1029 — Using natural language processing to identify firearm mentions in the electronic health record
Lead/Presenter: Lisa Brenner,
VISN 19 Rocky Mountain MIRECC for Suicide Prevention
All Authors: Simonetti JA (VISN 19 Rocky Mountain MIRECC for Suicide Prevention), Bangerter A (Center for Care Delivery and Outcomes Research, Veterans Affairs Medical Center, Minneapolis, MN) Niu Z (Center for Care Delivery and Outcomes Research, Veterans Affairs Medical Center, Minneapolis, MN; Institute for Health Informatics, University of Minnesota, Minneapolis, MN) Taylor O (Center for Care Delivery and Outcomes Research, Veterans Affairs Medical Center, Minneapolis, MN) Brenner LA (Rocky Mountain Mental Illness Research, Education, & Clinical Center for Suicide Prevention, Veterans Health Administration, Aurora, CO; Departments of Physical Medicine and Rehabilitation, Psychiatry & Neurology, University of Colorado Anschutz School of Medicine, Aurora, CO) Dudley RA (Center for Care Delivery and Outcomes Research, Veterans Affairs Medical Center, Minneapolis, MN; Institute for Health Informatics, University of Minnesota, Minneapolis, MN)
Access to a firearm is an important risk factor for suicide, and 69% of all veteran suicides are attributable to firearm injury. However, suicide risk prediction models, clinical interventions, and research efforts are limited because it is difficult to extract information about firearms from electronic health records (EHR). Our objective is to develop a natural language processing algorithm to identify mentions of firearms in the Veterans Health Administration (VHA) EHR.
We identified firearm-related terms using the Unified Medical Language System (UMLS) Metathesaurus and expanded the term list through two rounds of stakeholder engagement (round 1: academic researchers; round 2: veteran firearm owners). We then identified a national subsample of veterans who died by suicide from May 2020 through April 2022 (n = 2,175) and matched them 1:1 with control patients of the same gender who had no history of suicide attempt and were: alive on that index date; not identified within the EHR as having high suicide risk; born in the same month/year; and enrolled in the same VHA clinical center. We used Structure Query Language to identify and compare firearm mentions between groups using the three terminology sets in all clinical notes written within 90 days of the index date.
Using 25 UMLS terms, 51.9% of suicide decedents and 6.2% of controls had a firearm-related mention in their record. Identification increased using term lists expanded by researchers (65 terms; example term â€œAR-15;â€ 53.7% cases, 7.1% controls) and veteran firearm owners (89 terms; example term â€œRemington;â€ 56.3% cases, 7.7% controls).
Using stakeholders to expand on established firearm terminology, we identified firearm mentions in the EHR of one-half of veteran suicide decedents and only eight percent of living controls. Current work includes developing a word embedding model using a Continuous Bag of Words Model, and incorporating machine and deep learning methods to differentiate between mentions of firearms indicating access or lack thereof.
The availability of such an algorithm may accelerate research and clinical efforts around firearm injury prevention among veterans who receive care in VHA clinical settings.