Clustering of Emotional States under Different Task Difﬁculty Levels for

the Robot-assisted Rehabilitation system-RehabRoby

Yigit Can Aypar

, Yunus Palaska

, Ramazan Gokay

, Engin Masazade

, Duygun Erol Barkana

and Nilanjan Sarkar

Department of Electrical and Electronics Engineering, Yeditepe University, 34755, Istanbul, Turkey

Department of Mechanical Engineering, Vanderbilt University, Nashville, TN 37235, U.S.A.

Keywords:

Robot-assisted Rehabilitation System, Biofeedback Sensors, Unsupervised Learning.

Abstract:

In this paper, we study an unsupervised learning problem where the aim is to cluster the emotional state (ex-

citedness, boredom, or stress) using the biofeedback sensor data while subjects perform tasks under different

difﬁculty levels on the robot assisted rehabilitation system-RehabRoby. The dimension of the training vectors

has been reduced by using the Principal Component Analysis (PCA) algorithm after collecting the biofeedback

sensor measurements from different subjects under different task difﬁculty levels to better visualize the sensor

data. The reduced dimension vectors are fed into a K-means clustering algorithm. Numerical results have been

given to demonstrate that for each training vector, the emotional state decided by the clustering algorithm is

consistent with the subjects declaration of his/her emotional state obtained via surveys after performing the

task.

1 INTRODUCTION

Robot-assisted rehabilitation systems were ﬁrst used

in large scaled clinical tests in 1998 and since then

several robot-aided rehabilitation systems have been

developed. There are two kinds of robot-assisted re-

habilitation systems for the upper extremities in terms

of mechanical design, which are end-effector-based,

and exoskeleton type rehabilitation robots. MIT-

MANUS (Krebs et al., 2004), MIME (Lum et al.,

2006), GENTLE/S (Loureiro et al., 2003) and NeRe-

Bot are end-effector-based, and ARMin (Nef et al.,

2009), CADEN-7 (Perry et al., 2007), RUPERT IV

(Balasubramanian et al., 2008), WREX (Rahman

et al., 2006), SRE (Caldwell et al., 2007), RehabExos

(Vertechy et al., 2009), ExoRob (Rahman et al.,

2009), SUEFUL-7 (Kiguchi et al., 2008), IntelliArm

(Ren et al., 2009) are exoskeleton type robot-assisted

rehabilitation systems. Exoskeleton type robots re-

semble the anatomy of the human arm and each of

the robot’s joints can be controlled separately, which

reduces the control complexity. Recently, we have de-

veloped an exoskeleton type upper-extremity robot-

assisted rehabilitation system, which is called Re-

habRoby (Ozkul et al., 2012), (Ozkul and Barkana,

2013). Robot-assisted rehabilitation systems have

shown to be helpful in neuromotor rehabilitation be-

cause it is possible to deliver interactive and repeat-

able sensorimotor exercise, and monitor the actual

performance continuously. However, to obtain opti-

mal performance from such exercises, the task dif-

ﬁculty needs to be appropriately challenging. Note

that a rehabilitation task that is too easy or under chal-

lenging can be perceived as boring, while a task that

is too challenging can be frustrating. An optimally

challenging rehabilitation task can motivate and cause

maximum mental engagement for the patients (No-

vak et al., 2012). Mental engagement of patients have

shown to be a key factor to improve the outcome of

the rehabilitation (Maclean and Pound, 2000). Mo-

tor learning theory states that learning rate increases

when the rehabilitation task challenges and excites the

subjects (Guadagnoli and Lee, 2004). Thus, it is im-

portant that a robot- assisted rehabilitation system aid-

ing in rehabilitation tasks should be capable of detect-

ing that patient is either becoming bored or frustrated,

and then modifying the rehabilitation task to better

suit the patient’s abilities.

A recent survey paper presents the current state

of the art and the new challenges in automatic,

dimensional and continuous analysis and synthesis

of human emotional behavior (Gunes et al., 2011).

Aypar Y., Palaska Y., Gokay R., Masazade E., Erol Barkana D. and Sarkar N..

Clustering of Emotional States under Different Task Difﬁculty Levels for the Robot-assisted Rehabilitation system-RehabRoby.

DOI: 10.5220/0005052600340041

In Proceedings of the 11th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2014), pages 34-41

ISBN: 978-989-758-039-0

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

The most commonly employed strategy in auto-

matic dimensional emotional state detection is the

2D valence-arousal (V-A) emotion model (Russell,

1989). Various difﬁculty levels of a rehabilitation task

has previously been deﬁned according to the circum-

plex model of affect (Koenig et al., 2011b), (Koenig

et al., 2011a). Emotional state may become apparent

by subjective experiences (how the person feels), in-

ternal/inward expressions (psychophysiological mea-

surements) and external/outward expressions (audio/

visual signals) (Gunes et al., 2011). The great chal-

lenge of psychophysiological measurements is the

abundance of available data, and the hundreds of

features that can be extracted by considering all the

physiological responses. Psychophysiological mea-

surements such as heart and muscle activity, brain

activity, blood pressure, skin temperature, respira-

tion, are multichannel recordings from both the cen-

tral nervous system (CNS), and the autonomic ner-

vous systems (ANS). These measurements are all rich

sources of information concerning the physiological

responses of the human body. The transition from

one emotional state to another, for instance, from state

of boredom to state of anxiety is accompanied by

dynamic shifts in indicators of ANS activity. Fur-

thermore, it has previously been mentioned that the

signals from the autonomic nervous system (ANS)

can indicate mental engagement (Filipovic and An-

dreassi, 2001). Changes in psychophysiological mea-

surements can be used to assess mood and engage-

ment, and to understand emotions of people in a va-

riety of applications (Mandryk and Atkins, 2007),

(Rani et al., 2003b), (Rani et al., 2003a). Psychophys-

iological measurements can also be used to under-

stand emotions of patients during the execution of a

rehabilitation task using a robotic system. Heart rate,

skin conductance responses, and skin temperature

have previously been used to estimate and classify

a patient’s psychological state during robot-assisted

gait rehabilitation (Koenig et al., 2011b), (Koenig

et al., 2011a).The biofeedback sensors can be used

to measure the emotions. In this study, we use skin

conductance, electrocardiography (ECG), tempera-

ture, blood volume pulse (BVP) and respiratory rate

biofeedback sensors to detect if the subject has been

bored, overstressed or excited (emotional states) dur-

ing the execution of a rehabilitation task.

Since the raw data from the biofeedback sensors

are inadequate, several features from the raw sen-

sor data have previously been deﬁned to classify the

emotion states (Koenig et al., 2011b), (Koenig et al.,

2011a). Note that proper feature selection from the

sensory data determines the accuracy of the emo-

tional state classiﬁcation. The Pulse Transit Time

(PTT) and Heart Rate Variability (HRV) have previ-

ously been used as feature for Electrocardiography

(ECG), and delta(1-3Hz), theta(4-7Hz), alpha1(8-

9Hz), alpha2(10-12Hz), beta1(13-17Hz), beta2(18-

30Hz), gamma1(31-40Hz), gamma2(41-50Hz) have

been used as features for Electroencephalography

(EEG) sensor (Kandemir, 2013). High frequency

components, standard deviation, standard deviation

of coefﬁcients of Daubechie transform (3th and 4th)

and mean of the peak values of ECG signal have pre-

viously been used as features to detect the excite-

ment, anxiety, boredom, frustration and anger (Rani

et al., 2007). Mean values derived for phasic and

tonic components of raw skin conductance data, am-

plitude of hurbis, mean and median frequency of hur-

bis, and standard deviation of raw EMG and mean,

power spectrum density and slope of skin temper-

ature have also been used as features to detect the

emotions (Rani et al., 2007). In this study, we use

the features heart rate(HR), HR mean(beat/min), in-

terbeat interval(IBI), pulse transit time(with BVP)

(PTT(ms)), very low frequency(VLF) power(%),

low frequency(LF) power(%), high frequency(HF)

power(%), total VLF power, total LF power, total

HF power, and HR standard deviation (from ECG);

time between two successive beats of the heart (IBI)

(from BVP); skin conductance value(%), mean of

skin conductance value(uS), and percentage of mean

of the skin conductance level (from skin conduc-

tance); mean of the temperature (deg), percentage of

the mean of the temperature, and percentage value

of temperature (from temperature sensor), and respi-

ration rate(br/min) (from the respiratory rate) to de-

tect the emotional states (boredom, stress and excite-

ment).

In this paper, an unsupervised learning problem

has been considered. The biofeedback signals from

different subjects under different difﬁculty levels have

been collected. The aim is to distinguish the emo-

tional state of the subjects while performing a re-

habilitation task with a robot-assisted rehabilitation

system-RehabRoby. The emotional states that we

want to classify are excitedness, boredom, and over-

stress. Correct classiﬁcation of the emotional state is

an important step for the later steps of the rehabilita-

tion task. As an example, if it is determined that the

subject is bored, the difﬁculty level of the task can be

very easy for the subject, and the difﬁculty level may

need to be increased. If, on the other hand, it is de-

termined that the subject is excited, we can deduce

that the task difﬁculty level is optimal for the subject.

Finally, if the subject is found to be stressed, the dif-

ﬁculty level of the task can be very challenging for

the subject and the difﬁculty level may need to be de-

ClusteringofEmotionalStatesunderDifferentTaskDifficultyLevelsfortheRobot-assistedRehabilitation

system-RehabRoby

creased (Koenig et al., 2011b).The emotional states

can be classiﬁed solely based on the survey results af-

ter performing the task but the subjects may not be

sure about their feelings, may not want to reveal their

feelings or may declare a misleading answer (Novak

et al., 2012). Therefore, emotional state classiﬁcation

based on biofeedback sensor data may depend on the

surveys. On the other hand, survey results can still

be used to determine the accuracy of the classiﬁca-

tion algorithm (Koenig et al., 2011b). The output of

the classiﬁcation algorithm can be further used in the

later steps of the rehabilitation task. In this paper, af-

ter collecting biofeedback signals from subjects, we

develop a clustering problem by using principal com-

ponent analysis (PCA) and K-means clustering algo-

rithms (Bishop, 2006) to obtain the clusters related

with each emotional state. The training vectors re-

lated to each emotional state is then compared with

the survey results to test the accuracy of the cluster-

ing algorithm.

The rest of this paper is organized as follows, in

Section II, we present our materials, experiments and

unsupervised learning methods. In Section III, we

present our numerical results. In Section IV, we con-

clude our work and address future research directions.

2 MATERIALS AND METHODS

In this section, we present the details of the hardware

properties of the robot-assisted rehabilitation system,

RehabRoby, the task, and the unsupervised learning

method used for classiﬁcation of the emotional states.

2.1 Hardware

In this study, an exoskeleton type upper-extremity

robot-assisted rehabilitation system, which is called

RehabRoby, has been used (Fig. 1). RehabRoby has

been designed in such a way that i) it can implement

passive mode therapy, active-assisted mode therapy

and resistive mode therapy, ii) it can be easily adjusted

for people of different heights and with different arm

lengths, and iii) it can be used for both right and left

arm rehabilitation. RehabRoby is designed to pro-

vide extension, ﬂexion, abduction, adduction, rota-

tion, pronation and supination upper-extremity move-

ments and the combinations of these movements that

are necessary for the tasks and activities of daily liv-

ing. An arm splint, which has humeral and forearm

thermoplastic supports with Velcro straps and a sin-

gle axis free elbow joint, has been designed and at-

tached to RehabRoby (Figure1). The thermoplastic

arm splint designed for the RehabRoby has humeral

and forearm supports with Velcro straps, and a ther-

moplastic inner layer that is covered by a soft material

(Plastazote). Two force sensors (Kistler - 9313AA1;

Kistler France, Les Ulis, France) are placed in the in-

ner surface of the plate attached dorsally to the fore-

arm splint (Figure1). One of the force sensors mea-

sures the force applied during the elbow ﬂexion move-

ment and the other measures the force applied during

the shoulder ﬂexion movement. Ensuring the safety

of the subject is a critical issue for a robot-assisted

rehabilitation system. Thus, in the event of an emer-

gency situation, the therapist can press an emergency

button to stop the RehabRoby (Figure1), and the mo-

tor drivers of RehabRoby are disabled separately or

together by pressing the driver enable/disable buttons

without turning off RehabRoby. The system is pow-

ered by an uninterruptible power supply, thus, there

can be no power loss and RehabRoby will not col-

lapse at any time.

Figure 1: Robot-Assisted Rehabilitation System-

RehabRoby.

In this study, skin conductance, electrocardiogra-

phy (ECG), temperature, blood volume pulse (BVP)

and respiratory rate from Thought Technology Ltd

have been used for biofeedback sensory information

to detect three emotion states boredom, overstressed

and excited. Physiological signals have been sam-

pled at 256Hz using Procomp Inﬁnite Encoder. Skin

conductance has been measured using skin conduc-

tance sensor (sc-ﬂex pro) (Figure 2-left). The elec-

trodes have been placed on the middle phalanx of

index ﬁnger and ring ﬁnger. There are three ECG

electrodes used for the measurement. Two of them

are placed where the Deltoid anterior heads and Pec-

toralis major muscles intersect, several centimeters

below the clavical. The third ECG electrode is placed

near the xiphoid. Temperature has been recorded us-

ing a skin type sensor that is placed on the ﬁnger tip of

the thumb (Figure2-left). Blood volume pulse (BVP)

has been recorded using BVP-Pro ﬂex sensor that is

placed on the ﬁnger tip of the middle ﬁnger (Figure 2-

left). Respiration sensors (resp-ﬂex pro), which have

been placed around abdomen and the chest, have been

ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics

used to record the respiratory rate (Figure 2-right).

Skin conductance, temperature and respiratory rate

signals have been sampled at 256Hz and ECG and

BVP have been sampled at 2048Hz. The features

derived from the signals of these sensors have been

downsampled to 32Hz for learning algorithm because

of the hardware limitations of the computer.

BioGraph Inﬁnity is the interface software pro-

vided by Thought Technology Ltd to get both raw

sensor data and the features from the raw data.

19 features from the skin conductance, electrocar-

diography (ECG), temperature, blood volume pulse

(BVP) and respiratory rate sensory data have been

selected for this study. The physiological sig-

nals we examined were: various features of ECG,

including heart rate(HR), HR mean(beat/min), in-

terbeat interval(IBI), pulse transit time(with BVP)

(PTT(ms)), very low frequency(VLF) power(%),

low frequency(LF) power(%), high frequency(HF)

power(%), total VLF power, total LF power, total

HF power, and HR standard deviation; time between

two successive beats of the heart (IBI) (from BVP);

skin conductance value(%), mean of skin conduc-

tance value(uS), and percentage of mean of the skin

conductance level (from skin conductance); mean

of the temperature (deg), percentage of the mean

of the temperature, and percentage value of tem-

perature (from temperature sensor), and respiration

rate(br/min) (from the respiratory rate). These signals

were selected because they i) were shown to capture

important information about the underlying targeted

emotion states, (ii) could be measured non-invasively;

and iii) were relatively resistant to movement arti-

facts.

Figure 2: Sensor Placement on Subjects.

2.2 Task

A well-known rehabilitation tasks has been selected

in consultation with therapists at Yeditepe Univer-

sity’s Physiotherapy and Rehabilitation Department.

The task is the elbow ﬂexion and extension movement

(i.e., reaching up to the chest button up). Each task

takes one minute. Subjects are seated in the chair,

and their arms are placed in the splint and tightly se-

cured with Velcro straps (Figure1). The height of the

RehabRoby can be adjusted for each subject so that

they would start the task in the same arm conﬁgura-

tion. Initially, the subject’s shoulder is positioned at

an extension of 90

◦

, the elbow is it the neutral posi-

tion, the lower arm is at a pronation of 90

◦

, and the

hand and the wrist are free at the neutral position. A

computer monitor has been placed in front of the sub-

ject to provide visual feedback about his/her motion

trajectory during the execution of the task.

The subject is asked to catch a ball that is dis-

played on the computer monitor (Figure3). Black

ball demonstrates the reference trajectory and red ball

demonstrates the subject’s movement. Subjects are

expected to perform the same task with ﬁve different

difﬁculty levels following resting period. The differ-

ent difﬁculty level is deﬁned by changing the number

of repeated ﬂexion and extension movement in a cer-

tain amount of time. For example, the subject is asked

to ﬂex and extend his/her arm 5 times in 4 minutes

for the ﬁrst difﬁculty level and 25 times for the ﬁfth

difﬁculty level. Thus, the angular speeds of these ﬁve

different levels are 2.5

◦

/s,5

◦

/s,7.5

◦

/s,10

◦

/s and 12.5

◦

consideringthe repetition numbers. The reference tra-

jectories for each difﬁculty level are computed using

minimum-jerk trajectory.

Figure 3: The Reference Trajectory and the Subject’s Move-

ment.

2.3 Experiment Procedure

Initially, the biofeedback sensors were placed on

the subject as shown in (Figure 2). Then sub-

ject was asked to track the black ball as shown

in (Figure3) for all 5 different angular velocities

(2.5

◦

/s,5

◦

/s,7.5

◦

/s,10

◦

/s and 12.5

◦

/s). Subjects were

asked to complete The Self-Assessment Manikin

(SAM) survey, which has previously been used to

measure emotional responses (Bradley and Lang,

1994). We also used SAM to verify the difﬁculty lev-

els of the task really resulted in feelings of boredom,

stress and excitement.

2.4 Classiﬁcation Method

In unsupervised learning, we have m training vec-

tors x

, x

, . . . , x

to be clustered where a training

vector i (i ∈ {1, 2, . . . , m}), x

has n features. Then,

ClusteringofEmotionalStatesunderDifferentTaskDifficultyLevelsfortheRobot-assistedRehabilitation

system-RehabRoby

, [x

, x

, . . . , x

]

where T is the transpose opera-

tor.

The elements of each training vector is obtained from,

• x

, IBI from BVP: Used to compute the time be-

tween beats (period) from a BVP signal (Average

of every 15 samples).

• x

, Heart Rate(HR) from IBI: Number of heart

beats during a one minute period(from bvp sen-

sor) (Average of every 15 samples).

• x

, ECG VLF power(%): Percentage of very low

frequency power (Average of every 285 samples).

• x

, ECG LF power(%):Percentage of low fre-

quency power (Average of every 285 samples).

• x

, ECG HF power(%):Percentage of high fre-

quency power (Average of every 285 samples).

• x

, ECG VLF power(total): Very low frequency

components of total power (Average of every 285

samples).

• x

, ECG LF power(total): Low frequency compo-

nents of total power (Average of every 285 sam-

ples).

• x

, ECG HF power(total): High frequency com-

ponentsof total power(Averageof every 285 sam-

ples).

• x

, Skin Conductance value(%): Percentage value

of skin conductance (Average of every 400 sam-

ples).

• x

, Temperature value(%): Percentage value of

temperature (Average of every 400 samples).

• x

, ECG Heart Rate(HR) Mean:Mean of number

of heart beats during a one minute period(from

ecg sensor) (Average of every 380 samples).

• x

, ECG HR standart deviation: Standard devia-

tion of heart rate (Average of every 380 samples).

• x

, Skin Conductance mean:Mean of skin con-

ductance value(uS) (Average of every 400 sam-

ples).

• x

, Skin Conductance mean(%): Percentage of

mean of the skin conductance level (Average of

every 400 samples).

• x

, Temperature mean: Mean of the temperature

(Deg) (Average of every 400 samples)

• x

, Temperature mean(%):Percentage of the

mean of the temperature (Average of every 400

samples).

• x

, Respiration Rate: Number of the breaths dur-

ing a one minute period (Average of every 380

samples).

• x

, ECG IBI NN Intervals: Inter-beat interval be-

tween nearest neighbor beats (Average of every

380 samples).

• x

, Pulse Transit Time (PTT): Time it takes fort

he blood pumped by heart to reach arteries in the

arms (Average of every 15 samples).

In our experiments, we have m = 798 training vec-

tor which are obtained under 5 difﬁculty levels and

one resting period as explained in Section 2.2. Then

each training vector reﬂects the emotional state for

1.8 seconds. We ask the subject to express his/her

the emotional feelings after each 4 minute task exe-

cutions. Then SAM surveys are used to validate the

classiﬁcation results.

2.4.1 Principal Component Analysis (PCA)

Dimensionality reduction is a useful step for visual-

izing and processing high-dimensional data as in this

problem. In order to visualize the biofeedback sen-

sor data under different emotional state, the 19 di-

mensional training vector needs to be converted in

2-dimensional or 3-dimensional vectors. Principal

Component Analysis (PCA) (Bishop, 2006) is a very

useful dimensionality reduction method to convert

high dimensional data into lower dimensions while

still keeping as much as the variations of the original

data. In this paper, using the PCA method, the n = 19

dimensional biofeedback sensor data is mapped into

k = 2 dimensional data as follows (Ng, 2014):

• Initialize the training vectors x

, x

, . . . , x

• Feature scaling/mean normalization

∑

i=1

and s

∑

i=1

− µ

)

• Replace x

with

−µ

• Compute the covariance matrix

Σ =

∑

i=1

)

• Compute the eigenvectors of Σ using singular

value decomposition.

• Pick the ﬁrst k eigenvectors of Σ to create the ma-

trix U of size k× n.

• Multiply each training vector x

with U. So the

reduced dimension training vector is obtained as

= Ux

which has dimension k × 1.

2.4.2 K-means Clustering Algorithm

The K-means clustering algorithm partition m obser-

vations into K clusters where each observation be-

longs to the cluster with the nearest mean (Bishop,

ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics

2006). According to (Ng, 2014), the K-means clus-

tering algorithm can be summarized as follows,

• Set the number of clusters K (In this problem K =

3) and get the reduced dimension training vectors

as x

= z

, . . . , x

= z

• Randomly initialize K cluster centroids,

, . . . , µ

• Cluster assignment step: Assign each training

vector x

to the cluster c

according to

= arg min

k∈{1,2,3}

||x

− µ

• Centroid update step: Update the cluster centroid

based on the mean of the training vectors belong

to c

• Repeat Cluster assignment and Centroid update

steps until the cluster centroids converge.

3 EXPERIMENTS

In this section, we demonstrate the effectiveness of

the clustering algorithm with numerical results.

3.1 Ethical Approval

This study has been approved by the Institutional

Review Board of Yeditepe University Hospital (IRB

no.32). The subjects were informed of the experi-

ment protocol and an orientation has been given to

each subject.

3.2 Subjects

7 subjects (3 female and 4 male), whose ages are in

the range of 22-26, have participated in this study.

The biofeedback signal data of these 7 subjects have

been used for training the clustering algorithm. All

the subjects are healthy and have no background of

any diseases that might have affected the study. Total

task duration is about 30 min for each subject.

3.3 Results

Three different emotional states, which are classi-

ﬁed by plotting the ﬁrst two principal components,

are shown in Fig.4. We further highlight the data of

one particular subject who has reported her emotional

state in 6 different tasks for each run in 4 minute in-

tervals. The numericalresults showthat the emotional

states of a subject during the rehabilitation task is sep-

arable. We ﬁrst determined the emotional state by

Table 1: The number of training vectors where the emo-

tional states decided by the clustering algorithm match with

the subjects survey results

Accuracy

Subject 1 (Female) 82 %

Subject 2 (Male) 71 %

Subject 3 (Female) 83 %

Subject 4 (Female) 77 %

Subject 5 (Male) 84 %

Subject 6 (Male) 68 %

Subject 7 (Male) 81 %

Overall 78 %

Table 2: Percentage of training vectors under different emo-

tional states.

Stress Excitedness Boredom

Survey 12 % 33 % 52 %

Clustering 18.54 % 35.08 % 46.36 %

the clustering algorithm and compared this emotional

state with the emotional state reported by the subject

for each training sample (Table 1). The accuracy per

subject is then deﬁned as the number of training ex-

amples where the emotional states proposed by the

clustering algorithm match with the subjects’ own re-

port. Numerical results show that the clustering re-

sults and the subjects’ reports are consistent to each

other where the overall accuracy is around 80%. In

the survey case, we have assumed all training vec-

tors reﬂect the same emotional state as declared by

the subject during a 4 minute task. On the other hand,

subjects emotional state may change during the task

and in reality the subject may feel multiple emotional

states while performing the task and may declare only

one of them.

Finally, we compare the percentage of training

vectors under different emotional states obtained with

the clustering algorithm and the survey results. Table

2 shows that the ratio of solutions obtained with the

clustering algorithm for stress, excitement and bore-

dom are consistent with the survey results.

4 CONCLUSIONS

In this paper, we consider an unsupervised learning

problem where the task is to classify the emotional

states of a subject while in performing different difﬁ-

culty levels of a robot assisted rehabilitation task. We

have ﬁrst used dimensionality reduction using PCA to

reduce the dimension of each training example with

19 features to 2 dimensions to better visualize the

data. Then, we have used the K-means clustering al-

gorithm and decided the clusters associated with each

ClusteringofEmotionalStatesunderDifferentTaskDifficultyLevelsfortheRobot-assistedRehabilitation

system-RehabRoby

−6 −4 −2 0 2 4 6 8

−5

First Principal Component

Second Principal Component

Figure 4: Visualization of the biofeedback sensor data after dimensionality reduction and K-means clustering. Cyan dots

represent the boredom cluster, blue dots represent the excitedness cluster, and purple dots represent the stress cluster. One

particular subject (Subject 1) declared her emotional state as bored for the vectors represented with red squares, excited for

the vectors represented with blue squares and stressed for the vectors represented with black squares.

emotional state. Numerical results show that the clus-

ters offered by the clustering algorithm are consis-

tent with the survey results where subjects report their

own emotional state.

The proposed clustering algorithm provides a ba-

sis for the supervised learning problems for the robot-

assisted rehabilitation task. As a future work, we need

to determine the decision boundaries among clusters

reﬂecting different emotional states. Therefore, while

subjects perform the task, the learning algorithm will

be able to predict the emotional state of the subject,

and accordingly adjust the difﬁculty level of the task

in real-time. Furthermore, more comparative results

can be included when different learning algorithms

are used for classiﬁcation of emotional states. Addi-

tionally, it is also desirable to evaluate the accuracy of

the classiﬁcation algorithms for the disabled people.

ACKNOWLEDGEMENTS

The study was supported by the European Union Sev-

enth Framework Programme FP7/2007-2013 under

grant agreement n. 270396 (I-SUR). We gratefully

acknowledge the help of Prof. Serap Inal and Prof.

Sule Badilli Demirbas, who are in the Physiother-

apy and Rehabilitation Department of the Faculty of

Health Sciences Yeditepe University, for their valu-

able feedback concerning the task design. Addition-

ally, we acknowledge the help of Cenap Turan for his

help during the preparation of the surveys. We appre-

ciate the comments and suggestions provided by the

reviewers that signiﬁcantly improve the clarity and

content of this paper.

REFERENCES

Balasubramanian, S., Wei, R., Perez, M., Shepard, B.,

Koeneman, E., Koeneman, J., and He, J. (2008). Ru-

pert: an exoskeleton robot for assisting rehabilitation

of arm functions. In Virtual Rehabilitation, 2008,

pages 163–167. IEEE.

Bishop, C. M. (2006). Pattern Recognition and Machine

Learning, volume 1. Springer New York.

Bradley, M. M. and Lang, P. J. (1994). Measuring emotion:

the self-assessment manikin and the semantic differ-

ential. Journal of behavior therapy and experimental

psychiatry, 25(1):49–59.

Caldwell, D. G., Tsagarakis, N. G., Kousidou, S., Costa, N.,

and Sarakoglou, I. (2007). ” soft” exoskeletons for up-

per and lower body rehabilitationdesign, control and

testing. International Journal of Humanoid Robotics,

4(03):549–573.

Filipovic, S. R. and Andreassi, J. L. (2001). Psychophys-

iology: Human behavior and physiological response.

Journal of Psychophysiology, 15(3):210–212.

Guadagnoli, M. A. and Lee, T. D. (2004). Challenge point:

a framework for conceptualizing the effects of vari-

ous practice conditions in motor learning. Journal of

motor behavior, 36(2):212–224.

ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics

Gunes, H., Schuller, B., Pantic, M., and Cowie, R. (2011).

Emotion representation, analysis and synthesis in con-

tinuous space: A survey. In IEEE International Con-

ference on AutomaticFace & Gesture Recognition and

Workshops (FG 2011),, pages 827–834. IEEE.

Kandemir, M. (2013). Learning Mental States from Biosig-

nals. PhD thesis, Aalto University School of Science.

Kiguchi, K., Rahman, M. H., Sasaki, M., and Teramoto, K.

(2008). Development of a 3dof mobile exoskeleton

robot for human upper-limb motion assist. Robotics

and Autonomous Systems, 56(8):678–691.

Koenig, A., Novak, D., Omlin, X., Pulfer, M., Perreault,

E., Zimmerli, L., Mihelj, M., and Riener, R. (2011a).

Real-time closed-loop control of cognitive load in

neurological patients during robot-assisted gait train-

ing. IEEE Transactions on Neural Systems and Reha-

bilitation Engineering, 19(4):453–464.

Koenig, A., Omlin, X., Zimmerli, L., Sapa, M., Krewer, C.,

Bolliger, M., M¨uller, F., and Riener, R. (2011b). Psy-

chological state estimation from physiological record-

ings during robot-assisted gait rehabilitation. Journal

of Rehabilitation Research & Development, 48(4).

Krebs, H. I., Ferraro, M., Buerger, S. P., Newbery, M. J.,

Makiyama, A., Sandmann, M., Lynch, D., Volpe,

B. T., and Hogan, N. (2004). Rehabilitation robotics:

pilot trial of a spatial extension for mit-manus. Jour-

nal of NeuroEngineering and Rehabilitation, 1(1):5.

Loureiro, R., Amirabdollahian, F., Topping, M., Driessen,

B., and Harwin, W. (2003). Upper limb robot me-

diated stroke therapygentle/s approach. Autonomous

Robots, 15(1):35–51.

Lum, P. S., Burgar, C. G., Van der Loos, M., Shor, P. C.,

Majmundar, M., and Yap, R. (2006). Mime robotic

device for upper-limb neurorehabilitation in subacute

stroke subjects: A follow-up study. Journal of reha-

bilitation research & development, 43(5).

Maclean, N. and Pound, P. (2000). A critical review of the

concept of patient motivation in the literature on phys-

ical rehabilitation. Soc Sci Med, 50(4):495–506.

Mandryk, R. L. and Atkins, M. S. (2007). A fuzzy

physiological approach for continuously modeling

emotion during interaction with play technologies.

International Journal of Human-Computer Studies,

65(4):329–347.

Nef, T., Guidali, M., and Riener, R. (2009). Armin III–arm

therapy exoskeleton with an ergonomic shoulder actu-

ation. Applied Bionics and Biomechanics, 6(2):127–

142.

Ng, A. (2014). Machine Learning Lecture Notes.

http://coursera.org/ml.

Novak, D., Mihelj, M., and Munih, M. (2012). A survey of

methods for data fusion and system adaptation using

autonomic nervous system responses in physiological

computing. Interacting with Computers, 24(3):154–

172.

Ozkul, F. and Barkana, D. E. (2013). Upper-extremity reha-

bilitation robot rehabroby: Methodology, design, us-

ability and validation. International Journal of Ad-

vanced Robotic Systems.

Ozkul, F., Barkana, D. E., Demirbas, S. B., and Inal, S.

(2012). Evaluation of elbow joint proprioception with

rehabroby: a pilot study. Acta Orthop Traumatol Turc,

46(5):332–338.

Perry, J. C., Rosen, J., and Burns, S. (2007). Upper-limb

powered exoskeleton design. IEEE/ASME Transac-

tions on Mechatronics, 12(4):408–417.

Rahman, M. H., Saad, M., Kenne, J.-P., and Archambault,

P. S. (2009). Modeling and control of a 7dof exoskele-

ton robot for arm movements. In IEEE International

Conference on Robotics and Biomimetics (ROBIO),

pages 245–250. IEEE.

Rahman, T., Sample, W., Jayakumar, S., King, M. M., Wee,

J. Y., Seliktar, R., Alexander, M., Scavina, M., and

Clark, A. (2006). Passive exoskeletons for assisting

limb movement. Journal of rehabilitation research

and development, 43(5):583.

Rani, P., Sarkar, N., and Adams, J. (2007). Anxiety-

based affective communication for implicit human–

machine interaction. Advanced Engineering Informat-

ics, 21(3):323–334.

Rani, P., Sarkar, N., and Smith, C. A. (2003a). Affect-

sensitive human-robot cooperation-theory and experi-

ments. In Robotics and Automation, 2003. Proceed-

ings. ICRA’03. IEEE International Conference on,

volume 2, pages 2382–2387. IEEE.

Rani, P., Sarkar, N., Smith, C. A., and Adams, J. A. (2003b).

Affective communication for implicit human-machine

interaction. In Systems, Man and Cybernetics, 2003.

IEEE International Conference on, volume 5, pages

4896–4903. IEEE.

Ren, Y., Park, H.-S., and Zhang, L.-Q. (2009). Develop-

ing a whole-arm exoskeleton robot with hand opening

and closing mechanism for upper limb stroke rehabil-

itation. In IEEE International Conference on Rehabil-

itation Robotics, 2009. ICORR 2009., pages 761–765.

IEEE.

Russell, J. A. (1989). Measures of emotion. In R. Plutchik

and H. Kellerman (Eds.), Emotion: Theory, research,

and experience, 4:83–111.

Vertechy, R., Frisoli, A., Dettori, A., Solazzi, M., and

Bergamasco, M. (2009). Development of a new ex-

oskeleton for upper limb rehabilitation. In IEEE In-

ternational Conference on Rehabilitation Robotics,

ICORR 2009., pages 188–193. IEEE.

ClusteringofEmotionalStatesunderDifferentTaskDifficultyLevelsfortheRobot-assistedRehabilitation

system-RehabRoby