Vermont Lab Uses Machine Learning to Guide Palliative Care Conversations

Researchers at the University of Vermont’s Conversation Lab are using machine learning to develop algorithms designed to optimize clinician-patient discussions of serious illness, palliative care and end-of-life care.

The research team — led by Robert Gramling, M.D., associate professor of family medicine and the Miller Chair in Palliative Medicine at the University of Vermont’s Larner College of Medicine has conducted several studies into features of these difficult conversations to identify common characteristics of effective communication that can be replicated.

“The purpose of the lab is to understand and promote high quality communication in serious illness care, and the spirit of the group is such that we welcome scholars and thinkers from multiple disciplines to think about what an effective conversation is and what a good conversation might look like,” Gramling said. “We are looking into how might we measure these kinds of things; so we have a pretty robust group of scholars interested in communication has really added to our thinking about how we approach serious illness conversation.”


In the course of seeking best practices, the group has also identified some potential pitfalls. For instance, they recently determined that clinicians have the tendency to overestimate a patient’s life expectancy if the patient and family seem optimistic during discussions of care plans, disease trajectory and the patient’s health care goals and wishes.

For that project, the Conversation Lab research team enrolled 189 hospital patients with advanced cancer undergoing palliative care consultations at two geographically distant sites. A total of 41 palliative care clinicians participated in the recorded consultations.

The group calculated the frequency and distribution of such variables as “clinician overestimation of survival time,” “patient (trait) dispositional optimism,” and “patient prognostic (state) optimism” and tracked patient survival and date of death and correlated it to clinical judgement.


The group’s findings indicated a correlation between higher levels of patient optimism and clinicians’ greater likelihood of overestimating survival, even after adjusting for clinical markers of survival time.

“Clinicians are often imperfect in determining how long a person is likely to live, and similar to other disciplines we sometimes overestimate length of survival, Gramling told Hospice News. “Curiously, when patients self-assess with regards to optimism and to what degree to they tend to expect the best, that self-rated trait is associated with clinicians overestimating how long people will live, even controlling for all the different clinical factors that we take into account.”

This work has implications for hospice for which eligibility depends on a six-month terminal prognosis. Hospice stakeholders agree that too many patients are admitted to hospice too late in the course of their illness to receive the full benefit of those services. Improving the accuracy of prognoses can help hospice patients begin receiving services as early as possible.

Ongoing research in the Conversation Lab includes an examination of pauses during serious illness conversations and consideration of what they might indicate. The team developed an algorithm that identifies and collects data about pauses in these conversations, with ongoing follow-up research to determine whether a human could identify and interpret the same kind of pauses if they know what to look for.

“We are seeking to determine if, when people stop talking, are their measurable features to the sounds before [a pause], the sounds after, or maybe even during a pause that can help us know whether this might be a marker of listening and presence, or a time of pivotal change of dynamic in a conversation, or whether it indicates distraction or awkwardness or some other factor,” Gramling said. “Some of the preliminary work we did demonstrated that there is a musicality to the sounds that happen around these pauses. That leads us to believe that we could train an algorithm to find these acoustic features of pauses that identify connection.”

Companies featured in this article: