ArticlePDF Available

Custom-designed haptic training for restoring reaching ability to individuals with poststroke hemiparesis

Authors:
  • University of Illinois at Chicago & the Shirley Ryan AbilityLab

Abstract and Figures

We present an initial test of a technique for retraining reaching skills in patients with poststroke hemiparesis, in which errors are temporarily magnified to encourage learning and compensation. Individuals with poststroke hemiparesis held a horizontal plane robotic manipulandum that could exert a variety of forces while recording patients' movements. We measured how well the patients recovered movement straightness in a single visit to the laboratory (approximately 3 h). Following training, we returned forces to zero for an additional 50 movements to discern if aftereffects lasted. We found that all subjects showed immediate benefit from the training, although 3 of the 10 subjects did not retain these benefits for the remainder of the experiment. We discuss how these approaches demonstrate great potential for rehabilitation tools that augment error to facilitate functional recovery.
Content may be subject to copyright.
643
JRRD
JRRD
Volume 43, Number 5, Pages 643–656
August/September 2006
Journal of Rehabilitation Research & Development
Custom-designed haptic training for restoring reaching ability
to individuals with poststroke hemiparesis
James L. Patton, PhD;
1–2
*
Mark Kovic, BS, COTA/L;
1
Ferdinando A. Mussa-Ivaldi, PhD;
1–3
1
Sensory Motor Performance Program, Rehabilitation Institute of Chicago, Chicago, IL;
2
Departments of Physical
Medicine and Rehabilitation, Mechanical Engineering, and Biomedical Engineering and
3
Department of Physiology
and the Institute for Neuroscience, Northwestern University, Chicago, IL
Abstract—We present an initial test of a technique for retrain-
ing reaching skills in patients with poststroke hemiparesis, in
which errors are temporarily magnified to encourage learning
and compensation. Individuals with poststroke hemiparesis
held a horizontal plane robotic manipulandum that could exert
a variety of forces while recording patients’ movements. We
measured how well the patients recovered movement straight
-
ness in a single visit to the laboratory (~3 h). Following train-
ing, we returned forces to zero for an additional 50 movements
to discern if aftereffects lasted. We found that all subjects
showed immediate benefit from the training, although 3 of the
10 subjects did not retain these benefits for the remainder of
the experiment. We discuss how these approaches demonstrate
great potential for rehabilitation tools that augment error to
facilitate functional recovery.
Key words: adaptation, control, cortex, force fields, haptics,
hemiparesis, human, human-machine interface, impairment,
lesion, motor learning, rehabilitation, robots, stroke, teaching.
INTRODUCTION
The range of robotic possibilities for teaching and
rehabilitation has yet to be established, but the options do
go beyond what a therapist can do—robots are precise,
tireless devices that can measure progress with high
accuracy. We have been focusing on robotic forces that
may facilitate recovery from brain injuries such as stroke.
Some conventional therapeutic interventions use guid
-
ance and resistance principles to promote motor recovery
in the hemiparetic upper limb. Some traditional rehabili
-
tation sources recommend therapeutic intervention that
eliminates unwanted muscle activity and muscle tone and
then introduces normal movement patterns, which may
facilitate rehabilitation [1]. Other theories suggest that
facilitating reaching patterns promotes improvements in
motor function. One component of this approach is the
use of resistance in a direction opposite the movement
[2]. However, the most effective rehabilitation algorithms
have yet to be determined, which leaves a fertile area for
scientific inquiry.
Interestingly, several researchers are exploring
robotic techniques that are not necessarily designed to
imitate the conventional therapeutic process but to
instead uniquely probe new capabilities. For example,
one possible technique is to have the robot guide (pull)
the hand toward the desired trajectory and have the guid
-
ance transition to resistance as recovery progresses [3–4].
Another technique for hemiparetic stroke patients is pro
-
viding the patient with a bimanual master-slave robot
system, which guides the paretic arm by the actions of
Abbreviations: DOF = degrees of freedom, FM = Fugl-Meyer.
*
Address all correspondence to James L. Patton, PhD; Sen-
sory Motor Performance Program, Rehabilitation Institute
of Chicago, 345 East Superior St, Room 1406, Chicago, IL
60611; 312-238-1232; fax: 312-238-2208.
Email: j-patton@northwestern.edu
DOI: 10.1682/JRRD.2005.05.0088
644
JRRD, Volume 43, Number 5, 2006
the patient’s nonparetic limb [4]. Still another technique
is to provide force biofeedback during movement along a
constraining rail, which encourages the patient to push in
the appropriate direction [4–6]. Others have also
attempted sophisticated virtual reality techniques [7–15].
Our research and the work of others suggest that one
promising novel approach may be adaptive training [16–
20]. In this technique, we use the natural adaptive tenden
-
cies of the nervous system to facilitate motor recovery.
Motor adaptation studies have demonstrated that when peo
-
ple are repeatedly exposed to a force field that systemati-
cally disturbs arm motion, subjects learn to anticipate and
cancel out the forces and recover their original kinematic
patterns. After the disturbing force field is unexpectedly
removed, the subjects make erroneous movements in direc
-
tions opposite the perturbing forces (aftereffects). This
technique has recently been shown to alter and hasten the
learning process in nondisabled individuals [19,21].
*
Motor adaptation and its related aftereffects have been
demonstrated by many investigators under many condi
-
tions, ranging from simple position-, velocity-, and accel-
eration-dependent force fields [22–26] to Coriolis forces
caused by movement in a rotating room [27] and skew-
symmetric “curl” fields that produce forces in a direction
perpendicular to the velocity of the hand [25]. Similar
results have also been observed after manipulations of sub
-
jects’ visual perception that altered the visual feedback of
movement [28–31]. Recent results support the view that
subjects adapt by learning the appropriate internal model
of the perturbation rather than learning a temporal
sequence of muscle activations [25–26]. The most encour
-
aging result is that engineering techniques have been suc-
cessful in predicting both how the arm is disturbed by a
force field and the aftereffects of training [24,32–33]. Con
-
sequently, one possible rehabilitation method may be for
investigators to reverse-engineer the adaptation process by
using the models to design an appropriate force field that
will eventually result in the desired aftereffect.
Adaptive training will only work, however, if stroke
patients can adapt. Several studies have demonstrated
that at least a large subpopulation of stroke patients retain
their ability to adapt to a force field [16–20] or other dis
-
turbances [34–37]. However, severely affected individu-
als used atypical correction strategies [18], and the
amount of adaptation in individuals with more severe
impairment is somewhat diminished compared with non
-
disabled individuals [20]. Our recent work agrees with
these findings [38]. Furthermore, our preliminary studies
on stroke patients have revealed that aftereffects may
persist longer when the aftereffects resemble nondisabled
unperturbed movement [16–17,39].
We focused on adaptive training by using robot-applied
forces to restore function to hemiparetic stroke patients.
This investigation is a pilot study for determining the poten
-
tial of this approach to rehabilitation. We simply applied a
technique to stroke patients that has already proven effec
-
tive at causing desired results in nondisabled individuals
[19]. While straightening slightly curved movements is
generally not perceived as the most important clinical goal,
it represents our initial effort at testing the promise of this
approach in a well-known scientific framework. We
addressed two questions. First, we sought to determine
whether adaptation can be exploited for restoring move
-
ment ability. Second, we sought to determine whether the
benefit persists for the duration of the experiment [17].
METHODS
Subjects
Fifteen stroke patients without any other musculoskele-
tal injury volunteered to participate. Their demographic
details are listed in Table 1. The Northwestern University
Internal Review Board approved the research to conform to
ethical standards from the 1964 Declaration of Helsinki and
Federal mandates that protect research subjects. Before
beginning, each subject signed a consent form that con
-
formed to these Northwestern University guidelines. All
stroke participants were in the chronic stage, having suffered
a stroke 19 to 132 months prior to the experiment. Our
exclusion criteria were (1) bilateral impairment; (2) severe
sensory deficits in the limb; (3) aphasia, cognitive impair
-
ment, or affective dysfunction that would influence the abil-
ity to comprehend or to perform the experiment; (4) inability
to provide an informed consent; and (5) other current severe
medical problems. Subjects were randomly assigned to one
of two groups: a treatment group (n = 12) that received cus
-
tom-designed forces for part of the experiment or a control
group (n = 9) that received no forces but otherwise per
-
formed the same experimental protocol. We were able to
have six subjects return on a separate day for a second visit,
so some subjects served as their own controls. The order of
*
Wei Y, Patton JL. Forces that supplement visuomotor learning: A ‘sensory
crossover’ experiment. Exp Brain Res. Unpublished observations, 2006.
645
PATTON et al. Custom-designed haptic training for individuals with poststroke hemiparesis
presentation was randomized. The research therapy staff
was blinded to the type of forces the subjects received and
performed the modified Ashworth scale to assess spasticity
before beginning the experiment. We performed the upper-
limb portion of the Fugl-Meyer (FM) examination before
and after the robotic experiment to assess general changes in
motor capability.
Apparatus
Subjects held the free limb (here referred to as the “end-
point”) of a 2 degree-of-freedom DOF robot (Figure 1)
described elsewhere [26,40]. Endpoint forces and torques
were monitored with a 6 DOF load cell that was fixed to the
handle of the robot (ATI Industrial Automation, Inc, Apex,
North Carolina, model F/T Gamma 30/100). The robot was
equipped with position encoders that record the angular
position of the two robotic joints with a resolution exceeding
20'' of rotation (Gurley Precision Instruments, Inc, Troy,
New York, model 25/045-NB17-TA-PPA-QAR1S). The
position, velocity, and acceleration of the handle were
derived from these two position encoders. We used two
torque motors to apply programmed forces to the subjects’
hands (PMI Motor Technologies, Wood Dale, Illinois,
model JR24M4CH). Motion and force data were collected
at 100 Hz. At all times during the experiment, the software
generated an additional set of compensatory torques that
canceled the inertial effects of the robot-arm linkage and
resulted in the feeling of free movement on a slippery sur
-
face when the force field was not present.
Protocol
Subjects were seated so that the starting point of
the
targets was approximately at the center of their theo-
retical workspace, directly anterior from the shoulder
(Figure 1(b)). The experiment involved only the hemip
-
aretic limbs of the stroke subjects, which corresponded to
the dominant limb in 9 of the 11 subjects (Table 1). If
subjects had difficulty reaching the center point, we
adjusted their chair position slightly. To avoid fatigue,
subjects rested their elbow and forearm on a lightweight
frictionless linkage (Figure 1(a)), and they could choose
to rest between movements (subjects rarely rested longer
than a few seconds every hundred movements).
Starting from a point centered in front of the shoulder,
subjects were presented a target at one of two locations
10
cm distant (out and to the left or out and to the right).
Table 1.
Subject information for treatment and control subjects of this study (N = 15).
Subject Sex
Assisted
Vision
Dominant
Hand
Pathology
Affected
Hand
Affected Neural Region
1 Female Yes Right Hemorrhagic stroke Right Left intercerebral hemorrhage
2 Male Yes
*
Right Stroke Left Unknown
3 Male No Right TBI followed by 2 strokes Right Unknown, left CVA
4 Female Yes
*
Right Stroke Left Right frontal cortex
5 Male Yes
Right AVM hemorrhage Left Right frontal and parietal lobe
6 Female No Right Hemorrhagic stroke, RIND
1 yr previous
Left Right posterior internal capsule
infarct
7 Female No Right Hemorrhagic stroke Left Right intercerebral hemorrhage
8 Male Yes Right Stroke Left Right thalamus
9 Male Yes
*
Left Stroke Right Left parietal lobe
10 Male Yes Right Stroke Left Right carotid artery, thrombotic
11 Female Yes Left Stroke Right Pons and midbrain
12 Male Yes Right Stroke Left Unknown
13 Male Yes Right Stroke Left Left subcortical lacunar
14 Female No Right Thrombotic stroke Right Left MCA, posterior
15 Male Yes
*
Right Thrombotic stroke with
neurosurgery
Left Right subcortical
*
Vision not assisted for experiment.
Bifocals used for experiment.
AVM = arteriovenous malformation, CVA = cerebrovascular accident, MCA = middle cerebral artery, RIND = reversible ischemic neurological deficit, TBI = traumatic
brain injury.
646
JRRD, Volume 43, Number 5, 2006
These two “main” targets were 90° from each other and
formed a “v” pattern that was centered along the paramid
-
line extending from the shoulder (Figure 1(b)). Subjects
were given a cue to return to the center point after they had
either (1) initiated an attempt and 3 seconds had elapsed or
(2) reached and stayed in the target (1 cm radius) for at
least
0.5 seconds. Subjects’ arms were eclipsed by the projection
platform. Hence, they could only see the target and a cursor
that represented the instantaneous location of the hand.
Subjects were neither told about nor shown the desired
movement, although they were instructed to try to move to
the target at the appropriate speed in a straight line. Only
the outward movements were recorded for this experiment.
Additionally, we included two extra “generalization” tar
-
gets that were not practiced but only experienced briefly at
the beginning and end of the experiment. These targets
were only 30° from each other and formed a “v” in the area
between the main targets, also centered on the paramidline
extending from the shoulder. The generalization targets
were used to determine if any learning was carried to move
-
ment directions that were not practiced.
We controlled for a peak speed of 0.288 m/s by giving
subjects feedback at the end of each movement using col
-
ored dots and auditory tones. These cues let subjects know
if
they were going too fast, too slow, or within a range
of
±0.05 m/s. Consequently, subjects’ speeds remained
roughly constant across the entire experiment. Subjects were
instructed to initiate their movements at a self-determined
time after they saw the target appear. To prevent fatigue, we
instructed subjects to rest anytime they chose.
Machine Learning and Force-Field Design
As explained in greater detail elsewhere, a time-
record
of training forces was custom-designed with an
iterative machine-learning algorithm [19]. A machine-
learning phase iteratively determined the forces that shifted a
subject’s movement to the “desired” trajectory,
by
intermittently exerting forces (once every four movements,
randomly presented) and adjusting them based on the
response of the subject. For this experiment, we used a
smooth, minimum-jerk trajectory along a straight line to the
target for
which was believed to resemble a
“healthy” trajectory [41]. For each iteration, a force
was applied to the robot handle in the first 200 ms of the
movement, where
and (for i > 1) was
adjusted from one movement to the next based on the error
between the actual,
and the desired trajectories with
the simple machine-learning rule
Here, the parameter
μ
is a learning rate, which has
been heuristically found to work in the range from 10 to
Figure 1.
Experimental apparatus from (a) side view and (b) top view. (c) Han
d
strap and posture used. Ball handle was on low-friction axle, so it was
free to pivot in horizontal plane. (d) Subject seated at projection
p
latform with movement target displayed. Only target and cursor
representing hand were visible to subject.
x
D
t(),
x
D
t(),
F
D
i
t()
F
D
1
t() 0= F
D
i
t()
xt(),
F
D
i 1+
t() F
D
i
t()
μ
xt() x
D
t()().+=
647
PATTON et al. Custom-designed haptic training for individuals with poststroke hemiparesis
30 N
m
–1
. A
μ
that is too large leads to unstable learning,
and a
μ
that is too small results in a lengthy machine-
learning session. We chose
μ
= 20 N
m
–1
for our experi-
ments. The algorithm allowed the subject to initiate the
movement when they chose. Forces initiated when veloc
-
ity exceeded 0.1 m/s or the subject had exited the starting
window. All subjects performed a total of 744 movements
(trials), broken down into the following experimental
phases:
Unperturbed familiarization: 58 movements (~6.5 min)
for becoming familiar with the system and the task of
moving the manipulandum.
Unperturbed baseline: 10 movements (~1.5 min) for
establishing a baseline pattern of reaching movements.
Unperturbed baseline, generalization targets: 10 move-
ments (~1.5 min) for establishing a baseline pattern
of reaching movements on the generalization targets
described previously.
Machine learning: 200 movements (~25 min) with
forces exerted intermittently and randomly once in
every 4 movements. The robot gradually learned the
average forces necessary to push the subject to the
“desired” trajectory. Note that because these forces
occurred intermittently and randomly, these move-
ments did not lead to adaptation, because any small
amount of adaptation was washed out in the move-
ments between the movements with forces.
Second unperturbed baseline: 10 movements (~1.5 min)
for determining if the baseline pattern changed.
Second unperturbed baseline, generalization targets:
10 movements (~1.5 min) for determining if the base-
line pattern changed on the generalization targets.
Learning: 222 movements (~30 min) with constant
exposure to the training forces. These forces were the
vector opposite of the forces that were determined in
the machine-learning phase.
Aftereffects catch trials: 80 movements (~10 min)
with random, intermittent removal of the force field
for 1 in 8 of the movements (catch trials) for deter-
mining the aftereffects.
Aftereffects catch trials, generalization targets: 80 move-
ments (~10 min) with random, intermittent removal of
the force field for 1 in 8 of the movements (catch trials)
for determining the aftereffects.
Training refresher: 2 movements (~15 s) identical to the
learning phase.
Washout: 50 movements (~6 min) with no forces
applied.
Final movements, generalization targets: 10 move-
ments (~2 min) with no forces applied on the generali-
zation targets.
The movements in each direction were divided
equally in each phase. Subjects were also required to take
breaks (approximately 1–2 min) after movements 54,
278, 510, and 682 so they could rest and our data collec
-
tion equipment could be reset. We chose the instance of
these breaks to minimally disrupt the learning process
and provide the subject with a chance to rest. The sub
-
jects in the control group received no forces for the entire
experiment but otherwise experienced the same protocol.
The entire session lasted approximately 3 hours, which
included screening by therapist and pre- and postclinical
measures by the research occupational therapist.
Analysis
We restricted our focus in this study to the early part of
movements for two reasons. First, stroke patients often
make excessively large corrections later in their movements
that may depend on earlier errors [42–43]. Second, we were
primarily interested in the early phase of the movement that
best reflects the operation of a feed-forward controller
based on an internal model of the arm-environment dynam
-
ics. Our measure, the initial direction error, reflected this
early phase of movement by forming a vector from the start
point to 25 percent of the distance to the target (2.5 cm).
This measurement corresponded to approximately the first
200 to 300 ms of a movement that, if no error corrections
existed at the end of the movement, lasted about 1.1 s. Posi
-
tive error corresponded to a counter clockwise rotation from
the actual trajectory to the desired trajectory, and zero corre
-
sponded to a straight line to the target. Initial direction error
was used for testing our hypotheses on the feed-forward
controller and also was found to be highly correlated with
the perpendicular distance measure used in other adaptation
studies [26,44–45]. Even though we were studying the ini
-
tial part of movement, we were nonetheless also curious
about whether “fixing” earlier stages of movement will cor
-
rect (or improve) the latter part of movement.
All hypotheses were tested at an
α
level of 0.05. We
tested for a shift in initial direction from baseline to after
-
effects and for tendency of the aftereffects to disappear in
the washout phase.
648
JRRD, Volume 43, Number 5, 2006
RESULTS
For all the movements observed, unperturbed baseline
movements for stroke subjects were approximately twice
as variable as in nondisabled subjects for the same experi
-
mental conditions (Figure 2) [19]. All subjects showed
significant errors in one or more of the movement direc
-
tions in the unperturbed baseline phase before any treat-
ment (Figure 2(a)). Initial errors averaged 17° of initial
direction error. Intermittently in the machine-learning
phase, these errors were pushed by corrective forces that
evolved to eventually shift movements closer to a straight
line (Figure 2(b)). The forces resulting from the machine-
learning phase ranged from 0
to 12 N and differed from
subject to subject.
Subjects then began the learning phase in which the
forces learned by the robot were inverted and applied repeat
-
edly (Figure 2(c)). These training forces tended to amplify
(double) the initial errors we saw in the unperturbed baseline
phase. By the end of the learning phase, however, subjects
had reduced their errors (Figure 2(d)) to a level similar to
their unperturbed baseline phase. Beneficial aftereffects
were observed in some catch trials, where for a single
motion, the forces were removed and the subject was
returned to the “normal” world (Figure 2(e)). Finally, we
tested to see if the effects of the adaptation washed out by
leaving the forces off for 60 more movements. For most of
the subjects, the benefits were retained until the end, as was
the case for the subject shown in Figure 2(g).
However, in this initial investigation, we failed to
detect, from direct analysis of the data whether, as a
group, subjects statistically benefited from training by
having aftereffects lie closer to nondisabled movement.
Several subjects did not appear to adapt for one or both of
the movement directions (Figure 3). We suspected that
this was due to a combination of several factors. First, it
could have been because of a small (or no) effect size.
Second, it could have been because of a large amount of
variability in the data. Third, and most importantly, it
could have been because some subjects’ movements had
no room for improvement.
To test the hypothesis that benefit was difficult to
detect because there was nothing to improve upon, we sin
-
gled out subjects’ movement directions that actually
showed significant error before training (i.e., their baseline
movement errors were significantly different from zero).
This reduced the analysis to 13 movement directions for
the treatment group and 8 movement directions for the
control group (about half the data). By this criterion, data
from two subjects of the treatment group and four subjects
of the control group was not considered further because
their baseline movement errors were not significantly dif
-
ferent from zero (Tables 2 and 3).
Figure 2.
Typical patterns for successive phases of experiment for single stroke subject: (a) unperturbed baseline, (b) late machine learning, (c) early training,
(d) late training, (e) aftereffects, (f) early washout, and (g) late washout. Shown 315° movement direction for clarity only. Desired trajectories are
b
old dotted lines, average trajectories are bold solid lines, individual trajectories are thin lines, and shaded areas indicate running 95% confidence
intervals of ensemble. Note that although dotted tracing indicates an ideal trajectory, only target and cursor representing hand were visible to subject.
649
PATTON et al. Custom-designed haptic training for individuals with poststroke hemiparesis
Surprisingly, all but one of the treatment group move-
ments showed beneficial aftereffects (Figure 3, bars,
average reduction in error of –54%, p < 0.05, t-test on per
-
cent reduction in error). One subject’s error reduction was
300 percent because the training process caused a rotation
of the movement direction that went beyond the goal.
Nevertheless, if this outlying data point was not consid
-
ered, the error reduction for all the remaining subjects was
still significant (p < 0.01). Interestingly, this subject’s
overcorrection snapped back to almost zero error by the
end of the experiment.
The most critical question, however, was whether
these beneficial aftereffects could be retained. While we
chose this initial experiment not to extend our examina
-
tion beyond the single 3-hour visit to the lab oratory
(days or weeks would be necessary to be clinically con
-
vincing), we decided to check if beneficial aftereffects
were retained beyond the normal time that nondisabled
people wash out their aftereffects (about 20 movements).
Similar to the evaluation of aftereffects (Figure 3,
“Change to second baseline”), the final 10 movements of
the 50 movements of the washout phase were also evalu
-
ated for a reduction in error from baseline. We found that
10 of the 13 subjects’ error remained low and some were
even lower than in the aftereffects phase. As a group, this
reduction in error was marginally significant (p < 0.054,
t-test on percent reduction in error, Figure 3, diamond
symbols and bars for “Persistence”).
An additional evaluation was whether subjects could
generalize what they had learned and perform better in
directions that were not practiced. We inspected the
reduction in error movements that subjects made in the
unpracticed generalization-target trials. However, we
found no evidence of benefit either in the aftereffects or
by the end of the experiment (not displayed).
Two important intermediate questions were whether
(1) the machine-learning phase alone had any influence on
movement error and (2) mere practice without forces
might also result in some benefit (tested by the control
group). Both of these questions are related and relevant to
the assertion that a small amount of practice alone might
lead to benefits that could confound results from the
custom-designed force fields [46]. To test this, we first
evaluated the error reduction by the final 15 trials of
machine-learning phase and the corresponding trial num
-
bers for the control group (Figure 3, “Change to second
baseline”). Second, we evaluated the error reduction for
the control group for trials that corresponded to those eval
-
uated for the treatment group (Figure 3, circles). Although
we failed to detect any significant difference between the
control and the treatment groups, the treatment group
showed a significant benefit while the control group did
not (Figure 3, circles vs diamonds).
Figure 3.
Percent improvements from training only for movements in which
error existed before training. Reduction of errors are negative values
and were evaluated at second baseline phase (after end of machine
learning; left two columns of data), at aftereffects phase (after force-
field training; center two columns of data), or at end of experiment
(after forces had been off for 50 movements; right two columns o
f
data). Each subjects’ average improvement for each movement
direction is indicated by symbol: data from treatment group (
) an
d
control groups ({) are separated into columns. 95% and 90%
confidence intervals are indicated as thin and thick vertical shade
d
b
ars behind symbols, respectively, indicating where t-tests showe
d
significant benefit from training. Bars indicate significant benefit fo
r
treatment group in aftereffects trials and marginally significant
p
ersistence of benefit by experiment end. *Indicates significantly
different than zero at 0.05 level. (*)Indicates significantly different
than zero at 0.07 level.
650
JRRD, Volume 43, Number 5, 2006
Finally, the FM clinical scores also improved
slightly in the treatment group. For the treatment group,
FM scores marginally increased an average of 1.6
(Table 2, p = 0.06). No such improvement was seen in
the control group (Table 3, p > 0.27). This mild
improvement was also only loosely correlated to partic
-
ipants’ error reduction (i.e., a Pearson correlation coef-
ficient of 0.21).
Table 2.
Selected characteristics and experimental data for treatment group (n = 12).
Subject
Age
(yr)
Time Since
Stroke (mo)
Height
(m)
Mass
(kg)
Elbow
MAS
*
Fugl-Meyer Upper Extremity
Initial
Significant
Error
Movements
Final Mean Δ
Error
Pretreatment
Δ Pre- to
Posttreatment
1
§
30 129 1.60 70.45 26 0
2 48 55 1.73 75.45 37 2 R –114
3
§
37 132 1.75 71.82 52 4 R 15
4 76 19 1.73 75.00 44 5 R –143
5
56 94 1.80 106.82 2 43 6 R –58
6 49 54 1.47 60.00 2 32 –1 LR –124
7
51 73 1.60 47.73 2 51 –2 R –37
8 48 26 1.78 84.09 3 15 1
9
§
72 96 1.70 72.72 2 33 1 LR –53
10
53 113 1.85 90.89 2 22 2 L –63
11 40 26 1.65 68.18 4 LR –10
12 48 110 1.80 102.27 2 23 0 R 135
Mean 50.7 77.3 1.7 77.1 2 34 1.6 –45.2
*
Scale 0 to 4.
L indicates movement out and to left and R indicates movement out and to right.
Percent of baseline error.
§
Subjects served as their own control and were randomized into treatment group first.
Subjects served as their own control and were randomized into control group first.
MAS = modified Ashworth scale.
Table 3.
Selected characteristics and experimental data for control group (n = 9).
Subject
Age
(yr)
Time Since
Stroke (mo)
Height
(m)
Mass
(kg)
Elbow
MAS
*
Fugl-Meyer Upper Extremity
Initial
Significant
Error
Movements
Final Mean Δ
Error
Pretreatment
Δ Pre- to
Posttreatment
1
§
30 131 1.60 70.45 23 1
3
§
37 132 1.75 71.82 52 0 LR –46
5
56 93 1.80 106.82 2 47 0 L –45
7
51 72 1.60 47.73 2 46 2 R –24
9
§
72 97 1.70 72.72 2 30 1 R –28
10
53 111 1.85 90.89 2 23 1 R –5
13 55 11 1.83 79.09 2 43 0 LR 82
14 57 85 1.63 50.91 3 26 –2
15 46 167 1.80 81.82 25 1
Mean 50.8 99.9 1.7 74.7 2.1 35 0.4 –11.1
*
Scale 0 to 4.
L indicates movement out and to left and R indicates movement out and to right.
Percent of baseline error.
§
Subjects served as their own control and were randomized into treatment group first.
Subjects served as their own control and were randomized into control group first.
MAS = modified Ashworth scale.
651
PATTON et al. Custom-designed haptic training for individuals with poststroke hemiparesis
DISCUSSION
We conducted this initial pilot study to demonstrate
how adaptive training might be useful for restoring arm
movement. The stroke patients in this study showed less
conspicuous results compared with nondisabled subjects
exposed to the same algorithm [19]. Nevertheless, when
we restricted our analysis to movement directions that
were affected by significant error before training, results
were quite evident in the stroke group: movements
showed beneficial aftereffects after training (error
decreased) that persisted in all but three patients. This
persistence was twice as long as for nondisabled people.
While these results are only an encouraging hint at what
might be a possible therapeutic intervention, we believe
that these results suggest the need for a longer, more
comprehensive look at adaptive training as a means of
restoring function following brain injury.
A key assumption of our approach was that motion is
impaired because of an ineffective motor plan that can be
changed through adaptive training. However, one alterna
-
tive explanation of impairment is that passive contrac-
tures, commonly seen in chronic stroke patients, alter the
movement pattern. Indeed, this may explain the three sub
-
jects who deadapted to a pattern that was perhaps the most
biomechanically optimal for the characteristics of their
contracted limb. Another alternative explanation for the
three subjects is that some subjects may have less ability
to adapt. Indeed, other studies have reported diminished
ability to adapt in some stroke patients [18,20,38].
Another alternative explanation is that shifts in movement
patterns may have actually been present but were “buried
in the noise” of motor variability and therefore statistically
undetectable—a statistical power problem. Motor vari
-
ability is a commonly reported feature of stroke patients
[38,43,47–49]. Another possibility may be that subjects
could have been so highly functional that no improvement
was possible in the context of this experiment—the so
called “ceiling effect” in the learning process [50]. But
after all movements that did not exhibit significant error
before training were stripped away, the remaining move
-
ments showed signs of beneficial adaptation.
Whether this technique will lead to benefits that
might persist for days or weeks remains to be seen.
Another possibility is to consider prolonging this type of
training over many days to get ever closer to the desired
outcome. As rehabilitation training typically requires a
balance of repetitive practice, strengthening, and expert
guidance, we believe that the pilot results we present here
may inspire other new forms of rehabilitation.
One limitation of our approach is that the forces that
push the subject over to the desired trajectory x
D
(t) are
recorded in a Cartesian coordinate system. The assumption
is that these forces can be learned correctly even though
the hand is not moving along the desired positions with the
desired velocities. Because a nonlinear relationship exists
between joint torques and endpoint forces, the torques
applied along one trajectory are not necessarily appropriate
for another [51]. However, we argue that if that the desired
and expected trajectories are within a “domain of proper
generalization” [21], the forces applied are a good enough
approximation and can lead to desirable aftereffects.
Recent studies have presented evidence that motor learn
-
ing is broadly tuned so that training in one set of directions
can influence others [24,26,52–54]. Furthermore, this
learning process is quantifiable via an adaptation model
[45]. Still, more sophisticated machine-learning algorithms
would likely improve the performance by storing the
machine-learning forces in the intrinsic (joint or muscle)
coordinate system.
Most subject data (10 of 13) displayed persistent bene-
fits for the duration of the experiment (50 washout move-
ments)—beyond the time a nondisabled subject would
take to deadapt. Consequently, the data strongly suggests
that the benefits of the aftereffects are retained because
they are perceived to be an improvement. The motor con
-
trol system may respond to and retain the benefits of this
type of adaptive training for several reasons. One reason
may be that stroke patients have the confusing challenge of
being able to use only a few remaining motor pathways
after their injury. The brain attempts to send conventional
(preinjury) motor command signals, which are now inap
-
propriate because of injury, to the descending motor path-
ways. The training methods of this experiment may coax
the nervous system to attempt a new motor strategy that is
not intuitively obvious to the injured system but becomes a
“motor epiphany” following the removal of the training
forces. In this scenario, the nervous system is essentially
shown the right way to execute the task, much like a coach
may get an athlete to try a new strategy.
One might also speculate that change in reflex tone
leads to better movement. Reflex torque elicited by an
imposed stretch of the elbow has been shown to cause
decreases on average of 50 percent with tens of stretches
in a single experiment on flexor but not extensor muscles
[55]. However, if a spastic muscle pulls the limb to the
652
JRRD, Volume 43, Number 5, 2006
side and then a robot pushes to increase the error as is
done in this experiment, the spastic muscle would be
shortened. Therefore, this muscle would have less stimu
-
lation that might cause a spastic response.
Another reason adaptive training could lead to bene-
fits is that the impaired nervous system does not react to
nor does it try to learn from small errors in movement.
Our approach might promote learning by making errors
more noticeable. One can imagine many possible strate
-
gies for amplifying error, and recent research in our
group has shown promising results with several different
types [38,56]. Models of learning systems, such as neural
networks, suggest that error drives learning. As a conse
-
quence, these systems can learn better and faster if error
is larger [18,57–59]. Such error-driven learning processes
are believed to be central to the acquisition and adapta
-
tion of skill in human movement [59–60]. Augmenting
error may heighten motivation and attention or lead to
anxiety, which has been suggested to correlate with learn
-
ing [61]. Moreover, intensifying error can raise the sig-
nal-to-noise ratio for sensory feedback and self-
evaluation. Errors that are more noticeable may trigger
responses that would otherwise not be perceived. Other
studies agree with the hypothesis that error augmentation
can enhance learning and “trick” the nervous system into
certain behaviors by giving altered sensory feedback
[62–71]. Conversely, suppression of visual feedback can
slow down the deadaptive process [19]. However, not all
kinds of augmented feedback on practice conditions have
proven to be therapeutically beneficial in stroke [72].
Hence, limits may exist to the amount and type of error
augmentation that is useful [64,73].
Straightening movements may be considered a small
clinical goal compared with extending the functional
workspace or enhancing reach-and-grasp capability. This
initial effort merely tested the promise of this approach in
a well-known framework. More protracted studies lasting
weeks—currently underway in our laboratories—are
needed to provide more clinical significance to this
approach by demonstrating lasting and functional bene
-
fits of repeated treatments.
Encouraging evidence points to future studies that
exploit the natural adaptive tendencies in the nervous sys
-
tem for restoring function. Perturbation and electromyo-
graphic studies may challenge the hypothesis of reflex
modulation. Imaging and transcranial magnetic stimula
-
tion studies may determine if alternate motor pathways
are used or if the cortex is used differently after training.
Lesion site locations from magnetic resonance images
may indicate whether damage to certain functional areas
leads to limits in one’s ability to recover via adaptive
training.
While this research focuses on adaptation to forces
(kinetics), researchers have also observed similar adapta
-
tion to a more easily implemented visuomotor distortion
(kinematics). These distortions involve complex transfor
-
mations using prisms [74], nonlinear mappings [65], or
simple rotations or stretches [75–76]. All of these distor
-
tions appear to also induce an adaptation process and can
trigger rapid recovery from sensory disorders, such as
hemispatial neglect, seen in stroke patients [36], which
shortens the recovery process from months to hours.
Moreover, adaptation to both visual and mechanical dis
-
tortions appears to involve the same mechanism [77].
*
One sensory system can facilitate the other, and a combi-
nation is the most powerful. One might consider distor-
tions to induce an inappropriate and indirect form of
learning, but the addition of more sensory inputs, such as
the cutaneous sensors in the hand, the proprioceptive mus
-
cle spindles, and Golgi tendon organs may facilitate the
learning process by providing more signals. Combining
haptics (robotic forces) with sophisticated graphics (such
as virtual reality) may provide the most promising form of
rehabilitation for individuals with brain injury. However,
recent work suggests going beyond virtual reality to dis
-
torted reality in order to facilitate functional recovery.
This is currently of great interest to our group [14].
CONCLUSIONS
New opportunities for recovery after stroke are
offered by extending intensive therapy beyond present
inpatient rehabilitation stays, and robotic therapy may be
one way to economically accomplish this [78]. While
specialized forces are useful for inducing adaptive
responses, they are likely to be most effective if com
-
bined with other rehabilitation strategies. We believe that
the error-amplification approach presented here for indi
-
viduals with stroke provides a new pathway for augment-
ing motor relearning in individuals with brain injury.
*
Wei Y, Patton JL. Forces that supplement visuomotor learning: A ‘sensory
crossover’ experiment. Unpublished observations, 2006.
653
PATTON et al. Custom-designed haptic training for individuals with poststroke hemiparesis
ACKNOWLEDGMENTS
We thank Jamie Hitchens, Kathy Stubblefield, Elisa
Pelosin, and Yejun Wei for their insights later in the anal
-
ysis phase of this study.
Preliminary data for this study were presented in con-
ference proceedings [17]. For additional information on
this and other research in our laboratory, please see
http://
www.SMPP.northwestern.edu/RobotLab/.
This material was based on work supported by
the
American Heart Association (grant 0330411Z),
the
National Institutes of Health (grants R24 HD39627,
5
RO1 NS 35673, and F32HD08658), the National Science
Foundation (grant BES0238442), and the Dr. Ralph and
Marion Falk Medical Research Trust, Chicago, Illinois.
The authors have declared that no competing inter-
ests exist.
REFERENCES
1. Bobath B. Adult hemiplegia: Evaluation and treatment.
London (England): Heinemann; 1978. p. 1–160.
2. Voss DR, Ionta MK, Myers BJ. Proprioceptive neuromuscu-
lar facilitation. Philadelphia (PA): Harper and Row; 1985.
p. 1–370.
3. Volpe BT, Krebs HI, Hogan N, Edelsteinn L, Diels CM,
Aisen ML. Robot training enhanced motor outcome in
patients with stroke maintained over 3 years. Neurology.
1999;53(8):1874–76. [PMID: 10563646]
4. Lum PS, Burgar CG , Shor PC, Majmundar M, Van der
Loos M. Robot-assisted movement training compared with
conventional therapy techniques for the rehabilitation of
upper-limb motor function following stroke. Arch Phys
Med Rehabil. 2002;83(7):952–59. [PMID: 12098155]
5. Reinkensmeyer DJ, Kahn LE, Averbuch M, McKenna-
Cole A, Schmit BD, Rymer WZ. Understanding and treat-
ing arm movement impairment after chronic brain injury:
Progress with the ARM guide. J Rehabil Res Dev. 2000;
37(6):653–62. [PMID: 11321001]
6. Kahn LE, Zygman ML, Rymer WZ, Reinkensmeyer DJ.
Effect of robot-assisted and unassisted exercise on func-
tional reaching in chronic hemiparesis. In: Proceedings of
the 23rd Annual International IEEE Engineering in Medi-
cine and Biology Society Conference; 2001 Oct 25–28;
Istanbul, Turkey. New York: IEEE; 2001. p. 1344–47.
7. Jack D, Boian R, Merians AS, Tremaine M, Burdea GC,
Adamovich SV, Recce M, Poizner H. Virtual reality-
enhanced stroke rehabilitation. IEEE Trans Neural Syst
Rehabil Eng. 2001;9(13):308–18. [PMID: 11561668]
8. Schultheis MT, Rizzo AA. The application of virtual reality
technology for rehabilitation. Rehabil Psychol. 2001;46(3):
1–16.
9. Boian R, Sharma A, Han C, Merians A, Burdea G , Adamo-
vich S, Recce M, Tremaine M, Poizner H. Virtual reality-
based post-stroke hand rehabilitation. Stud Health Technol
Inform. 2002;85:64–70. [PMID: 15458061]
10. Rydmark M, Broeren J, Pascher R. Stroke rehabilitation at
home using virtual reality, haptics and telemedicine. Stud
Health Technol Inform. 2002;85:434–37. [PMID: 15458128]
11. Zhang L, Abreu BC, Seale GS, Masel B, Christiansen CH,
Ottenbacher KJ. A virtual reality environment for evaluation
of a daily living skill in brain injury rehabilitation: Reliability
and validity. Arch Phys Med Rehabil. 2003;84(8):1118–24.
[PMID: 12917848]
12. Broeren J, Rydmark M, Sunnerhagen KS. Virtual reality
and haptics as a training device for movement rehabilita-
tion after stroke: A single-case study. Arch Phys Med
Rehabil. 2004;85(8):1247–50. [PMID: 15295748]
13. Deutsch JE, Merians AS, Adamovich S, Poizner H, Burdea
GC. Development and application of virtual reality tech-
nology to improve hand use and gait of individuals post-
stroke. Restor Neurol Neurosci. 2004;22(3–5):371–86.
[PMID: 15502277]
14. Patton JL, Dawe G , Scharver C, Mussa-Ivaldi FA, Kenyon
R. Robotics and virtual reality: A perfect marriage for motor
control research and rehabilitation. In: Proceedings of the
26th Annual International IEEE Engineering in Medicine
and Biology Society Conference; 2004 Sep 1–5; San Fran-
cisco, CA. New York: IEEE; 2004. p. 4840–43.
15. Sveistrup H. Motor rehabilitation using virtual reality. J Neu-
roengineering Rehabil. 2004;1(1):1–10. [PMID: 15679945]
16. Patton JL, Rymer WZ, Mussa-Ivaldi FA. Robotic-induced
improvement of movements in hemiparetics via an
implicit learning technique. In: Society for Neuroscience;
2001 Nov 10–15; San Diego, CA.
17. Patton JL, Mussa-Ivaldi FA, Rymer WZ. Altering move-
ment patterns in healthy and brain-injured subjects via cus-
tom designed robotic forces. In: Proceedings of the 23rd
Annual International IEEE Engineering in Medicine and
Biology Society Conference; 2001 Oct 25–28; Istanbul,
Turkey. New York: IEEE; 2001. p. 1356–59.
18. Dancause N, Ptito A, Levin MF. Error correction strategies
for motor behavior after unilateral brain damage: Short-term
motor learning processes. Neuropsychologia. 2002;40(8):
1313–23. [PMID: 11931934]
19. Patton JL, Mussa-Ivaldi FA. Robot-assisted adaptive training:
Custom force fields for teaching movement patterns. IEEE
Trans Biomed Eng. 2004;51(4):636–46. [PMID: 15072218]
20. Takahashi CD, Reinkensmeyer DJ. Hemiparetic stroke
impairs anticipatory control of arm movement. Exp Brain Res.
2003;149(2):131–40. [PMID: 12610680]
654
JRRD, Volume 43, Number 5, 2006
21. Mussa-Ivaldi FA, Patton JL. Robots can teach people how
to move their arm. In: Proceedings of the IEEE International
Conference on Robotics and Automation; 2000 Apr 24–28;
San Francisco, CA. New York: IEEE; 2000. p. 300–305.
22. Bock O. Load compensation in human goal-directed arm
movements. Behav Brain Res. 1990;41(3):167–77.
[PMID: 2288670]
23. Flash T, Gurevitch F. Arm movement and stiffness adapta-
tion to external loads. In: Proceedings of the 13th Annual
International IEEE Engineering in Medicine and Biology
Society Conference; 1991 Oct 31–Nov 3; Orlando, FL.
New York: IEEE; 1991. p. 885–86.
24. Shadmehr R, Mussa-Ivaldi FA. Adaptive representation of
dynamics during learning of a motor task. J Neurosci.
1994;14(5 Pt 2):3208–24. [PMID: 8182467]
25. Gandolfo F, Mussa-Ivaldi FA, Bizzi E. Motor learning
by field approximation. Proc Natl Acad Sci U S A. 1996;
93(9): 3843–46. [PMID: 8632977]
26. Conditt MA, Gandolfo F, Mussa-Ivaldi FA. The motor sys-
tem does not learn the dynamics of the arm by rote memo-
rization of past experience. J Neurophysiol. 1997;78(1):
554–60. [PMID: 9242306]
27. Lackner JR, Dizio P. Rapid adaptation to Coriolis force per-
turbations of arm trajectories. J Neurophysiol. 1994;72(1):
299–313. [PMID: 7965013]
28. Held R, Freedman SJ. Plasticity in human sensorimotor
control. Science. 1963;142:455–61. [PMID: 14064442]
29. Miall RC, Weir DJ, Wolpert DM, Stein JF. Is the cerebel-
lum a Smith Predictor? J Mot Behav. 1993;25(3):203–16.
[PMID: 12581990]
30. Pine ZM, Krakauer JW, Gordon J, Ghez C. Learning of
scaling factors and reference axes for reaching movements.
Neuroreport. 1996;7(14):2357–61. [PMID: 8951852]
31. Krakauer JW, Ghilardi MF, Ghez C. Independent learning of
internal models for kinematic and dynamic control of reach-
ing. Nat Neurosci. 1999;2(11):1026–31. [PMID: 10526344]
32. Kawato M, Wolpert D. Internal models for motor control.
Novartis Found Symp. 1998;218:291–304; discussion 304–7.
[PMID: 9949827]
33. Bhushan N, Shadmehr R. Computational nature of human
adaptive control during learning of reaching movements in
force fields. Biol Cybern. 1999;81(1):39–60.
[PMID: 10434390]
34. Weiner MJ, Hallett M, Funkenstein HH. Adaptation to lat-
eral displacement of vision in patients with lesions of the
central nervous system. Neurology. 1983;33(6):766–72.
[PMID: 6682520]
35. Cohen LG , Ziemann U, Chen R, Classen J, Hallett M, Ger-
loff C, Butefisch C. Studies of neuroplasticity with transcra-
nial magnetic stimulation. J Clin Neurophysiol. 1998;15(4):
305–24. [PMID: 9736465]
36. Rossetti Y, Rode G , Pisella L, Farne A, Li L, Boisson D,
Perenin MT. Prism adaptation to a rightward optical devia-
tion rehabilitates left hemispatial neglect. Nature. 1998;
395(6698):166–69. [PMID: 9744273]
37. Nudo RJ, Friel KM. Cortical plasticity after stroke: Impli-
cations for rehabilitation. Rev Neurol (Paris). 1999;155(9):
713–17. [PMID: 10528355]
38. Patton JL, Stoykov ME, Kovic M, Mussa-Ivaldi FA. Evalu-
ation of robotic training forces that either enhance or
reduce error in chronic hemiparetic stroke survivors. Exp
Brain Res. 2006;168(3):368–83. [PMID: 16249912]
39. Raasch CC, Mussa-Ivaldi FA, Rymer WZ. Motor learning
in reaching movements by hemiparetic subjects [abstract].
Soc Neurosci Abstr. 1997;23:2374.
40. Scheidt RA, Reinkensmeyer DJ, Conditt MA, Rymer WZ,
Mussa-Ivaldi FA. Persistence of motor adaptation during con-
strained, multi-joint, arm movements. J Neurophysiol. 2000;
84(2):853–62. [PMID: 10938312]
41. Flash T, Hogan N. The coordination of arm movements:
An experimentally confirmed mathematical model. J Neu-
rosci. 1985;5(7):1688–1703. [PMID: 4020415]
42. Krebs HI, Aisen ML, Volpe BT, Hogan N. Quantization of
continuous arm movements in humans with brain injury.
Proc Natl Acad Sci U S A. 1999;96(8):4645–49.
[PMID: 10200316]
43. Beer RF, Dewald JP, Rymer WZ. Deficits in the coordina-
tion of multijoint arm movements in patients with hemi-
paresis: Evidence for disturbed control of limb dynamics.
Exp Brain Res. 2000;131(3):305–19. [PMID: 10789946]
44. Scheidt RA, Rymer WZ. Control strategies for the transition
from multijoint to single-joint arm movements studied
using a simple mechanical constraint. J Neurophysiol. 2000;
83(1):1–12. [PMID: 10634848]
45. Thoroughman KA, Shadmehr R. Learning of action through
adaptive combination of motor primitives. Nature. 2000;
407(6805):742–47. [PMID: 11048720]
46. Cirstea MC, Ptito A, Levin MF. Arm reaching improve-
ments with short-term practice depend on the severity of
the motor deficit in stroke. Exp Brain Res. 2003;152(4):
476–88. [PMID: 12928760]
47. Levin MF. Interjoint coordination during pointing movements
is disrupted in spastic hemiparesis. Brain. 1996;119(Pt 1):
281–93. [PMID: 8624689]
48. Fisher BE, Winstein CJ, Velicki MR. Deficits in compensatory
trajectory adjustments after unilateral sensorimotor stroke.
Exp Brain Res. 2000;132(3):328–44. [PMID: 10883381]
Erra-
tum in: Exp Brain Res. 2000;132(3):417.
49. Reinkensmeyer DJ, Iobbi MG , Kahn LE, Kamper DG , Taka-
hashi CD. Modeling reaching impairment after stroke using a
population vector model of movement control that incorpo-
rates neural firing-rate variability. Neural Comput. 2003;
15(11):2619–42. [PMID: 14577856]
655
PATTON et al. Custom-designed haptic training for individuals with poststroke hemiparesis
50. Schmidt RA. Motor control and learning. Champaign (IL):
Human Kinetics Publishers; 1988. p. 352–53.
51. Hollerbach MJ, Flash T. Dynamic interactions between
limb segments during planar arm movement. Biol Cybern.
1982;44(1):67–77. [PMID: 7093370]
52. Ghahramani Z, Wolpert DM, Jordan MI. Generalization to
local remappings of the visuomotor coordinate transforma-
tion. J Neurosci. 1996;16(21):7085–96. [PMID: 8824344]
53. Schaal S, Atkeson CG. Constructive incremental learning
from only local information. Neural Comput. 1998;10(8):
2047–84. [PMID: 9804671]
54. Mussa-Ivaldi FA. Modular features of motor control and
learning. Curr Opin Neurobiol. 1999;9(6):713–17.
[PMID: 10607638]
55. Schmit BD, Dewald JP, Rymer WZ. Stretch reflex adapta-
tion in elbow flexors during repeated passive movements in
unilateral brain-injured patients. Arch Phys Med Rehabil.
2000;81(3):269–78. [PMID: 10724069]
56. Wei Y, Bajaj P, Scheidt RA, Patton JL. Visual error aug-
mentation for enhancing motor learning and rehabilitative
relearning. In: Proceedings of the 9th International Confer-
ence on Rehabilitation Robotics; 2005 Jun 28–Jul 1; Chi-
cago, IL. New York: IEEE; 2005. p. 505–10.
57. Rumelhart DE, Hinton GE, Williams RJ. Learning repre-
sentations by back-propagating errors. Nature. 1986;323:
533–36.
58. Lisberger S. The neural basis for the learning of simple motor
skills. Science. 1988;242(4879):728–35. [PMID: 3055293]
59. Kawato M. Feedback-error-learning neural network for
supervised learning. In: Eckmiller R, editor. Advanced neu-
ral computers. Amsterdam (the Netherlands): Elsevier;
1990. p. 365–72.
60. Wolpert DM, Ghahramani Z, Jordan MI. An internal model
for sensorimotor integration. Science. 1995;269(5232):
1880–82. [PMID: 7569931]
61. Alleva E, Santucci D. Psychosocial vs. “physical” stress
situations in rodents and humans: Role of neurotrophins.
Physiol Behav. 2001;73(3):313–20. [PMID: 11438356]
62. Emken JL, Reinkensmeyer DJ. Robot-enhanced motor learn-
ing: Accelerating internal model formation during locomo-
tion by transient dynamic amplification. IEEE Trans Neural
Syst Rehabil Eng. 2005;13(1):33–39. [PMID: 15813404]
63. Brewer B, Fagan M, Klatzky R, Matsuoka Y. Perceptual
limits for a robotic rehabilitation environment using visual
feedback distortion. IEEE Trans Neural Syst Rehabil Eng.
2005;13(1):1–11. [PMID: 15813400]
64. Wei Y, Patton JL, Bajaj P, Scheidt RA. A real-time haptic/
graphic demonstration of how error augmentation can enhance
learning. In: Proceedings of the 2005 IEEE International Con-
ference on Robotics and Automation; 2005 Apr 18–22; Barce-
lona, Spain. New York: IEEE; 2005. p. 4406–11.
65. Flanagan JR, Rao AK. Trajectory adaptation to a nonlinear
visuomotor transformation: Evidence of motion planning
in visually perceived space. J Neurophysiol. 1995;74(5):
2174–78. [PMID: 8592205]
66. Srinivasan MA, LaMotte RH. Tactual discrimination of soft-
ness. J Neurophysiol. 1995;73(1):88–101. [PMID: 7714593]
67. Robles-De-La-Torre G , Hayward V. Force can overcome
object geometry in the perception of shape through active
touch. Nature. 2001;412(6845):445–48. [PMID: 11473320]
68. Ernst MO, Banks MS. Humans integrate visual and haptic
information in a statistically optimal fashion. Nature. 2002;
415(6870):429–33. [PMID: 11807554]
69. Sainburg RL, Lateiner JE, Latash ML, Bagesteiro LB. Effects
of altering initial position on movement direction and extent.
J Neurophysiol. 2003;89(1):401–15. [PMID: 12522189]
70. Brewer BR, Klatzky R, Matsuoka Y. Effects of visual feed-
back distortion for the elderly and the motor-impaired in a
robotic rehabilitation environment. In: Proceedings of the
2004 IEEE International Conference on Robotics and
Automation; 2004 Apr 26–May 1; New Orleans, LA. New
York: IEEE; 2004. p. 2080–85.
71. Kording KP, Wolpert DM. Bayesian integration in sen-
sorimotor learning. Nature. 2004;427(6971):244–47.
[PMID: 14724638]
72. Winstein CJ, Merians AS, Sullivan KJ. Motor learning
after unilateral brain damage. Neuropsychologia. 1999;
37(8):975–87. [PMID: 10426521]
73. Kording KP, Wolpert DM. The loss function of sensorimotor
learning. Proc Natl Acad Sci U S A. 2004;101(26):9839–42.
[PMID: 15210973]
74. Miles FA, Eighmy BB. Long-term adaptive changes in pri-
mate vestibuloocular reflex. I: Behavioral observations.
J Neurophysiol. 1980;43(5):1406–25. [PMID: 6768851]
75. Imamizu H, Miyauchi S, Tamada T, Sasaki Y, Takino R,
Putz B, Yoshioka T, Kawato M. Human cerebellar activity
reflecting an acquired internal model of a new tool. Nature.
2000;403(6766):192–95. [PMID: 10646603]
76. Krakauer JW, Pine ZM, Ghilardi MF, Ghez C. Learning of
visuomotor transformations for vectorial planning of reach-
ing trajectories. J Neurosci. 2000;20(23):8916–24.
[PMID: 11102502]
77. Tong C, Wolpert DM, Flanagan JR. Kinematics and dynam-
ics are not represented independently in motor working
memory: Evidence from an interference study. J Neurosci.
2000;22(3):1108–13. [PMID: 11826139]
78. Fasoli SE, Krebs HI, Ferraro M, Hogan N, Volpe BT. Does
shorter rehabilitation limit potential recovery poststroke?
Neurorehabil Neural Repair. 2004;18(2):88–94.
[PMID: 15228804]
Submitted for publication May 23, 2005. Accepted in
revised form November 17, 2005.
... One of the initial discussions on bi-lateral training by Burgar et al. [62] showed evidence that the corticospinal ipsilateral pathways, which are involved in recovery from hemiplegia are also found to be active in bilateral movements, thus, potentially beneficial for motor recovery of upper limbs. Few works [63][64][65] discuss bi-manual rehabilitation as a form of physical coupling where the unimpaired limb assists humans in rehabilitation of the impaired limb. This is because both arms receive the same neural signal from the brain and they tend to move together in symmetry. ...
... Generally, force-feedback enabled robotic prosthetic have been effective tools in physical rehabilitation training for stroke survivors [78][79][80][81][82][83][84][85]. Few works [63,86] discuss the portability aspect of rehabilitation training in order to use robotic systems away from medical centers after the initial treatment phase. These devices can be economical as well as a relatively quicker way for patient-recovery as more time can be spent towards post-recovery treatments such as therapy. ...
Article
Full-text available
Bi-manual (two-handed) actions have shown notable success in rehabilitative and therapeutic applications from the point of motor symmetry. Recent studies have shown that symmetry in actions is attributed to sensorimotor perception than mere co-activation of homologous muscles. In this paper, we present a study of symmetric and asymmetric haptic (specifically force) feedback on human perception and motor action during bi-manual spatial tasks. To the best of our knowledge, ours is the first procedure to specifically test the perceptual aspect of bi-manual actions in contrast to other works that typically characterize the physical/bio-mechanical aspects. Thereby in our experiment, healthy individuals were tasked with stretching a virtual spring using two symmetrically located haptics devices that provide an equal amount of resistive force on each hand while pulling the spring. In this experiment, we implement four kinesthetic conditions, namely (1) feedback on both hands, (2) feedback only on dominant hand, (3) feedback only on non-dominant hand, and (4) no feedback as our control. Our first goal was to determine if there exists a range of spring stiffness in which the individual incorrectly perceives bi-manual forces when the feedback is deactivated on one hand. Subsequently, we also wanted to investigate what range of spring stiffness would lead to such perceptual illusions. Our analysis shows that not only does such a range exist, it is wide enough so as to be potentially utilized in future rehabilitative applications.
... Although the user performance improves while haptic guidance is provided, the user performance worsens after the haptic guidance is withdrawn [3], [7]. In contrast, gross resistance is a method of increasing the task difficulty over the training session by providing haptic feedback that interferes with the movement [8]. Through this approach, the trainee is encouraged to reduce the existence of artificial resistance and comes to understand how to perform the same task better without resistance after practice. ...
... Error augmentation, which is an example of gross resistance, allows healthy adults to adapt with the viscous force field by amplifying task dynamics [9]. Several studies found that the use of haptic feedback in motor learning is effective in terms of acquiring new motor skills [7], [8], [10], [11]. However, the increase in feedback due to gross assistance can cause the learner to rely on the haptic feedback with less motivation and lead to worse performance in a retention test showing long-term improvement [3], [7], whereas larger errors in gross resistance tend to increase the motivation of the learner to learn by trying to reduce the error. ...
Article
To learn motor skills using a finger in many haptic training systems, a user places his or her finger in a holder to communicate with a haptic interface. It is known that the force-detection capability at the fingertip is reduced when the user's finger is enclosed in a holder. The learning performance might therefore be impaired in training. Stochastic resonance is known to improve sensitivity at the fingertip. We thus propose a fingertip force learning method using stochastic resonance. We first examine the effect of stochastic resonance on the user haptic performance in a two-dimensional task when the fingertip is within the finger holder. The outcomes indicate that the user fingertip sensitivity increases even when the finger is placed within a holder. We next perform force learning tasks in one- and two-dimensional space. Our proposed method that combines haptic feedback and stochastic resonance is compared with the method in which only haptic feedback is provided to the user in force leaning tasks. Results obtained using the proposed method indicate that the performance outcome of the proposed method is higher than that of the comparison method. This study demonstrates the potential of the proposed method for motor learning tasks.
... One approach to affect Page 2 of 9 Abdel Majeed et al. J NeuroEngineering Rehabil (2020) 17:156 movement speed is to directly increase it with a negative viscous field; previous work [13][14][15][16] showed that training with negative viscosity can improve participant movement and movement generalization abilities. Another possibility is to leverage the motor control mechanisms of error augmentation and after effects. ...
... Since we were conditioning the arm, repeated exposure to a stimulus that causes spastic responses may have a therapeutic effect. Training with velocity-enhancing fields showed that negative viscosity can improve participant movement and movement generalization abilities [13][14][15][16]. Our work here also did not definitively address the responsiveness to prolonged speed training, where spastic post-stroke patients may respond differently to such therapy. ...
Article
Full-text available
Background Our previous work showed that speed is linked to the ability to recover in chronic stroke survivors. Participants moving faster on the first day of a 3-week study had greater improvements on the Wolf Motor Function Test. Methods We examined the effects of three candidate speed-modifying fields in a crossover design: negative viscosity, positive viscosity, and a “breakthrough” force that vanishes after speed exceeds an individualized threshold. Results Negative viscosity resulted in a significant speed increase when it was on. No lasting after effects on movement speed were observed from any of these treatments, however, training with negative viscosity led to significant improvements in movement accuracy and smoothness. Conclusions Our results suggest that negative viscosity could be used as a treatment to augment the training process while still allowing participants to make their own volitional motions in practice. Trial registration This study was approved by the Institutional Review Boards at Northwestern University (STU00206579) and the University of Illinois at Chicago (2018-1251).
... There are multiple possible training conditions that may achieve this increase, and here we compare three candidate classes of conditions. One approach to affect movement speed is to directly increase it with a negative viscous field; previous work [13,14,15,16] showed that training with negative viscosity can improve participant movement and movement generalization abilities. Another possibility is to leverage the motor control mechanisms of error augmentation and after effects. ...
... Since we were conditioning the arm, repeated exposure to a stimulus that causes spastic responses may have a therapeutic effect. Training with velocity-enhancing fields showed that negative viscosity can improve participant movement and movement generalization abilities [13,14,15,16]. Our work here also did not definitively address the responsiveness to prolonged speed training, where spastic post-stroke patients may respond differently to such therapy. ...
Preprint
Full-text available
Background: Our previous work showed that speed is linked to the ability to recover in chronic stroke survivors. Participants moving faster on the first day of a three-week study had greater improvements on the Wolf Motor Function Test. Methods: We examined the effects of three candidate speed-modifying fields in a crossover design: negative viscosity, positive viscosity, and a "breakthrough" force that vanishes after speed exceeds an individualized threshold. Results: Negative viscosity resulted in a significant speed increase when it was on. No lasting after effects on movement speed were observed from any of these treatments, however, training with negative viscosity led to significant improvements in movement accuracy and smoothness. Conclusions: Our results suggest that negative viscosity could be used as a treatment to augment the training process while still allowing participants to make their own volitional motions in practice. Trial registration: This study was approved by the Institutional Review Boards at Northwestern University (STU00206579) and the University of Illinois at Chicago (2018-1251).
... Rehabilitation after a stroke using interactive robotic technologies provides automatic, repetitive training, with high volume and accuracy that can be harnessed to objectively measure patients' motor performance. 1 Robotic devices can also be used in rehabilitation to deliberately interfere with patients' movements by providing mechanical perturbation or by distorting visual feedback during movement execution. 2,3 While both of these methods involve a large amount of practice, their physiological rationales differ considerably. ...
Article
Full-text available
Stroke is a major cause of long-term functional disability and requires physical rehabilitation. Due to population aging, the number of people post stroke is going to rise. Robotically neurorehabilitation has a great potential to improve outcome measurements. Error Augmentation training using a robotic interface is thought to promote motor recovery by enhancing proprioceptive feedback, which motivates and challenges patients to optimize their performance during training. Here, we investigated the effectiveness of robotic Error Augmentation training on motor recovery after a stroke, compared to standard robotic training in a null field. Post-stroke patients were randomly assigned to one of two groups: a study group (n = 9) that was trained on a 3D robotic system applying Error Augmentation forces, and a control group (n = 7) that carried out the same protocol in null field conditions. The robotic rehabilitation intervention was applied in addition to the standard rehabilitation protocol of the rehabilitation center. Error Augmentation training increased clinical scores compared to standard robotic training by 266% on the Motor Assessment Scale, and 88% on the Fugl-Meyer scale. The Motor Assessment Scale scores were significantly correlated with the Fugl-Meyer scores (p = 0.03, r = 0.541). There were more movement errors on the initial trials of the game sequence using the DeXtreme robotic device with Error Augmentation compared to trials with no force field. This difference vanished however after 10 trials. Error Augmentation training decreased the number of movement units and jerkiness compared to the control treatment. There was a robust effect of magnifying the acceleration component of movement using EA on the smoothness of the movement. These findings suggest that EA training may enhance motor performance possibly through motor adaptation. Future study should include EMG to better elucidate the neural mechanisms involved in motor learning post CNS injury.
... Este modelo propone que la derivada del tiempo y la aceleración para comprobar un tirón. El modelo de mínimos de Jerk es la tercera derivada de la posición, la cual comprobara la teoría, y el esquema de trayectoria más utilizado en entornos de terapia robótica (Patton, Kovic & Mussa-Ivaldi, 2006;Loureiro et al., 2003). La teoría ha sido planteada por Hogan y Flash en 1984 (Flash & Hogan, 1985). ...
Article
Full-text available
Se presentará un estado del arte de los exoesqueletos más relevantes a nivel mundial, y a partir de ello, se realiza-rá un análisis del movimiento de la extremidad superior, enfocado en los movimientos del antebrazo y la muñeca, teniendo en cuenta los puntos y ángulos con sus grados de libertad para evaluar en la extensión y la flexión. Se dan a conocer los tipos de movimientos para los músculos distales y proximales, así como la función del ejercicio más adecuada para la recuperación de acuerdo con la señal de electromiografía (EMG) se define la mejor terapia, pasiva o activa, mediante un modelado cinético del movimiento en el tipo de terapia, a partir del modelo matemático JERK a través de las funciones de MATLAB. Abstract A state of the art of the most relevant exoskeletons worldwide, will be presented, and from this, an analysis of the movement of the upper extremity will be performed, focusing on the movements of the forearm and the wrist, taking into account the points and angles with their Degrees of freedom to evaluate in extension and flexion. The types of movements for distal and proximal muscles are known, as well as the most appropriate function of exercise for recovery, according to the electromyography signal (EMG); there is defined the best therapy, passive or active, by means of a kinetic modeling of the movement in the type of therapy, from the mathematical model JERK through the MATLAB functions.
Chapter
Brain injury often results in a partial loss of the neural resources communicating to the periphery that controls movements. Consequently, the signals that were employed prior to injury may no longer be appropriate for controlling the muscles for the intended movement. Hence, a new pattern of signals may need to be learned that appropriately uses the residual resources. The learning required in these circumstances might in fact share features with sports, music performance, surgery, teleoperation, piloting, and child development. Our lab has leveraged key findings in neural adaptation as well as established principles in engineering control theory to develop and test new interactive environments that enhance learning (or re-learning). Successful application comes from the use of robotics and video feedback technology to augment error signals. These applications test standing hypotheses about error-mediated neuroplasticity and illustrate an exciting prospect for rehabilitation environments of tomorrow. This chapter highlights our works, identifies our acquired knowledge, and outlines some of the successful pathways for restoring function to brain-injured individuals.
Article
Full-text available
The ability to accurately complete goal-directed actions, such as reaching for a glass of water, requires coordination between sensory, cognitive and motor systems. When these systems are impaired, like in people with multiple sclerosis (PwMS), deficits in movement arise. To date, the characterization of upper limb performance in PwMS has typically been limited to results attained from self-reported questionnaires or clinical tools. Our aim was to characterize visually guided reaching performance in PwMS. Thirty-six participants (12 PwMS who reported upper limb impairment (MS-R), 12 PwMS who reported not experiencing upper limb impairment (MS-NR), and 12 age- and sex-matched control participants without MS (CTL)) reached to 8 targets in a virtual environment while seeing a visual representation of their hand in the form of a cursor on the screen. Reaches were completed with both the dominant and non-dominant hands. All participants were able to complete the visually guided reaching task, such that their hand landed on the target. However, PwMS showed noticeably more atypical reaching profiles when compared to control participants. In accordance with these observations, analyses of reaching performance revealed that the MS-R group was more variable with respect to the time it took to initiate and complete their movements compared to the CTL group. While performance of the MS-NR group did not differ significantly from either the CTL or MS-R groups, individuals in the MS-NR group were less consistent in their performance compared to the CTL group. Together these findings suggest that PwMS with and without self-reported upper limb impairment have deficits in the planning and/or control of their movements. We further argue that deficits observed during movement in PwMS who report upper limb impairment may arise due to participants compensating for impaired movement planning processes.
Conference Paper
Stroke rehabilitation is often terminated once a plateau in motor recovery is observed, but new training modalities have demonstrated that further functional improvement is possible after the onset of the chronic phase. In particular, feedback technologies augmenting error proved to foster the relearning process. Here we explore the possibility of a robot-free implementation of Error-Augmentation (EA), where only visual feedback is distorted. We present the interim results from our ongoing blinded, randomized, controlled clinical trial testing the efficacy of parallel bimanual reaching with visual EA. Subjects trained in the virtual environment in 45-minute sessions, three times a week, for three weeks, half with and half without EA. A blinded therapist performed clinical evaluations before, 1 week after, and two months after training. Available results showed that both groups significantly improved. An advantage in the treatment group could be tracked at all time points, but no statistical significance was detectable between groups. Gains in the two groups were found to be compatible with the results of previous studies using robots and may prove to have similar effectiveness without the need for a costly and complicated robotic device. One new finding was that EA caused significantly higher inter-trial variability.
Article
Full-text available
Virtual reality (VR) is an emerging technology with a variety of potential benefits for many aspects of rehabilitation assessment, treatment, and research. Through its capacity to allow the creation and control of dynamic 3-dimensional, ecologically valid stimulus environments within which behavioral responding can be recorded and measured, VR offers clinical assessment and rehabilitation options that are not available with traditional methods. Initial applications of VR in other aspects of medicine and psychology have yielded encouraging results, but continued research and understanding of this evolving technology will be crucial for its effective integration into rehabilitation. This article provides a brief introduction to VR technology, examines the specific benefits VR offers consumers and providers of rehabilitation services and discusses potential areas of application and important considerations in applying this technology. Finally, 2 examples of current Vr applications are presented. (PsycINFO Database Record (c) 2012 APA, all rights reserved)
Conference Paper
Full-text available
In order to design a robotic rehabilitation environment using visual feedback distortion, we investigated in this study the limits and effects of visual feedback distortion for the elderly and the motor-impaired. To determine the minimum imperceptible amount of visual distortion, we measured the Just Noticeable Differences (JNDs) for force and position for elderly, unimpaired subjects; values of 31.0% (0.619 N) and 16.1% (5.01 mm), respectively, were obtained. JNDs of 46.0% (0.920 N) and 45.0% (14.8 mm) were measured for a motor-impaired individual. These JNDs were larger than corresponding measurements previously taken with young subjects, showing a decrease in discrimination ability with age and impairment. Visual distortion based on these values caused elderly subjects and the motor-impaired individual to increase their force production levels by 72.5% and 97.7%, respectively. These results were similar to those obtained with young subjects, but differences were observed on interspersed trials with no visual feedback. Poor discrimination abilities in elderly and impaired subjects and visual dominance in our environment for this subject group support our hypothesis that visual distortion can be an effective tool for rehabilitation in a robotic environment.
Conference Paper
Full-text available
First Page of the Article
Conference Paper
While adaptive processes in the cerebral cortex have long been thought to contribute to functional recovery after stroke, the precise neuronal structures and mechanisms underlying these processes have been difficult to identify. Over the past 15 years, a large number of studies conducted in human stroke patients and in experimental animal models have contributed to a more coherent picture of the brain's adaptive capacity after injury. These studies suggest that the cerebral cortex undergoes significant functional and structural plasticity for at least several weeks to months following injury. Adaptive changes have been demonstrated in the intact tissue surrounding the lesion, as well as in other cortical motor areas remote from the site of injury. Recent results from non-human primate studies of cortical reorganization after stroke demonstrate marked functional changes in the intact cortical tissue adjacent to the infarct in the weeks following an ischemic lesion. Further, intensive task-specific practice with the impaired limb has a modulatory effect on the inevitable cortical plasticity. Taken together with parallel studies of forced use in human stroke patients, it is likely that use of the impaired limb can influence adaptive reorganizational mechanisms in the intact cerebral cortex, and thus, promote functional recovery.
Conference Paper
Objective: To evaluate the effects of repeated, externally imposed, flexion-extension movements of the elbow on the resulting stretch reflex response in hemiparetic spastic brain-injured patients. These effects were compared within a recording session and across sessions for the same subject to determine the impact of movement history on the quantification of spastic hypertonia using the stretch reflex response. Design: Twenty to 30 sequential, constant velocity flexion-extension movements were applied to the impaired elbow of our cohort, with a 10-second hold interposed between flexion and extension. Movements were applied regularly at 1-minute intervals. Changes in stretch reflex responses were monitored during the applied movements. Participants: We examined a convenience sample of seven hemiparetic brain-injured subjects between the ages of 26 and 60 yrs, with moderate-to-severe spastic hypertonia of elbow muscles (Ashworth score 2-4/4). Subjects participated in 2 to 9 sessions. Measures: Elbow torque, position, velocity, and electromyograms of the biceps, brachioradialis, and triceps muscles were recorded for each flexion and extension movement. Stretch reflex torque was calculated by subtracting-passive torque from total elbow torque, recorded over large amplitude movements. A linear regression analysis quantified both the initial torque response of the stretch reflex and the ensuing adaptation of the stretch reflex during sequential movements. Intersession variability was characterized both for spastic hypertonia measures and for stretch reflex adaptation. Results: Repeated, externally imposed, sequential flexion-extension movements of the elbow decreased the elbow flexor stretch reflex in six of seven subjects. The mean reduction in reflex torque after 30 movements was 50% of the initial torque values (p = .001, t test vs 0% change). Intersession stretch reflex responses for each subject were found to vary greatly (SDs of reflex torque ranged from 0.1 to 4.0Nm), and there were also significant variations in the degree of adaptation between subjects. Conclusions: Stretch reflex adaptation must be taken into consideration when spastic hypertonia is quantified using repeated joint motion, as is often the case. The magnitude of intersession variation in spastic hypertonia measures suggests that ideally, such measurements should be made across multiple sessions before conclusions are made regarding the efficacy of spastic hypertonia interventions. This study provides quantitative evidence that repeated joint movements may have a significant short-term beneficial effect on spastic hypertonia.
Article
The directional control of reaching after stroke was simulated by including cell death and firing-rate noise in a population vector model of movement control. In this model, cortical activity was assumed to cause the hand to move in the direction of a population vector, defined by a summation of responses from neurons with cosine directional tuning. Two types of directional error were analyzed: the between-target variability, defined as the standard deviation of the directional error across a wide range of target directions, and the within-target variability, defined as the standard deviation of the directional error for many reaches to a single target. Both between and within-target variability increased with increasing cell death. The increase in between-target variability arose because cell death caused a nonuniform distribution of preferred directions. The increase in within-target variability arose because the magnitude of the population vector decreased more quickly than its standard deviation for increasing cell death, provided appropriate levels of firing-rate noise were present. Comparisons to reaching data from 29 stroke subjects revealed similar increases in between and within-target variability as clinical impairment severity increased. Relationships between simulated cell death and impairment severity were derived using the between and within-target variability results. For both relationships, impairment severity increased similarly with decreasing percentage of surviving cells, consistent with results from previous imaging studies. These results demonstrate that a population vector model of movement control that incorporates cosine tuning, linear summation of unitary responses, firing-rate noise, and random cell death can account for some features of impaired arm movement after stroke.
Article
We describe a new learning procedure, back-propagation, for networks of neurone-like units. The procedure repeatedly adjusts the weights of the connections in the network so as to minimize a measure of the difference between the actual output vector of the net and the desired output vector. As a result of the weight adjustments, internal 'hidden' units which are not part of the input or output come to represent important features of the task domain, and the regularities in the task are captured by the interactions of these units. The ability to create useful new features distinguishes back-propagation from earlier, simpler methods such as the perceptron-convergence procedure.