# Evaluating the accuracy of the Veterans Health Administration’s REACH VET suicide prediction model for legal involved veterans

## TL;DR

A suicide prediction model used by the VA has very low accuracy, especially for veterans with legal system involvement.

## Contribution

This study evaluates the accuracy of the REACH VET model for suicide prediction in legally-involved veterans.

## Key findings

- The model has a 0.05% positive predictive value for predicting death by suicide overall.
- False negative rates are extremely high, at 98.1% overall and 82.4% for legally-involved veterans.
- Accuracy for the combined outcome of suicide attempt or death is better but still low.

## Abstract

The Veterans Health Administration’s (VA) Recovery Engagement and Coordination for Health–Veterans Enhanced Treatment (REACH VET) suicide prevention program uses a model that estimates every patient’s suicide risk. We evaluated predictive accuracy of that model for death by suicide, and separately for the combined outcome of suicide attempt or death, for all VA patients and for patients with criminal legal system involvement who have one of the highest suicide rates among VA users. We found that the model has low accuracy for the prediction of death by suicide (positive predictive value = 0.05% overall, 0.10% for legal-involved; false negative rate = 98.1% overall, 82.4% for legal-involved). For the combined outcome, model accuracy was better and more so for legal-involved veterans, but overall still low. While the REACH VET program has been shown to have benefits, the underlying model that identifies patients at high risk of suicide needs improvement.