review: Issues in Designing Agents for Long Term Behavior Change

Introduction
Some people are not ready to change their health
behavior, such as stopping smoking, or improving
exercising or dieting, even though they may know all of
the reasons for change. Such individuals will not likely
take the first step toward change, even when asked to
do so by a counselor or therapist. These individuals,
who admit to no plans or willingness to change, are
referred to as “precontemplators” within the framework
of the transtheoretical (stages of change) model of
health behavior change [20]. They represent a
particular challenge for human and automated health
counselors alike, and have not received much attention
in the literature on automated health behavior change
to date.


Challenges
Minimizing Repetitiveness
One surprising finding from prior longitudinal studies of
health counseling agents we have conducted was that,
even though dialogue scripts had been authored to
provide significant variability in each days' interaction,
most participants found the conversations repetitive at
some point during the month, and because of this
many lost motivation to follow the agent’s advice [4,5].
As one participant put it, “It would be great if Laura
could just change her clothes sometimes.” This
repetitiveness was more than an annoyance; some
subjects indicated that it negatively impacted their
motivation to exercise (e.g., “In the beginning I was
extremely motivated to do whatever Laura asked of
me, because I thought that every response was a new
response.”). The amount of behavioral, linguistic and
visual variability required to avoid the perception of
robotic repetitiveness remains an open research
question.

Establishing Therapeutic Alliance
The therapeutic alliance – the strength of the bond
between counselor and patient, and their mutual
agreement on the goals and tasks of therapy – is a key
component of successful change across a wide variety
of different counseling methods and strategies [6]. We
believe that establishing a strong therapeutic alliance
between an agent and its user will, similarly, be a key
requirement in maintaining engagement through the
course of a long-term behavior change intervention.
The patient’s assessment of the therapeutic alliance
tends to be established early, within the first few
sessions; this early assessment is relatively stable over
time, and predictive of successful outcomes [17]. The
behavior of the agent within the first few interactions
with the user must be carefully designed, as a failure to
develop a strong alliance may be difficult to correct
later. The alliance also tends to develop through a
cycle of short-term ruptures and repairs [21]; a
successful agent must be able to assess and respond to
these variations in its working relationship with users.
Maintaining Persistence Across Counseling Sessions
In order to perform counseling actions that span more
than one session, in addition to demonstrating
continuity in the working relationship [10], the agent
must remember something about its past encounters
with users. Many schools of psychotherapy involve
giving patients some form of “homework” to do in
between counseling sessions, and most behavioral
techniques (e.g., shaping and positive reinforcement)
require that patients’ past behavior and/or goals be
remembered for comparison purposes. At a minimum,
the fact that the agent has interacted with a given user
before, and perhaps the number and/or duration of
such interactions must be remembered between
sessions. Persistent memory should ultimately be
represented as an episodic store recording details of all
past interactions with users. A useful middle ground is
to record specific facts that can be referenced in future
conversations. Examples in the physical activity
coaching domain include remembering the name of a
user’s walking buddy or favorite walking location, as
well as purely social (off-task) facts, such as the user’s
favorite television program and whether they had any
big plans for the upcoming weekend. In our system, a
user model is loaded from a relational database at the
start of each counseling session, and saved back out at
the end of the session, in order to provide persistence
across sessions.
Authoring Counseling Dialogue
Devising an efficient and effective process for authoring
large quantities of dialogue for longitudinal interaction
is a significant challenge of this project. We must
provide sufficient content for up to 30 conversations
per user, encompassing as large a range of user
situations as possible. The content should be reviewed
by experienced counselors or other domain experts.
Finally, the content should be modular and reusable, to
ease the implementation effort of future systems. In
previous work [5], we have used dialogue systems
based on augmented transition networks. This simple
formalism has been usable by domain experts with
relatively little training. Thus far we have found that
our current hierarchical task modeling approach is
significantly more difficult to understand and author.
Whether these increases in authoring difficulty result in
commensurate improvements in reusability remains to
be seen.
Eliciting Open-Ended Responses
Open-ended questions and continuation prompts (“tell
me more about that”) are used extensively in
motivational interviewing to get clients talking about
their own motivations for change. Eliciting this kind of
information through multiple-choice menus represents
one of the biggest hurdles to our use of hierarchical
task models to emulate this style of counseling.
Approaches we have taken include: keeping user
responses very abstract; providing “drill down” trees to
index desired statements from general categories to
specific responses; and using knowledge about users
gleaned from other sources (e.g., enrollment web
forms) that they can simply endorse during a
counseling session. None of these approaches is
entirely satisfactory, and this remains an open area of
research.
Conclusion
We have described many challenges and open research
problems in building a re-usable health counseling
system for longitudinal health behavior change
interventions. We plan to conduct initial testing of an
exercise promotion intervention based on this
framework early in 2009 and then proceed to porting
the framework for use in a diet intervention (fruit and
vegetable.

From:

Timothy Bickmore
Daniel Schulman
Northeastern University
College of Computer and
Information Science
360 Huntington Ave, WVH202
Boston, MA 02115
bickmore@ccs.neu.edu
laurap@ccs.neu.edu
schulman@ccs.neu.edu