Physiological-based Dynamic Difficulty Adaptation in a Theragame

for Children with Cerebral Palsy

Adrien Verhulst

, Takehiko Yamaguchi

and Paul Richard

LARIS Laboratory, University of Angers, Angers, France

Harada Laboratory, Tokyo University of Science, Tokyo, Japan

Keywords: Dynamic Difficulty Adaptation, Physiological Signal, Emotion Recognition, Theragame, Cerebral Palsy.

Abstract: The purpose of this research is to provide a physiological-based Dynamic Difficulty Adaptation (DDA) for

rehabilitation of children with Cerebral Palsy (CP). In this paper, we present all the steps of the DDA devel-

opment by going through (1) the acquisition of physiological signals, (2) the extraction of the physiological

signals’ features, (3) the training of a learning classifier of physiological signals' features, and (4) the im-

plementation of the DDA in a game-based rehabilitation system. As a result, we successfully implement a

physiological-based DDA based on the user affective state (anxiety and boredom).

1 INTRODUCTION

Research and development in game-based stroke

rehabilitation (theragame) spreads increasingly. The

main advantages of the gaming approach are: (1) a

more effective rehabilitation through an increase of

the motivation (Hocine and Gouaich, 2011), (2) an

access to quantified data, (3) the possibility to repeat

and to adapt the exercises, and (4) the possibility to

scale the rehabilitation (McCue et al., 2010).

Children are most likely to lose motivation in

front of a repetitive exercise than adults (Liebert et

al., 2006). Since the repetition is necessary in the

rehabilitation process, it is important to keep them

motivated. A way to do it is by keeping them in the

flow state (Huang et al., 2010).

The flow (refer Figs. 1 and 2) is “a state in which

people are so involved in their activity that nothing

else seems to matter; the experience itself is so

enjoyable that people will do it even at great cost,

for the sheer sake of doing it” (Csikszentmihalyi,

1990).

The flow has been used in several researches to

measure the enjoyment of a user in interactive

applications. The idea behind the flow theory can be

summarized as follows: (1) the user is more

susceptible to be in the flow state when the

challenge equals the skill, (2) if the challenge is less

important than the skill, the user is bored (too easy

for him/her), and (3) if the challenge is more

important than the skill, the user is frustrated (too

hard for him/her).

There are several reasons of why a simple

selection of the level of difficulty (easy-medium-

hard), at the beginning of the application, may not be

sufficient: (1) there is a limited difficulty variation,

(2) there is a difficulty gap between levels, (3) it is

not enough responsive to player learning, (4) it is

time-consuming to implement those variations and

(5), the user has to «guess» his/her level regarding

the levels of difficulty. The DDA’s main advantage

over non-dynamic difficulty adaptation is the

possibility to automatically adapt itself from the user

actual performance.

In this context, we are developing the Children

Rehabilitation Project (CRP) which is based on a

new modular adaptive system dedicated to

rehabilitation of children having CP. The system has

been designed as a framework permitting to easily

implement rehabilitation applications. The first

theragame developed in the context of the CRP is

the Rehab-Island theragame (Fig. 3). In this

application, the user controls the avatar through a

Kinect

, and must touch objects by moving his/her

left arm. The application parameters (such as

objects’ velocity, size, etc.) are decided by the

therapist. In the next section, we survey the state of

the art concerning Dynamic Difficulty Adaptation

(DDA). In section 3, we present the proposed

approach. Section 4 is dedicated to the software

implementation of our system. In section 5 we

164

Verhulst A., Yamaguchi T. and Richard P..

Physiological-based Dynamic Difﬁculty Adaptation in a Theragame for Children with Cerebral Palsy.

DOI: 10.5220/0005271501640171

In Proceedings of the 2nd International Conference on Physiological Computing Systems (PhyCS-2015), pages 164-171

ISBN: 978-989-758-085-7

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

describe the feature extraction protocol. Section 6

presents the training of the classifier. Section 7

presents the implementation of the DDA. Section 8

provides a short discussion of our method. The paper

ends by a conclusion and gives some tracks for

future works.

2 STATE OF THE ART

Cerebral Palsy (CP) refers to various motor impair-

ments caused by damage to the motor control center

of the developing brain, and can occurs during preg-

nancy, during childbirth, or after birth up to about

age three (Aisen et al., 2011). Children having im-

paired motor movements after CP need physical

therapy, which can be enhance by rehabilitation

applications (Parsons et al., 2009).

There are only a few papers presenting a

Dynamic Difficulty Adaptation (DDA) for stroke or

CP rehabilitation systems. Moreover, in all those

papers, DDA is only based on user’s performance

(mainly score). For example, Parnandi et al., (2013)

proposed an approach based on control theory’s

principles. They used variation of actual and desired

arousal of the user as the variation error to minimize.

Huang et al., (2010), following the work of Hao

et al., (2010); and Li et al., (2010), presented Real

Time Computational Intelligence (RTCI) and

Adaptive ANN Computational Intelligence

(ANNCI) approaches for NPC, both being based on

Monte-Carlo Tree Search (MCTS).

Wong (2008) also presented an ANNCI

approach, but not based on MCTS. Hocine and

Gouaich (2011) described an approach based on

prior assessments of the capability of the user. The

adaptation is done through an ability zone, which

contains information on the difficulty to do a given

task at given position. They conducted an

experiment on 8 subjects through a reaching-task

application.

Gouaïch et al., (2012), following the work of

Hocine, proposed a digital pheromone approach

based on the ant algorithm introduced by Dorigo and

Stützle (2004). The adaptation is done through an

ability zone updated regarding user’s performance.

They conducted an experiment on 10 subjects

through a reaching-task application.

Arulraj (2010) proposed a differential learning

approach for NPC. The agent learning-rate reduces

with time, while being impacted by user’s

performance. The approach feasibility has been

tested using the Minigate game. Andrade and

Ramalho (2005) and Tan et al., (2011) both proposed

a Reinforcement Learning (RL) approach for NPC.

G. Andrade implemented it by using the Q-learning

algorithm, the adaptation being done by choosing

the action-value which fit the level of the user. C.

Tan implemented it by using Adaptive Uni-

Chromosome Controller (AUC) and Adaptive Duo-

Chromosome Controller (ADC) algorithms, the

adaptation being done by activating controller’s

behaviour which fit the level of the user.

Figure 1: Out-dated version of the flow theory. Researches

on affect-based DDA mainly use it, since it is easier to

apply than the updated version of the flow theory.

Figure 2: Updated version of the flow theory. apathy,

anxiety and boredom are still used.

Regarding physiological based DDA, the DDA is

usually handled through the flow theory, where the

emotional states captured orient the DDA to move

the challenge to emotional state.

Liu (2009), showed that a player's affective state

can be deciphered from his physiological state

during gaming, and that a DDA can be based on the

affective state. They conducted an experiment on 9

subjects, and investigate several classification

algorithms to test the affective models (Regression

Tree, K Nearest Neighbours, Bayesian Network

Physiological-basedDynamicDifficultyAdaptationinaTheragameforChildrenwithCerebralPalsy

165

Technique, and Support Vector Machines (SVM)).

They showed that the majority of the participants

improved during the affect-based DDA session, and

that they perceived the game more challenging with

an affect-based DDA than a performance based

DDA.

Figure 3: Illustration of the Rehab-Island application. The

avatar's arms follow the user's arms through the Kinect

Chanel and Rebetez (2011) developed a DDA to

maintain player in flow through an affective model

based on EEG signals. They showed that the

affective state was changing when playing Tetris.

They then trained a classifier to recognize the

affective state in order to adapt Tetris.

Guillaume et al., (2012) proposed a method to

implement an affect-DDA in games. They recorded

the player's physiological signals and then extracted

the signals' features when the player was (1) playing

the game and was (2) in a given affective state (such

as boredom). They then trained a learning classifier

to recognize the affective state from the

physiological features. Finally, they used the

learning classifier to know the affective state of the

player (boredom or not) and use that information to

adapt the game in real-time.

Parnandi and Ahmed (2014) presented Chill-Out,

an adaptive biofeedback game monitoring the

breathing rate of the player and adapting the game in

a way that encourages relaxing behaviours. They

showed that such an adaptation led to improved

performances.

3 APPROACH

We decided to use a simple approach to do our

DDA. Based on the user's physiological state a

learning classifier identifies the user's affective state.

This affective state is then passed to our DDA algo-

rithm (refer Figs. 4 and 5) which adapts accordingly

the level of difficulty of the theragame. This ap-

proach allows us to use the flow theory, which de-

pends of the user's affective state.

The flow theory uses several affective states. The

out-dated flow theory use anxiety, boredom and

flow, while the updated flow theory uses apathy,

worry, anxiety, boredom, arousal, relaxation, control

and flow. In our work, we decided to use only

anxiety and boredom, since (1) they are very

relevant to the flow theory (already present in the

out-dated version), (2) they have already been used

in physiological-based DDA (Liu et al., 2009),

(Giakoumis et al., ).

In previous works anxiety and boredom have

been already classified using (1) ICG, BVP, EMG,

ECG, EDA (or GSR) and heart sound features (Liu

et al., 2009); (2) ECG and EDA (or GSR) features

for boredom (Giakoumis et al., 2011).

Since anxiety is closely related to the stress

affective state, which can be detected from ECG and

EDA features, we decided to only use ECG and

EDA. It allow us (1) to have less constraint in our

material choice (since it is then supposed to be used

with children having CP in rehabilitation, we had to

limit the captors), (2) to use the same signals for

both the anxiety and boredom.

In summary, from the choice of detecting anxiety

and boredom, we decided to use ECG and EDA

signals. We decided to use BITalino

in order to

record them. Concerning the features choice and

extraction, we based our work on the previous work

done by Lieu C. et al and Giakoumis D. et al. We

selected 14 different features, based on ECG and

EDA signals (see table 1).

Table 1: Features extracted from EDA and ECG.

Signals Features

EDA

EDA Mean;

EDA SD;

EDA 1 derivation average;

EDA 1 derivation RMS;

EDA SCR;

EDA 1 difference raw data;

EDA 1 difference normalized data;

EDA 2 difference normalized data.

ECG

IBI RMSSD;

IBI Pnn50;

IBI 1 difference raw data;

IBI 1 difference normalized data.

We then used those features to train a SVM.

While others learning classifier (such as Regression

Tree) may give better results, we felt more confident

about the SVM, which usually gives good result in

emotion classification (Yoo et al., 2007).

We decided to train 2 SVM, one trained to

recognized anxiety, and the other one to recognized

PhyCS2015-2ndInternationalConferenceonPhysiologicalComputingSystems

166

boredom. Since, to the best of our knowledge, no

method exists to recognize flow, we thought that

detecting the absence of anxiety or the absence of

boredom would indicate that the user was in flow.

We decided that if there was no anxiety and

boredom detected, and that if the level of difficulty

was above or equal to 10%, the user was in flow. If

there was no anxiety and no boredom detected, and

that if the level of difficulty was below 10%, the

user was in apathy. If the user was in flow, then the

difficulty will not change. If the user was in apathy,

then the difficulty will increase (to stimulate the

user). If there was anxiety and no boredom, then the

difficulty will decrease. If there was no anxiety and

boredom, then the difficulty will increase. If there

was anxiety and boredom, then the difficulty will not

change, since it may be an error in the signals

reading. The figures 4 and 5 summarize the idea.

The DDA algorithm can be explained in those

few lines: the application change the difficulty until

anxiety/boredom is not detected and the difficulty is

above a given threshold.

4 IMPLEMENTATION

We used the BITalino Board Kit

, a low cost de-

vice able to record EMG, ECG, EDA (Guerreiro and

Martins, 2013) in order to record ECG and EDA.

While fairly new, it is already world-wide used.We

decided to use it since it was: (1) Cheap, others ma-

terials commonly used in research (such as Pro-

comp5) are above the thousands of €; (2) Described

in the literature; (3) Delivered with a SDK and (4)

Wearable.

Figure 4: When anxiety is detected, the level of difficulty

decreases over time, until it reaches another affective state.

We developed a BITalino C# SDK (.NET 2.0)

based on the BITalino Java SDK in order to

implement BITalino in Unity3D (v4.5x).

By doing so, we successfully implemented

BITalino in Unity3D (see Fig. 6). It allowed us to

record in real-time ECG and EDA signals.

We also developed several Unity scripts to allow

us to set the BITalinoTM parameters from the

application (such as the frequency). The BITalino C#

SDK as well as our Unity's scripts is available online

in the API section of the BITalino official website

Figure 5: If the level of difficulty is to low (< 10%), we

consider that the user is in the apathy state.

Figure 6: BITalino

working inside Unity3D. We can see

in red the ECG signal. In the left panel, we can select

which signals to display.

5 FEATURES EXTRACTION

BITalino

did not provide software to extract phys-

iological features. Therefore we had to implement a

feature extraction system ourselves. While the calcu-

lations were fairly easy to do, we decided to do it

using a MATLAB C-shared DLL, to be able to use

some built-in MATLAB functions. We did it in order

to be able to easily change and adapt the algorithms.

Physiological-basedDynamicDifficultyAdaptationinaTheragameforChildrenwithCerebralPalsy

167

It could however also have been done directly in C#

through mathematical libraries. We developed sever-

al MATLAB scripts able to extract the features we

were interested in (see table 1). We then exported the

scripts through a C-shared DLL, to be able to call

them from Unity3D.

Concerning the features extraction calculation,

we followed the work previously done by

Giakoumis, D. et al. The scripts were able, from the

ECG and EDA raw data, to automatically extract the

features of the table 1. The raw data were sampled

(at 255 Hz, from the 1000 Hz raw data), then

normalized and smoothed with a local regression

filter.

The ECG signal was measured using three

electrodes placed in lead II (negative electrode on

the right arm, positive electrode on the left leg). This

placement allowed the subjects' arms to move freely.

The EDA signal was measured using two electrodes

in bipolar placement on the left (or right) hand

(index and middle finger). Even if it was as less

restrictive as possible, this placement requires the

subjects' left (or right) arm to not move. By doing

so, we successfully managed to extract the features

within Unity3D, from the signals recorded by

BITalino, in real-time.

6 CLASSIFIER TRAINING

We integrated LibSVM (Chang and Lin, 2013) in

Unity3D; through one of its open-source implemen-

tation available in C# (We slightly modified it to be

able to use it in .NET 2.0). To train the SVM, we

developed an application (ProvokeAffect) able to put

the user in boredom affective state and anxiety affec-

tive state.

We developed three scenes in ProvokeAffect: (1)

boredom scene, (2) anxiety scene, (3) excitement

scene. Each of those scenes follows the same rules:

A message displays the name of a colour (Red; Blue;

Green). The subject have to click (one time) on the

sphere having the named colour. After few seconds

(or milliseconds) the colour’s name changes. If the

subject clicks on the sphere before the time limit, he

gains 1 point. If he does not click on the sphere

before the time limit, or if he misses, he loses 1

point. The differences between each scene depend of

the: (1) overall difficulty, (2) environment (sounds,

background). To provoke boredom, the interval of

time between each message change is set to 4

seconds, without difference between (1) colour’s

spheres, (2) colour’s message. Moreover, the

sphere’s number is 2.

To provoke anxiety, the interval of time between

each message changes is set to 0.85 seconds, with a

difference between (1) colour’s spheres, (2) colour’s

message. Moreover, the sphere’s number is 3. In top

of that, the anxiety scene has a score-DDA to

increase difficulty if the user is doing too well. The

DDA (1) changes the interval of time between each

message (minimum: 0.5s; maximum: 0.85s), (2)

changes the shape of the sphere, (3) changes the

background colour, (4) produces heart-beat sound.

Figure 7: Anxiety scene. We can see that the message is

coloured in a different way from what is intuitively sup-

posed to be (red instead of green).

To provoke excitement, the interval of time

between each message changes is set to 1.5 seconds,

with a difference between (1) colour’s spheres, (2)

colour’s message. Moreover, the sphere’s number is

3. In top of that, the excitement scene has a score-

DDA to increase difficulty if the user is doing too

well. The DDA changes the interval of time between

each message (minimum: 0.75s; maximum: 1.5s).

Finally, it is possible to do several series of combo.

Successful combo produces (1) score increase, (2)

sounds effects. Note that we did not use the

excitement scene to train a SVM dedicated to

excitement (the preliminary results did not show a

great difference between anxiety and excitement, so

we preferred to not use it), unlike the anxiety and

boredom scene.

Having the ProvokeAffect application, we did a

preliminary test to see if there were noticeable

differences between anxiety and boredom scene

when looking at the physiological signals.

We can see that, even without looking at the

features, there are important differences between

“EDA boredom” and “EDA anxiety” signals (see

Figs 8 and 9). Concerning the experimentation

relative to the training of the learning classifier, we

used all three scenes of ProvokeAffect. All the

experimentation (including the Likert-scale test) was

automatized, in order to not induce change(s) in the

affective state of the user. The experiment was 15

minutes long. After an explanation of the rules

PhyCS2015-2ndInternationalConferenceonPhysiologicalComputingSystems

168

(automated), the user had to pass 3 trials of the

boredom scene, then 3 trials of the anxiety scene,

then 3 trials of the excitement scene. The change

between the scenes was automated. Between each

trials, there was a one minute during which the

subject had to fill a 5 Likert-scale self-report (still

automated).

Figure 8: EDA when doing the Boredom scene of Pro-

vokeAffect.

The Likert-scale comported several non-relevant

questions to our experiment in order to not orient the

subject in his answers. As relevant questions, we

asked if he was bored (1-5) or if he was anxious (1-

5). After the experiment, we labelled the results as

anxiety / boredom if the user answered 4 or 5 and

not anxiety / boredom if it was 1 or 2. If it was 3,

then we did not include the result. Finally, we had 10

subjects who did the experiment (and therefore a

total 90 trials). From those labelled features, we

trained the two learning classifier (one for anxiety

recognition and one for boredom recognition).

Figure 9: EDA when doing the Anxiety scene. We can

clearly see that there are a lot more of SCR than with the

Boredom scene of ProvokeAffect.

By training our SVMs (after grid-search) we

obtained a cross-validation result up to 75% for both

anxiety and boredom SVM. While this result is

below other studies (probably because there was not

enough trials), it was enough to continue our work

on our affect-DDA.

7 IMPLEMENTATION OF THE

DDA

Since the DDA is supposed to be used by different

theragames, it was important to decouple it from the

theragames into a DDA module. The DDA module

comports the reading of physiological signals, the

features extraction, the SVM classifiers and the

implementation of the DDA algorithm. The DDA

module gives to the theragames a difficulty level

between 0-100, which may change over time (unless

the user is in flow). 0% correspond to the easiest

difficulty while 100% correspond to the hardest

difficulty.

It is then up to the developer of each theragame,

according to a value of difficulty in percentage

(which may change over time) to implement the

“meaning” of such difficulty. For example, in our

Rehab-Island theragame, we observed that the

difficulty was mainly dependant of 4 main different

variables. We gave to each of those variables a

minimum and maximum value, which follow a

linear, exponential or logarithmic curve. By doing

so, when the difficulty is updated, each of our

difficulty variables is automatically updated.

However, this is a pretty straightforward example.

There are others ways to implement the difficulty,

which depend mainly of the kind of game

developed.

8 DISCUSSION

By decoupling our DDA from our theragame, we

successfully managed to implement it in the CRP. In

summary, we implemented a physiological-based

DDA, which, through the affective-state (boredom

and anxiety states) of the user, adapts the difficulty.

At the best of our knowledge, there is non-yet a

physiological-based DDA using boredom and anxie-

ty state together. There is non-yet either a physiolog-

ical-based DDA in a theragame.

Even if our work present a working solution to a

physiological-based DDA implementation in an

adaptive interface, its purpose was to be the base of

researches in rehabilitation using DDA. We would

Physiological-basedDynamicDifficultyAdaptationinaTheragameforChildrenwithCerebralPalsy

169

like to see if a physiological-based DDA is more

effective than score-based DDA in CP's

rehabilitation, and in what ways it is more effective.

We would also like to try others devices than

BITalino (such as the Kinect 2, able to track the

blood flow of the user). Eventually, we would also

like to improve the existing system, by training a

more accurate classifier (Liu et al., suggested to use

a Regression Tree classifier).

9 CONCLUSION AND FUTURE

WORK

We successfully managed to implement a philologi-

cal-based DDA in an adaptive interface (the CRP).

This DDA uses the EDA and ECG signals of the

user. From those signals, it extract its important

features (such as GSR and HR), in order to accurate-

ly classify user's affective state. Using 2 trained

SVM (one for boredom affect state, one for anxiety

affect state), the DDA is able to know if the user is

bored, anxious or likely to be in the flow affect state,

and therefore to increase, decrease, or not change the

difficulty. By providing such a DDA, we would like

to propose a more effective DDA in order (1) to

improve rehabilitation, (2), to allow the patient to be

less dependent of the patrician.

In our future study, we will see if the

physiological-based DDA presented in this work is

more effective than a straightforward score-based

DDA in CP's rehabilitation. To do so, it will be

necessary to retrain the SVMs, since the SVMs data

in our experiment came from adults, and CP

rehabilitation is for children. During the training, it

will be also necessary to propose a “boring”

application and an “anxious” application which ask

the user to perform a physical effort on his upper-

body. Luckily enough, Rehab-Island seem adapted

to propose such applications.

REFERENCES

Aisen, M. L. et al., 2011. Cerebral palsy : clinical care and

neurological rehabilitation. The Lancet Neurology,

10(9), pp.844–852.

Andrade, G., Ramalho, G., and Coruble, V., 2005. Chal-

lenge-sensitive action selection: an application to

game balancing, International Conference on Intelli-

gent Agent Technology, pp. 194-200.

Arulraj, J. J. P., 2010. Adaptive agent generation using

machine learning for dynamic difficulty adjustment.

International Conference on Computer and Communi

cation Technology (ICCCT), pp. 746 - 751.

Chanel, G. & Rebetez, C., 2011. Emotion assessment from

physiological signals for adaptation of game difficulty.

Systems, Man and Cybernetics, 41(6), pp.1052–1063.

Chang, C. & Lin, C., 2011. LIBSVM : A Library for Sup-

port Vector Machines. ACM Trans. Intelligent Systems

and Technology, (2), pp. 1-27.

Csikszentmihalyi, M., 1990. Flow : The Psychology of

Optimal Experience, Harper and Row, New York.

Dorigo, M. & Stützle, T., 2004. Ant Colony Optimization,

Giakoumis, D. et al., 2011. Automatic Recognition of

Boredom in Video Games Using Novel Biosignals

Moment-Based Features. IEEE Transactions on Affec-

tive Computing, 2(3), pp.119–133.

Gouaïch, A. et al., 2012. Digital-pheromone based diffi-

culty adaptation in post-stroke therapeutic games. Pro-

ceedings of the 2

ACM SIGHIT symposium on In-

ternational health informatics – IHI’12, p.5.

Guerreiro, J., Martins, R., Silva, H., Lourenço, A., and

Fred A., 2013, “BITalino: A Multimodal Platform for

Physiological Computing” International Conference

on Informatics in Control, Automation and Robotics

(ICINCO) (1), pp.500-506.

Guillaume, C., Konstantina, K. & Thierry, P., 2012. Ga-

mEMO: how physiological signals show your emo-

tions and enhance your game experience. Proceedings

of the 14

ACM, pp. 1–3.

Hao, Y. H. Y. et al., 2010. Dynamic Difficulty Adjustment

of Game AI by MCTS for the game Pac-Man. Sixth

International Conference on Natural Computation

(ICNC), 8, pp. 3918-3922.

Hocine, N. & Gouaich, A., 2011. Motivation based diffi-

culty adaptation for therapeutic games. Serious

Games.

Huang, W. et al., 2010. Dynamic difficulty adjustment

realization based on adaptive neuro-controlled game

opponent. Intelligence (IWACI), 2010.

Li, X. et al., 2010. To create DDA by the approach of

ANN from UCT-created data. International Confer-

ence on Computer Application and System Modeling

(ICCASM), pp.475–478.

Liebert, M. A. et al., 2006. Feasibility, Motivation, and

Selective Motor Control : Virtual Reality Compared to

Conventional Home Exercise in Children with Cere-

bral Palsy., 9(2), pp.123–129.

Liu, C. et al., 2009. Dynamic Difficulty Adjustment in

Computer Games Through Real-Time Anxiety-Based

Affective Feedback. International Journal of Human-

Computer Interaction, 25(6), pp. 506–529.

McCue, M., Fairman, A. & Pramuka, M., 2010. Enhancing

quality of life through telerehabilitation. Physical med-

icine and rehabilitation clinics of North America,

21(1), pp.195–205.

Parnandi, A. & Ahmed, B., 2014. Chill-Out: Relaxation

Training through Respiratory Biofeedback in a Mobile

Casual Game. Mobile Computing, pp.252–260.

Parnandi, A., Son, Y. and Gutierrez-Osuna, R., 2013. A

Control-Theoretic Approach to Adaptive Physiological

Games. Humaine Association Conference on Affective

Computing and Intelligent Interaction, pp.7–12.

PhyCS2015-2ndInternationalConferenceonPhysiologicalComputingSystems

170

Parsons, T. D. et al., 2009. Virtual reality in paediatric

rehabilitation : A review. , 12(August), pp.224–238.

Tan, C., Tan, K. & Tay, A., 2011. Dynamic game difficulty

scaling using adaptive behavior-based AI. IEEE

Transactions on AI in Games, 3(4), pp.289–301.

Wong, K., 2008. Adaptive computer game system using

artificial neural networks. Neural Information Pro-

cessing, pp.675–682.

Yoo, S., Lee, C. & Park, Y., 2007. Determination of Bio-

logical Signal for Emotion Identification. World Con-

gress on Medical Physics, 14, pp.4047–4049.

Physiological-basedDynamicDifficultyAdaptationinaTheragameforChildrenwithCerebralPalsy

171