Predicting suicides after outpatient mental health visits in the Army Study to Assess Risk and Resilience in Servicemembers (Army STARRS)

Abstract

The 2013 U.S. Veterans Administration/Department of Defense Clinical Practice Guidelines (VA/DoD CPG) require comprehensive suicide risk assessments for VA/DoD patients with mental disorders but provide minimal guidance on how to carry out these assessments. Given that clinician-based assessments are known not to be strong predictors of suicide, we investigated whether a precision medicine model using administrative data after outpatient mental health specialty visits could be developed to predict suicides among outpatients. We focused on male non-deployed Regular U.S. Army soldiers because they account for the vast majority of such suicides. Four machine learning classifiers (naïve Bayes, random forests, support vector regression, elastic net penalized regression) were explored. 41.5% of Army suicides in 2004-2009 occurred among the 12.0% of soldiers seen as outpatient by mental health specialists, with risk especially high within 26 weeks of visits. An elastic net classifier with 10-14 predictors optimized sensitivity (45.6% of suicide deaths occurring after the 15% of visits with highest predicted risk). Good model stability was found for a model using 2004-2007 data to predict 2008-2009 suicides, although stability decreased in a model using 2008-2009 data to predict 2010-2012 suicides. The 5% of visits with highest risk included only 0.1% of soldiers (1047.1 suicides/100,000 person-years in the 5 weeks after the visit). This is a high enough concentration of risk to have implications for targeting preventive interventions. An even better model might be developed in the future by including the enriched information on clinician-evaluated suicide risk mandated by the VA/DoD CPG to be recorded.

Spatial Proximity to Incidents of Community Violence Is Associated with Fewer Suicides in Urban California

“…This study offers numerous strengths over previous studies of the association between exposure to community violence and suicide [8][9][10][11][23][24][25] by estimating easily interpretable risk difference parameters, employing objective exposure and outcome measures, and utilizing comprehensive statewide data at fine geographic and temporal resolution. A primary limitation of previous studies is that exposure to community violence is highly correlated with other life adversity (poverty, neighborhood deprivation, domestic violence, lack of access to social services) at the individual and community levels, and disentangling these effects is challenging.…”

Letter to the Editor: Suicide as a complex classification problem: machine learning and related techniques can advance suicide prediction - a reply to Roaldset (2016)

“…Applying this approach to suicide prediction has already shown promise. Recently, Kessler et al (2015) applied machine learning methods to a large sample of Army soldiers, which yielded an area under the curve (AUC) of 0.85 -a figure considerably stronger than what the literature has been able to produce over the last five decades (i.e. AUC = 0.56; J. C. Franklin et al unpublished data).…”

Machine learning in mental health: A systematic scoping review of methods and applications

“…[266,267]); and creating models of risk to improve health system delivery (e.g. [268,269]) (see Table 3). Public health applications typically used social media data (n=11), electronic health records (n=6), and clinical data (e.g., diagnostic surveys and tools; n=9).…”

“…ML applied to electronic health records was demonstrated to predict suicide risk with an accuracy similar to clinician assessment [268,271], as well as predict dementia and its risk factors with high accuracy [272]. Research has also investigated the use of ML with clinical data to improve variable selection in epidemiological data analysis [273], and to better understand the relationship between complex risk factors for mental health conditions such as depression [274].…”

“…Regression [281,282], DT [282], SVM [282], RF [281], Super Learner [281] Interview [282], Survey [281] Psychiatric Emergency BN [269], DT [269], SVM [269] Social Media [269] Psychiatric Stressors NLP [283], Named-entity recognition [283] Clinical Notes [283] Psychosis Regression [284], RF [285] Clinical Assessment [285], Electronic Health Records [284] Social Support LIWC [267], SVM [267] Social Media [267] Stress Cluster analysis [286], Sentiment Analysis [270], SVM [270] Clinical Assessment [286], Social Media [270] Substance Use PCA [265], NLP [265], RF [285] Social Media [265], Clinical Assessment [285] Suicide/Self-Harm ARM [271], DT [271], Genetic Algorithm [287], NB [268,271], RF [268,271], Regression [268,271,[288][289][290], SVM …”