MODELLING THERAPEUTIC
EMPATHY IN A VIRTUAL AGENT TO SUPPORT
THE REMOTE TREATMENT OF MAJOR DEPRESSION
Juan Martínez-Miranda, Adrián Bresó and Juan Miguel García-Gómez
ITACA Institute, Biomedical Informatics Group, Universitat Politècnica de València, Camino de Vera s/n, Valencia, Spain
Keywords: Virtual agents, Modelling emotions, Therapeutic empathy.
Abstract: The use of computer based psychotherapy has been one of the fields that have attracted the interest of
practitioners and computer scientists in the last years given the initial and promising results. In particular the
use of computerised cognitive behavioural therapy for the treatment of major depression has been supported
the evidence that psychological therapies can be delivered effectively without face to face contact.
However, the value of these tools for patients is limited by the difficulty of staying engaged during the long-
time periods of the treatment. The use of Virtual Agents as enhanced human-computer interaction brings the
opportunity to overcome this limitation by establishing effective long-term social relationships with
patients. We introduce the main ideas behind the design of a cognitive-emotional model aimed to generate
therapeutic empathic responses and support the remote treatment of major depression.
1 INTRODUCTION
The recent advances in the development of non-
invasive, wearable sensors and wireless
communications, have contributed to create
environments where hardware devices are
transparent, seamlessly integrated and connected in
everyday life people’s situations. One of the main
fields that has benefited from these novel
environments is healthcare through the development
of systems that provide telemedicine/pervasive care
capabilities for a higher patient’s empowerment.
Several applications have been developed in the last
decade where different physiological, physical and
environmental data are obtained through sensors that
continuously and remotely monitor the patients’
condition (Baga et al., 2009), (Farré et al., 2009).
The main advantage of these systems is that the
real-time generated data through monitoring can be
used to provide immediate feedback to patients,
enabling a more active involvement of the person in
his/her treatment. The treatment of mood disorders
is one of the cases where important part of the
success is due to the active participation of the
patient. In particular, during the treatment of major
depression each person should develop their own
personalised strategies for staying well and avoiding
relapses, strategies that are tailored to their
individual needs and characteristics (Kelly, 2000).
The use of telemedicine applied to mental health
(and particularly in the treatment of major
depression) is a relatively new area under research.
One of its main aims is to promote a participative
attitude of patients over the course of the treatment
and minimise the risk of a premature discontinuation
of it which appear one of the main reasons behind
relapsing. Some initiatives already exist in this
context such as computerised cognitive behavioural
therapy (CCBT) which uses stand-alone computer
software or a web application to encourage patients
to complete self-help tasks that involve altering
behaviour and reflecting on and reframing
cognitions. Examples of these applications include
Beating the Blues
(http://www.beatingtheblues.co.uk/), MoodGYM
(http://moodgym.anu.edu.au/) and FearFighter
(http://www.fearfighter.com/) among others. CCBT
appears only weakly effective when the motivation
to continue is left to the patient, but has greater
benefit when use is augmented by support
(Kaltenthaler, 2006); to date this has been provided
either face to face or by telephone.
A complementary approach is a system that
supports the remote patient’s treatment through the
264
Martínez-Miranda J., Bresó A. and García-Gómez J..
MODELLING THERAPEUTIC EMPATHY IN A VIRTUAL AGENT TO SUPPORT THE REMOTE TREATMENT OF MAJOR DEPRESSION .
DOI: 10.5220/0003833302640269
In Proceedings of the 4th International Conference on Agents and Artificial Intelligence (ICAART-2012), pages 264-269
ISBN: 978-989-8425-96-6
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
provision of a continuous and personalised
monitoring over specific behaviour patterns to
collect relevant data allowing the identification of
current patients’ condition. The collected
information can then be used to provide specific
information and recommendations to the patient
complementing the treatment and more important,
prevent the recurrence of symptoms in the future.
Equally important to the suitable collection and
interpretation of patient’s data, is the manner to
provide the information and recommendations to
the patient in order to increase his/her engagement
and help him/her to form an alliance with the
system. The work presented in this paper describes
the first steps towards the design of a virtual agent
(VA) as the main user interface within a system to
support the remote treatment of patients diagnosed
with major depression. In particular, we present our
main ideas for the development of the internal
mechanism in the virtual agent that will lead an
effective therapeutic empathic behaviour to engage
patients and motivate them to keep interacting with
the system and in consequence, promote treatment
adherence.
The following section briefly presents some
current works where synthetic characters have been
developed for mental health interventions. Section 3
makes a general description of the system to support
the remote treatment of major depresssion in which
our VA is embedded. Section 4 describes the main
ideas behind the design of the VA mainly discussing
the differences between natural empathy and
therapeutic empathy and how the conceptualisation
of the second will be integrated in the VA’s
cognitive-emotional model to generate empathic
responses. Finally section 5 presents the conclusions
by describing the current and further work.
2 SYNTHETIC CHARACTERS
IN MENTAL HEALTH
Semi-automated computer-driven interactions have
long been used in healthcare for reminder messages,
disseminating health information, administering
questionnaires, and even conducting interventions
(Glasgow et al., 2004). As the technology to create
and control believable synthetic characters has
become more widely available, the user interface
design has changed to include embodied
conversational agents. The key to successful
interfaces that are structured around an agent is the
relationship between user and agent (Bickmore and
Mauer, 2006). If patients build a rapport with their
agents, they are more likely to use the system, and
more motivated to complete longer sessions. Well-
designed VA’s have been shown to motivate users
(Baylor, 2009) and lead to higher user engagement
than traditional static visual user interfaces (Dehn
and van Mulken, 2000).
There are currently some promising examples of
VA-based interfaces in mental health applications.
The work presented in (Lisetti and Wagner, 2008)
discusses the issues and potential of animated
characters to promote healthy behaviours for helping
teenagers with alcohol abuse through motivational
interventions. Tartaro and Cassell (2008) describe
the development and evaluation of a virtual
character to engage children suffering from autism
spectrum disorders (ASD) in collaborative narrative
to produce contingent discourse. Bickmore et al.
(2010a) developed an agent which encouraged
people with schizophrenia to adhere to their
medication regime, as well as an agent who provided
discharge information to people with depression
who were about to leave the hospital where they had
been admitted for treatment (Bickmore et al.,
2010b). In (Pontier and Siddiqui, 2008) a VA that
guides the user through the Beck Depression
Inventory (BDI), –a questionnaire used to measure
the severity of depression– is presented. Through a
basic emotional model, the VA shows an empathic
behaviour (through its facial expression representing
sadness and/or happiness) to the user depending on
his/her assessed depression level.
All these previous efforts have contributed to
identify a set of key characteristics that these
interfaces should adopt. Such features include the
appearance, verbal and non-verbal communication,
coherent personality, referencing knowledge of prior
interactions, variability in agent behaviour and the
adoption of an empathic behaviour during
interaction. In this paper we focus on our current
activities towards the design of a VA able to produce
effective therapeutic empathy interactions.
3 HELP4MOOD: SUPPORTING
THE TREATMENT OF MAJOR
DEPRESSION
Our proposed VA is a key component of a FP7-EU
research project aimed to support the remote
treatment of people with major depression
(www.help4mood.info). The main aim of the project
is to support people with major depression in its
MODELLING THERAPEUTIC EMPATHY IN A VIRTUAL AGENT TO SUPPORT THE REMOTE TREATMENT OF
MAJOR DEPRESSION
265
mild to moderate form (those people who do not
require hospitalisation but need to follow the
treatment as part of their daily activities). The
proposed system will collect a relevant set of
parameters that allow the detection of specific
behaviour patterns in the patient, prompting
adherence to computerised cognitive behavioural
therapy, and promoting healthy behaviours in
response to monitored inputs.
The three main components of the Help4Mood
system include a Personal Monitoring System
(PMS), a Virtual Agent-based user interface, and a
Decision Support System that will support clinicians
in interpreting self-reported and monitoring data.
The PMS will collect two types of data: i)
behavioural data including patterns of sleeping,
motor activity and speech; and ii) subjective data
including brief validated scales to measure mood,
cognition and behaviours. The identified behaviour
patterns will then be used by the VA to prompt the
patient (when appropriate) to carry out potentially
helpful activities, such as relaxation or exercise,
offer the opportunity to add entries to a spoken diary
or, if the collected data suggest a potential treatment
failure and/or a suicide risk, urgently alert the
clinical site leading to a direct communication.
The specific feedback provided by the VA will
be mainly through a combination of dialogue
interaction and a basic set of body movements and
facial expressions designed to maintain the attention
of the patient and help him/her to effectively manage
important stages of his/her treatment. The VA is
composed of three main modules: the graphical
appearance, a dialogue manager system which will
implement and manage the interaction dialogues
between the VA and the patient, and the module
which generates the VA’s cognitive-emotional
behaviour which is introduced in this paper.
A patient-side Decision Support System will
facilitate intelligent, adaptive interpretation of self-
reported data. A clinician-side Decision Support
System will distil both self-reported and objective
data into a succinct textual and graphical overview
that will help clinicians in the assessment of the
patient’s current state and guide further treatment
decisions.
4 DESIGNING
THE VIRTUAL AGENT
An important aspect in the development of the
Help4Mood’s VA is to motivate a long-term use and
interaction between the user and the agent. This is an
especially important issue when the main objective
of a VA is to promote the use of the system and
complete longer sessions (Bickmore and Mauer,
2006). The key characteristics that highly contribute
to maintain the motivation of users while interacting
with virtual characters include the believability of
the VA’s behaviour and a good verbal and non-
verbal communication. To successfully obtain these
two characteristics in a VA, there should be an
internal mechanism that produces a coherent
emotional behaviour and communication that
complements the cognitive actions and decisions
taken during interaction.
A coherent and consistent emotional behaviour
(usually referred as emotional competence with
respect of two domains: emotion production and
emotion perception (Scherer, 2010)) has been
identified as particularly important in VAs used for
psychotherapy (Bickmore and Gruber, 2010),
(Lisetti, 2008) and it is still an open research line
where several considerations (including theoretical,
technical and ethical) need to be addressed. In the
last years, several computational models of emotions
have emerged trying to cover specific (and
sometimes theoretically incompatible) emotional
mechanisms (Marsella et al., 2010). The differences
in many of the existent computational models of
emotions are a logical consequence of the different
emotion theories where these models have their
roots (Scherer, 2010).
From the current existent theories of emotion, the
one that predominates above the others in the efforts
dedicated to implement computational models of
emotions is the cognitive appraisal theory of
emotions (Scherer et al., 2001). The core concept of
appraisal theories refer that the events in a person’s
environment are constantly identified and evaluated
by the individual. This cognitive evaluation (or
appraisal) process leads to an emotional response
(according to the event’s relevance for the person)
which in turn generates a specific behaviour to cope
with the appraised events. The high success in
choosing this theory as the theoretical background in
several computational models seems in part due to
the emphasis and explanation of the connection
between cognition and emotion which help in the
construction of artificial systems that simulate
complex human-like behaviours. Moreover,
appraisal theories of emotion appear the most
comprehensive way to represent the complexity of
the emotion process, covering the whole path from
low-level appraisals of the eliciting event to high-
level influence over behaviour (Scherer, 2010).
ICAART 2012 - International Conference on Agents and Artificial Intelligence
266
Several computational architectures based on the
appraisal theory of emotions have been developed
since the early 90’s (Elliott, 1992) and as reflected in
(Marsella et al., 2010), virtual characters based on
this type of architectures has allowed the creation of
real-time interactive characters that exhibit emotions
in order to make them more compelling, more
realistic, or more able to induce desirable social
effects in the users.
4.1 Natural Empathy vs. Therapeutic
Empathy
The promising research results reported in the
literature such as the works mentioned above,
encourage us to consider the adoption of a
computational architecture based on appraisal
theories of emotion for the implementation of the
internal model in our Help4Mood’s VA. We of
course need to consider the particular characteristics
of our target users to adopt the most suitable
solution. As suggested by the clinicians of the
Help4Mood consortium, two of the important
characteristics in our VA should be an adaptable
empathic behaviour and a personality style that
motivates the use of the system.
A key difference in the empathic behaviour of
our proposed VA regarding existent similar works
such as e.g. the presented in Paiva et al., (2004) and
Prendinger and Ishizuka (2005), is that the empathic
responses needs to be modulated according to the
special characteristics of the patients. Although the
VA should be able to show empathy during
interaction, it is highly important from a clinical
perspective that the range of emotions displayed
should be restricted to a neutral stance or positive
emotions (such as joy and happiness) and not to
produce an empathic behaviour by just adopting the
same (frequently negative) emotions or affective
state reported by the patient.
In this sense, it is important to distinguish the
difference between natural empathy (experienced by
every people in every-day situations) and therapeutic
empathy in order to provide to the patients with a
useful feedback for their particular treatment and
promoting an effective therapeutic alliance. In
(Hoffman, 2000) natural empathy is associated to
psychological processes that make an individual to
have feelings that are more congruent with another’s
situation that with his/her own situation.
From the psychotherapeutic and counselling
perspective the term therapeutic empathy is defined
as when the therapist is sensing the feelings and
personal meanings which the client is experiencing
in each moment, when he can perceive these from
’inside’, as they seem to the client, and when he can
successfully communicate something of that
understanding to his client (Rogers, 1961 p. 62).
This particular type of empathic behaviour is the one
that it is clinically relevant and should include the
mechanism to modulate those therapeutic useless or
inadequate emotions.
There is currently some effort to identify and
dissect the theoretical key components of therapeutic
empathy. Thwaites and Bennett-Levy (2007) argue
that a therapeutic empathy system should contain the
following four components:
1. Empathic attitude/stance: infuses other aspects
of empathic skill with a sense of benevolence,
curiosity and interest.
2. Empathic attunement: a perceptual skill referred
to as an active on-going effort to stay attuned on
a moment-to-moment basis with the client’s
communication and unfolding process
3. Empathic communication skills: are the skills to
explicitly communicate empathic understanding
as well as emotions not only through the content
of the speech itself, but also by non-verbal
behaviour and tone of voice providing a sense
of safety, warmth, understanding and
acceptance.
4. Empathy knowledge: defined as what therapists
learn from teachers and from reading during
training and professional development and it is
one of the key factors that differentiate
therapeutic empathy from natural empathy.
Although our VA will not have the role of a
therapist (but more related to a helper/guide during
the treatment’s short every-day sessions), it is
interesting to analyse how the conceptualisation of
therapeutic empathy can be modelled to implement a
computational therapeutic empathy system. A
central aspect is the integration of the four concepts
at the different components of the computational
appraisal architecture (Figure 1).
An empathic attunement can be modelled in the
VA when specific events are detected in the
environment. These events are inferred using the
data collected from the personal monitoring system
and the patient’s self-reports. Following the VA’s
goals (which are activated according to the particular
objectives of each treatment’s session), the event
could be appraised as desirable or not in terms of the
patient’s perspective. Every event needs to be
understood in these terms for the VA remains
empathically attuned on a moment-to-moment basis.
MODELLING THERAPEUTIC EMPATHY IN A VIRTUAL AGENT TO SUPPORT THE REMOTE TREATMENT OF
MAJOR DEPRESSION
267
The result of this appraisal process is used to
select the specific emotion in the VA, which is also
influenced by the modelled personality in the VA
and the empathy knowledge layer. For the modelling
of personality we are considering those trends
related to an agreeableness personality to facilitate
an empathic attitude in the VA. One of the main
functions in the empathy knowledge component will
be the modulation of the emotion triggered by the
emotion selection mechanism. The empathy
knowledge modulates the negative emotions to suit
the clinical perspective for depression treatment, i.e.
adopting a neutral stance when patient indicates
negative moods, thoughts and feelings, and display
positive emotions in the rest of situations.
Figure 1: The Help4Mood VA’s internal model.
The specific triggered emotion will be used to
select the action for coping with the detected event.
The empathy knowledge will also be used by the
action-selection mechanism to choose the most
adequate action during the interaction with the
patient. As already introduced, the empathy
knowledge is one of the key factors to distinguish
therapeutic empathy from natural empathy and its
inclusion in both, the emotion and action selection
mechanisms, intends that the VA emulates the
emotional involvement during specific interactions
with the patient and an emotional detachment that
allows for a more objective appraisal of the situation
(Clark, 2007).
The actions take the form of high level
commands (such as SUGGEST_ACTIVITY(x) or
APPLY_QUESTIONNAIRE(y)) which are passed
jointly with the emotion as a <task, emotion> pair to
be implemented through the concrete utterances by
the dialogue manager system and the corresponding
facial expression in the graphical component of the
VA. The <task, emotion> pair will affect the verbal
and non-verbal communication of the VA by
affecting the voice, the actual wording of the prompt
and the facial expressions of the VA leading to an
empathic communication.
5 CONCLUSIONS
The model presented in the past section is currently
under development and we are analysing existent
computational architectures of emotion that can be
extended with our proposal. A good candidate is the
known as the FAtiMA (FearNot Affective Mind
Architecture) architecture (Paiva et al., 2004).
FAtiMA is an open source software and it has been
an evolving (appraisal theory based) architecture
developed in the context of three EU projects: the
FP-5 VICTEC (http://www.macs.hw.ac.uk/victec/
index_geral.html), the FP-6 ECIRCUS
(http://www.macs.hw.ac.uk/EcircusWeb/) and the
FP-7 LIREC (http://lirec.eu/) projects. The original
objective of FAtiMA was the creation of empathic
agents within a computer application to tackle and
eventually help to reduce bullying problems in
schools. The core of the architecture has been
subsequently extended to include additional
components of the emotion phenomenon such as the
implementation of a double-appraisal mechanism in
order to evaluate the emotional impact of possible
actions (Aylett and Louchart, 2008); the memory
retrieval (and re-appraisal) of emotionally past
episodes (Gomes et al., 2011); or the generation of
emergent empathic reactions in social agents (Paiva,
2011).
Once implemented and integrated with the rest
of the system, the outcome behaviour produced in
the VA can be clinically evaluated in the further
stages of Help4Mood and assess its suitability in the
building of VA – patient interactions.
ACKNOWLEDGEMENTS
This paper reflects only the author’s views. The
European Community is not liable for any use that
may be made of the information contained herein.
This research is carried out within the EU FP7
Project “Help4Mood – A Computational Distributed
System to Support the Treatment of Patients with
Major Depression” [ICT-248765].
ICAART 2012 - International Conference on Agents and Artificial Intelligence
268
REFERENCES
Aylett, R. and Louchart, S. (2008). If I were you: double
appraisal in affective agents. In Proc. of 7
th
Int. Conf.
on Autonomous Agents and Multiagent Systems
(AAMAS 2008), Padgham, Parkes, Müller and
Parsons (eds.), Estoril, Portugal, pp. 1233-1236.
Baga, D., Fotiadis, D. I., Konitsiotis, S., Tsoulli, S.,
Diakou, M., Arredondo, M. T., Estrada, J. J., Pansera,
and M., Akay, M. (2009). Perform: A platform for
monitoring and management of chronic
neurodegenerative diseases: The Parkinson and
amyotrophic lateral sclerosis case. In Proc. of the 4
th
Int. IEEE EMBS Conf. on Neural Engineering.
Baylor, A. L. (2009) Promoting motivation with virtual
agents and avatars: role of visual presence and
appearance. Philosophical transactions of the Royal
Society of London. Series B, Biological sciences,
364(1535), 3559–65.
Bickmore T. and Gruber, A. (2010). Relational agents in
clinical psychiatry. Harvard review of psychiatry,
18(2), 119-130.
Bickmore, T., Puskar, K., Schlenk, E., Pfeifer, L. and
Sereika, S. (2010a). Maintaining reality: Relational
agents for antipsychotic medication adherence.
Interacting with Computers 22(4), 276-288.
Bickmore, T., Mitchell, S., Jack, B., Paasche-Orlow, M.,
Pfeifer, L. and Odonnell, J. (2010b). Response to a
Relational Agent by Hospital Patients with Depressive
Symptoms. Interacting with computers 22(4), 289-298.
Bickmore, T. and Mauer, D. (2006). Modalities for
building relationships with handheld computer agents.
In Proceedings of The Conference on Human Factors
in Computing Systems (CHI-2006).
Clark A. J. (2007). Empathy in Counseling and
Psychotherapy. Perspectives and Practices. Lawrence
Erlbaum Associates. London.
Dehn, D. M. and van Mulken, S. (2000). The impact of
animated interface agents: a review of empirical
research. International Journal of Human-Computer
Studies, 52(1), 1–22.
Elliott, C. (1992). The Affective Reasoner: A Process
Model of Emotions in a Multi-Agent System. PhD
thesis, The Institute for the Learning Sciences,
Northwestern University.
Farré, R., Papadopoulus, A., Munaro, G., and Rosso, G.
(2009). An open, ubiquitous and adaptive chronic
disease management platform for chronic respiratory
and renal diseases (chronious). In the 2009
International Conference on eHealth, Telemedicine,
and Social Medicine, pp. 184-189.
Glasgow, R., Bull, S., Piette, D. and Steiner. J. (2004).
Interactive behavior change technology. A partial
solution to the competing demands of primary care.
American J. of Preventive Medicine, 27(2), 80–87.
Gomes, P., Martinho, C. and Paiva. A. (2011). I’ve Been
Here Before! In Proc. of 10
th
Int. Conf. on
Autonomous Agents and Multiagent Systems
(AAMAS 2011), Tumer, Yolum, Sonenberg and Stone
(eds.), May, 2–6, 2011, Taipei, Taiwan, 1039-1046.
Hoffman, M. (2000). Empathy and Moral Development.
Cambridge University Press.
Kaltenthaler, E., Brazier, J., De Nigris, E., Tumur, I.,
Ferriter, M., Beverley, C., Parry, G., Rooney, G. and
Sutcliffe, P. (2006). Computerised cognitive behaviour
therapy for depression and anxiety update: a
systematic review and economic evaluation. Health
Technology Assessment, 10(33), 1 – 168.
Kelly, M. (2000). Life on a roller-coaster: living well with
depression and maniac depression. Simon and
Schuster.
Lisetti, C. (2008). Embodied conversational agents for
psychotherapy. In CHI 2008 Workshop on
Technology in Mental Health.
Lisetti, C., Wagner, E. (2008). Mental Health Promotion
with Animated Characters: Exploring Issues and
Potential. In Proceedings of the AAAI Spring
Symposium on Emotions, Personality and Behavior
(AAAI SSS'08), 29(3), 107-115.
Marsella, S., Gratch, J. and P. Petta. (2010).
Computational Models of Emotion. In Blueprint for
Affective Computing: A Sourcebook. Scherer, K.,
Bänziger, T. & Roesch, E. (Eds.), Chapter 1.2, 21-46,
Oxford University Press.
Paiva A. (2011). Empathy in Social Agents. The
International Journal of Virtual Reality, 10(1), 65-68.
Paiva, A., Dias, J., Sobral, D., Aylett, R., Sobreperez, P.,
Woods, S., Zoll, C. and Hall, L. (2004). Caring for
agents and agents that care: Building empathic
relations with synthetic agents. In Proc. of the Third
Int. Conf. on Autonomous Agents and Multi Agent
Systems (AAMAS 2004), 194 – 201, ACM Press.
Pontier, M., and Siddiqui G. (2008). A Virtual Therapist
That Responds Empathically to Your Answers. In 8
th
Int. Conf. on Intelligent Virtual Agents, Prendinger,
H., Lester, J. and Ishizuka, M. (Eds.), 417-425.
Prendinger, H. and Ishizuka, M. (2005). The empathic
companion: A character-based interface that
addresses users’ affective states. Applied Artificial
Intelligence, 19, 267–285.
Rogers, C. R. (1961). On Becoming a Person: a
therapist’s view of psychotherapy. London: Constable
and Company.
Scherer, K. (2010). Emotion and emotional competence:
conceptual and theoretical issues for modelling
agents. In Blueprint for Affective Computing: A
Sourcebook. Scherer, K., Bänziger T. & Roesch E. B.
(Eds.), Chapter 1.1, 3-20, Oxford University Press.
Scherer, K., Schorr, A. and Johnstone, T. (Eds.). (2001).
Appraisal Processes in Emotion: Theory, Methods,
Research. Oxford University Press.
Tartaro A., Cassell J. (2008). Playing with virtual peers:
Bootstrapping contingent discourse in children with
autism. In proceedings of International Conference of
the Learning Sciences (ICLS).
Thwaites, R. and Bennett-Levy, J. (2007).
Conceptualizing empathy in cognitive behaviour
therapy: Making the implicit explicit. Behavioural and
Cognitive Psychotherapy, 35, 291–612.
MODELLING THERAPEUTIC EMPATHY IN A VIRTUAL AGENT TO SUPPORT THE REMOTE TREATMENT OF
MAJOR DEPRESSION
269