Predicting firearm suicide in military veterans outside the VA health system using linked civilian electronic health record data

ABSTRACT. The ongoing epidemic of suicide among former U.S. military personnel—17 deaths every day—
lies at the core of a 20-year trend of increasing suicide rates in the U.S. The rate of suicide in veterans is about
1.5 times that of the civilian population, due to veterans' unique burden of medical, psychological, and social-
environmental risk factors compounded by easy access to lethal means. To date, veteran suicide research and
prevention efforts have focused almost entirely on the population served by the Veterans Health Administration
(VHA). Meanwhile, most veterans do not seek VHA care but prefer private-sector health services. Since 2005,
suicides among veterans outside the reach of VHA have increased at more than double the rate seen among
VHA users (57% vs. 28%, respectively). This study's primary objective is to develop efficient longitudinal
predictive algorithms for suicide and firearm-related suicide among military veterans who utilized non-
VHA health care, by analyzing the largest database ever assembled of linked civilian medical record
data pertinent to veteran suicide risk. Too little is known about veterans receiving care outside the VHA,
including the nature and severity of their health conditions, their patterns of healthcare utilization, and their
unique risk factors for all suicide and firearm-related suicide. Filling these gaps in knowledge is crucial to the
goal of meaningfully reducing suicide in the veteran population overall. To that end, our multi-disciplinary team
of nationally distinguished researchers will assemble and analyze an unprecedented longitudinal database of
linked VA and Department of Defense data, health records, social indicators, and death records of veterans
receiving health care from 5 large civilian health systems in North Carolina and Utah. The database will yield
an estimated 3.8 million person-year observations, including approximately 900 firearm-involved suicides and
1,190 total suicide deaths. We will analyze these data to describe the demographic and health characteristics
of veterans who utilize non-VHA healthcare services, their patterns of healthcare utilization, their mortality
outcomes, and their incidence of suicide deaths, by method. We will use machine learning methods to develop
specific risk algorithms for predicting all suicides and firearm-related suicides among veterans who utilize non-
VHA healthcare, to identify veterans at elevated risk of suicide. Utah-based collaborators will use linked VHA
data to identify and describe risk patterns for veterans who combine VHA and non-VHA healthcare. Finally, we
will conduct a series of key informant interviews to better understand barriers and facilitators to integrating this
type of algorithm into civilian health system workflows. In summary, the proposed work will fill critical gaps in
the literature by leveraging large-scale, real-world data sources to yield novel knowledge of suicide risks, while
informing prevention efforts aimed at reducing veteran suicide. We will also gather implementation information
to inform how large civilian health systems will be able to use this information to identify and intervene with the
veterans who are at greatest risk of suicide within their patient populations.