Machine Learning Analysis of Handgun Transactions to Predict Firearm Suicide Risk

Abstract

Key Points Question Can handgun purchasing records, coupled with machine learning techniques, be used to forecast firearm suicide risk? Findings In this prognostic study of nearly 2 million individuals with handgun transaction records, among transactions classified in the riskiest 5%, close to 40% were associated with a purchaser who died by firearm suicide within 1 year. Among the small number of transactions with a random forest score of 0.95 and above, more than two-thirds were affiliated with a purchaser who died by firearm suicide within 1 year (24 of 35). Meaning This study suggests that passively collected administrative data on handgun transactions may be used to inform targeted interventions based on risk stratification.

Predicting Firearm Suicide—Small Steps Forward With Big Data

“…It is also imperative to consider the potential uses of this model in relation to its accuracy. The model by Laqueur et al 1 shows that, depending on the random forest threshold, sensitivity varied from 0.39 to 0.75, specificity varied from 0.70 to 0.95, the negative predictive value was steady at 0.99, and the positive predictive value (PPV) was low at 0.002 to 0.145 (area under the curve, 0.81).…”

“…Many who die by firearm suicide may not self-disclose or seek help from a medical professional. Laqueur et al look instead to the question of whether firearm suicide risk might be predicted—and those at risk helped—at or around the time of firearm purchase. Using California’s database of nearly 5 million handgun purchases, they applied machine learning techniques to predict subsequent firearm suicide deaths among purchasers.…”

“…Machine learning models can provide insights where clinical trials or meta-analyses may be lacking. Laqueur et al chose a supervised machine learning approach using a random forest model that has previously been applied in prior risk prediction of suicide or suicide behavior . Benefits of such models are that they provide both classification and regression, helping identify those at risk while also showing important variables .…”

“…It is also imperative to consider the potential uses of this model in relation to its accuracy. The model by Laqueur et al shows that, depending on the random forest threshold, sensitivity varied from 0.39 to 0.75, specificity varied from 0.70 to 0.95, the negative predictive value was steady at 0.99, and the positive predictive value (PPV) was low at 0.002 to 0.145 (area under the curve, 0.81). A low PPV does not mean that the model is not useful, as this is common among most suicide prevention models, but a low PPV does warrant careful practice considerations to avoid inappropriate overidentification of risk.…”

“…A low PPV does not mean that the model is not useful, as this is common among most suicide prevention models, but a low PPV does warrant careful practice considerations to avoid inappropriate overidentification of risk. Laqueur et al focus on the proportion of suicides that fall within high-risk ventiles, but this ignores the size of the pool of purchasers within each ventile. A feasible and cost-effective prevention program using the proposed model would need to incorporate information on the number of purchases within each ventile.…”