol. 100, no. 4, pp.1311-1323, 2006

[6] S. Marchesotti, M. Bassolino, A. Serino, H. Bleuler, and O. Blanke,”Quantifying the role of motor imagery in brain-machine interfaces,”Scientific Reports, 2016;6:24076. doi:10.1038/srep24076..