HEARD & UNDERSTOOD™

The Vermont Conversation Lab team of interprofessional scientists have established international expertise in scalable conversation analytics, particularly those that focus on markers of human connection and meaning-making. We continue to create, refine and apply our HEARD & UNDERSTOOD ^TM suite of scalable conversation analytic methods for many natural healthcare environments, including hospital based palliative care consultations, telehealth consultations and psychedelic assisted therapy settings.

We support open science by publishing our work in peer reviewed outlets and, when we do, by releasing the prototype programming code or described approach in sufficient detail to allow other scientists the opportunity to understand, evaluate and replicate our methods. We further curate these algorithms over time to improve their performance in previously validated clinical contexts, to retrain and validate them in additional clinical environments, and to engineer the prototype programming to software ready standards.

We would love to collaborate with you and your team in one of two ways:

1) You send us your data and we analyze it for you

• We work with you to complete Data Use Agreements and Service Contracts prior to any data sharing

• We prepare and analyze your data using our state-of-the-art, PHI compliant machine learning computational infrastructure

• We provide you with analytic output data files and summary reports for your team to analyze within your study

2)  We send you our algorithms and you analyze your data yourself

• We work with you to complete a licensing agreement for use of the algorithms and trademark

• We send you algorithms in their most updated forms. We are also working to offer “containerized” versions of select algorithms (starting with CONSERT) for research teams who prefer not to use the native programming languages to run our algorithms via a prototype open-source user interface, the HEARD & UNDERSTOOD ^TM App

• We offer optional consultation services to help your team install and apply the algorithm(s) to your data

Interest, inquiries, and to get started: please email gramlinglab@med.uvm.edu

Available Analytic Tools

CONSERT (CONnectional Silence Ensemble-beRT): This is a supervised machine learning pipeline that incorporates multiple representations of de-identified conversational data to find and sub-classify moments of Connectional Silence in serious illness conversations. Our previous work identifies that the presence of these moments predicts patient reported outcomes.

Example peer reviewed manuscripts using CONSERT:

Matt JE, Rizzo DM, Javed A, Eppstein MJ, Gramling CJ, Manukyan V, Dewoolkar A, Gramling R. An Acoustical and Lexical Machine Learning Ensemble to Detect Connectional Silence. Journal of Palliative Medicine. 2023. Dec;26(12):1627-1633.

SOMTimeS (Self-Organizing Maps for Time Series): This is an unsupervised machine learning clustering algorithm that combines the conceptual structure of conversational storytelling that humans use to find and share meaning from our life experiences with the computational methods of Self-Organizing Maps. SOMTimeS processes time-series data and scales linearly, thus overcoming previous computational demands for clustering temporal phenomena.

Example peer reviewed manuscripts using SOMTimeS:

Javed A, Rizzo DM, Suk Lee B, Gramling R. SOMTimeS: Self Organizing Maps for Time Series Clustering and its Application to Serious Illness Conversations. Data Mining & Knowledge Discovery. 2024. 38(3):813-839.

Gramling R, Javed A, Durieux BN, Clarfeld LA, Matt JE, Rizzo DM, Wong A, Braddish T, Gramling CJ, Wills J, Arnoldy FL, Straton J, Cheney N, Eppstein MJ, Gramling D. Conversational Stories & Self Organizing Maps: Innovations for the Study of Uncertainty in Healthcare Communication. Patient Education & Counseling. 2021. Nov;104(11):2616-2621.

CODYM (COnversational DYnamics Model): This Markov model analyzes patterns of information flow in conversations based on sequential dependencies in the lengths of speaker turns and changes in speakers.

Example peer reviewed manuscripts using CODYM:

Clarfeld L, Gramling R, Rizzo D, Eppstein M. A general [Markov] model of conversational dynamics with application in serious illness communication. PLoS One. 2021. Jul 1;16(7):e0253124.

TEVA (Tandem Evolutionary Algorithm): This is a non-linear optimization method for identifying the importance of conversational features that best distinguish complex outcomes, including persons feeling Heard & Understood in healthcare environments. Among TEVA’s strengths is attention to equifinality – that different patterns in complex phenomena, like conversations, may lead to the same outcome.

Example peer reviewed manuscripts using TEVA:

Hanley JP, Rizzo DM, Buzas JS, Eppstein MJ. A Tandem Evolutionary Algorithm for Identifying Causal Rules from Complex Data. Evol Comput. 2020 Spring;28(1):87-114. doi: 10.1162/evco_a_00252. Epub 2019 Feb 28. PMID: 30817200.

 NLP Uncertainty Dictionary

Description: The Uncertainty Dictionary is a list of terms that naturally occur in clinical serious illness conversations indicating the presence and sub-types of uncertainty.

List of terms here: https://vermontconversationlab.com/uncertainty-corpus/)

Citation: Gramling R, Javed A, Durieux BN, Clarfeld LA, Matt JE, Rizzo DM, Wong A, Braddish T, Gramling CJ, Wills J, Arnoldy FL, Straton J, Cheney N, Eppstein MJ, Gramling D. Conversational Stories & Self Organizing Maps: Innovations for the Study of Uncertainty in Healthcare Communication. Patient, Education & Counseling. 2021. Nov;104(11):2616-2621.

NLP Palliative Care Clinical Talk (PCCT) Dictionary

Description: The PCCT Dictionary is a list of symptom, treatment and prognosis terms that commonly occur in palliative care consultations with people who are hospitalized with advanced cancer. 

List of terms here (as described in Ross et. al. below):

Symptoms: comfortable, worried, tired, painful, symptom, shortness, hurting, confused, uncomfortable, weak, happy, comfort, sleepy, depressed, hurts, symptoms, pain, breathing, cough, constipation, dry, energy, appetite, awake, hurt, coughing, sleep, breathe, strength, breath, sleeping, bothering, nausea, strong, anxiety, wake, scary, depression, worry, stronger, anxious

Treatment: morphine, patch, medications, drug, trial, CPR, line, tylenol, button, doses, drugs, medical, feeding, oxygen, ativan, oxycodone, therapy, dilaudid, chemotherapy, machine, antibiotics, treatment, radiation, surgery, treat, dose, meds, medicines, fluids, tube, hospice, medicine, dialysis, methadone, oral, ventilator, milligrams, management, resuscitation, fentanyl, chemo, pill, nutrition, ICU, milligram, medication, procedure, liquid, treatments, IV, pills

Prognosis: cure, future, dying, die, prognosis, probably, hope, risk, hoping, death

Citation:  Ross L, Danforth C, Eppstein MJ, Clarfeld LA, Durieux BN, Gramling CJ, Hirsch L, Rizzo DM, Gramling R. Story Arcs in Serious Illness: Natural Language Processing Features of Palliative Care Conversations. Patient Education & Counseling. 2020. April;103(4):826-832.

NLP Temporal Reference Tagger

Description: The Temporal Reference Tagger uses conversational lexicon to distinguish when speakers are alluding to the past, present or future. (more description here)

Link to Python source code: https://vermontconversationlab.com/sdm_downloads/vcl-temporal-reference-tagger-code/

Citation: Ross L, Danforth C, Eppstein MJ, Clarfeld LA, Durieux BN, Gramling CJ, Hirsch L, Rizzo DM, Gramling R. Story Arcs in Serious Illness: Natural Language Processing Features of Palliative Care Conversations. Patient, Education and Counseling. 2020. April;103(4):826-832.

Coming soon