PROJECT SUMMARY/ABSTRACT
Promoting high quality communication in serious illness is a pressing national priority. However, traditional
approaches to measuring communication in the natural clinical setting are cumbersome, expensive, not
confidential and time-consuming. In order to understand, disseminate, support and incentivize high quality
communication, we need measurement methods that can be scaled for larger clinical studies and routine
health service quality reporting. This proposal leverages extraordinary data that we have at our disposal and
our team's expertise in machine learning to automatically measure features of serious illness communication
quality in the natural clinical setting. We focus on episodes of Connectional Silence that occur between
clinicians and patients because these moments are markers of effective listening, engaged shared decision-
making and compassionate presence. The Palliative Care Communication Research Initiative (PCCRI) is a
multi-site cohort study of a diverse population of 240 hospitalized patients with advanced cancer and 54
palliative care clinicians. Participants represent diversity in age, gender, racial and ethnic backgrounds,
religious affiliations, levels of financial insecurity, linguistic repertoire, personality and treatment preferences.
The cohort data include 363 audio-recorded palliative care inpatient consultations comprising more than
10,000 minutes of conversation, 180,000 speaker turns, and 1.4 million words. We have completed the
following Preliminary Studies in preparation for this proposed work: (a) developed a supervised machine
learning algorithm to identify over 28,000 conversational pauses in speaking amid the noisy hospital
environment (e.g., oxygen masks, IV pumps, televisions, hallway voices, etc.),(b) developed a reliable Human
Coding approach for identifying Connectional Silence as a distinct type of conversational pause, and (c)
successfully trained preliminary deep neural networks with both lexical and sound data surrounding pauses in
the PCCRI data. Our team is well positioned to overcome traditional barriers to direct measurement of
communication quality in healthcare settings, advance machine learning technology for analysis of clinical
conversations, and ultimately, improve communication for all seriously ill people. We propose to complete the
following over the two-year award period: Aim One) to train a Deep Neural Network that uses lexical features
of the immediate surrounding conversation to accurately distinguish moments of Connectional Silence from
other types of pauses; Aim Two) Using de-identified spectral images of sound surrounding conversational
pauses, to develop an integrated deep learning algorithm that uses both lexicon and prosody of speech data to
distinguish moments of Connectional Silence from other types of conversational pauses. Exploratory Aim) to
identify the features of communication surrounding Connectional Silence that are associated with patients
feeling more Heard & Understood.