Talk to the Veterans Crisis Line now
U.S. flag
An official website of the United States government

VA Health Systems Research

Go to the VA ORD website
Go to the QUERI website
HSRD Conference Logo



2023 HSR&D/QUERI National Conference Abstract

Printable View

1048 — Developing a domain-specific vocabulary for extraction of 3-step Theory of Suicide concepts from clinical progress notes

Lead/Presenter: Alex Harris,  COIN - Palo Alto
All Authors: Meerwijk EL (Center for Innovation to Implementation, VA Palo Alto Health Care System, Menlo Park), Jones GA (Center for Innovation to Implementation, VA Palo Alto Health Care System, Menlo Park) Shotqara AS (Center for Innovation to Implementation, VA Palo Alto Health Care System, Menlo Park) Reyes SI (Center for Innovation to Implementation, VA Palo Alto Health Care System, Menlo Park) Eddington HS (Center for Innovation to Implementation, VA Palo Alto Health Care System, Menlo Park; Department of Surgery, Stanford University, Stanford) Tamang SR (Center for Innovation to Implementation, VA Palo Alto Health Care System, Menlo Park; Department of Biomedical Data Science, Stanford University, Stanford) Reeves RM (VA Tennessee Valley Healthcare System, Nashville; Department of Biomedical Informatics, Vanderbilt University Medical Center, Nashville) Finlay AK (Center for Innovation to Implementation, VA Palo Alto Health Care System, Menlo Park; National Center on Homelessness Among Veterans; Schar School of Policy and Government, George Mason University, Arlington) Ilgen MA (Center for Clinical Management Research, Ann Arbor; Department of Psychiatry, University of Michigan, Ann Arbor) Harris ASH (Center for Innovation to Implementation, VA Palo Alto Health Care System, Menlo Park, CA; Department of Surgery, Stanford University, Stanford)

Objectives:
Despite our knowledge of risk and protective factors for suicidal behavior, accurately predicting a Veteran’s suicide or suicide attempt remains a challenge. To improve this situation, we oriented our investigation toward the state-of-the-art 3-step Theory of Suicide (3ST), which describes why people consider suicide and who will act on their suicidal thoughts and attempt suicide. The central concepts of 3ST - psychological pain, hopelessness, connectedness, and capacity for suicide - are missing from VHA’s clinical suicide risk prediction models. As these four concepts are not systematically recorded in structured fields of VHA’s electronic healthcare records, we developed a domain-specific vocabulary that enables automated extraction of these concepts from clinical progress notes using natural language processing. Here, we report on our methods to develop the vocabulary and initial extraction results obtained with a preliminary vocabulary.

Methods:
We purposefully selected 40 VHA stations across the US and obtained clinical progress notes for Veterans at those stations who attempted suicide or who died by suicide. We focused on notes in temporal proximity of the suicidal event of encounters in mental health care, primary care, general (acute) medicine, and urgent and emergency care. Unique mentions in those progress notes related to psychological pain, hopelessness, connectedness, and capacity for suicide were annotated and added to the domain-specific vocabulary. Annotations for each concept were guided by a schema that was based on suicide theory initially and expanded as the annotation effort progressed. The schema defined the concept and its child concepts.

Results:
Based on clinical notes from 160 Veterans annotated so far, the preliminary vocabulary for 3ST concepts contains about 1400 entries for psychological pain divided across 15 child concepts (e.g. pain exceeds tolerance, avoid shame or guilt, releasing pain), 350 entries for hopelessness divided across three child concepts (feelings, motivation, and expectation), 900 entries for connectedness divided across 12 child concepts (e.g. loneliness, support, purpose), and 1000 entries for capacity for suicide divided across four child concepts (acquired, practical, dispositional, and situational). Early piloting of this vocabulary involved automated concept extraction from a test corpus of 215,000 clinical notes of 650 randomly selected Veterans from each VA station who attempted suicide or who died by suicide. Among frequently mentioned terms from the vocabulary were grief, overwhelmed, tearful, feelings of guilt, chronic pain, judged, upset, feelings of hopelessness, despair, supportive family, engaged in treatment, grateful, lives with, religious, suicide attempt, jump off, cut his, access to firearms, and substance abuse.

Implications:
VHA clinical progress notes contain ample mentions related to psychological pain, hopelessness, connectedness, and capacity for suicide, to develop a domain-specific vocabulary for automated extraction of these concepts from clinical progress notes.

Impacts:
We demonstrate that automated extraction of psychological pain, hopelessness, connectedness, and capacity for suicide from VHA clinical progress notes is possible. Validation of these extractions and their ability to improve current suicide risk prediction models used in mental health operations are part of subsequent aims of this study.