LEARNING FROM BIOFEEDBACK

Patient-specific Games for Neuromuscular Rehabilitation

Ouriel Barzilay and Alon Wolf

Biorobotics and biomechanics Laboratory, Technion - Israel Institute of Technology, Haifa, Israel

Keywords: Artificial neural networks, Patient-specific rehabilitation, Virtual reality, Biofeedback.

Abstract: Rehabilitation tasks are generally subjected to the physiotherapist’s qualitative interpretation of the patient’s

pathology and needs. Motivated by the recently increasing use of virtual reality in rehabilitation, we propose

a novel approach for the design of those biomechanical tasks for an improved patient-specific and

entertaining rehabilitation. During training, the subject wears 3D goggles in which virtual tasks are

displayed to him. His kinematics and muscles activation are tracked in real time and an inverse model is

estimated by artificial neural networks. The resulting inverse model produces a physical exercise according

to the observed abilities of the subject and to the expected performance dictated by the physiotherapist. The

system offers several advantages to both the patient and the physiotherapist: the tasks can be presented in

the form of interactive personalized 3D games with augmented feedback, stimulating the patient’s

motivation and reducing the need of constant monitoring from the therapist. Additionally, offline

quantitative data from every training session can be stored for further analysis. The results of our study on

arm movements suggest an improvement in the training efficiency by 10% for the biceps and by 32%

(p=0.02) for the triceps.

1 INTRODUCTION

Physiotherapy aims at helping patients recover

maximal movement and functionality after surgical

operations, injuries or strokes. Neuromuscular

rehabilitation is generally performed in the form of

biomechanical tasks, designed to restore a cognitive

or mechanical function within the patient. These

tasks are elaborated by the physiotherapist, based on

his diagnosis of the patient’s pathology. Success of

the rehabilitation training relies on the adequate

design of these tasks, on the repetition of the

physical exercises by the patient, on the subject’s

motivation and on the feedback to the patient

(Holden, 2005). Furthermore, for good results, the

task must be adapted to the actual performance of

the patient. This adaptive physiotherapy is very

difficult to perform online and is subjected to the

trainer’s interpretation of the patient's performance.

Today, rehabilitation in virtual reality has

become a large field of research and several studies

have been published on the efficiency and the

advantages offered by this approach. Virtual

rehabilitation consists in the execution of

biomechanical tasks in virtual environments,

generally by the means of display devices,

biofeedback or haptic instrumentation and adapted

software. Virtual rehabilitation has proved efficient

in the treatment of neurological diseases (e.g. in Jack

et al., 2001 or Holden et al., 2005) for patients with

balance disorders (Jacobson et al., 2001) or sports

medicine. Studies have shown scientific evidence

that motor skills can be learned in virtual

environments (Regian et al., 1992) and transferred to

the real world (Holden and Dyar, 2002).

Furthermore, the augmented feedback on

performance offered by virtual reality improves the

results of rehabilitation (Shea and Wulf, 1999). It is

also likely to increase the motivation of the patient

during training (Maclean et al., 2000, Rizzo and

Kim, 2005). Some researchers even claim that motor

learning in virtual environments can surpass training

in the real world (e.g. Todorov et al., 1997).

Training in virtual environments also permits to

enhance the rehabilitation platform with

computational models and learning systems. In this

paper, we introduce a virtual adaptive biofeedback

rehabilitation approach aimed at improving

neuromuscular training using artificial neural

networks able to learn from biofeedback and to

168

Barzilay O. and Wolf A..

LEARNING FROM BIOFEEDBACK - Patient-speciﬁc Games for Neuromuscular Rehabilitation.

DOI: 10.5220/0003679801680174

In Proceedings of the International Conference on Neural Computation Theory and Applications (NCTA-2011), pages 168-174

ISBN: 978-989-8425-84-3

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

produce online new patient-specific virtual

physiotherapy missions. With the help of a motion

capture system and electromyograms (EMG), our

rehabilitation system tracks at any time the

kinematics of a subject and his muscle activation.

The subject is exposed, via a head mounted display

unit, to virtual tasks, which he is asked to perform.

After a calibration phase, a neural network is trained

to respond to the subject's biofeedback information,

based on the desired muscle activation and motions

prescribed by the physiotherapist. Once training is

complete, the network calculates a new trajectory as

biomechanical exercise, subjected to the previous

performance of the subject. This adaptive loop is

repeated continuously, resulting in an online

biofeedback-based adaptive virtual rehabilitation

system.

We exploit in this study the ability of artificial

neural networks to accurately model systems on

which little information is available or complex

systems where a computational model is preferred

over an explicit model based on arbitrary

assumptions and constraints. Biological systems

undoubtedly are the most difficult systems to model,

as they often involve several functional elements.

For instance, virtual upper-limb rehabilitation not

only involves motor learning and control of the

subject, but also his interpretation of the virtual

environment or his hand-eye coordination faculty.

Our research hypothesis is that an artificial neural

network can be utilized to model biological systems

by defining only the input and output signals to the

system. This contrasts with works aiming at building

sophisticated internal models of human motor

control (e.g. Kawato 1990, 1999). We validate our

hypothesis on a simple case of neuromuscular

rehabilitation. To the best of our knowledge, the

approach presented in this paper is novel and has no

antecedent in the literature. The results of this study

can encourage others in this field to further explore

the ability of learning systems to model human

functions by the means of biofeedback

instrumentation.

2 METHODS

2.1 Experimental Setup

Using the Vicon™ motion capture system

capabilities, we track the subject's motions in real

time and continuously gather kinematic data.

Markers are placed on the subject's body or on part

of it (we first focused our study on the arms). During

training, the subject is immersed in a virtual

environment in which he is shown floating targets

(Figure 1). The subject is then asked, for instance, to

follow the motions of a target with his pointing

finger (virtual ball application). His motions are

continuously recorded by the system while

following the virtual missions presented to him.

Moreover in order to record the subject muscular

activation, we place electromyograms sensors on

key muscles associated related to the motion.

Figure 1: Experimental Setup.

The user receives, in real time, augmented

feedback on his performance during training. For

instance in the virtual ball application, in which the

subject must continuously get his hand as near as

possible to the ball in motion, the feedback on the

distance to the target is provided in several manners.

On the virtual replica of the subject's hand is drawn

a ball, having the same radius as the virtual ball to

reach. The color of this ball changes with respect to

the distance to the target, according to a pre-defined

color code. The subject can also read his score rising

proportionally to the distance to the target. An

additional indication on the distance to the virtual

ball in the horizontal plane is provided by the

shadows of both balls, projected vertically onto the

virtual ground. Those indicators allow the user to

correct his gestures while performing the exercise.

For an improved accuracy yet, when the subject is

relatively close to the target, an additional feedback

is given to the subject, as the volume of a musical

background varies according to the distance to the

target. This last estimator offers the subject the

opportunity to perform fine-tuning on his hand's

position. The activation of a specific muscle is also

displayed as sweat drops coming out the virtual

sleeve, proportionally to the produced effort.

2.2 An Inverse Model of the Subject

The subject can be seen as a model receiving a

biomechanical exercise as an input and producing a

performance, that may include kinematic signals

LEARNING FROM BIOFEEDBACK - Patient-specific Games for Neuromuscular Rehabilitation

169

(e.g. the trajectory of a limb), as well as EMG

signals as an output. We wish to develop a system

able to generate a patient-specific physiotherapeutic

task, given the kinematic and/or muscular

performance of the subject. This goal may be

attained from an estimation of the inverse model of

the subject, with the desired performance as the

input and the exercise trajectory at the output

(Figure 2).

Figure 2: Best Estimated Inverse of the Subject.

2.3 The Learning System

Our goal is to train an artificial neural network

capable of producing a subject-specific

biomechanical task, given a desired subject’s

performance. The definition of the subject’s

performance is arbitrary and may include for

example kinematic, kinetic, or muscular parameters.

2.3.1 Network with Kinematic Input Only

In the first phase of our study, we tracked the

kinematics of the subject’s pointing finger. We use

this data to train a neural network. However, we first

have to determine the network architecture.

The universal approximation theorem for neural

networks states that every continuous function that

maps intervals of real numbers to an output interval

of real numbers can be approximated at any level of

desired accuracy by a multi-layer feed-forward

neural network with a single hidden layer having a

sigmoid activation function. In our case, the network

is designed to map, at each instant, the desired

position and velocity of a marker placed on the

subject, to another spatial point and velocity

corresponding to the displayed exercise.

Consequently, we need a system capable of

modeling the mapping from R

to R

. We assume the

mapping to be a continuous function and use the

approximation theorem to build a multi-layer feed-

forward network with one hidden layer. This

network has a generic architecture for all subjects

and is specific to this definition of the performance.

Nonetheless, each subject will have his own tuned

network.

To define the network’s architecture, we start

with a known exercise trajectory. The subject is

presented with this task and is asked to follow the

target trajectory displayed to him. Concurrently, the

tracking system records his motions at given

timestamps. The recorded kinematic data serves as

an input set to the network, while the corresponding

exercises displayed to him serve as target outputs.

We use Levenberg-Marquardt error back-

propagation learning method (Moré, 1977) so that

the error between the actual output of the network

and the target output is minimized. The network’s

weights are initialized according to Nguyen-

Widrow’s method (Nguyen & Widrow, 1990).

The network used for a single marker contains

six input neurons and six output neurons:

corresponding to the spatial position and velocity

vectors. Next, the number of neurons in the hidden

layer needs to be determined. On the one hand, this

parameter affects the runtime and should therefore

be minimized. On the other hand, it also influences

the system’s accuracy and balance between

precision and efficiency must be attained. We

performed a set of tuning experiments where several

healthy subjects were given a cyclic trajectory as a

task to follow while their motions were tracked

(Figure 3).

Figure 3: Three-dimensional task and subject’s

performance.

The network was trained on the training set

composed by the exercise and the average

performance over the number of cycles. We used

three criteria to evaluate the network’s performance

for each subject and exercise: the output error, the

positional output error and the average of the

NCTA 2011 - International Conference on Neural Computation Theory and Applications

170

positional output errors over each of the ten different

cycles performed by the subject. The average of the

output errors reflected the network’s ability to

extrapolate its results on samples that were not

directly included in the training set. After iteratively

testing different sigmoid activation functions for the

hidden layer of the network, numbers of epochs and

numbers of hidden neurons, the final configuration

of the network was determined.

The resulting network comprised seven hidden

neurons with the hyperbolic tangent as activation

function and the number of epochs was set to 50.

The reader is referred to previous publication

(Barzilay and Wolf, 2009) for a detailed explanation

on the setting of the network’s architecture.

2.3.2 Network with Kinematic and EMG

signals as Input

In the second step of this research, we added the

patient's biceps and triceps EMG signals to the input

of the network, such that the new exercise would be

designed with respect to the information on the

muscles of the subject as well as the knowledge on

the kinematic data of his limb.

The data provided by the electromyograms

contain useful information that can be deciphered by

signal processing. There are numerous ways

described in the literature to extract this information

from the EMG signal, including analysis in time or

frequency domains. We use the workflow described

in Hodges and Bui (1996) to compute the linear

envelope of the signal by processing it in the time

domain. The processed signal is needed at every

instant in our application and the processing time

has to be minimized, all the more since several

signals are needed simultaneously. We accelerated

this operation by using the processed signals from

the precedent instant and reduced the processing

time by approximately 96% (Barzilay and Wolf,

2011). This fast implementation allows providing

the subject with continuous visual feedback on his

own muscular performance during the training.

The same parameters that were described in

section 2.3.1 are used to evaluate the network, but

now in addition to the 3D curve, the desired EMG

performance specific to that trajectory should also be

designed. We therefore determined a few desired

cyclic trajectories for the limb of the subject and

recorded the EMG performances of a dozen of

healthy subjects. The average of this set of data is

then used as the desired EMG performance over a

specific trajectory, and fed as input to the neural

network together with the trajectory of the desired

kinematic performance.

The number of neurons in the hidden layer has

been set to 17, according to the evaluation criteria

which were previously used.

2.3.3 System Evaluation

The first network, described in section 2.3.1,

considers only the endpoint kinematics of the subject

and has obviously less physiotherapeutic interest

than the network involving the subject’s muscular

performance (section 2.3.2). Nevertheless, the

optimistic results (section 3.1 and Barzilay and

Wolf, 2009) suggested evidence of the feasibility of

modeling human motor control with neural networks

and brought us to expand the subject model to

include muscular performance as well.

From a therapeutic perspective, the muscles

activation of the patient is more significant than his

ability to accurately reproduce specific trajectories.

For that reason, we focus our efforts on minimizing

the error in the EMG performance, whereas the

kinematics error is considered more moderately.

Although the EMG signals are calibrated from

measurement of the maximal voluntary contraction

prior to the training, the signals’ amplitudes tend to

differ between different subjects. Furthermore, we

focus on the rhythmical patterns of the muscles more

than on the activation intensity. To do that, we

consider the error between the desired and actual

EMG performance in the frequency domain.

For the evaluation of the adaptive system, we

thus consider the root mean squared deviation, in the

frequency domain, between the desired EMG

performance and the smoothed EMG performance of

the subject. Each participant (n = 16) performed

motor training on two exercises: the patient-specific

exercise produced by the trained neural network

(adapted training), and a general exercise having for

trajectory the desired kinematic performance. The

latter resembles a standard physiotherapeutic session

where the physiotherapist demonstrates to the

patient, for example with his hand, the desired

gesture to reproduce (conventional training). The

primary criterion for the system evaluation was

defined as the ratio of the errors obtained in the

performances in both cases.

3 RESULTS

3.1 Network with Kinematic Input

The exercise trajectory, designed by the network,

LEARNING FROM BIOFEEDBACK - Patient-specific Games for Neuromuscular Rehabilitation

171

deviates by 15 millimeters per point in average from

the exact trajectory. This average deviation is

reduced to 3-5 millimeters per point when a

smoothing filter is applied to the trajectory produced

by the network. The network succeeds by such to

estimate the inverse model of the subject.

In Figure 3, one can see that, due to the relative

location between the planar target and the subject's

eye, the task (in blue) can be perceived as a

projection on a plane normal to the subject's line of

sight. Nevertheless, the neural network system

learned and corrected the projection, although far

from being a linear phenomenon.

Figure 4: The exercise projection: side (left) and front

(right) views.

Once trained, the network is capable of

generating a patient-specific exercise, given a

desired patient performance. The inverse model of

the patient can be evaluated by comparing the

measured performance of the subject with respect to

desired one. Given in figure 5 are the desired

performance, the patient-specific exercise created by

the network, and the performance of the subject on

this adapted task.

Figure 5: Desired Vs. actual performance: front (left) and

side (right) views.

Let us recall that, during rehabilitation session,

the subject does not see the whole exercise

trajectory, but only the virtual ball, in motion along

that trajectory. Moreover, he is not exposed to the

desired trajectory. In the depicted case, the average

distance between the subject's performed trajectory

and the desired performance was approximately

equal to the average distance between the hand

trajectory and the trajectory of the displayed

exercise. However, in several sections of the task,

the subject's trajectory was noticeably closer to the

desired trajectory than to the virtual ball, as can be

observed on the left side of the side view in Figure

5. It is also notable that, in this section, the network

was able to predict that, in order to cause the

subject's hand to follow the desired trajectory (in

black), the virtual ball had to be displayed a bit

farther along the y axis (farther from the subject).

This observation suggests that the model created by

the network was able to detect some of the subject's

behavioral patterns. This phenomenon was observed

in several sessions and for different subjects. It is

also notable, in Figure 5, that the system’s prediction

was effected twice on the same portion of the

exercise trajectory, while the subject’s hand had

different velocities.

3.2 Network with Kinematic and EMG

Signals as Input

Figure 6 demonstrates the capability of the network

to adapt itself to the muscular information recorded

from the electromyograms. Patterns characteristic to

the desired signal appear within the performance of

the subject on the exercise designed by the system.

Frequency-domain analysis shows how close the

spectra of the desired and the actual performance

signals are. We found that in many cases the subject

obtained a better performance on the network-

designed exercise than if he is directly shown the

desired trajectory of his limb as an exercise (Figure

6).

Figure 6: EMG performance: time (left) and frequency

(right) domains.

The deviations from the desired EMG

performance in the frequency domain for the

adapted and conventional training are presented in

Table 1. These results are summarized in Table 2.

NCTA 2011 - International Conference on Neural Computation Theory and Applications

172

Table 1: Adapted and conventional training comparison.

Suhject

Adapted Training

Conventional

Training

Error Ratio

Biceps Triceps Bi. Tri. Bi. Tri.

53.13 66.56 73.84 77.13 1.39 1.16

36.25 85.10 76.95 79.76 2.12 0.94

59.49 32.24 78.75 58.90 1.32 1.83

85.00 85.64 40.44 92.74 0.48 1.08

88.92 60.19 67.09 83.63 0.75 1.39

63.85 97.61 56.58 84.20 0.89 0.86

92.61 94.98 62.03 89.72 0.67 0.94

62.94 39.51 18.53 68.19 0.29 1.73

72.05 59.18 81.32 88.28 1.13 1.49

#10

74.67 42.19 148.20 58.98 1.98 1.40

#11

72.05 59.18 81.32 88.28 1.13 1.49

#12

65.00 29.28 74.27 86.06 1.14 2.94

#13

92.78 64.54 103.48 48.57 1.12 0.75

#14

56.12 55.39 71.95 69.94 1.28 1.26

#15

73.25 47.07 54.98 52.16 0.75 1.11

#16

44.71 93.98 54.85 72.04 1.23 0.77

Table 2: Adapted and conventional training comparison ˗

Summary.

Adapted Training

Conventional

Training

Error Ratio

Biceps Triceps Bi. Tri. Bi. Tri.

Average

68.30 63.29 71.54 74.91 1.10 1.32

Std. Dev.

16.50 22.53 28.10 14.16 0.48 0.54

Most subjects (n = 14, 87.5 %) benefitted from

our system for at least one of the muscles, and

almost half of them improved the accuracy of their

muscular performance for both biceps and triceps (n

= 7, 43.75%). In summary, the results indicate that

the average muscular performance of the subjects is

closer to the desired performance when the exercise

is generated by the system, rather than set as the

desired kinematic performance like in conventional

physiotherapy. This is indicated by a 10% increase

for the biceps performance and by 32% for the

triceps performance.

A one-tailed Student T-test shows that the

improvement in the triceps performance is attained

with statistical significance (p = 0.02). However, the

improvement in the biceps performance is lesser in

magnitude and in statistical significance (p = 0.34

and p = 0.09 with omission of two subjects). We

believe that this is due to the fact that the physical

exercise stimulated by the system involves the

biceps in a smaller measure than the triceps or the

shoulder muscles.

4 DISCUSSION

We introduce, in this study, a platform for motor and

cognitive rehabilitation, able to model the subject's

kinematics and to generate a subject-specific

physiotherapeutic exercise. The system requires no

prior knowledge on the patient, nor any model of his

motor control or trajectory planning. It only involves

the desired performance dictated by the

physiotherapist and a training reference set, recorded

in situ from the patient's performance prior to

rehabilitation. To date, and to the best of our

knowledge, no study combining virtual reality

rehabilitation and learning algorithms for patient-

specific training has been reported.

The developed system offers several

opportunities to both the physiotherapist and the

patient. The virtual tasks can be designed as

interactive games and stimulate the motivation of the

patient during rehabilitation. We have developed

several applications where the subjects are enjoined

to pop bubbles, stop soccer balls, or whack objects

with their hands in a controlled way. Most of the

participants expressed their enthusiasm after having

performed motor training in our virtual applications.

In every session, all the kinematic and EMG data

are stored and may be further analyzed offline by the

physiotherapist. Furthermore, the system proved to

emphasize some kinematic and muscular patterns in

motor training, and may contribute to a better

diagnosis of the subject. This system may find its

application in patients after stroke, with cerebral

palsy, dyslexia or other developmental coordination

disorders.

At this time, we have tested the system on

healthy subjects only, choosing to generalize and

validate it before testing it on pathological subjects.

The success of the system in learning some of the

subject’s behavioral patterns leads us to expect good

results in the modeling of motor patterns in patients

with pronounced pathology. Besides clinical trials,

we would like to expand the system to combine

more markers and EMG sensors, and other biometric

sensors.

Besides physiotherapy, this system could prove

useful in sportive performance enhancement, in the

development of new types of human-machine

interfaces, in entertainment, and in the training of

any kind of motor skills.

While a description of the subject model,

including his motor control and trajectory planning,

LEARNING FROM BIOFEEDBACK - Patient-specific Games for Neuromuscular Rehabilitation

173

hand-eye coordination and probably many additional

features, would be very difficult to elaborate, the

computational power of the simplest form of feed-

forward neural networks provided very optimistic

results in the modeling of the subject. First to

combine virtual rehabilitation with machine learning

of human models, the positive results of this study

encourage carrying on the use of biofeedback-based

artificial intelligence and virtual reality, for

applications in therapy and other diverse areas.

REFERENCES

Barzilay, O. and Wolf, A. (2009). An adaptive virtual

system for neuromuscular rehabilitation. In: IFMBE

Proc. World Congress on Medical Physics and

Biomedical Engineering, Munich, Germany. 25(4):

1291–94.

Barzilay, O. and Wolf, A. (2011). A fast implementation

for EMG signal linear envelope computation. Journal

of Electromyography and Kinesiology, 21: 678-682.

Hodges, P. and Bui B. (1996). A comparison of computer-

based methods for the determination of onset of

muscle contraction using electromyography.

Electroenc. Clin. Neurophysiol, 101, 511–519.

Holden, M. K. (2005). Virtual environments for motor

rehabilitation: review. Cyberpsychology & Behavior,

8(3),187–211.

Holden, M. K. and Dyar, T. (2002). Virtual environment

training: a new tool for neurorehabilitation. Neurology

Report, 26, 62–71.

Holden, M. K., Dyar, T., Schwamm L. and Bizzi, E.

(2005). Virtual Environment-Based Telerehabilitation

in Patients with Stroke. Presence: Teleoperators and

Virtual Environments, 14(2), April, 214-233.

Jack, D., Boian R., Merians, A. S., Tremaine, M., Burdea,

G. C., Adamovich, S. V., Recce, M. and Poizner, H.

(2001). Virtual reality–enhanced stroke rehabilitation.

IEEE Transactions on Neural Systems and

Rehabilitation Engineering, 9, 308–318.

Jacobson, J., Redfern, M. S., Whitney, S. L. and Wilson, J.

B. (2001). Balance NAVE: A Virtual Reality Facility

for Research and Rehabilitation of Balance Disorders.

In Proc. Virtual Reality Software and Technology

Meeting, Banff, Canada.

Kawato, M. (1990). Computational schemes and neural

network models for formulation and control of

multijoint arm trajectories. In Neural networks for

control, eds. Miller WT, Sutton RS and Werbos PJ.

MIT press.

Kawato, M. (1999). Internal models for motor control and

trajectory planning. Current Opinion in Neurobiology

9:718–727.

Maclean, N., Pound, P., Wolfe, C. and Rudd, A. (2000). A

qualitative analysis of stroke patients' motivation for

rehabilitation. BMJ, 321, 1051–1054.

Moré, J. J. 1977. The Levenberg-Marquardt algorithm:

Implementation and Theory. In: Numerical Analysis.

Lecture Notes in Mathematics 630, ed. Watson GA,

105–116. Berlin: Springer-Verlag.

Nguyen, D., Widrow, B. (1990). Improving the learning

speed of 2-layer neural networks by choosing initial

values of the adaptive weights. In Proc.International

Joint Conference on Neural Networks, July, 3:21–26.

Regian, J. W., Shebilske, W. L. and Monk, J. M. (1992).

Virtual reality: an instructional medium for visual

spatial tasks. Journal of Communication, 7, 131–145.

Rizzo, A. and Kim, G. J. (2005). A SWOT Analysis of the

Field of Virtual-Reality Rehabilitation and Therapy.

Presence: Teleoperators and Virtual Environments,

14(2), April, 119-146.

Shea, C. H. and Wulf, G. (1999). Enhancing motor

learning through external-focus instructions and

feedback. Human Movement Science, 18, 553–571.

Todorov, E., Shadmer, R. and Bizzi, E. (1997).

Augmented Feedback Presented in a Virtual

Environment Accelerates Learning of a Difficult

Motor Task. Journal of Motor Behavior, 29, 147–158.

VICON Motion Systems at http://www.vicon.com/

NCTA 2011 - International Conference on Neural Computation Theory and Applications

174