Predicting Firearm Suicide among US Army Veterans Transitioning from Active Service

Area of Science:

• Public Health
• Veterans Affairs
• Machine Learning Applications

Background:

• United States (US) Veterans have higher suicide rates than civilians.
• Machine learning (ML) models aid in identifying high-risk individuals for suicide prevention.
• Firearms are the primary method in most Veteran suicides, necessitating targeted interventions.

Purpose of the Study:

• To develop and compare firearm-specific ML models against a method-agnostic model.
• To predict firearm suicides among transitioning US Army Veterans up to 10 years post-discharge.
• To evaluate model performance and fairness across different intervention thresholds.

Main Methods:

• Utilized data from US Army Veterans (2010-2019, N=800,579).
• Developed and compared the predictive accuracy of firearm-specific ML models versus a method-agnostic model.
• Assessed model performance using AU-ROC and ICI metrics, considering fairness across sex and race/ethnicity.

Main Results:

• Both firearm-specific and method-agnostic models demonstrated comparable overall performance (AU-ROC=0.710-0.708).
• The method-agnostic model excelled at predicting firearm suicides at higher intervention thresholds.
• The firearm-specific model performed better at lower thresholds and demonstrated superior fairness across demographic groups.

Conclusions:

• Model selection for Veteran suicide prevention depends on balancing predictive accuracy and fairness.
• Optimal thresholds may differ for coordinated firearm-specific and method-agnostic suicide prevention interventions.
• Firearm-specific models offer advantages in fairness, crucial for equitable suicide risk assessment in Veterans.