Vital Signs: Assessing Conversation Quality
How do you gauge the quality of a conversation?
Even a chat between two friends could be
considered constructive by one but difficult for the
other. A good talk is subjective.
Bob Gramling, M.D., D.Sc., wants to
pinpoint the attributes of an effective conversation,
specifically one between a palliative medicine
specialist and a patient with a serious illness. That
way, the healthcare system could measure the
value of serious illness conversations for patient
care — the same way it would a drug treatment
or procedure, says Gramling, chief of palliative
medicine at the UVM Medical Center and the Holly
and Bob Miller Chair of Palliative Medicine in the
Robert Larner, M.D. College of Medicine.
To assess something as complex and dynamic
as a conversation, Gramling solicited the help
of Maggie Eppstein, Ph.D., chair of the UVM
Department of Computer Science and a founder
of the University’s Complex Systems Center,
which applies high-level mathematical modeling
techniques to real-world challenges, and Donna
Rizzo, Ph.D., UVM professor of engineering and
computer science, who specializes in machinelearning
tools for complex environments.
Together with national scholars from
linguistics, communication science, anthropology,
nursing, health services research, psychology and
epidemiology, they comprise the new Vermont
Conversation Lab at the Larner College of
Medicine. The mission of the VCL is “to understand
and promote high-quality communication in
serious illness,” Gramling says.
They are using audio recordings of almost 400
palliative care consultations collected from around
the country, funded by the American Cancer Society,
for which Gramling is the principal investigator.
Currently, the team isn’t focusing on the
actual words in the conversation but, rather, the
times when the speaking stops. Some types of
silence indicate “moments of connection” in a
conversation, Gramling explains.
Usually, discussions between doctors and
patients are rushed, he says. “Oftentimes there
isn’t space for people’s voices to be heard. So what
silence can offer is a recognition that what you’ve
just told me is important, and if you want to tell
me more, I’m going to give you space to tell me.”
With discussions of illness and death, a
palliative specialist might just sit for a moment
when a patient expresses fear or sadness, he says.
“The pause and that space provides a potential
moment of connection. And it’s not just in
palliative care. It’s in any conversation.”
Capturing silence amid the constant
din of a hospital setting poses a challenge.
Machines beep. Nurses talk in the hallway.
TVs blare. The computer must learn to focus
only on the conversation and the moments
that conversation stops. It also must discern
the difference between “distracted” silence,
perhaps when a physician pauses to take notes,
and purposeful or “contemplative” silence.
Eppstein, Rizzo and Viktoria Manukyan,
a Complex Systems graduate student doing
her thesis on computer analysis of silence, are
using different techniques of machine learning
— including decision trees and artificial neural
networks — to pinpoint the characteristics of
the silence they seek. From there, they’ll develop
an algorithm for computers to identify that
silence with high accuracy.
“The point of it is being able to create
a tool that would aid in potentially training
or assessment of quality of conversations in
a variety of applications,” even beyond the
medical setting, Eppstein says. “To date, there is
no way of assessing the quality of conversations
in medicine, even though it’s critical to good
care, especially to palliative medicine.”
Without methods to seamlessly and
meaningfully measure actual conversation
quality, researchers and health policy makers
generally have relied on patient feedback to
evaluate a conversation. These questionnaires
can be quite valuable as important outcomes
of communication (eg. “How much do you feel
heard and understood by the doctors, nurses
and hospital staff caring for you?”) but tend
not to accurately reflect the content or process
of the actual conversation.
The Vermont Conversation Lab aims to
create an objective, automated measure that
could be used routinely in natural healthcare
settings. Voice-recognition and machinelearning
technology can pick up clues in real
time without recording the conversation or
collecting identifying information in order to
maintain confidentiality.
The technology would allow collection of
data for research on conversations and assist
with palliative care training, so students can learn
techniques and receive immediate feedback.
Perhaps most important, it would give
hospitals and other medical practices a means
to evaluate conversational quality and to
support — with financial incentives and other
means — those systems that do it best.