User Models for Post-stroke Therapies and Consequences for the

Interaction with a Humanoid Robot

Alexandru Bundea

and Peter Forbrig

Universität Rostock, Fakultät für Informatik, Lehrstuhl für Softwaretechnik, Albert-Einstein-Str. 22, 18059, Germany

Keywords: Social Robot, Human-Robot-Interaction, User Model, Therapy.

Abstract: In the future, social robots may become a viable help to provide more patient therapy sessions for post-stroke

patients. Certain therapies require still another person to work with the patient. If a non-medically skilled

helper person could be used, this could dampen the lack of available therapists worldwide. But for this,

strategies how robots may best advise a patient and the helper must be found. This paper aims to find a

common basis, where to formulate rules how a social robot could react and act in a therapy session. We

present the focused therapy and propose our model and give an exemplary idea how to use this framework.

1 INTRODUCTION

A worldwide shortage of medical personnel in all

areas is forcing medicine to innovate. In order to cope

with the increasing number of stroke patients,

research is being carried out in some places on how

robots can help patients in neurorehabilitation

therapy. Further subdivided, there is the possibility of

doing this with a physical aid, in which a mechanical

system moves the limbs by a robot and socially

assistive robots (SAR). Their task is, among other

things, to accompany the patient through verbal

instructions and feedback through a therapy session.

The main goal of the robot is to let patients perform

their training exercises in a good way and to keep the

patient motivated throughout the therapy.

An important aspect is a continuously verbal

feedback during therapy. This was observed during

sessions with a human therapist, but becomes even

more important in the case of SAR as the robot cannot

provide physical assistance and body contact to the

patient. In addition, an intelligent recognition of

parameters must already be available for the feedback

in order to optimally adapt the feedback with regard

to the current therapy situation. While the recognition

of a specific therapy session situation represents a

multimodal recognition challenge with the use of, for

example, several sensors like cameras, sensors for

https://orcid.org/0000-0001-8315-3405

https://orcid.org/0000-0003-3427-0909

sound or devices analyzing the interaction of a

patient.

Based on the sensed situation actions have to be

specified that should be activated. Rules can be used

for this purpose. They can e.g. specify when the robot

should say something to a patient. This aspect of rules

will become a topic that we want to discuss in this

paper. We will focus on cases with 2 people present,

the patient and a helper. Because of the lack of

professional therapists, a helper that might come from

the family will support the patient.

Most social robot therapies deal with robot-to-

patient cases. This works in cases in which the patient

can carry out the therapy alone with the robot. In

certain forms of stroke therapy, however, the main

feature is that the severely handicapped patient's

limbs have to be moved by a therapist. Here now adds

an idea, to substitute this therapist, with a non-

medically skilled “helper”, to move the limbs of the

patient. For this approach, the robot would then not

only have to instruct and motivate a patient, but also

this other person.

In order to carry out a suitable spontaneous,

helpful feedback strategy, one has to define rules of

what to do. Here we look into finding a common basis

and a structure, on where rules could be applied to.

Thereby we aim to incorporate a way, how we can use

a user model and the personality of the therapy

participants to consider how the robot provides his

Bundea, A. and Forbrig, P.

User Models for Post-stroke Therapies and Consequences for the Interaction with a Humanoid Robot.

DOI: 10.5220/0010722100003060

In Proceedings of the 5th International Conference on Computer-Human Interaction Research and Applications (CHIRA 2021), pages 283-290

ISBN: 978-989-758-538-8; ISSN: 2184-3244

283

feedback. The corresponding user models were based

on interviews using the Repertory Grid Technique

(George A. Kelly, 2001).

The further structure of this paper first shows the

related works and then the stroke therapy considered

here and then presents a model of when feedback

from the robot is appropriate. This is followed by a

discussion and a summary.

2 RELATED WORKS

Winkle researched in (Katie Winkle, 2020) how a

SAR can be used as a treadmill running instructor and

how participants would accept it. She uses an input

space of 20 features with different dimensions to

monitor a runner approach. It includes features like

heart rate, speed, activity level and also psychological

traits combined in the Big 5 model.

This is one of the commonly used model to assess

the personality of patients as shown by Dwan et al.

(Toni Dwan and Tamara Ownsworth, 2019). Coming

from the field of psychology, the five personality

factors are “Neuroticism”, “Extraversion”,”

Openness to experience”, “Agreeableness”,

“Conscientiousness”. A person’s factors will be

determined through questionnaires.

The Robot has many different actions as a coach

for a treadmill like reminding about time or how to

correct the pose of the runner. Winkle uses an

interactive machine learning approach, whereby a

trainer enters during a live-session with a runner new

state-action pairs into the system. The robot then

improves his next autonomous action from the

previously gathered examples from the trainer.

Casas et al. (Jonathan A. Casas, et al., 2019) use a

NAO Robot in a therapy setup to increase cardiac

function. Patients run on a treadmill and the nearby a

robot provides verbal and gestural feedback. Inside

their patient-robot interaction they monitor the level

of exertion with the “Borg scale” (Bahar Irfan, et al.,

2020) and the patient’s heart rate. The Borg scale is a

method to quantify perceived exertion and aims to

make it comparable between individuals. Even with

its subjective nature, this method proved to have a

good correlation with the level of more measurable

effort metrics in diverse application.

Additionally, they watch several cases with these

two metrics and switch between different interaction

modes. Depending on a “high” or “low” exertion

and/or heart rate they ask the patient if he is feeling

alright. For cases whereby, the exceeds the heart rate

exceeds for too long the normal values, the robot will

e.g. be saying: “Doctor, the patient needs help.”. This

interaction mode then can only be switched off, when

the medical staff touches the head of the robot.

A lately, different approach has been done by

Irfan et al. (Bahar Irfan, et al., 2020) by measurements

with an ECG, Borg Scale entered thorugh the patient

via a tablet, a laser range finder and an inertial

measurement unit and a tablet camera. Next to

predefined phrases at certain points in the session, and

alerting the patient and medical staff about a high

exertion or Heart rate, the robot will provide prompts

if the patient should improve his cervical posture.

A rather different perspective to look on the

provision of feedback to the patients comes directly

from medical literature. Bachelor et al. (Alexandra

Bachelor, 1991) state, that the success of the therapy

outcome and the “alliance” between patient and

therapist is that the patients perception yields a better

prediction of the success than the therapist

perception. Furthermore, from a patient’s perspective

therapist-provided aid, warmth, caring, emotional

involvement are factors which truly helped in

therapies.

Duncan et al. (Barry L. Duncan, et al., 1994) adds

to these findings and states, that in the “alliance”

literature, the therapy progresses the most, if the

therapist talks about what the patients see as

important for them. Additionally the therapist should

have chosen methods, that the patient will think, that

it helps him to reach the goal specified by the patient.

Because of these finding, we would primarily

target the patient’s needs and we later present robot

actions, which we try to tailor to the findings of

Bachelor and Duncan et al.

3 ARM BASIS TRAINING (ABT)

3.1 Classic ABT with a Therapist

ABT has been designed for severe arm impairment.

The focus lies on improving the patient’s capabilities

for selected movements of the patients’ arms as

displayed in figure 1. The ABT will be done in a

structured repetitive training manner with each

session to complete a set of arm and hand exercises

(Thomas Platz, Bernhard Elsner, and Jan Mehrholz.

2015). The therapy’s’ movements are starting with

joint movements “without” the factor of gravity (step

1), meaning that the therapist is holding the extremity

of the patient up and manually assists the movement

of the patients arm, hand or finger during the exercise.

The next step -after the patient acquires the full range

of motion of the movement- is to let the patient do the

exercises with gravity (step 2), whereby the patient is

Humanoid 2021 - Special Session on Interaction with Humanoid Robots

284

supposed to do single joint movements, while the

therapist is still holding the hand or arm, letting the

patient focus only on selected joint movements of the

rest of his arm. Ideally, the patient performs multiple-

joint-movements on his own (step 3). Until then, the

supporting therapist needs to hold the weight of the

arm and to ensure that the patient does not need a

postural control by himself during the exercise.

There is no specific performance-related feedback

from the therapist to the patient predefined, so the

therapist may engage in feedback as he wants.

In the first session of a patients’ therapy, a human

therapist will do an intensive assessment and

introduce him to the therapy. Only after this, the next

therapy session will include the robot.

Figure 1: Three Arm Basis Training exercises with a human

therapist. The start- and endpoints for each movement are

here displayed (Thomas Platz, 2019).

3.2 ABT with Helper and Robot

This will be the later setup for our robot study and at

first adds only the robot to the therapy session. Of

course, there needs to be more technical measurement

systems like cameras to recognize the movements, the

emotional state, body positions and verbal

communication.

Figure 2: Exemplary execution of the ABT therapy (non-

medical envirionment).

It is the idea to check with sensors whether the

execution of the exercises is correct. This means e.g.

the speed and the extend of the movement.

Figure 3 gives an impression of the models that

our analysis software provides. The (upper) image

shows what the system calculates with the angles of

the limbs. In this one the patient is supposed to do an

arm curl. Thereby we analyse the degree of the angle

and this may come into consideration of what we

would express as feedback e.g. “Move your hand

closer to your shoulder.”

The lower image displays the situation with a

helper from a different perspective. We will provide

similar hints

Figure 3: Examples of the visualisation of sensor tracking

of arm movements.

Additionally, it is intended to identify the

emotions of the patient and the helper. The humanoid

robot Pepper provides a quite good solution based on

facial expression and voice that works quite well for

one person. We did not have a solution for two

persons yet. However, in this paper we want to focus

on a different aspect, the rule model.

User Models for Post-stroke Therapies and Consequences for the Interaction with a Humanoid Robot

285

4 PROPOSED RULE MODEL

FOR

4.1 Basic Patient/Helper Model

Starting with the basic form of one robot to one

patient, we also use an exertion metric with binary

values, in this case “low” and “high” engagement. For

emotional states we use “neutral”, “unhappy” and

“happy.

Table 1: Basic state table.

Emotion

High

ement

Low

ement

Unha

ppy

Neutral

Happy

Table 1 presents 6 different states of a patient, to

which a robot could react and provide spontaneous

feedback. Note that this is only a general framework

for such an arm rehabilitation scenario. More specific

feedback for arm posture is not the main focus of this

model.

Currently with this table alone, one would only

describe the person’s state. We now need to introduce

the robot’s possible actions. After this, we can

combine both the patient/helper state with the

possible robot actions to design a “grid”. With this

grid, we have one simple way how to define rules for

the interaction. So, in the next subsection we will

focus on a model for the action space of the robot.

4.2 Robot Actions

From the robot side we need to model the possible

robot actions according to what the robot should

achieve. The selection of goals and actions, depicted

in table 1 are only a subset of what could be perceived

as useful. These actions can be done in different ways,

such as “positive”, e.g. to generate a good, confirming

way, so that the patient gets confirmed, that what he

already does, is good. Otherwise a “challenging” style

e.g. tries to tell the person, that he is already ok, but

he can do it even better. Styles will be more

highlighted in a moment.

For our work here, we identify four important

feedbacks for the patient.

“Provide task-related feedback” (PTR) is directly

linked how the patient performs his task. “Provide

social Feedback” (PSF) aims to motivate the person.

Then to make a conversation a bit more memorisable,

the robot may “provide small talk” (PST), this is

mainly for cases, to improve a therapy scenario,

where no one talks and the atmosphere between the

two persons may become too uncomfortable. Finally,

the robot may ask a person, e.g. to enter how the

session is going on today or the current exertion level,

we call this action “Request information” (RI). This

is the only robot action, where we would not include

a positive or challenging style, as the person should

only focus on entering data as directly as possible.

We intent to use the systems decision making to

determine, if a certain action/feedback should be

provided by the robot. If an action has been triggered,

the style, “how the robot” should formulate his verbal

prompt is determined by the psychological profile of

this person.

Table 2: Social robot action space.

Action goal Action Style Example

Improve

Performance

PTR

“You do well, but you

should move the arm

higher up in each

repetition.”

N “move the arm higher”

“Dear <name>, are your

arms tiring? Move the

patients arm higher up”

Improve

Motivation

PSF

“Great results with your 40

points, you’re approaching

your goal fast!

“You reached 40 points,

remember your goal to be

able to play again.”

“With 40 points, I’m sure

you’re still holding back

your performance”

relax

interaction

PST

“We’re halfway through

the exercises, you are a

good team!”

“We’re halfway through

the exercise.”

“I think we as a team can

work harder to achieve a

better performance! There

is only half of the exercises

left to prove it!”

Ask for user

data (only to

Patient)

P -

“Please enter your exertion

level.”

C -

The column “style” means “feedback style”,

which will be either “neutral” (n), “positive” (p) and

“challenging” (c). We use the “Big 5” or “OCEAN”

Model as the foundation to decide in which

conversation style the robot’s action should be

formulated.

In this manner, for a person being less “open to

experience” new things according to the “Big 5”

Humanoid 2021 - Special Session on Interaction with Humanoid Robots

286

model, we may not use the challenging style often,

whereby a more open person may find this style more

engaging than others.

The idea, is, that a robot will always say the same,

when providing a certain action, but the patient

profile will determine, which feedback style will be

used.

4.3 Patient + Helper Model

The main point here is that we try to look on both

persons separately. Therefore, we have two user

models. However, in this paper we focus on the

models for patients only. For helpers the structure of

the model looks similar. However, the spoken phrases

of the robot are different.

If we try to summarize the possible states of a

participant’s states (table1) with all possible robot

actions (table 2), we have 72 state and action

combinations. This means, that the robot says one or

several phrases for each of these states. Additionally,

for the aspect of engagement of either person, the

robot could integrate a more precise answer for task-

related feedbacks, since the patient is the more

passive one and the helper will be the more active

executor in the first stage of the therapy. But this

would result in even more sentences and more

detailed feedback, which need to be prepared

beforehand and also be recognized.

4.4 Patient and Helper Attributes and

Metrics and Its Role for the

Feedback Style

As mentioned before we base our decision which

feedback style the robot should use on the Big 5

Model, as briefly explained in section 2. The arm and

hand speed of the participants detected by devices are

what we specifically want to use for the low/high

engagement. Since the whole therapy is based that the

patient’s arm and hands should be moved by the

helper and later by the patient alone, this will be the

main metric for engagement. This will be a dynamic

value recognized by sensors and cameras.

For the static values of the Big 5 factors, both the

patient and helper should fill out a questionnaire

before the first therapy session.

4.5 History of Previous Commands

The previously performed actions as “chain” of

actions can be added. In theory, the robot should take

his previous actions into account, of what to do next.

Saying e.g. for the fourth consecutive time, that the

helper should better his performance, might be an

annoying for him and the patient.

To prevent this, a chain of previous actions can be

added to each patient or helper state and add rules

basing on them as well. On the other hand, this

temporal aspect could become unnecessary if a robot

action is triggered at a certain time.

A verbal utterance could only be triggered once

e.g. 10 or 15 seconds. It needs to be evaluated how

the frequency of feedback should be.

5 DEFINING THE RULES

The next step would now to define the rules that,

when a certain action happens, which style of the

feedback should be used. This will mostly due

determined by the style of the clinicians and or system

designer when to use what and when a high or low

engagement has been detected.

For the exemplary case we defined an exemplary

list of rules:

1. The positive Style will not be used on

already happy participants with a high

engagement.

2. The “challenging” style should only be used

carefully and when appropriate. (This

feedback style can be perceived as “social

pressure”.)

3. If the hand/arm movement speed moves

below a certain speed X, the engagement

level becomes “low”.

4. The robot should ask less often for more

information about the participant, if he is

more introverted (has not so much

extraversion).

This list should only show an idea, how the Table

3 can be used, especially if this framework would be

used for other therapies, different therapies involving

different tasks for patients and helpers. Only rule 1 &

2 are visually shown in Table 3.

Phrases within a style and the patient's

engagement are the same for the sake of simplicity.

In contrast to our example, individual phrases can be

determined for each different patient state. On the

other hand, depending on the application, there are

many other opportunities where spontaneous

feedback can be given. In the case of the ABT, we

have 17 different exercises, which all can be executed

wrong in several ways and therefore could integrate

multiple correctional feedback options.

User Models for Post-stroke Therapies and Consequences for the Interaction with a Humanoid Robot

287

Table 3: Exemplary table for an exemplary patient for which we do not want to challenge him/her verbally. In this example

we would not use the "challenging" style (1

rule with purple text). Additionally, we wouldn't use the positive style, if the

person is already in a happy state (2

rule with blue lines) and highly engaged. Here we regard ABT step 3 (patient trains

by himself). Crossed out sentences means that the robot will not use this feedback option.

Patient/Helper

High Engagement Low Engagement

Action goal Action Style Happy Neutral Unhappy Happy Neutral Unhappy

Improve

Performance

Provide

Task-related

Feedback

Positive

“You do well, but

you should move

the arm higher up

in each

repetition.”

“You do well, but

you should move

the arm higher up

in each

repetition.”

“You do well, but

you should move

the arm higher up

in each

repetition.”

“You’re on a

good track, but

can you maybe

move your arm

higher up?”

“You’re on a

good track, but

can you maybe

move your arm

higher up?”

“You’re on a

good track, but

can you maybe

move your arm

higher up?”

Neutral

“Move your arm

higher”

“Move your arm

higher”

“Move your arm

higher”

“Move your

arm higher”

“Move your

arm higher”

“Move your

arm higher”

Challenging

“Dear [name], are

your arms tiring?

Move your arm

higher up”

“Dear [name], are

your arms tiring?

Move your arm

higher up”

“Dear [name], are

your arms tiring?

Move your arm

higher up”

“[name] you

are supposed

to move your

arm higher up”

“[name] you

are supposed

to move your

arm higher up”

“[name] you

are supposed

to move your

arm higher up”

Keep/Improve

Motivation

Provide

social

Feedback

Positive

“Great result with

your 40 points,

you’re

approaching your

goal fast!

“Great result with

your 40 points,

you’re

approaching your

goal fast!

“Great result with

your 40 points,

you’re

approaching your

goal fast!

“Awesome

results with

your 40

points!”

“Awesome

results with

your 40

points!”

“Awesome

results with

your 40

points!”

Neutral

“You reached 40

points, remember

your goal to be

able to play

again.”

“You reached 40

points, remember

your goal to be

able to play

again.”

“You reached 40

points, remember

your goal to be

able to play

again.”

“You reached

40 points,

remember your

goal to be able

to play again.”

“You reached

40 points,

remember your

goal to be able

to play again.”

“You reached

40 points,

remember your

goal to be able

to play again.”

Challenging

“With 40 points,

I’m sure you’re

still holding back

your

performance”

“With 40 points,

I’m sure you’re

still holding back

your

performance”

“With 40 points,

I’m sure you’re

still holding back

your

performance”

“I’m sure in

the next round,

you can

achieve more

than 40

oints.”

“I’m sure in

the next round,

you can

achieve more

than 40

oints.”

“I’m sure in

the next round,

you can

achieve more

than 40

oints.”

Change

interaction

climate

Small talk/

Tell a joke

Positive

“We’re halfway

through the

exercises, we are

a good team!”

“We’re halfway

through the

exercises, we are

a good team!”

“We’re halfway

through the

exercises, we are

a good team!”

“We’re already

in the 2

half

of the

exercises we

are a good

team!”

“We’re already

in the 2

half

of the

exercises we

are a good

team!”

“We’re already

in the 2

half

of the

exercises we

are a good

team!”

Neutral

“We’re halfway

through the

exercise.”

“We’re halfway

through the

exercise.”

“We’re halfway

through the

exercise.”

“We’re

halfway

through the

exercise.”

“We’re

halfway

through the

exercise.”

“We’re

halfway

through the

exercise.”

Challenging

“There is only

half of the

exercises left to

show everything

what you got!”

“There is only

half of the

exercises left to

show everything

what you got!”

“There is only

half of the

exercises left to

show everything

what you got!”

“Come on,

only 8

exercises left

to show

everything

what you got!”

“Come on,

only 8

exercises left

to show

everything

what you got!”

“Come on,

only 8

exercises left

to show

everything

what you got!”

Ask for

wellbeing

Request for

information

Positive - - - - - -

Neutral

“Please enter

your exertion

level.”

“Please enter

your exertion

level.”

“Please enter

your exertion

level.”

“Please enter

your exertion

level.”

“Please enter

your exertion

level.”

“Please enter

your exertion

level.”

Challenging - - - - - -

6 DISCUSSION AND

LIMITATION

Finding a common basis to further discuss when to

intervene and how that intervention should look like

for two persons is a challenging task. We think of our

presented work as a starting point of how to formalize

the provision of feedback to patient and helper.

Additionally, we have to track two persons

simultaneously. We will focus on the ideas mentioned

by Winkle et. al and the ideas discussed in section 2.

From the work of Duncan et al. and Bachelor and

other related works on the subject of the strategic

work alliance (section 2), we have taken the paradigm

of focusing on the patient. Due to lack of paper space,

only one example targeting the patient alone was

shown in this work in therapy. Table 3 would also

Humanoid 2021 - Special Session on Interaction with Humanoid Robots

288

exist in a similar form for the helper. However, it can

be argued that the "Request information" action is less

important and will not be carried out often, since the

helper should be fit for the role as helper and it slows

down the therapy session, but the benefit to measure

the helper’s exertion is rather small. The "relax

interaction" action no longer needs to be used,

because it is primarily intended for situations when

both people no longer interact with each other. If this

happens, the robot action will automatically affect the

helper. Consequently, the monitoring for this action

does not have to be done for the helper.

The actions to motivate the helper and to correct

his performance will be changed to suit to his role as

supporter. But other than there will be no systematic

changes. Also, the feedback styles are used and their

usage is derived from the psychological Big 5 profile.

For a practical implementation, one could follow

the idea from (Katie Winkle, 2020) and construct an

interactive-machine-learning setup. Patient and

helper states can be stored as “tuples” of a

<(multi-)state, robot action, (assigned) reward>.

At a given t, one would pick the next best robot

action. In this way, a state could then contain the

whole patient’s profile, engagement level, and the so

far elapsed therapy session.

Looking into the limitations, it is clear, that

integrating personality traits into the process of

choosing the appropriate feedback is a very

challenging task. Due to the complexity of this

problem, the upcoming evaluation could be viewed

only as a preliminary study with the ideas presented

here. The results of that study might show that a

different approach might be more suitable.

Furthermore, the introduction of more states in

the user model can describe the therapy situation

better, but adds more complexity for the robot’s

action space. It is yet not clear whether this is really

necessary. Currently, we have the opinion that six

states are appropriate. However, this has to be

evaluated with patients and helpers.

If we use the same framework for helper and

patient, we would also need to consider to add a

suitable robot action and or phrase for him as well.

More importantly, there should be variation inside

every robot action to avoid a therapy experience and

situations where patients and helper perceive the

robot as boring or monotonous, especially, if these

robot actions get triggered more often than others.

It might be necessary to specify more than one

sentence in each situation and implement an arbitrary

selection of one of those phrases. Thus, the actual

number of predefined sentences or general

vocabulary increases and also the work behind that.

In table 3 we briefly showed the case with 2 examples

per robot action, depending on low and high

engagement. But in reality, one should aim for several

utterances more, since the therapy may take several

weeks of daily sessions.

One may argue, that certain sentences or phrases

displayed here are not particularly positive. But for a

first evaluation, we wanted to let the whole utterance

sound “positive” and focus on exact words later.

Finally, we have to consider the extent to which a

regular big 5 questionnaire can be used for stroke

patients and may instead use a big 5 questionnaire

specifically for stroke survivors.

7 CONCLUSIONS

In this paper we introduced a framework of

observable states of patient and a non-medically

skilled helper to execute a post-stroke therapy for arm

and hand rehabilitation. The training tasks are guided

by a social humanoid robot Pepper. This framework

is intended to be used as a user model of a patient and

a helper. The frameworks are based on the Big 5

Model and further therapy-specific attributes. The

model supports the decision when to take an action by

the robot and also how the feedback should be

expressed. We concentrated on six states for each

person, determined by a high or low engagement that

is analysed from the data of the patient that were

created after an initial interview with a human

therapist. Additionally, the emotions unhappy,

neutral or happy are identified from facial

expressions.

Moreover, the robot may take spontaneous

actions in different conversation styles, to e.g.

challenge or to compliment the users. How the style

will be chosen, depends on pre-defined rules to better

fulfil the patient’s wishes.

We intend to use the framework and define a

detailed list of rules, for an ABT-study inside our E-

BRAiN (Peter Forbrig, et al., 2020) project. In this

project we mainly focus to research how to motivate

patients with a social robot in post-stroke therapies.

It is our goal to start with evaluations by patients

very soon. This had unfortunately to be postponed

several times because of contact restrictions. Even

when the restrictions were lowered, patients did not

want to come to the hospital. Hopefully, this will

change in the near future.

User Models for Post-stroke Therapies and Consequences for the Interaction with a Humanoid Robot

289

ACKNOWLEDGEMENTS

This joint research project “E-BRAiN - Evidence-

based Robot Assistance in Neurorehabilitation” is

supported by the European Social Fund (ESF),

reference: ESF/14-BM-A55-0001/19-A01, and the

Ministry of Education, Science and Culture of

Mecklenburg-Vorpommern, Germany. The sponsors

had no role in the decision to publish or any content

of the publication.

REFERENCES

George A. Kelly. 2001. Clinical diagnosis and

psychotherapy (Reprinted.). The psychology of

personal constructs, / George A. Kelly; Vol. 2.

Routledge, London.

Katie Winkle. 2020. Expert-Informed Design and

Automation of Persuasive, Socially Assistive Robots.

Dissertation. The University of Bristol.

Toni Dwan and Tamara Ownsworth. 2019. The Big Five

personality factors and psychological well-being

following stroke: a systematic review. Disability and

rehabilitation 41, 10, 1119–1130. DOI: https://doi.org/

10.1080/09638288.2017.1419382.

Jonathan A. Casas, Nathalia Céspedes, Carlos A. Cifuentes,

Luisa F. Gutierrez, Mónica Rincón-Roncancio, and

Marcela Múnera. 2019. Expectation vs. Reality:

Attitudes Towards a Socially Assistive Robot in

Cardiac Rehabilitation. Applied Sciences 9, 21, 4651.

DOI: https://doi.org/10.3390/app9214651.

Stephen Hampton, Gina Armstrong, Monika S. Ayyar, and

Sheng Li. 2014. Quantification of perceived exertion

during isometric force production with the Borg scale

in healthy individuals and patients with chronic stroke.

Topics in stroke rehabilitation 21, 1, 33–39. DOI:

https://doi.org/10.1310/tsr2101-33.

Bahar Irfan, Nathalia C. Gomez, Jonathan Casas,

Emmanuel Senft, Luisa F. Gutierrez, Monica Rincon-

Roncancio, Marcela Munera, Tony Belpaeme, and

Carlos A. Cifuentes. 2020. Using a Personalised

Socially Assistive Robot for Cardiac Rehabilitation: A

Long-Term Case Study. In. IEEE. DOI: https://doi.org/

10.1109/ro-man47096.2020.9223491.

Alexandra Bachelor. 1991. Comparison and relationship to

outcome of diverse dimensions of the helping alliance

as seen by client and therapist. Psychotherapy: Theory,

Research, Practice, Training 28, 4, 534–549. DOI:

https://doi.org/10.1037/0033-3204.28.4.534.

Barry L. Duncan, Mark A. Hubble, and Greg Rusk. 1994.

To Intervene or Not to Intervene? That is Not the

Question. Journal of Systemic Therapies 13, 4, 22–30.

DOI: https://doi.org/10.1521/jsyt.1994.13.4.22.

Thomas Platz, Bernhard Elsner, and Jan Mehrholz. 2015.

Arm basis training and arm ability training: two

impairment-oriented exercise training techniques for

improving arm function after stroke. Cochrane

Database of Systematic Reviews 17, 6, 666. DOI:

https://doi.org/10.1002/14651858.CD011854.

Thomas Platz. 2019. Impairment-Oriented Training -

Official Homepage (May 2019). Retrieved March 25,

2020 from http://www.iotraining.eu/abt.html.

Peter Forbrig, Alexandru Bundea, Ann Pedersen, and

Thomas Platz. 2020. Digitalization of Training Tasks

and Specification of the Behaviour of a Social

Humanoid Robot as Coach. In Human-Centered

Software Engineering. Springer International

Publishing, Cham, 45–57.

Humanoid 2021 - Special Session on Interaction with Humanoid Robots

290