Building an On-Demand Avatar-Based Health Intervention for Behavior Change
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Proceedings of the Twenty-Fifth International Florida Artificial Intelligence Research Society Conference
Building an On-Demand Avatar-Based Health Intervention for Behavior Change
Christine Lisetti, Ugan Yasavur, Ubbo Visser
Claudia de Leon, Reza Amini, Napthali Rishe Department of Computer Science
School of Computing and Information Sciences University of Miami
Florida International University Coral Gables, FL, 33146, USA
Miami, FL, 33199, USA
Abstract 2002; Miller and Rose 2009). MI is a recently developed
style of interpersonal communication which was originally
We discuss the design and implementation of the pro-
totype of an avatar-based health system aimed at pro- conceived for clinicians in the field of addiction treatment,
viding people access to an effective behavior change and which has been since then proven useful for medical
intervention which can help them to find and cultivate and public health settings (Emmons and Rollnick 2001), as
motivation to change unhealthy lifestyles. An empathic well as for concerned others. It is aimed at helping clients
Embodied Conversational Agent (ECA) delivers the in- (or people in general) to find the motivation to change their
tervention. The health dialog is directed by a compu- lifestyle with respect to a specific unhealthy pattern of be-
tational model of Motivational Interviewing, a novel havior. Although originally conceived as a preparation for
effective face-to-face patient-centered counseling style further treatment, enhancing motivation and adherence, re-
which respects an individual’s pace toward behavior search revealed that change often occurred soon after one or
change. Although conducted on a small sample size, re-
two session of MI, without further treatment, relative to con-
sults of a preliminary user study to asses users’ accep-
tance of the avatar counselor indicate that the system trol groups receiving no counseling at all (Miller and Roll-
prototype is well accepted by 75% of users. nick 2002).
1 Introduction At the same time as novel behavior change interventions
have developed, simulated human characters have become
Recent epidemics of behavioral issues such as excessive al- increasingly common elements of user interfaces for a wide
cohol, tobacco, or drug use, overeating, and lack of exercise range of applications such as interactive learning environ-
place people at risk of serious health problems (e.g. obe- ments, e-commerce and entertainment applications such as
sity, diabetes, cardiovascular troubles). Medicine and health- video games, virtual worlds, and interactive drama. Simu-
care have therefore started to move toward finding ways of lated characters that focus on dialog-based interactions are
preventively promoting wellness rather than solely treating called Embodied Conversational Agents (ECAs). ECAs are
already established illness. Health promotion interventions digital systems created with a physical anthropomorphic
aimed at helping people to change lifestyle are being de- representation (be it graphical or robotic), and capable of
ployed, but the epidemic nature of these problems calls for having a conversation (albeit still limited) with a human
drastic measures to rapidly increase access to effective be- counterpart, using some artificial intelligence broadly re-
havior change interventions for diverse populations. ferred to as an “agent“ (Cassell and J. Sullivan, S. Prevost,
Lifestyle changes toward good health often involve rela- & E. Churchill 2000). With their anthropomorphic features
tively simple behavioral changes that can easily be targeted and capabilities, they provide computer users with a some-
(as opposed to major life crisis that usually require psy- what natural interface, in that they interact with humans
chotherapy): for example lower alcohol consumption, ad- using humans’ natural and innate communication modali-
just a diet, or implement an exercise program. One of the ties such as facial expressions, body language, speech, and
main obstacle to conquer change of unhealthy habits, is of- natural language understanding (Magnenat-Thalmann, N.
ten ’just’ a matter of finding enough motivation. There has and Thalman 2004; Predinger and Ishizuka 2004; Allen et
recently been a growing effort in the health promotion sec- al. 2006; Becker-Asano 2008; Boukricha and Wachsmuth
tor to help people find motivation to change, and a number 2011; Pelachaud 2009). We posit that given the well-defined
of successful evidence-based health promotion interventions structure of MI (discussed next), MI-based interventions can
have been designed. be automated and delivered with Embodied Conversational
Our current research interests have brought us to consider Agent (ECA). We describe the design and implementation of
one such counseling style for motivating people to change – an MI-based automated intervention aimed at helping people
namely Motivational Interviewing (MI) (Miller and Rollnick with some excessive alcohol consumption to become aware
Copyright c 2012, Association for the Advancement of Artificial and potentially motivated to change unhealthy patterns of
Intelligence (www.aaai.org). All rights reserved. behavior.
1752 Related Research Hester, R. K., Skire, D.D., and Delaney 2005). The check-
Several avatar-based health intervention or counseling up (along with modifications of it) is the most widely used
applications have been developed and evaluated. Silverman brief motivational interviewing adaptation, where the client
and colleagues coupled a simulation package with an an- (usually alcohol or drug addicted) is given feedback, in a
imated pedagogical heart-shaped avatar to promote health MI “style“based on individual results from standardized as-
behavior change in heart attacks (Silverman et al. 2001). sessment measures (Burke, B. L. and Arkowitz, H. and and
Anthropomorphic pedagogical agents are used in a system Dunn 2002).
called DESIA for psychosocial intervention deployed on The On-demand Health Counselor System Architecture
hand-held computers in order to teach caregivers problem we developed can deliver personalized and tailored behav-
solving skills (Johnson, LaBore, and Chiu 2004). FitTrack ior change interventions using the MI communication style
uses an ECA to investigate the ability of an avatar-based sys- via multimodal verbal and non-verbal channels. The system
tem to establish and maintain a long-term working alliance is developed in the .NET framework as a three-tier architec-
with users (Bickmore and Picard 2005). The Virtual Hospi- ture which uses third party components and services.
tal Discharge Nurse has been developed to explain written The system architecture is composed of the main mod-
hospital discharge instructions to patients with low health ules shown in Figure 1, which we describe in details later. In
literacy. It is available to patients on their hospital beds and short, during any interaction, the user’s utterances are pro-
helps to review material before discharging them from the cessed by the Dialog Module which directs the MI sessions
hospital. It indicates any unresolved issues for the human and elicits information from the user; the user’s facial ex-
nurse to clarify (Bickmore, Pfeifer, and Jack 2009). pressions are captured and processed by the Affective Mod-
More recently MI has been identified as particu- ule, in an attempt to assess the user’s most probable affective
larly useful for health dialog systems (Lisetti 2008; states and convey an ongoing sense of empathy via a Mul-
Lisetti and Wagner 2008). Since then some researchers have timodal Avatar-based Interface (using 3D animated virtual
explored the use of 2-D cartoon or comic-like characters to character with verbal and nonverbal communication); Psy-
provide MI-related material to patients or to train health care chometric Analysis is performed based on the information
personnel in MI (Schulman, Bickmore, and Sidner 2011; collected by the Dialog Module, and its results are stored in
Magerko et al. 2011). Although the avatars employed in the User model which is maintained over multiple sessions
these mentioned systems have shown some promising to offer a dynamically Tailored Intervention in the form of
acceptance by their users, because they are 2D, they lack sensitive feedback or specific behavior change plans.
dynamic expressiveness which is key for communicating
facial affect (Ehrlich, Schiano, and Sheridan 2000), itself
essential in establishing the MI requirement of empathic Tailored ntervention
communicative style (Miller and Rollnick 2002). Behavior Tailored
User
Change Feedback
Model
Planner Engine
Motivational Interviewing (MI) has been defined by
Miller and Rollnick (Miller and Rollnick 2002) as a direc-
tive, client-centered counseling style for eliciting behavior Psychometric
change by helping clients to explore and resolve ambiva- Analysis
lence. One of MI central goals is to magnify discrepancies Score
Evaluator
Psychometric
Data
that exist between someone’s goals and current behavior.
MI basic tenets are that a) if there is no discrepancy, there is
no motivation; b) one way to develop discrepancy is to be- Dialog
come ambivalent; c) as discrepancy increases, ambivalence Module
first intensifies; if discrepancy continues to grow, ambiva- Psychometric Dialog
MI-based
Dialog
lence can be resolved toward change. In the past few years, Instruments Planner
Engine
adaptations of motivational interviewing (AMIs) have mush- User
Utterances
roomed with the purpose to meet the need for motivational
interventions within medical and health care settings (Burke, Affective Multimodal Avatar-based
nterface
B. L. and Arkowitz, H. and and Dunn 2002). In such set- User's
Facial
Module Empathy
Engine
tings, motivational issues in patient behavior are quite com- Expression Text-To-Speech
Engine
mon, but practitioners may only have a few minutes to deal
with them. Some of these AMIs can be as short as one 40- Facial
Avatar System
minute session. Processing
3 Health Counselor System Architecture
We based our system on one such AMI referred to as the
Check-Up that specifically targets excessive drinking behav- Figure 1: On-demand Avatar-based Counselor Architecture.
iors and with which people can reduce their drinking by an
average of 50% (Miller, W., Sovereign, R. and Krege 1988;
176Dialog Module have missed work about once every two weeks, you have
The Dialog Module evaluates and generates dialog utter- gotten into arguments with friends frequently at parties, ...”);
ances using three key components: a Dialog Planner, an and 3) reflective listening.
MI-based Dialog Engine and a collection of Psychometric Reflective Listening is a client-centered communication
Instruments. The interaction of the client with the system is strategy involving two key steps: seeking to understand a
based on a series of dialog sessions, each of which having speaker’s thoughts or feelings, then conveying the idea back
a specific assessment goal to identify different aspects of to the speaker to confirm that the idea has been understood
the user’s behavioral problem (if any) and in this article we correctly (Rogers 1959). Reflective listening also includes
focus on excessive intake of alcohol. Each session is based mirroring the mood of the speaker, reflecting the emotional
on psychometric instruments which the Dialog Planner state with words and nonverbal communication (which we
uses as a flexible plan so that the dialog-based interaction discuss in the Affect Module below). Reflective listening is
can be dynamically conducted based on the knowledge that an important MI technique which the system simulates to
the system has acquired about the user (maintained in the help the client notice discrepancies between his or her stated
user model), and on the user’s current entries and affective beliefs and values, and his or her current behavior. We cur-
states. rently implement simple reflection to elicit ambivalence (e.g.
– Client: “I only have a couple of drinks a day.” – Counselor:
Psychometric Instruments: We use a collection of well- “ So you have about 14 drinks per week.”).
validated Psychometric Instruments commonly used by We also use Double-sided reflection to summarize am-
therapists to assess an individual’s alcohol use in order to bivalence, which can help to challenge patients to work
design each assessment session (Miller and Rollnick 2002). through their ambivalence about change. Statements such as
Each psychometric instrument contains a set of questions “on the one hand, you like to drink and party, but on the other
which represent the plan for an assessment session. For hand, you are then less likely to practice safe sex” can clearly
example one instrument (DrInC, 50 items) is used to summarize a patient’s ambivalence toward change (Botelho
identify consequences of the client’s drinking problem with 2004). Other double-sided reflections such as “Ok, so your
five different parameters (e.g. interpersonal relationships, wife and children complain daily about your drinking but
work) and these are stored in the user model (see below). you do not feel that your ability to be a good parent is at
stake, is that right?”, can magnify the discrepancies between
Dialog Planner: This component of the dialog module the user’s current situation and what they would like it to be.
decides what type of utterance is to be generated based on
the previous interaction in that session, i.e. it decides the Psychometric Analysis
next stage of the session. For instance, when the system As shown in Figure 1, psychometric analysis is concerned
needs to assess the client’s awareness of the consequences with processing Psychometric Data collected in the dia-
of his/her drinking, the The Drinker Inventory of Conse- log module using the collection of psychometric instruments
quences (DrInC) psychometric instrument is used as the (described earlier). Based on instructions in each instru-
plan for the session, and each question is assigned one of the ment, the Score Evaluator module calculates the score of
five topics of the DrInC (e.g. inter-personal, intra-personal). the user for a particular measuring instrument, according to
To measure consequences in each area, there are sets of normative data (MATCH 1998). The result is stored in the
non-sequentially positioned questions for each area and user model (see next) in order to identify specific aspects of
these questions are conceptually related with each other. the drinking problem. For instance, the instrument assessing
The dialog planner aims at detecting discrepancies between drinking patterns can locate the user on a spectrum of al-
received answers for questions in the same area. If the cohol consumption (low, mild, medium, high, or very high,
planner detects discrepancy in the client’s answers, then where a score above 8 indicates mild or higher problems).
the MI-based Dialog Engine is invoked to generate some When provided feedback about this score (by the Tailored
MI-style dialog, such as reflective listening (explained Feedback module), the user can make an informed decision
next). If no such discrepancy nor ambivalence in the user’s about whether or not to continue with the offered self-help
answers is detected, then the dialog planner continues with program. Respecting the MI spirit, the user is never coerced
the assessment session. and can at any moment exit the intervention and the system
gracefully terminates the conversation.
MI-based Dialog Engine: This engine applies a finite set of
well-documented MI techniques known under the acronym Tailored Intervention
(OARS): Open-ended questions, Affirmations, Reflective lis- User Model: The user model enables the system to tailor
tening, and Summaries, which are applied toward goals con- its intervention to a specific user. Tailoring can be achieved
cerning specific behaviors (e.g. excessive alcohol use, drug in multiple ways: by using the user’s name referred to as
use, overeating). The engine adapts each static psychometric personalization, by using known characteristics of the user
instrument to MI style interaction and currently implements such as gender referred to as adaptiveness, and by using self-
1) affirmations (e.g.“You have much to be proud off”); 2) identified needs of the user referred to feedback-provision.
Summaries (“Ok, so let me summarize what you’ve told me The user model collects the age, gender, weight, height
today about your experience during the last 90 days. You and first name or nickname to preserve anonymity. It also
have had more than 20 drinks, during the last 90 days you gathers the user’s ethnicity to implement patient-physician
177ethnic concordance with diverse avatars matching the eth-
nicity of the user because it is known that racial concor-
dance increases the quality of patient-physician communica-
tion and some patient outcomes (e.g. patient positive affect)
(Cooper et al. 2003). In addition to static demographic infor-
mation, the user model builds and dynamically maintains the
profile of each user as it changes over time. The user model
includes different variables in six different areas: motivation,
personal attributes, frequency of drinking, risk factors, con-
sequences and dependence. Each area consists of multiple
variables to assess specific aspects of the problem from dif-
ferent view points. These psychometric data are calculated Figure 2: Dynamic Facial Expression Animations
based on psychometric instruments (see above).
Because the intervention is designed to assist the user
to change unhealthy behavioral patterns, it can be used (ECA) or avatar, which can deliver verbal communication
over multiple sessions as on-demand follow-up sessions. with automatic lip synchronization, as well as non-verbal
As the user continues to interact with the system over communication cues (e.g. facial expressions).
time, the user model keeps track of the fluctuating user’s
progress toward change in terms of: the user’s current Avatar System: We integrated a set of features that we
level of ambivalence, by identifying and weighing positive consider necessary for health promotion interventions in our
and negative things about drinking; the user’s readiness to current prototype (using a third party avatar system called
change, from pre-contemplation to maintenance (Prochaska Haptek): 1) a 3-dimensional (3D) graphical avatar well-
and Velicer 1997); as well as the result of other specialized accepted by users as documented in an earlier study (Lisetti
self-report questionnaires (e.g. the user’s awareness of the et al. 2004); 2) a selection of different avatars with different
consequences of drinking). ethnicity (e.g. skin color, facial proportions) shown in Fig-
ure 3 (Nasoz, F. and Lisetti 2006); 3) a set of custom-made
Tailored Feedback Engine: This engine generates tailored facial expressions capable of animating the 3D avatar with
feedback based on the user model. Several studies showed dynamical believable facial expressions(e.g. pleased, sad,
that tailored health communication and information can be annoyed shown in Figure 2) (Paleari, Grizard, and Lisetti
much more effective than non-tailored one because it is 2007); 4) a Text-To-Speech (TTS) engine able to read text
considered more interesting and relevant, e.g. (Noar, Benac, with a natural voice with lip synchronization; and 5) a set
and Harris 2007). Because the feedback module reveals of mark-ups to control the avatar’s TTS vocal intonation and
discrepancies between the user’s desired state and present facial expressions in order to add to the agent’s believability
state, it is likely to raise resistance and defensiveness. The (Predinger and Ishizuka 2004).
style in which the feedback is given therefore needs to be
empathetic enough to avoid generating defensiveness. A Affective Module
number of pre-identified sub-dialog scripts are used for Our on-demand avatar counselor aims at delivering health
the various topics of feedback, e.g. self-reported drinking, promotion messages that are not only tailored to the user (via
perceived current and future consequences of drinking, the user-model), but also empathetic. Because discussing is-
pros and cons of current consumption patterns, personal sues about at-risk behaviors are highly emotional for people
goals and their relation to alcohol use, readiness to change to talk about (e.g. embarrassment, discouragement, hope-
drinking behavior. The user receives feedback about the fulness), our context or domain application is particularly
assessed self-check information, which is put in perspective suited to inform the design of empathic characters. Empathy
with respect to normative data (e.g. “according to the is considered the most important core condition in terms of
current national norm, 10% of people your age drink as promoting positive outcomes in psychotherapeutic contexts
much as you do”). (Bellet and Maloney 1991) and it can account for as much
as 25%-100% rate of improvement among patients (Miller
Behavior Change Planner: This component helps to create and Rollnick 2002; Rogers 1959).
personalized behavior change plans for clients who are ready Currently, the Empathy Engine is limited to deciding
to change, and is only entered by users who have received which facial expressions the avatar needs to display. Our cur-
feedback from the Tailored Feedback Engine. The first step rent system focusses on mimicking the user’s facial expres-
of the behavior change planner is to assess the user’s readi- sions, which is considered as a type of primitive empathy
ness to change along a continuum from “Not at all ready” to (Boukricha and Wachsmuth 2011). In the Facial Process-
“Unsure” to “Really ready to change”. Depending upon the ing component, pictures of the user’s face are taken at fixed
level of readiness, different sub-modules are performed. intervals during the user’s interaction with our avatar-based
system, and these are analyzed and labeled with categori-
Avatar-based Multimodal User Interface cal emotion labels (neutral, sad, happy, surprised, and an-
The user interface of the system is web-based with an em- gry) using the face recognition engine from (Wolf, Hassner,
bedded anthropomorphic Embodied Conversational Agent and Taigman 2011). Since health counselors only carry out
178avatar to interact with, building that way, a patient-physician
concordance that could lead to a strong working alliance
(Cooper et al. 2003).
Lastly, modeling empathy for the patient was a key point
when individuals were asked how they felt about the level of
expression of emotions of the avatars. A 45% agreed on the
suitability of the expression of emotions while a 36% said
the avatar’s facial expressions were not intense enough.
When the last group of individuals was asked why the level
of expressions was not appropriate, they argued the need for
more positive emotions such as smiles and body movement.
Although the number of subjects was small, these results
seem to indicate that using avatars for health promoting in-
terventions might lead to user’s higher engagement, lower
Figure 3: User can choose race-concordant avatar.
attrition rates than in text-only computer-based health sys-
tems, and overall increased divulgation of sensitive issues
due to increased confidentiality.
partial mimicry to retain a positive overtone during the inter-
action, we model semi-mimicry so that the avatar displays a 5 Conclusion
sad expression if the user’s expression is identified as sad,
and smiling or neutral otherwise. The demand for health promotion interventions that are tai-
lored to an individual’s specific needs is real, and the tech-
4 Avatar Evaluation Results nological advance that we described to deliver MI behavior
change interventions with avatars is an attempt to address
In a recent survey study, eleven individuals with different
that need. From our current promising results, we anticipate
levels of technicality were asked to interact with the system
that avatar health counselors will be very well accepted by
and evaluate their experience based on functionality and on
users once the field has matured from further research on the
how much the users liked interacting with the avatar.
topic. We hope this research will contribute to promoting
Specific results such as the willingness of individuals to
user’s engagement with health intervention program, and,
discuss alcohol consumption with an embodied conversa-
ideally, increases positive outcomes.
tional agent (ECA) showed that 75% of users felt as com-
fortable or more comfortable interacting with the avatar
counselor than they would with a human counselor: 6 Acknowledgements
37.5% felt as comfortable, 37.5% felt even more comfort- We would would like to thank Eric Wagner for introducing
able, and a remaining 25% said it was less comfortable than us to MI and how it is used in his NIAAA-funded adolescent
with a human. behaviors and lifestyle evaluation (ABLE) program. This re-
Although the small group that was uncomfortable with the search was supported in part by NSF grants HRD-0833093.
ECA argued that body language – currently lacking – was
important to obtain feedback, the groups in favor pointed References
out that because the avatar could not judge nor embarrass
them, they felt more positive and comfortable. These results Allen, J.; Ferguson, G.; Blaylock, N.; Byron, D.; Cham-
are in line with research findings that patients that engage in bers, N.; Dzikovska, M.; Galescu, L.; and Swift, M. 2006.
behavior that can put them at risk (e.g. excessive drinking, Chester: Towards a personal medication advisor. Journal of
unsafe sex, over-eating) tend to report more information to a Biomedical Informatics 39(5):500–513.
computer interviewer than to its human counterpart (Servan- Becker-Asano, C. 2008. WASABI: Affect simulation for
Schreiber 1986). The subjects in favor of the ECA interac- agents with believable interactivity. Ph.D. Dissertation, Uni-
tion further pointed out the benefit of the system in terms versity of Bielefeld, Faculty of Technology.
of availability when human counselors are not present, or Bellet, P., and Maloney, M. 1991. The importance of em-
simply not needed at all times. pathy as an interviewing skill in medicine. Journal of the
In a general context, the statistic of whether individuals American Medical Association 266:1831–1832.
would consider an ECA to discuss personal topics was
73% positive. According to 64% of individuals, the imple- Bickmore, T., and Picard, R. 2005. Establishing and
mentation of this system using animated avatars made the maintaining long-term human-computer relationships. ACM
interaction engaging. This fact can become substantial to Transactions on Computer-Human Interaction (TOCHI)
address high attrition rates found in textual computer-based 12(2):293–327.
health intervention (Pallonen et al. 1998). Bickmore, T.; Pfeifer, L.; and Jack, B. 2009. Taking the time
The availability of a variety of avatars based on race, to care: empowering low health literacy hospital patients
gender and physical appearance caused a positive impact with virtual nurse agents. In Proceedings of the 27th interna-
on individuals with a 91% rate of acceptance of diverse tional conference on Human factors in computing systems,
avatars. People liked the fact that they could choose which 1265–1274. ACM.
179Botelho, R. 2004. Motivational Practice: Promoting Miller, W. R., and Rollnick, S., eds. 2002. Motivational
Healthy Habits And Self-care Of Chronic Diseases. MHH interviewing: 2nd Ed. Preparing people to change addictive
Publications. behavior. New York: Guilford Press.
Boukricha, H., and Wachsmuth, I. 2011. Empathy-Based Miller, W., and Rose, G. 2009. Toward a theory of motiva-
Emotional Alignment for a Virtual Human: A Three-Step tional interviewing. American Psychologist 64(6):527.
Approach. KI - Künstliche Intelligenz 25(3):195–204. Miller, W., Sovereign, R. and Krege, B. 1988. Motiva-
Burke, B. L. and Arkowitz, H. and and Dunn, C. 2002. The tional Interviewing with Problem Drinkers: II. The Drinker’s
Efficacy of Motivational Interviewing and Its Adaptation. Check-Up as a Preventive Intervention. Behavioural Psy-
In Motivational Interviewing: Preparing People for Change. chotherapy 16:251–268.
New-York, NY: Guilford Press, 2nd editio edition. 217–250. Nasoz, F. and Lisetti, C. L. 2006. MAUI avatars: Mirror-
Cassell, J., and J. Sullivan, S. Prevost, & E. Churchill, E. ing the user’s sensed emotions via expressive multi-ethnic
2000. Embodied Conversational Agents. Social Psychology facial avatars. Journal of Visual Languages & Computing
40(1):26–36. 15 Issue(5):430–444.
Cooper, L.; Roter, D.; Johnson, R.; Ford, D.; Steinwachs, Noar, S.; Benac, C.; and Harris, M. 2007. Does tailoring
D.; and Powe, N. 2003. Patient-centered communication, matter? Meta-analytic review of tailored print health behav-
ratings of care, and concordance of patient and physician ior change interventions. Psychological bulletin 133(4):673.
race. Annals of internal medicine 139(11):907–15. Paleari, M.; Grizard, A.; and Lisetti, C. L. 2007. Adapting
Psychologically Grounded Facial Emotional Expressions to
Ehrlich, S.; Schiano, D.; and Sheridan, K. 2000. Commu-
Different Anthropomorphic Embodiment Platforms. In Pro-
nicating facial affect: It’s not the realism, it’s the motion. In
ceedings of the FLAIRS Conference.
CHI’00 extended abstracts on Human factors in computing
systems. NY: ACM Press. Pallonen, U. E.; Velicer, W. F.; Prochaska, J. O.; Rossi, J. S.;
Bellis, J. M.; Tsoh, J. Y.; Migneault, J. P.; Smith, N. F.;
Emmons, K. M., and Rollnick, S. 2001. Motivational in- and Prokhorov, A. V. 1998. Computer-based smoking
terviewing in health care settings. Opportunities and limita- cessation interventions in adolescents: Description, feasibil-
tions. American journal of preventive medicine 20(1):68–74. ity, and six-month follow-up findings. Informa Healthcare
Hester, R. K., Skire, D.D., and Delaney, H. 2005. The 33(4):935–965.
Drinker’s Check-up: 12-month outcomes of a controlled Pelachaud, C. 2009. Modelling multimodal expression of
clinical trial of a stand- alone software program for problem emotion in a virtual agent. Philosophical transactions of
drinkers. Journal of Substance Abuse Treatment 28(2):159– the Royal Society of London. Series B, Biological sciences
169. 364(1535):3539–48.
Johnson, W.; LaBore, C.; and Chiu, Y. 2004. A pedagogical Predinger, H., and Ishizuka, M. 2004. Life-Like Characters.
agent for psychosocial intervention on a handheld computer. Springer, NY.
In AAAI Fall Symposium on Dialogue Systems for Health Prochaska, J. O., and Velicer, W. F. 1997. The transtheoret-
Communication, 22–24. ical model of health behavior change. American journal of
Lisetti, C. L., and Wagner, E. 2008. Mental health pro- health promotion AJHP 12(1):38–48.
motion with animated characters: Exploring issues and po- Rogers, C. 1959. A Theory of Therapy, Personality
tential. In Proceedings of the AAAI Spring Symposium on and Interpersonal Relationships as Developed in the Client-
Emotion, Personality and Social Behavior. Centered Framework. In Koch, S., ed., Psychology: the
Lisetti, C. L.; Brown, S.; Alvarez, K.; and Marpaung, A. Study of a Science, Vol. 3. New-York: McGraw-Hill. 184–
2004. A social informatics approach to human-robot inter- 256.
action with an office service robot. IEEE Transactions on Schulman, D.; Bickmore, T.; and Sidner, C. 2011. An Intelli-
Systems, Man, and Cybernetics - Special Issue on Human gent Conversational Agent for Promoting Long-term Health
Robot Interaction. 34(2). Behavior Change using Motivational Interviewing. In 2011
Lisetti, C. L. 2008. Embodied Conversational Agents for AAAI Spring Symposium Series, 61–64.
Psychotherapy. In Proceedings of the CHI 2008 Conference Servan-Schreiber, D. 1986. Artificial intelligence and psy-
Workshop on Technology in Mental Health. chiatry. Journal of Nervous and Mental Disease 174:191–
Magerko, B.; Dean, J.; Idnani, A.; Pantalon, M.; and 202.
D’Onofrio, G. 2011. Dr. Vicky: A Virtual Coach for Silverman, B.; Holmes, J.; Kimmel, S.; Branas, C.; Ivins,
Learning Brief Negotiated Interview Techniques for Treat- D.; Weaver, R.; and Chen, Y. 2001. Modeling emotion and
ing Emergency Room Patients. In AAAI Spring Symposium. behavior in animated personas to facilitate human behavior
Magnenat-Thalmann, N. and Thalman, D. 2004. Handbook change: the case of the HEART-SENSE game. Health Care
of Virtual Humans. Wiley and Sons, Ltd. Management Science 4(3):213–228.
MATCH, G. P. R. 1998. Matching Alcoholism Treatments to Wolf, L.; Hassner, T.; and Taigman, Y. 2011. Effective un-
Client Heterogeneity: Project MATCH Three-year Drinking constrained face recognition by combining multiple descrip-
Outcomes. Alcoholism: Clinical and Experimental Research tors and learned background statistics. PAMI, IEEE Trans-
23(60):1300–1311. actions on 33(10):1978–1990.
180You can also read
