A Haptic Robot Reveals the Adaptation Capability of Individuals with Multiple Sclerosis

http://ijr.sagepub.com

Research The International Journal of Robotics

DOI: 10.1177/0278364907084981 2007; 26; 1225 The International Journal of Robotics Research

Maura Casadio, Vittorio Sanguineti, Claudio Solaro and Pietro G. Morasso A Haptic Robot Reveals the Adaptation Capability of Individuals with Multiple Sclerosis

http://ijr.sagepub.com/cgi/content/abstract/26/11-12/1225 The online version of this article can be found at:

Published by:

http://www.sagepublications.com

On behalf of:

Multimedia Archives

can be found at:The International Journal of Robotics Research Additional services and information for

http://ijr.sagepub.com/cgi/alerts Email Alerts:

http://ijr.sagepub.com/subscriptions Subscriptions:

http://www.sagepub.com/journalsReprints.navReprints:

http://www.sagepub.com/journalsPermissions.navPermissions:

http://ijr.sagepub.com/cgi/content/refs/26/11-12/1225SAGE Journals Online and HighWire Press platforms):

(this article cites 34 articles hosted on the Citations

© 2007 SAGE Publications. All rights reserved. Not for commercial use or unauthorized distribution. at PENNSYLVANIA STATE UNIV on April 10, 2008 http://ijr.sagepub.comDownloaded from

Maura CasadioDepartment of Informatics, Systems and TelematicsUniversity of GenovaVia Opera Pia 13, 16145 Genova GE, ItalyandFoundation ‘Don Carlo Gnocchi’Via Cisa Vecchia, 19038 Sarzana SP, [email protected]

Vittorio SanguinetiDepartment of Informatics, Systems and Telematics andResearch Center for Neuroscience and NeuroengineeringUniversity of GenovaVia Opera Pia 13, 16145 Genova GE, [email protected]

Claudio SolaroNeurology Department, Hospital “P. Antero Micone”Via D. Oliva 22, 16158 Genova GE, [email protected]

Pietro G. MorassoDepartment of Informatics, Systems and TelematicsUniversity of GenovaVia Opera Pia 13, 16145 Genova GE, [email protected]

A Haptic Robot Revealsthe AdaptationCapability of Individualswith Multiple Sclerosis

AbstractA prerequisite for rehabilitation is that patients preserve their abil-ity to adapt to novel dynamic environments, an ability that has beenassociated with the cerebellar system. In this study, we use a robotmanipulandum to assess the ability of multiple sclerosis (MS) sub-jects in the early phase of the disease to adapt to a speed-dependentforce field. Their performance is compared with an equal number ofage-matched controls. We found that MS subjects display subtle in-coordination problems but do not significantly differ from controls intheir ability to adapt to the force field. These findings are discussedin terms of the possible benefits that MS subjects might receive fromrobot-assisted therapy that is specifically aimed at impaired visuomo-tor coordination.

KEY WORDS—rehabilitation robotics, motor learning, mul-tiple sclerosis

The International Journal of Robotics ResearchVol. 26, No. 11–12, November/December 2007, pp. 1225–1233DOI: 10.1177/0278364907084981c�SAGE Publications 2007 Los Angeles, London, New Delhi and Singapore

1. Introduction

Over the last 20 years, robots have been used widely in theexperimental investigation of the mechanisms underlying theneural control of movement. In a typical application, robotsgenerate controlled perturbations to ongoing movements. Thisallows the response of the motor system to such perturbations,as well as the underlying control modalities, to be quantified(Mussa-Ivaldi et al. 1985� Gomi and Kawato 1996). Robotictechnology has also shown promise recently in using special-ized forces that stimulate adaptation in the nervous system asa mode of recovery.

The adaptive properties of the motor system have been stud-ied in experiments in which robots deliver forces that may bemade dependent on position, speed and/or acceleration, thusallowing specific dynamic environments (or “force fields”) tobe simulated. Most of these studies involved planar arm move-ments (Shadmehr and Mussa-Ivaldi 1994), but applications toorofacial control have been reported (Tremblay et al. 2003).Adaptation to force fields may be seen as a generalization of

1225

© 2007 SAGE Publications. All rights reserved. Not for commercial use or unauthorized distribution. at PENNSYLVANIA STATE UNIV on April 10, 2008 http://ijr.sagepub.comDownloaded from

1226 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / November/December 2007

perturbation studies. In fact, the patterns of sensorimotor adap-tation to unfamiliar dynamic environments may indirectly pro-vide information about the body’s mechanical impedance ata specific activation, configuration and speed (Scheidt et al.2001� Takahashi et al. 2001� Donchin et al. 2003). Adaptationstudies have provided a large body of knowledge on the mech-anisms underlying the way the brain reacts to novel dynamicenvironments (Shadmehr and Wise 2005).

Robots are being recognized as promising tools for the re-habilitation of upper-limb impairment. They may be used toguide or assist the movements of a patient in close interaction,in much the same way as a human physical therapist. Clinicaltrials (Prange et al. 2006) suggest that robot therapy is quiteeffective: for instance, the MIT-MANUS system (Krebs et al.1999) was demonstrated to accelerate the recovery of strokepatients (Volpe et al. 1999).

Most applications of robot therapy involve acute or chronicstroke patients and often mimic the exercises delivered by ahuman therapist. However, robot therapy would be even moreappealing in the treatment of deficits for which no rehabilita-tion methods have proved effective, such as those involving thecoordination of movements with multiple degrees of freedom,caused by lesions to the cerebellum and/or the basal ganglia.

A prerequisite for rehabilitation, whether robot- ortherapist-assisted, is that patients preserve their ability to adaptto novel dynamic environments, an ability related to the feed-forward component of control. Recently, an experimental pro-tocol that is widely used to investigate motor learning ca-pabilities (Shadmehr and Mussa-Ivaldi 1994) was applied tosubjects with degenerative cerebellar atrophy (Maschke et al.2004� Smith and Shadmehr 2005) and it was shown that thesesubjects completely lost their ability to adapt. In contrast, thissame ability is preserved in subjects with Huntington’s dis-ease (Smith and Shadmehr 2005) as well as in stroke survivors(Takahashi and Reinkensmeyer 2003� Patton et al. 2006).

Multiple sclerosis (MS) is the second biggest cause of neu-rological disability among young adults. Therefore, it seemsnatural to explore the potential of robots in the assessmentand treatment of these patients. MS is usually described as achronic autoimmune disease, characterized by inflammatorydemyelination, affecting the white matter in the central ner-vous system. This results in the impairment of multiple func-tional systems, in proportions that change widely from patientto patient. Common symptoms include weakness of one ormore extremities, muscle spasticity, double vision, urinary in-continence and loss of coordination. About 85% of the sub-jects show a relapsing–remitting course, i.e. acute phases (“re-lapses”) alternating with intervals in which the patient’s con-ditions remain stable. In the early stages of MS, although ax-onal injury is not reversible, relapses are often followed bypartial or complete functional recovery (“remissions”). Clini-cal recovery is determined by many factors, such as increasedexpression of sodium channels, recruitment of silent pathwaysand remyelination. Cortical reorganization may take place as

Fig. 1. Experimental set-up.

well (Rocca et al. 2003), thus playing a role in limiting theimpact of axonal loss. Reorganization may leave performancerelatively unaffected, but the compensatory strategies that areexploited in controlling the movements may be unveiled byhighly sensitive experimental and analytic techniques, aimedat investigating sensorimotor control and adaptation.

In this paper, we describe an application of the above-mentioned adaptation protocol to the assessment of sensorimo-tor performance and adaptation in subjects with a confirmeddiagnosis of MS. As functional brain imaging studies havesuggested that MS subjects with no disability display a sub-stantial cortical and sub-cortical reorganization (Rocca et al.2005), we asked whether MS subjects with sub-clinical sen-sorimotor symptoms display signs of compensatory strategies,and whether and to what extent these subjects preserve theirability to reorganize their sensorimotor behavior in order toadapt to an unfamiliar dynamical environment.

We also discuss the diagnostic implications and the per-spectives for robot-assisted rehabilitation with these patientsand MS patients in general.

2. The Robotic Manipulandum

The robotic manipulandum that we used in this study (see Fig-ure 1) was specifically designed for the evaluation of motorlearning and control, and for robot therapy.

2.1. Robot Architecture

The manipulandum (details are reported in Casadio et al.(2006)) has a planar, 80 � 40 cm elliptic workspace that canbe rotated around a horizontal axis in order to work on non-horizontal planes (this feature was not used in the reported ex-periments). The geometry of the robot was specified as theresult of the optimization of a global isotropy index (Stocco etal. 1998).

© 2007 SAGE Publications. All rights reserved. Not for commercial use or unauthorized distribution. at PENNSYLVANIA STATE UNIV on April 10, 2008 http://ijr.sagepub.comDownloaded from

Casadio, Sanguineti, Solaro, and Morasso/ A Haptic Robot Reveals the Adaptation Capability of Individuals with MS 1227

The robot is actuated by two direct-drive brushless motors,mounted proximally to minimize the overall inertia at the han-dle (less than 1 kg) as well as the frictional forces (less than0.06 N)� the manipulability index and the force/torque ratioare quite uniform over the whole workspace (23 � 2 cm and2�2 � 0�2 N Nm�1, respectively). The force available at thehandle is greater than 50 N (continuous) and greater than 200N (peak), in all directions.

Rotations of the motors were estimated by a pair of sin/cosencoders (equivalent resolution: 17 bits), which allowed a spa-tial resolution of less than 0.1 mm over the whole workspace.We did not use a force sensor at the handle, because the ex-tremely low friction and the low inertia allow to the interac-tion forces from the motor currents to be estimated directly.Although the robot is in many respects similar to the MIT-MANUS system, there are important differences in the me-chanical design (the ability to rotate the plane of movements)and in the available power (steady hand forces are almost threetimes greater).

2.2. Control Architecture

The control architecture consists of an inner current loop (run-ning at 16 kHz) and an outer impedance control loop (at1 kHz). The current loop is implemented by two control units(one for each motor). The impedance control loop runs on adedicated personal computer, under a real-time operating sys-tem (QNX). An additional computer is used for configurationand command and also runs the graphical user interface. Theimpedance control scheme is defined by the following equa-tion:

Tm � J �q�T � Fh�x� �x� �x�� (1)

where Tm is the torque vector to be generated by the two mo-tors, J �q� is the Jacobian matrix of the manipulator and q isthe vector of joint angles. In general, the specified impedancefunction at the handle, Fh, is a function of the position of thehandle (x) and of its time derivatives.

The software environment for the control and design of theexercise protocols is based on a Simulink fast-prototyping en-vironment, RT-Lab

R�(Opal-RT Technologies Inc.). For a given

exercise, a set of visual objects is specified and displayed ona computer screen. Visual objects are represented in termsof the Virtual Reality Modeling Language (VRML), by us-ing Simulink’s Virtual Reality toolset. The exercise protocolspecifies the interaction between the robot and visual objects.This is specified as a finite-state machine, implemented bymeans of a standard Matlab tool, Stateflow

R�.

3. Quantifying Sensorimotor Performance

3.1. Experimental Protocol and Task

In experiments aimed at the assessment of sensorimotor per-formance, the robot is used as a passive device. Motors are

turned off and the encoders are used to record hand trajecto-ries.

Subjects sit on a chair, with their torso and wrist restrainedby means of suitable holders, and grasp the handle of the ma-nipulandum with their dominant hand. The forearm is sup-ported by a low-friction sled on the horizontal surface of atable. The height of the seat is adjusted so that the arm canbe kept horizontal at the level of the shoulder joint. Therefore,only the shoulder and elbow could move and motion was re-stricted to the horizontal plane, with no influence of gravity.The position of the seat is also adjusted in such a way thatwhen the cursor is pointing at the center of the workspace, theelbow and the shoulder joints are flexed about 90 and 45,respectively.

The task consists of making movements in eight differentdirections, starting from the same initial position at the centerof the workspace. The targets were presented on a 19” LCDcomputer screen, placed in front of the subjects, about 1 maway, at eye level. Targets are displayed as round green cir-cles (2 cm diameter) against a black background. The currentposition of the hand in the workspace is also continuously dis-played as a yellow circle (0.4 cm diameter). The amplitude ofthe movements (distance of the targets from the center) was 10cm. The sequence of target presentations alternated the centraltarget and one of four peripheral targets, generated in randomorder. In order to decrease movement variability, the subjectswere encouraged to keep an approximately constant timing.We set the desired duration to 500 � 50 ms. If the estimatedduration was inside this range, a positive feedback/reward tothe subject (a pleasant sound) was provided. If the measurewas below or above that range, no sound was generated andthe color of the target was changed to red or white, respec-tively.

We also informed the subjects that the reaction time wasnot important—they could wait as long as they wanted aftertarget appearance before starting each movement—but whenready, they had to perform a single, rapid movement towardthe target.

The experiment was organized into target sets, each consist-ing of a sequence of target presentations in which each periph-eral target occurred 12 times, for a total 12 � 4 � 48 center–out movements (corresponding to the directions: 0, 45, 90and 135), plus the corresponding 48 return movements (180,225, 270 and 315). The endpoint of each movement wasused as the starting point for the subsequent movement.

3.2. Data Analysis

Hand trajectories were sampled at 100 Hz. The x and y compo-nents were smoothed with a sixth-order Savitzky–Golay filter(window size 270 ms, equivalent cut-off frequency of around7 Hz), which also allowed us to estimate the first three timederivatives ( �x� �x� ...

x � �y� �y� ...y ). We then estimated the following

indicators.

© 2007 SAGE Publications. All rights reserved. Not for commercial use or unauthorized distribution. at PENNSYLVANIA STATE UNIV on April 10, 2008 http://ijr.sagepub.comDownloaded from

1228 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / November/December 2007

Movement Duration

This is the time elapsed between movement onset and termi-nation. Movement onset is computed as the first time instantthat the hand speed exceeds a threshold equal to 12% of thepeak velocity, at least for 200 ms. Movement termination iscomputed as the first time instant after onset when movementspeed goes below the threshold and stays there for at least 200ms.

Linearity

This is the percentage increment of the length of the trajectorytraced by the hand, between the onset and termination times,with respect to the straight line (the distance between the ini-tial and final points of the trajectory). It is a measure of pathcurvature.

Aiming Error

This is the difference between the target direction and the ac-tual movement direction in the early phase of the movement(until 100 and 300 ms after movement onset). The 100 ms aim-ing error is indicative of the performance of the feed-forwardcomponent of control. In contrast, the 300 ms error is a generalmeasure of curvature, because at this time of the movement thelateral deviation is largest (Smith et al. 2000� Smith and Shad-mehr 2005).

Symmetry

This is the ratio between the duration of the acceleration anddeceleration phases. If the speed profile is bell-shaped, withequal durations of the acceleration and deceleration phase,symmetry is around 1.

Jerk Index

This is the square root of the jerk (norm of the third timederivative of the trajectory), averaged over the overall move-ment duration and normalized with respect to duration andpath length (Teulings et al. 1997):

jerk index � 12

���J �t�2 dt

�� T 5

L2� (2)

where T is the movement duration and L is the path length. Itis a measure of smoothness: large jerk indexes correspond toless smoothness. The square root is used to compress the largerange of variation of the jerk integral.

Jerk Ratio

This is the ratio of the jerk indexes (see the previous section)calculated during the deceleration (after the peak in the speedprofile) and acceleration phase of the movement. It indicates adifficulty in compensating for self-generated errors and there-fore an abnormal ability to exploit sensory information, as inthis case trajectories would be expected to be less smooth to-ward their end (Smith et al. 2000).

4. Quantifying Sensorimotor Adaptation

4.1. Experimental Protocol and Task

In adaptation experiments, we used the well-known force-field adaptation paradigm (Shadmehr and Mussa-Ivaldi 1994),which allows the study of both unperturbed reaching move-ments and the modifications induced by an unfamiliar artificialdynamic environment.

The experimental protocol was organized into three phases:(i) null field, in which the robot generates no force (five targetsets)� (ii) force field, in which the force field was turned on(five target sets)� and (iii) the after effect (two target sets). Eachset lasted approximately 5 min and the subjects were allowedto rest between sets.

The purpose of the null-field phase was to set a backgroundlevel of performance. During the force-field phase, the manip-ulandum generated a viscous curl field (Shadmehr and Mussa-Ivaldi 1994), i.e. a force field that perturbs the movements bygenerating forces that are perpendicular to the instantaneousvelocity vector of the hand and have a magnitude proportionalto the speed. In the impedance control framework, this can beachieved by using the following control equation:

F ��� 0 �B

B 0

�� � �x� (3)

where B is a viscous coefficient that we set to B �13 N m�1 s�1. This value corresponds to peak forces of 4–6 N. The hand velocity vector �x was estimated online as fol-lows: (1) read the joint angles provided by the high-precisionencoders� (2) compute the hand position by means of the directkinematic equations� (3) estimate hand velocity by means of anumerical differentiation technique.

During the field sets, we randomly inserted “catch trials”,i.e. reaching movements in which the force field was unex-pectedly turned off. The purpose of these trials is to sample theprogress of the internal representation that is supposedly learntby the subjects during field adaptation. In our experiments theprobability of catch trials was one in six, corresponding to onecatch trial per direction per trial set.

© 2007 SAGE Publications. All rights reserved. Not for commercial use or unauthorized distribution. at PENNSYLVANIA STATE UNIV on April 10, 2008 http://ijr.sagepub.comDownloaded from

Casadio, Sanguineti, Solaro, and Morasso/ A Haptic Robot Reveals the Adaptation Capability of Individuals with MS 1229

4.2. Data Analysis

The nervous system may react to the perturbations introducedby the curl field in different ways: (1) it may just ignore it,and accept the distortion introduced by the field, a viable al-ternative if we consider that this kind of perturbation does notspecifically affect the achievement of the target� (2) it may re-sist the perturbation by increasing the joint stiffness (a strategyof coactivation), without changing the underlying motor com-mands� (3) it may compensate for the perturbation by means ofa suitable internal model that learns to modify the backgroundreaching patterns by producing appropriate feed-forward com-mands.

The first alternative can be ruled out if we can demonstratethat the response patterns to the curl field do change duringthe experimental sessions, a finding that is well established fornormal subjects but needs to be demonstrated for the MS pa-tients. The second alternative can be ruled out if we observethe catch trials. In fact, if the reduction in the number of errorsfollowing the introduction of a force field is due to an increasein the joint stiffness alone, we should only observe a reducedscatter and an improved straightness of the “catch trial” paths(Burdet et al. 2006). In contrast, if error reduction is a conse-quence of learned feed-forward control, the unexpected disap-pearance of the field in the catch trials should generate errorsin the opposite direction. The problem is to define a learningindex, which should describe the learning process in a quan-titative way independently of the magnitude of the force fieldand user-specific parameters such as the net compliance of thearm. The solution that we used in this study is similar to thatproposed by Criscimagna-Hemminger et al. (2003) and usedby Smith and Shadmehr (2005):

learning index � yc

�yc � yf� (4)

where yf and yc are the 300 ms aiming errors in the field tri-als and catch trials, respectively. Both error measures were ad-justed for any bias that may have been present during the lastnull-field set. Therefore, errors in a field set always refer tochanges from errors in the null set.

5. Experimental Results

Eleven subjects with clinically definite, relapsing–remittingMS, according to Poser criteria (Poser et al. 1983), participatedin this study. The inclusion criteria were: (i) Extended Disabil-ity Status Scale (EDSS) at most 1 (the presence of only neu-rological signs, but no sign or symptoms at upper limbs)� (ii)“normal” score for the “arm” portion of the Scripps Neurolog-ical Rating Scale (NRS) (Sipe et al. 1984) for the sensory, mo-tor and cerebellar systems. The exclusion criteria were: (i) re-lapses within the last three months� (ii) treatment with corti-costeroids within the previous three months� (iii) Mini Mental

Fig. 2. Trajectories in the different phases of the experiments,for typical control (left) and MS subjects (right).

State Examination (MMSE) less than 24. This would corre-spond to below-normal cognitive abilities. The performanceof these subjects was compared with 11 age-matched controls.

A sample of the movements recorded during the variousexperimental phases (null field, force field and after effect) isdisplayed in Figure 2 for typical control and MS subjects. Thisfigure suggests that the performance of MS subjects differsfrom that of controls. Based on the above indicators, we com-pared the performance of MS subjects with that of controls.

5.1. Sensorimotor Performance

To test whether and how the motor performance of MS sub-jects differs from that of controls, we took the null-field por-tion of the adaptation experiment, when subjects had to be-come familiarized with the robotic device. More specifically,we looked at the effects of disease (control, MS) in the motorperformance, quantified through the indicators defined above,and at how disease affects their time course during the null-field phase. We found that in MS subjects, movements lastedlonger (0�59 � 0�02 s for controls versus 0�77 � 0�04 s forMS subjects, mean � SE� p � 0�00027, two-way repeated-measures analysis of variance (ANOVA)) and their speedprofile was significantly less symmetric with respect to con-trols (0�79� 0�03 versus 0�70� 0�03� p � 0�041). Moreover,in MS subjects, movements are more curved (4�8 � 0�4 % vs9�8 � 2�3 %� p � 0�046) and less smooth (7�4 � 0�5 versus11�1 � 0�7� p � 0�00054) than those of controls. Finally, MS

© 2007 SAGE Publications. All rights reserved. Not for commercial use or unauthorized distribution. at PENNSYLVANIA STATE UNIV on April 10, 2008 http://ijr.sagepub.comDownloaded from

1230 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / November/December 2007

subjects revealed a significantly greater 100 ms aiming error(11�4� 0�5 versus 15�5� 0�6� p � 0�00003).

During the familiarization phase, both control and MS sub-jects gradually improved their performance, as reflected by thefact that all indicators displayed a significant time effect. Weasked whether there are differences in the rate of improvementin MS subjects and controls (interaction between disease andtime). However, none of the above-mentioned indicators dis-played significant disease–time interactions.

5.2. Force-�eld Adaptation

We then focused on force-field adaptation and asked whetherMS subjects could learn to compensate for the force field.For each performance indicator, we ran a repeated-measuresANOVA with two factors: disease and time (early, late) overthe force field target sets. We found significant effects (p �0�05) of disease (all indicators excluding the jerk ratio) andhighly significant (p � 0�01) effects of time (all indicatorsexcluding the 100 ms aiming error).

The lack of significance of disease in the jerk ratio suggeststhat under the effect of the force field, although their sensori-motor performance is distinctively different, MS and controlsubjects have similar performance in making feedback correc-tions. The lack of time effect in the 100 ms aiming error is dueto the fact that the speed-dependent field does not affect theearly portion of the movement, in which the speed is low� seeFigure 3.

As regards the interactions between disease and time dur-ing field adaptation, only the jerk index displayed a significanteffect (p � 0�029). A closer look at Figure 3 suggests that thisis due to the fact that MS subjects have a greater improvement.In conclusion, MS subjects and controls are basically similarin their ability to adapt and in their rate of adaptation. This isclearly shown in Figure 3 for all indicators.

We finally asked whether MS subjects differ from controlsin their rate of adaptation. To do this, we compared the learningindex (Equation (4)) estimated from MS and control subjects.Even in this case, we found no significant differences. This isclearly shown in Figure 4.

6. Discussion

6.1. Abnormal Sensorimotor Performance and PreservedAdaptation Capability

Analysis of the motor performance during the null-field phasesuggests that MS patients who are asymptomatic on clini-cal examination indeed have subtle incoordination problems.Their movements are more curved, less smooth and have agreater aiming error, i.e. they start in the wrong direction. Thisis consistent with previous findings (Solaro et al. 2007), in

which hand trajectories were recorded with a digitizing tablet.In the present experiment, the observed abnormalities may beenhanced by the non-neglectable dynamics of the manipulan-dum.

The greater aiming error suggests that MS subjects generateinappropriate motor commands. This may reflect an inaccurateaccount for the anisotropy of the inertia of the arm and the ro-bot system. Normal subjects make errors as well (Gordon et al.1994), but MS subjects display an error of greater magnitude.Nevertheless, such errors do not prevent subjects from reach-ing the targets. This is likely because these errors are compen-sated for, at least in part, by making online corrections basedon the available visual and/or proprioceptive information. Thismay explain the increased magnitude of the jerk index.

As regards the ability to adapt to the force field, we foundno significant differences between MS subjects and controlsin the actual ability to adapt and in the magnitude and rateof adaptation. Therefore, like controls, MS subjects effectivelylearn to predict the force field generated by the robot ratherthan just trying to resist to perturbations by stiffening theirarm.

It should be noted, however, that the failure to detect a dif-ference between groups does not rule out the possibility thata more sensitive test might actually detect a difference in theway adaptation is achieved.

6.2. Fatigue in MS Patients

One major concern with MS subjects is that they become eas-ily fatigued. In the present experiments, subjects were allowedto rest between consecutive target sets� however, only one MSsubject actually did. This suggests that the task was well tol-erated. Moreover, Figure 3 and statistical analysis shows thatthere was no degradation of performance at the end of the re-covery phase as compared with the final portion of the nullphase. This, again, suggests that subjects displayed no prob-lems in performing the task in the late phase of the experiment.Indeed, a small improvement was often observed.

6.3. Implications for Robot Therapy in MS Subjects

What are the implications of these results for rehabilitation,considering that no consensus has been reached so far on themost effective approaches to robot therapy? The proposed ap-proaches fall into three basic categories, i.e. assistive, perturb-ing and adaptive.

In assistive approaches, the robot provides assistive orrestoring forces, which may vary in proportion to the instanta-neous difference between desired and actual movement (Krebset al. 1998) or on the basis of a position-dependent force field,directed to the intended target (Casadio et al. 2006).

In perturbing approaches, the robot generates disturbanceswhile the subject is performing a movement or maintaining a

© 2007 SAGE Publications. All rights reserved. Not for commercial use or unauthorized distribution. at PENNSYLVANIA STATE UNIV on April 10, 2008 http://ijr.sagepub.comDownloaded from

Casadio, Sanguineti, Solaro, and Morasso/ A Haptic Robot Reveals the Adaptation Capability of Individuals with MS 1231

Fig. 3. Time course of movement performance indicators during the different phases of the experiment (N: null field, F: forcefield). Thin curves indicate � SE.

particular posture. The notion that perturbations promote re-covery is supported by clinical and physiological observations(Takahashi et al. 2001� Patton et al. 2006).

Adaptive approaches rely on observations that are simi-lar to those reported here: exposition to a force field inducesan adaptation, which results in a gradual recovery of straightmovements. This occurs by building an internal representa-

tion of dynamics, as demonstrated by the presence of “aftereffects” when the field is removed. Such force fields maybe customized to the individual subject (Patton and Mussa-Ivaldi 2004). However, the very same motor adaptation pro-tocol described here may also be used as an adaptive ther-apy (Patton et al. 2006). In chronic stroke survivors, force-field training resulted in an improvement of motor perfor-

© 2007 SAGE Publications. All rights reserved. Not for commercial use or unauthorized distribution. at PENNSYLVANIA STATE UNIV on April 10, 2008 http://ijr.sagepub.comDownloaded from

1232 THE INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH / November/December 2007

Fig. 4. Top: aiming error (300 ms) for control and MS subjects.The dashed curves refer to the catch trials. Bottom: learningindex estimated from the indicator above.

mance during the late after-effect phase with respect to thebaseline.

In our experiments, impairment is too small to allow anyobservation of improvement of this kind, but this same pro-tocol might be beneficial for MS patients with a more severeimpairment.

In the case of cerebellar lesions, there is some indication(Bastian et al. 1996� Sanguineti et al. 2003) that defective coor-dination may depend on an abnormal capability of anticipatingthe muscle activations needed to compensate for arm dynam-ics (feed-forward control). The associated symptoms includeataxia (an inability to organize coordinated movements thatinvolve multiple joints) and kinetic tremor. At present, no ef-fective treatments are available for this condition, which is ex-tremely disabling and is common in patients with MS: impair-ment of the cerebellar system has been estimated to be presentin about 80% of MS patients.

Therefore, to answer the question that we formulated atthe beginning of the section, we suggest that techniques thatspecifically train the ability to predict the dynamics of body

and external environment would seem appropriate for thesepatients.

6.4. Future of Robotics Technology in the Rehabilitation ofMS Subjects

Even though the majority of studies regarding the evidence forefficacy of rehabilitation in MS are on patients with chronicprogressive MS (Freeman et al. 1997), there is growing ev-idence that patients with MS can benefit from rehabilitationinterventions after an acute relapse with incomplete recovery(Liu et al. 2003). Some studies suggest that some cortical re-organization in patients with MS may occur, but it is unclearwhether this plays a role in MS rehabilitation (Rasova et al.2005). The main effect is likely to be a consequence of im-proved compensation, adaptation and reconditioning (Rocca etal. 2002).

The above considerations suggest that in the rehabilita-tion of MS subjects, robots should primarily aim at promotingthe development and the maintenance of compensatory and/oradaptive capabilities. Moreover, owing to the peculiar variabil-ity of types and degrees of impairment found in MS subjects,the timing and mode of rehabilitation treatment in MS patientsshould be set individually, taking into account the degree andextent of impairment.

This points to the need for integration between rehabilita-tion exercises and the continuous monitoring of the control andadaptation performance. Robots should be able to discriminatebetween the ways in which different pathological conditionscan affect the control patterns and/or the learning capability.In particular, it makes sense to independently assess the de-gree of impairment of the two aspects (control and learning) inthe early phases of the disease in which the functional impair-ment is still mild and results in little or no disability, and thereis ground for very focused interventions.

This study suggests that robotic technology is potentiallyuseful in both types of assessment. It remains to be seen infurther research whether robotic devices may be effective inrestoring movement ability in these patients.

References

Bastian, A. J. et al. (1996). Cerebellar ataxia: abnormal con-trol of interaction torques across multiple joints. Journal ofNeurophysiology, 76(1): 492–509.

Burdet, E. et al. (2006). Stability and motor adaptation in hu-man arm movements. Biological Cybernetics, 94(1): 20–32.

Casadio, M. et al. (2006). Impedance controlled, minimallyassistive robotic training of severely impaired hemipareticpatients. Proceedings of the 1st IEEE/RAS-EMBS Inter-national Conference on Biomedical Robotics and Bio-mechatronics, Pisa, Italy.

© 2007 SAGE Publications. All rights reserved. Not for commercial use or unauthorized distribution. at PENNSYLVANIA STATE UNIV on April 10, 2008 http://ijr.sagepub.comDownloaded from

Casadio, Sanguineti, Solaro, and Morasso/ A Haptic Robot Reveals the Adaptation Capability of Individuals with MS 1233

Casadio, M. et al. (2006). Braccio di Ferro: a new hapticworkstation for neuromotor rehabilitation. Technology andHealth Care, 13: 1–20.

Criscimagna-Hemminger, S. E. et al. (2003). Learned dynam-ics of reaching movements generalize from dominant tonondominant arm. Journal of Neurophysiology, 89(1): 168–176.

Donchin, O. et al. (2003). Quantifying generalization fromtrial-by-trial behavior of adaptive systems that learn withbasis functions: theory and experiments in human motorcontrol. Journal of Neuroscience, 23(27): 9032–9045.

Freeman, J. A. et al. (1997). The impact of inpatient rehabilita-tion on progressive multiple sclerosis. Annals of Neurology,42(2): 236–244.

Gomi, H. and Kawato, M. (1996). Equilibrium-point con-trol hypothesis examined by measured arm stiffness duringmultijoint movement. Science, 272(5258): 117–120.

Gordon, J. et al. (1994). Accuracy of planar reaching move-ments. II. Systematic extent errors resulting from inertialanisotropy. Experimental Brain Research, 99(1): 112–130.

Krebs, H. I. et al. (1998). Robot-aided neurorehabilitation.IEEE Transactions on Rehabilitation Engineering, 6(1):75–87.

Krebs, H. I. et al. (1999). Overview of clinical trials with MIT-MANUS: a robot-aided neuro-rehabilitation facility. Tech-nology and Health Care, 7(6): 419–423.

Liu, C. et al. (2003). Does neurorehabilitation have a role inrelapsing–remitting multiple sclerosis? Journal of Neurol-ogy, 250(10): 1214–1218.

Maschke, M. et al. (2004). Hereditary cerebellar ataxia pro-gressively impairs force adaptation during goal-directedarm movements. Journal of Neurophysiology, 91(1): 230–238.

Mussa-Ivaldi, F. A. et al. (1985). Neural, mechanical, and geo-metric factors subserving arm posture in humans. Journalof Neuroscience, 5(10): 2732–2743.

Patton, J. L. and Mussa-Ivaldi, F. A. (2004). Robot-assisted adaptive training: custom force fields for teach-ing movement patterns. IEEE Transactions on BiomedicalEngineering, 51(4): 636–646.

Patton, J. L. et al. (2006). Evaluation of robotic training forcesthat either enhance or reduce error in chronic hemipareticstroke survivors. Experimental Brain Research, 168(3):368–383.

Poser, C. M. et al. (1983). New diagnostic criteria for multi-ple sclerosis: guidelines for research protocols. Annals ofNeurology, 13(3): 227–231

Prange, G. B. et al. (2006). Systematic review of the effectof robot-aided therapy on recovery of the hemiparetic armafter stroke. Journal of Rehabilitation Research and Devel-opment, 43(2): 171–184.

Rasova, K. et al. (2005). Is it possible to actively and purposelymake use of plasticity and adaptability in the neuroreha-

bilitation treatment of multiple sclerosis patients? A pilotproject. Clinical Rehabilitation, 19(2): 170–181.

Rocca, M. A. et al. (2002). Adaptive functional changes in thecerebral cortex of patients with nondisabling multiple scle-rosis correlate with the extent of brain structural damage.Annals of Neurology, 51(3): 330–339.

Rocca, M. A. et al. (2003). Functional cortical changes inpatients with multiple sclerosis and nonspecific findingson conventional magnetic resonance imaging scans of thebrain. Neuroimage, 19(3): 826–836.

Rocca, M. A. et al. (2005). Cortical adaptation in patients withMS: a cross-sectional functional MRI study of disease phe-notypes. The Lancet Neurology, 4(10): 618–626.

Sanguineti, V. et al. (2003). Cerebellar ataxia: quantitative as-sessment and cybernetic interpretation. Human MovementScience, 22(2): 189–205.

Scheidt, R. A. et al. (2001). Learning to move amid uncer-tainty. Journal of Neurophysiology, 86(2): 971–985.

Shadmehr, R. and Mussa-Ivaldi, F. A. (1994). Adaptive rep-resentation of dynamics during learning of a motor task.Journal of Neuroscience, 14(5 Pt 2): 3208–3224.

Shadmehr, R. and Wise, S. P. (2005). The ComputationalNeurobiology of Reaching and Pointing: a Foundation forMotor Learning. Cambridge, MA, MIT Press.

Sipe, J. C. et al. (1984). A neurologic rating scale (NRS) foruse in multiple sclerosis. Neurology, 34(10): 1368–1372.

Smith, M. A. et al. (2000). Motor disorder in Huntington’s dis-ease begins as a dysfunction in error feedback control. Na-ture, 403(6769): 544–549.

Smith, M. A. and Shadmehr, R. (2005). Intact ability to learninternal models of arm dynamics in Huntington’s diseasebut not cerebellar degeneration. Journal of Neurophysiol-ogy, 93(5): 2809–2821.

Solaro, C. et al. (2007). Subtle upper limb impairmentin asymptomatic multiple sclerosis subjects. MultipleSclerosis 13(3): 428–432.

Stocco, L. et al. (1998). Past constrained global minimax opti-mization of robot parameters. Robotica, 16: 595–605.

Takahashi, C. D. and Reinkensmeyer, D. J. (2003). Hemi-paretic stroke impairs anticipatory control of arm move-ment. Experimental Brain Research, 149(2): 131–140.

Takahashi, C. D. et al. (2001). Impedance control and internalmodel formation when reaching in a randomly varying dy-namical environment. Journal of Neurophysiology, 86(2):1047–1051.

Teulings, H. L. et al. (1997). Parkinsonism reduces coordina-tion of finger, wrist, and arm fine motor control. Experi-mental Neurology, 146: 159–170.

Tremblay, S. et al. (2003). Somatosensory basis of speech pro-duction. Nature 423(6942): 866–869.

Volpe, B. T. et al. (1999). Robot training enhanced motor out-come in patients with stroke maintained over 3 years. Neu-rology 53(8): 1874–1876.

© 2007 SAGE Publications. All rights reserved. Not for commercial use or unauthorized distribution. at PENNSYLVANIA STATE UNIV on April 10, 2008 http://ijr.sagepub.comDownloaded from