Content uploaded by Konstantinos Michmizos
Author content
All content in this area was uploaded by Konstantinos Michmizos on Nov 19, 2017
Content may be subject to copyright.
Uncorrected Author Proof
NeuroRehabilitation xx (20xx) x–xx
DOI:10.3233/NRE-171458
IOS Press
1
Pediatric robotic rehabilitation: Current
knowledge and future trends in treating
children with sensorimotor impairments
1
2
3
Konstantinos P. Michmizosa,∗and Hermano Igo Krebsb
4
aDepartment of Computer Science, Rutgers University, Piscataway, NJ, USA5
bDepartment of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA, USA6
Abstract.7
BACKGROUND: Robot-aided sensorimotor therapy imposes highly repetitivetasks that can translate to substantial improve-
ment when patients remain cognitively engaged into the clinical procedure, a goal that most children find hard to pursue.
Knowing that the child’s brain is much more plastic than an adult’s, it is reasonable to expect that the clinical gains observed
in the adult population during the last two decades would be followed up by even greater gains in children. Nonetheless, and
despite the multitude of adult studies, in children we are just getting started: There is scarcity of pediatric robotic rehabilitation
devices that are currently available and the number of clinical studies that employ them is also very limited.
8
9
10
11
12
13
PURPOSE: We have recently developed the MIT’s pedi-Anklebot, an adaptive habilitation robotic device that continuously
motivates physically impaired children to do their best by tracking the child’s performance and modifying their therapy
accordingly. The robot’s design is based on a multitude of studies we conducted focusing on the ankle sensorimotor control.
In this paper, we briefly describe the device and the adaptive environment we built around the impaired children, present the
initial clinical results and discuss how they could steer future trends in pediatric robotic therapy.
14
15
16
17
18
CONCLUSIONS: The results support the potential for future interventions to account for the differences in the sensorimotor
control of the targeted limbs and their functional use (rhythmic vs. discrete movements and mechanical impedance training)
and explore how the new technological advancements such as the augmented reality would employ new knowledge from
neuroscience.
19
20
21
22
Keywords: Rehabilitation robotics, robot-aided therapy, robot-aided neurorehabilitation, pediatric, cerebral palsy, adaptive
robotic therapy
23
24
1. Introduction25
Robot-aided neurorehabilitation is a form of phys-
26
ical or occupational therapy that uses a robotic device
27
to educate or re-educate the brain on how to move an
28
∗Address for correspondence: Konstantinos P. Michmizos,
Laboratory for Computational Brain, Department of Computer
Science, Rutgers University, 110 Frelinghuysen Road, Piscataway,
NJ 08854, USA. Tel.: +1 848 445 8841; Fax: +1 732 445 0537;
E-mail: konstantinos.michmizos@cs.rutgers.edu.
impaired limb in the presence of an acquired neuro- 29
logical disease, such as stroke, or a developmental 30
motor disability, such as cerebral palsy. The clinical 31
evaluation of rehabilitation devices provided strong 32
evidence that robots can replicate, even augment, the 33
functional improvement achieved by human thera- 34
pists (Kwakkel, Kollen et al., 2008) with comparable 35
cost (Lo, Guarino et al., 2010; Wagner, Lo et al., 36
2011). That is why the American Heart Associa- 37
tion now recommends robotic therapy as a standard 38
1053-8135/17/$35.00 © 2017 – IOS Press and the authors. All rights reserved
Uncorrected Author Proof
2K.P. Michmizos and H.I. Krebs / Pediatric robotic rehabilitation
post-stroke treatment, at least for the upper extrem-39
ities (UE) (Miller, Murray et al., 2010; Winstein,
40
Stein et al., 2016). A striking finding over the last41
few years is that a successful sensorimotor recovery42
requires more than limb motion: In fact, the first43
effort to employ the patient’s active participation into
44
robotic therapy (Krebs, Palazzolo et al., 2003) led45
to a series of adaptive treatments aiming to con-46
tinuously challenge and assist a patient to improve;47
This has signified the assist-as-needed robotic ther-
48
apy as a key mechanism for improving motor function49
(Volpe, Krebs et al., 2000; Patton & Mussa-Ivaldi50
2004; Patton, Stoykov et al., 2006).
51
Especially for children with sensorimotor impair-52
ments, motor function has been found to improve
53
by purposeful activity and task-specific training that54
adapts to each patient’s skills and needs (Valvano55
2004; Michmizos, Rossi et al., 2015). Although other,
56
highly intensive, approaches such as constraint-57
induced movement therapy have given positive58
results in children with mild upper limb impairments59
(Huang, Fetters et al., 2009; Gordon, Hung et al.,60
2011), what constitutes the appropriate treatment,61
especially for children with more severe sensori-62
motor impairments, remains to be found. For as
63
long as motor learning remains the major working
64
hypothesis for sensorimotor rehabilitation (Hogan,65
Krebs et al., 2006), a well-designed robotic ther-
66
apy, for either adults or children, should follow the67
principles of motor learning, namely massed prac-68
tice, cognitive engagement and functional relevance
69
(Damiano 2006). However, massed practice is dif-70
ficult to achieve in children whereas the number
71
of executed movements by itself cannot be fully
72
accounted for the improvement achieved through73
robots. What is more, cognitive engagement requires74
strong concentration, which is very difficult to pre-75
serve in children. Even the functional relevance76
of a rehabilitation task is not always appealing or77
even apparent to most children who rarely use their
78
impaired limb to execute the targeted task. These79
are some of the biggest challenges that any pediatric
80
robotic habilitation device should aim to address.81
We have recently introduced the MIT’s pedi-82
Anklebot, a robotic device that provides an intensive83
task-specific sensorimotor therapy to the ankle of84
children with motor disabilities (Michmizos, Rossi85
et al., 2015). The device lives in a human-robot sym-86
biotic system that assesses the ability of the child
87
to move and adapts the difficulty of the movement
88
(Michmizos & Krebs 2012; Michmizos & Krebs
89
2012) by changing its speed and accuracy constraints
90
(Michmizos & Krebs 2014; Michmizos & Krebs 91
2014; Michmizos, Vaisman et al., 2014). In this paper, 92
we briefly present the device, as well as the clinical 93
results upon its employment in two pediatric Hos- 94
pitals, Blythedale Children’s Hospital in Valhalla, 95
NY, USA and Bambino Ges´
u Children’s Hospital in 96
Rome, Italy, and show how these results can influence 97
future trends in pediatric robotic therapy. 98
2. The MIT’s pedi-Anklebot 99
The MIT’s pedi-Anklebot is an impedance- 100
controlled low-friction, back-driveable habilitation 101
device that targets the ankle joints and aims to pro- 102
mote motor learning in children of ages 6–10 years 103
old (Michmizos, Rossi et al., 2015). While the design 104
of its hardware follows that of the adult version of 105
the device, its software (controller and a set of seri- 106
ous games) aims to tackle the distinct challenges 107
in pediatric habilitation. To this end, we expanded 108
the performance-based adaptive robotic therapeutic 109
scheme that we had previously introduced for our UE 110
robots (Krebs, Palazzolo et al., 2003) to include motor 111
learning principles (Michmizos & Krebs 2012). In 112
the following sections, we briefly describe how the 113
device adapts the therapy to meet the needs and 114
the special challenges associated with the pediatric 115
habilitation, demonstrate its potential as a therapeu- 116
tic device that can induce motor learning in discrete 117
ankle movements and show evidence that its use can 118
result to notable changes in the walking pattern. 119
2.1. Hardware design 120
The pedi-Anklebot provides active assistance to 121
the ankle in 2 degrees-of-freedom (DOF), namely 122
dorsi-plantar flexion (DP) and inversion-eversion 123
(IE); it also has a passive DOF for internal-external 124
rotation. Overall, it allows 25◦dorsiflexion, 45◦125
plantar flexion, 25◦inversion, 15◦eversion, and 126
15◦internal/external rotation, which are near the 127
maximum range of comfortable motion for normal 128
subjects and beyond what is required for typical gait 129
(Weiss, Kearney et al., 1986; Weiss, Kearney et al., 130
1986). The device consists of two linear actuators 131
mounted in parallel so that if both push or pull in the 132
same (opposite) direction, a DP flexion (IE) torque 133
is produced (Fig. 1a). The actuation is provided by 134
2 brushless DC motors and a Rohlix linear traction 135
drive that can deliver a maximum torque ∼7.21 Nm in 136
DP flexion and ∼4.38 Nm in IE. The torque capability 137
Uncorrected Author Proof
K.P. Michmizos and H.I. Krebs / Pediatric robotic rehabilitation 3
Fig. 1. The MIT’s pediatric ankle robotic system (pedi-Anklebot). (A) The mechanical design of the robotic device showing the components
of the device. (B) An 8-year old child wearing the robot in seated position; the weight of the robot is supported from the chair, through a
screw-bolt. (C) The set of the adaptive serious games that were developed for the device based on motor learning principle (up) and the
relative positioning of the patient with respect to the projected games (bottom).
lifts approximately 25% of a child’s weight and there-138
fore can be used as a supplemental support to the
139
paretic ankle plantarflexors during walking; Alterna-140
tively, the robot can be used in a seated position where141
discrete ankle movements can be trained (Fig. 1b).142
The robot is also equipped with two motion sensors,
143
a mini-rail linear encoder (Schneeberger) and a Gur-144
ley rotary encoder, as well as load cells (Futek) at145
each actuator output.146
2.2. Software design
147
We have developed a set of 3 goal-directed serious148
games (SGs), namely a race (Noah’s Ark), Soccer
149
and Shipwreck games (Fig. 1c), to address motor150
impairments including poor coordination, impaired151
motor speed or accuracy, and diminished strength,152
as well as to address cognitive or perceptual impair-153
ments (Michmizos & Krebs 2012). To do so, all our154
SGs follow the same design principles that include155
meaningful play and challenge and have: a) an inter-156
esting concept, to support the level of perceptual157
joy throughout the therapeutic sessions; b) a simple158
visual interface, to communicate easily the game con-159
cept; c) easy controls, to facilitate guidance around160
the visual interface and focus on the game concept;161
and d) simple rules, to minimize learning period.162
The SGs are seamlessly integrated with the hard-163
ware and the controller of the pedi-Anklebot as they164
require discrete or rhythmic movements which can be165
assisted-as-needed by the robot. In addition, the task
166
objectives of the race, the soccer and the shipwreck
167
games were blocked, serial, and random, respec-
168
tively, covering the entire spectrum of the structured 169
practice. The predictability of the available game- 170
play environments also varied greatly. The race game 171
could provide a closed environment whereas the 172
other two game environments were open, with that 173
of the Shipwreck varying the most. As the player 174
plays the game and his or her skills and familiarity 175
increases, the game offers a higher level of chal- 176
lenge to retain attention and motivation; nonetheless, 177
since a too-difficult-to-play game will make a child to 178
become frustrated, the gameplay could become easier 179
if needed. Overall, the set of the SGs can promote all 180
three stages of motor learning: cognitive, associative, 181
and autonomous (Fitt & Posner 1967). 182
To retain active participation during therapy, we 183
translated to the ankle movements the concept of 184
adaptive assist-as-needed robotic therapy introduced 185
for the UE. Specifically, our recent finding that 186
the performance in visually evoked, visually guided 187
ankle pointing movements can be described by a lin- 188
ear function, as predicted by Fitts’ law, supported the 189
idea that the speed-accuracy trade-off (SAT) could 190
be incorporated into an assist-as-needed therapeu- 191
tic intervention for the ankle (Michmizos & Krebs 192
2014). In that sense, the SGs trained the child’s ankle 193
while challenging his/her ability to move fast and 194
accurately: Depending on one’s ability to aim, the 195
target of the game became smaller or larger; Depend- 196
ing on one’s ability to move fast or slow, the speed of 197
the game also changed; for details see (Michmizos & 198
Krebs 2012; Michmizos, Rossi et al., 2015). 199
To challenge children to improve while keep-
ing them motivated, each therapeutic session was
Uncorrected Author Proof
4K.P. Michmizos and H.I. Krebs / Pediatric robotic rehabilitation
grouped into sections of nrepetitions. For each
section, we defined the performance level (PL) as
follows:
PL =⎧
⎪
⎨
⎪
⎩
−1|PM <−0.1
0|−0.1≤PM ≤+0.1
+1|PM >+0.1
(1)
where the value of PL indicates whether patients per-
200
formed worse (PL = –1) or better (PL = 1) than their
201
expected ability (PL = 0 denotes when patients per-
202
form approximately the same); and PM is one of the 4
203
performance metrics (PMs) that we used, namely the
204
ability to initiate movement, the power to move from
205
the starting position to the target, the ability to reach206
the target efficiently and in a timely manner, and to
207
reach the target accurately; for details see (Michmi-208
zos, Rossi et al., 2015). By averaging over PM values209
and a weighted sum of the PL values in 3 consecutive210
sections (Fig. 2), the controller adapted to children’s
211
performance and variability, and challenged them to212
continue to improve. The performance-based adap-213
tive algorithm for the speed, s, and the accuracy, w,
214
in section J+1, is as follows:
215
s[J+1]=s[J]+λs·a(PLsum)·PMs[J](2)216
w[J+1]=w[J]+λw·a(PLsum)·PMw[J],(3)217
where λs>0,λ
w<0 were the gains of the simple
control loop and:
a(PLsum)=
⎧
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
1.00
0.75
0.50
0.25
0.25
0.50
0.75
1.00
1.50
2.00
−14 ≤PLsum ≤−1
−10 < PLsum ≤−8
−8< PLsum ≤−6
−6< PLsum ≤−3
−3< PLsum ≤3
3< PLsum ≤6
6< PLsum ≤8
8< PLsum ≤10
10 < PLsum ≤12
12 < PLsum ≤14
.(4)
The challenging component of our games was the
218
asymmetry of a (PLsum), which aimed to trigger219
patients to improve further but made the task eas-220
ier, yet to a lesser extent, when the performance was221
worsening.
Fig. 2. Weighted sum of three consecutive PL values for adapting
the game speed and accuracy constraints. PLsum is used for chal-
lenging performance. A window of size 3 is adjusted to each PL
so that the current PL value is weighted by 4, and the previous two
PL values are weighted by 2 and 1, respectively.
2.3. Experimental setup 222
The children wore a modified shoe and a knee 223
brace, to which the robot was attached. Subjects were 224
seated and the knee brace was securely fastened to the 225
chair to fully support the weight of the robot and to 226
ensure that measurements were made in a repeatable 227
posture. A screen was positioned at eye level (Fig. 1c). 228
Ankle position kinematics, with respect to the zero- 229
angle (neutral position), were recorded at 200 Hz 230
sampling frequency and converted to screen pixels 231
for visualization. A DP (IE) movement of the ankle 232
moved the cursor vertically (horizontally); hence the 233
cursor moved in a 2D coordinate system with the 234
origin corresponding to the ankle’s neutral position 235
defined as the sole being at a right angle to the tibia. 236
2.4. Experimental protocols 237
We have so far completed 2 pilot pedi-Anklebot 238
studies that are described in detail elsewhere (Mich- 239
mizos, Rossi et al., 2015; Krebs, Michmizos et al., 240
2016). Briefly, in the first study (discrete movements 241
study - DMS), we recruited 3 impaired children 242
(average age 9 years old) at the “Bambino Ges´
u” 243
Children’s Hospital, Rome, Italy; 2 of the children 244
were diagnosed with cerebral palsy (CP) and the other 245
was diagnosed with a lesion of the common peroneal 246
nerve. In the second study (rhythmic movement – 247
walking pattern study – WPS), we recruited 4 CP 248
children (average age 9 years old) at Blythedale Chil- 249
dren’s Hospital, Valhalla, NY. Enrollment criteria for 250
this study were a) age 6–10 years (based on child 251
size); b) congenital hemiplegia or acquired hemi- 252
plegia with at least 6 months from the acute phase; 253
c) cognitive and visual abilities adequate to under- 254
stand and perform the interactive robotic training; 255
Uncorrected Author Proof
K.P. Michmizos and H.I. Krebs / Pediatric robotic rehabilitation 5
Table 1
Discrete movement kinematics improvement (DMS). Reported numbers are % changes in average values between admission and discharge
in 3 children (CD-1, CD-2 and CD-3). Metrics are defined in (Michmizos, Rossi et al. 2015)
% Speed Accuracy Robot Robot Smooth Slot Dwell Min Min
change Initiated Assistance Movement Time Time Trajectory Distance
Movements
DP IE DP IE DP IE DP IE DP IE DP IE DP IE DP IE DP IE
CD-1 51 24 45 80 83 24 100 80 317 205 4 2 476 107 6 2.5 750 100
CD-2 33 29 30 61 73 29 100 61 3850 28 26 2 766 98 8 0 460 55
CD-3 70 5 46 81 100 5 100 81 4700 1000 26 3 53 382 3 4 450 60
d) Tardieu spasticity grade ≤3 at ankle plantar flexor256
muscles; e) ability to walk independently without the
257
use of aids or orthoses, f) agreed not to receive focal258
treatment with Botulinum toxin serotype A (BoNT-259
A) for muscle spasticity during the study; g) signed260
consent and assent. Children were excluded if they261
had a) functional surgery; or b) focal treatment with
262
BoNTA for muscle spasticity in the prior 6 months,263
in their lower extremities. For both studies, the chil-264
dren and their parents gave informed assent and
265
consent according to the procedure approved by the266
Ethics and Institutional Review Board committee of267
“Bambino Ges´
u” Children’s Hospital, Rome, Italy268
(DMS), the Blythedale’s Institutional Review Board
269
(WPS) and the Massachusetts Institute of Technology
270
Committee on the Use of Humans as Experimental271
Subjects.272
For both studies, the children were trained twice273
per week for 6 weeks in seated position for a total
274
of 12 sessions. Training was unilateral focusing on
275
the most impaired side (even in the case of bilat-
276
eral impairment). Each robotic training started with
277
an active-Range of Movement (A-ROM) measure-278
ment by a therapist and 80% of the A-ROM became279
the maximum ROM for the pedi-Anklebot. For the280
DMS study, each child conducted 44 movements per
281
direction (DP and IE). For the WPS, each child exe-282
cuted 6 blocks of 40 movements each for a total283
of 240 targeted ankle movements employing Noah’s284
Ark or Shipwreck games in DP flexion followed or
285
preceded by the same number of repetitions for IE
286
movements. All games were played with the adaptive287
robotic assistance as described elsewhere (Michmi-
288
zos, Rossi et al., 2015). Since all subjects were naive289
to the task upon enrollment, we selected Noah’s Ark290
and Shipwreck game because they were designed for
291
the first stages of motor learning, namely the cogni-292
tive, and associative stages, where one may have a293
vague idea of the movement required for a task but294
might not be confident on how to execute that move-295
ment. To ensure a continuous engagement into the296
rehabilitation session, the cognitive status of the chil- 297
dren was continuously assessed by the clinicians and 298
nurses. 299
2.5. Experimental results 300
The DMS examined a large number of metrics 301
assessing the discrete movement kinematics. Inter- 302
estingly, all metrics were improved in all kids. Table 1 303
shows the changes in the average values of the kine- 304
matics metrics between admission and discharge. 305
The results were consistent for both trained direc- 306
tions (DP and IE) with the percent improvements 307
being larger in DP direction than in IE. Overall, at 308
discharge, all kids consistently exhibited less robot- 309
initiated movements, less assistive power from the 310
robot, movements that were smoother and took con- 311
siderably less time to complete in addition to being 312
more accurate. 313
The WPS examined functional tasks that are 314
required in walking. Specifically, it measured the 6- 315
minute walk (6-MWT), which measures the distance 316
the patient can walk in 6 minutes on a flat surface. 317
and the timed-up-and-go (TUG) test, which assesses 318
mobility and balance by measuring the time it takes 319
for a person to stand and walk 3 meters away, turn 320
back, and sit down. Table 2 shows the changes in these 321
two tests at admission and discharge of the protocol. 322
Table 2
Walking Metrics Improvement (WPS). Reported numbers are %
changes in functional walking tests between admission and
discharge in 4 children (CW-1, CW-2, CW-3 and CW-4).
Negative value indicates worsening. The functional tests
were the 6-minute walk (6-MWT) and timed-up-and–go
(TUG) tests. Of notice, CW-1 has bilateral
impairments and training was limited to unilateral
% change 6-MWT TUG
CW-1 –4.2 10.6
CW-2 8.2 9.3
CW-3 21.2 10
CW-4 4.7 11
Uncorrected Author Proof
6K.P. Michmizos and H.I. Krebs / Pediatric robotic rehabilitation
3. Discussion323
In this paper, we provide a proof of concept for the324
pedi-Anklebot used in two different clinical settings,325
with two different therapeutic targets: discrete ankle
326
movements and functional walking. Our approach327
for applying therapy to the ankle has a number of328
unique characteristics: First, the therapeutic scheme329
adapts to each kid separately, based on quantitative
330
measurements of kinematic and kinetic performance.
331
Second, the therapeutic environment employs con-332
cepts of motor learning, initially designed for the333
upper extremity but recently proved to hold for the
334
lower limbs as well (Michmizos & Krebs 2014; Mich-335
mizos & Krebs, 2014; Michmizos, Vaisman et al.,336
2014). Third, the robotic therapy targeted specific337
ankle joints and went beyond rhythmic training, con-
338
trary to the current school of thought that favors339
rhythmic patterns of whole-body movements dur-340
ing partial body weight support treadmill walking;341
a technique that was found to produce limited results
342
(Duncan, Sullivan et al., 2011). Fourth, targeting of343
discrete movements is a different rehabilitation strat-
344
egy, compared to the traditional view that a robot345
should imitate the manual therapy and, therefore,346
impose continuous rhythmic walking movements347
(Colombo, Joerg et al., 2000). The results pre-
348
sented here substantiate new approaches in designing349
the next generation of pediatric neuro-rehabilitation350
robots, as discussed below.351
3.1. The difference between working and
352
walking our way through the world
353
The pedi-Anklebot targeted the ankle because of its354
crucial role in human walking and because a deficit355
in foot control is the most common and debilitating356
condition in upper motor neuron disorders involv-
357
ing the corticospinal tract, such as CP that affects 1358
to 4 per 1,000 live births (Odding, Roebroeck et al.,
359
2006). However, the extent to which the sensorimo-360
tor control of the LE, in general, and of the ankle,
361
in particular, resembles that of the UE will influence362
the translation of any method, working hypothesis or363
technology used from the UE to the LE. Following364
this line of research, we have conducted three studies365
on the sensorimotor control of ankle pointing move-366
ments at 3 modeling levels. We first demonstrated367
that a macroscopic law of gross UE motor behavior,
368
the Fitts’ law, can also be used to described the ankle
369
movements in both DP and IE direction (Michmi-
370
zos & Krebs 2014). We then showed that the speed
371
profiles of the normal ankle movements are remark- 372
ably similar to the speed profiles of the arm and wrist 373
point movements (Michmizos, Vaisman et al., 2014). 374
Interestingly enough, we also found that the reaction 375
time, a metric that can be used for assessing implicit 376
learning, also increases linearly with the number of 377
stimuli, as would be predicted by the Hick-Hyman 378
law in UE (Michmizos & Krebs 2014). 379
However, any straight analogy between the LE and 380
the UE seems to be an oversimplification, especially 381
when one does not take into account the special char- 382
acteristics of the lower limb movements. Since the 383
functionality is not the same, the same neurological 384
cause will have a different functional outcome. In the 385
LE, a common condition that occurs in stroke and CP 386
is a weakness of the dorsiflexor muscles that lift the 387
foot during walking, commonly referred to as “drop 388
foot.” The two major complications of drop foot – 389
slapping of the foot after heel strike (foot slap) and 390
dragging of the toe during swing (toe drag) – present 391
a major challenge to efficient gait since clearing the 392
ground during the swing phase and maintaining ankle 393
stability during the stance phase are essential for effi- 394
cient gait. Another striking observation is that the 395
ankle sensorimotor control in DP and IE is far from 396
being regarded as a natural expansion of controlling 397
the arm or the wrist. We have indicated a statistically 398
significant difference in the lag of ankle response 399
between the DP and IE directions and have attributed 400
this to the stronger cortical projections that the tib- 401
ialis anterior (a main muscle for DP movement) has 402
(Michmizos & Krebs 2014). Therefore, one should 403
carefully study the anatomy, the neurophysiology 404
and the behavioral characteristics of the specific limb 405
prior to designing a therapeutic intervention targeting 406
at it. 407
3.2. Training discrete vs. rhythmic movements 408
The working hypothesis of our rehabilitation 409
robots is based on the model of dynamic primitives, 410
according to which the sensorimotor control of UE 411
and LE can be broken down into three primitives: 412
submovements (i.e., discrete movements), oscilla- 413
tions (i.e., rhythmic movements), and mechanical 414
impedances that are needed for interaction with 415
the physical environment (Hogan & Sternad 2013). 416
Since different brain areas are involved in controlling 417
discrete and rhythmic movements (Schaal, Sternad 418
et al., 2004), it seems reasonable to assume that a 419
rehabilitation device should target discrete and rhyth- 420
mic movements, separately. At least for walking, 421
K.P. Michmizos and H.I. Krebs / Pediatric robotic rehabilitation 7
any normal step includes collision with the pave-422
ment as well as balance and that is why mechanical
423
impedance should also be targeted as a potential424
therapeutic component. Built upon our rehabilitation425
model, the MIT-Skywalker was recently introduced426
to offer three distinct training modes: discrete move-
427
ments, rhythmic movements, and balance (Susko,428
Swaminathan et al., 2016). Although a larger clinical429
trial is still pending, there is already accumulating430
evidence that the device can improve clinical and
431
kinematic measurements, at least for the impaired432
population that has been tested upon.433
3.3. Robotic therapy integrated into an
434
augmented reality environment435
Despite the promising clinical results with the
436
robot “assisting as needed” even when movement437
is significantly impaired, the therapeutic intervention438
does not account for the cognitive and perceptual defi-
439
ciencies that accompany poor coordination and motor440
disorders. Therefore, the optimal habilitation recipe
441
and how this can be given while keeping the chil-442
dren engaged into the task remain to be determined.
443
Recently, we have introduced the idea of integrat-444
ing a rehabilitation robot to an augmented reality445
environment where the relevant movement is mir-446
rored to an animated character projected in front of
447
the child (Kommalapati & Michmizos 2016). We
448
speculate that the visual observation of one’s own
449
movements will activate the “mirror neuron system”,450
a brain system underlying the human capacity to learn
451
by imitation (Gallese, Fadiga et al., 1996; Rizzolatti,
452
Fadiga et al., 1996; Rizzolatti & Craighero 2004).453
Our rehabilitation algorithm personalizes the diffi-454
culty of the tasks by adapting the difficulty of reaching455
virtual targets on the animated environment through
456
changing the visual gain between real and animated
457
movements. Whether this will result to a measurable458
and significant therapeutic outcome remains to be
459
seen.460
In summary, the initial results that we report
461
here support the potential for new neurorehabilita-462
tion methods and robotic devices that augment the463
impaired child’s ability towards an independent and464
productive life. The introduction of adaptive mech-465
anisms into the pedi-Anklebot seems to increase466
the child’s motor response quality in both discrete467
and rhythmic tasks (functional walking). While more
468
thorough studies are still needed, the results have
469
raised a number of open questions: What kind of
470
computational features should we optimize to find the
471
right responders or get the maximum improvement? 472
What dose intensity should we give to each patient? 473
When should we target a whole body training or a 474
modular approach? How one should decide on the 475
type of a virtual environment? What kind of feedback 476
should we give to the kids? Should we use computer 477
animation to provide proprioceptive feedback during 478
therapy? What constitutes an appropriate set of diffi- 479
culty metrics for our games? Would a type of social 480
interaction during a gameplay benefit the youngsters? 481
Answering these and other questions would also help 482
clinicians and scientists studying the origin of motor 483
learning across the spatial and temporal scales of 484
brain function, and offer novel and highly testable 485
hypotheses for neuro-rehabilitation. The latter would 486
hopefully add new therapeutic avenues to the ones 487
currently pursued. 488
4. Conclusion 489
The robotic rehabilitation of pediatric patients 490
involves unique constraints and, therefore, creates 491
unique challenges in comparison to adult rehabilita- 492
tion. Our robotic device aims to provide an effective 493
therapy to children, by engaging them and fostering 494
their desire to put forth maximal effort in rehabil- 495
itation, as well as growing with them as progress is 496
made. To fully harness the therapeutic power of adap- 497
tation, we need to continue our research on the special 498
characteristics of the sensorimotor control of the tar- 499
geted joints as well as on the differences between 500
rhythmic and discrete movements, as well as mechan- 501
ical impedance training. Being an iterative procedure, 502
this evidence-based design of the robotic therapy will 503
continue to improve our expectations, from chang- 504
ing the behavioral and clinical metrics to making a 505
difference in a child’s life. 506
Conflict of interest 507
None to report. 508
References 509
Colombo, G., et al. (2000). Treadmill training of paraplegic 510
patients using a robotic orthosis. Journal of rehabilitation 511
research and development,37(6), 693. 512
Damiano, D. L. (2006). Activity, activity, activity: Rethinking our 513
physical therapy approach to cerebral palsy. Physical therapy,514
86(11), 1534-1540. 515
8K.P. Michmizos and H.I. Krebs / Pediatric robotic rehabilitation
Duncan, P. W., et al. (2011). Body-weight–supported treadmill
516
rehabilitation after stroke. New England Journal of Medicine,517
364(21), 2026-2036.518
Fitt, P., & Posner, M. (1967). Learning and skilled performance in
519
human performance. Belmont, CA: Brock-Cole.520
Gallese, V., et al. (1996). Action recognition in the premotor cortex.521
Brain,119(2), 593-609.522
Gordon, A. M., et al. (2011). Bimanual Training and Constraint-
523
Induced Movement Therapy in Children With Hemiplegic
524
Cerebral Palsy A Randomized Trial. Neurorehabilitation and
525
Neural Repair,25(8), 692-702.
526
Hogan, N., et al. (2006). Motions or muscles? Some behavioral fac-527
tors underlying robotic assistance of motor recovery. J Rehabil528
Res Dev,43(5), 605.529
Hogan, N., & Sternad, D. (2013). Dynamic primitives in the control530
of locomotion. Frontiers in computational neuroscience,7, 71.
531
Huang, H.-H., et al. (2009). Bound for success: A systematic532
review of constraint-induced movement therapy in children
533
with cerebral palsy supports improved arm and hand use. Phys-
534
ical therapy.535
Kommalapati, R., & Michmizos, K. P. (2016). Virtual reality536
for pediatric neuro-rehabilitation: Adaptive visual feedback of537
movement to engage the mirror neuron system. Engineering538
in Medicine and Biology Society (EMBC), 2016 IEEE 38th539
Annual International Conference of the, IEEE.540
Krebs, H. I., et al. (2016). Pediatric Anklebot: Pilot clinical trial.
541
Biomedical Robotics and Biomechatronics (BioRob), 2016 6th542
IEEE International Conference on, IEEE.543
Krebs, H. I., et al. (2003). Rehabilitation robotics: Performance-544
based progressive robot-assisted therapy. Autonomous Robots,
545
15(1), 7-20.546
Kwakkel, G., et al. (2008). Effects of robot-assisted therapy on547
upper limb recovery after stroke: A systematic review. Neu-548
rorehabil Neural Repair,22(2), 111-121.549
Lo, A. C., et al. (2010). Robot-assisted therapy for long-term550
upper-limb impairment after stroke. New England Journal of551
Medicine,362(19), 1772-1783.
552
Michmizos, K., et al. (2015). Robot-aided neurorehabilitation: A
553
pediatric robot for ankle rehabilitation. Neural Systems and
554
Rehabilitation Engineering, IEEE Transactions on,PP(99),555
1-1.
556
Michmizos, K. P. & Krebs, H. I. (2012). Assist-as-needed in lower557
extremity robotic therapy for children with cerebral palsy.558
Biomedical robotics and biomechatronics (BioRob), 2012 4th559
IEEE RAS & EMBS international conference on, IEEE.560
Michmizos, K. P. & Krebs, H. I. (2012). Serious Games for the561
Pediatric Anklebot. Biomedical Robotics and Biomechatronics562
(BioRob), 2012 4th IEEE RAS & EMBS International Confer-563
ence on.564
Michmizos, K. P., & Krebs, H. I. (2014). Pointing with the ankle:565
The speed-accuracy trade-off. Experimental brain research,566
232(2), 647-657.
567
Michmizos, K. P., & Krebs, H. I. (2014). Reaction time in ankle568
movements: A diffusion model analysis. Experimental brain
569
research,232(11), 3475-3488.570
Michmizos, K. P., et al. (2015). Robot-aided neurorehabilitation: A 571
pediatric robot for ankle rehabilitation. IEEE Transactions on 572
Neural Systems and Rehabilitation Engineering,23(6), 1056- 573
1067. 574
Michmizos, K. P., et al. (2014). A comparative analysis of speed 575
profile models for ankle pointing movements: evidence that 576
lower and upper extremity discrete movements are controlled 577
by a single invariant strategy. Frontiers in Human Neuro- 578
science,8(962). 579
Miller, E. L., et al. (2010). Comprehensiveoverview of nursing and 580
interdisciplinary rehabilitation care of the stroke patient a sci- 581
entific statement from the american heart association. Stroke,582
41(10), 2402-2448. 583
Odding, E., et al. (2006). The epidemiology of cerebral palsy: 584
Incidence, impairments and risk factors. Disability and Reha- 585
bilitation,28(4), 183-191. 586
Patton, J. L., & Mussa-Ivaldi, F. A. (2004). Robot-assisted adaptive 587
training: Custom force fields for teaching movement patterns. 588
Biomedical Engineering, IEEE Transactions on,51(4), 636- 589
646. 590
Patton, J. L., et al. (2006). Evaluation of robotic training forces that 591
either enhance or reduce error in chronic hemiparetic stroke 592
survivors. Experimental Brain Research,168(3), 368-383. 593
Rizzolatti, G., & Craighero, L. (2004). The mirror-neuron system. 594
Annu Rev Neurosci,27, 169-192. 595
Rizzolatti, G., et al. (1996). Premotor cortex and the recognition 596
of motor actions. Cognitive Brain Research,3(2), 131-141. 597
Schaal, S., et al. (2004). Rhythmic arm movement is not discrete. 598
Nature Neuroscience,7(10), 1136-1143. 599
Susko, T.,et al. (2016). MIT-skywalker: A novel gait neurorehabil- 600
itation robot for stroke and cerebral palsy. IEEE Transactions 601
on Neural Systems and Rehabilitation Engineering,24(10), 602
1089-1099. 603
Valvano, J. (2004). Activity-focused motor interventions for chil- 604
dren with neurological conditions. Physical & Occupational 605
Therapy in Pediatrics,24(1-2), 79-107. 606
Volpe, B., et al. (2000). A novel approach to stroke rehabilita- 607
tion Robot-aided sensorimotor stimulation. Neurology,54(10), 608
1938-1944. 609
Wagner, T. H., et al. (2011). An economic analysis of robot- 610
assisted therapy for long-term upper-limb impairment after 611
stroke. Stroke,42(9), 2630-2632. 612
Weiss, P., et al. (1986). Position dependence of ankle joint 613
dynamics—I. Passive mechanics. Journal of Biomechanics,614
19(9), 727-735. 615
Weiss, P. L., et al. (1986). Position dependence of ankle joint 616
dynamics—II. Active mechanics. Journal of Biomechanics,617
19(9), 737-751. 618
Winstein, C. J., et al. (2016). Guidelines for adult stroke rehabil- 619
itation and recovery: A guideline for healthcare professionals 620
from the american heart association/american stroke associa- 621
tion. Stroke,47(6), e98-e169. 622
