Firearm injury risk detection and prevention

The 2023 Annual Report from the Johns Hopkins Center for Gun Violence Solutions reports an analysis of the 2021 firearm fatality data released from the Centers for Disease Control and Prevention in January of 2023.¹ The report highlighted alarming facts:

• 48 830 lives lost to gun violence—an all-time record

• 81% of all homicides and 55% of suicides were by firearm

• Guns were the leading cause of death among children and teens accounting for more deaths than COVID-19, car accidents, or cancers

• Gun violence was the cause of 51% of deaths for Black teens, with Black teens and young men (ages 15-34 years) accounting for 2% of the population but 36% of all gun homicide fatalities

• Gun suicides reached record levels increasing 8.3% from 2020

• Those 75 and older are at the greatest risk of dying by gun suicide with a rate twice the national average

• Gun death rates vary widely by location across the United States

In the area of school shootings, the Washington Post reports that 352 000 children have experienced gun violence at school since Columbine in 1999 (https://www.washingtonpost.com/education/interactive/school-shootings-database/). Moreover, the median age of a school shooter is 16 years, and children were responsible for more than half the US school shootings. The Post also found that when the source of the gun could be determined in school shootings by children, 86% of the weapons were found in the homes of friends, relatives, or parents. A 2022 study reports that 4.6 million children live in homes with at least 1 gun that is loaded and unlocked. Thus, increasing the risk of gun violence among children and youth.²

There is no doubt that there is a need for wide-scale adoption and implementation of policies that support firearm injury risk detection and prevention, and a 2023 survey by the Johns Hopkins Center for Gun Violence Solutions reports strong support (72%) across political lines for requiring a license from a law enforcement agency before purchasing a firearm and laws that support locking up firearms when not in use.³ Moreover, 69% of Americans support funding community-based gun violence prevention programs that provide outreach, conflict mediation, and social support for people at high risk of gun violence. Complementary to policy strategies, innovative health informatics and data science approaches have a role to play in this alarming public health crisis. I hope that the 5 papers highlighted in this editorial will motivate further team science and methods innovation from our community.

Zhou et al. developed a predictive model to identify healthcare encounters of adult patients at increased risk of firearm injury to target screening and prevention efforts.⁴ Using electronic health record data of cases (fatal and non-fatal firearm injuries) and matched controls from Kaiser Permanente Southern California, they identified more than 170 predictors, including diagnoses, healthcare utilization, and neighborhood characteristics. They used extreme gradient boosting (XGBoost) and a split sample design to train and test a model that predicted risk for firearm injury within the next 3 years at the encounter level. Firearm injury prevalence at the healthcare encounter level was 0.01%. The 15 most important predictors in the model included demographics, healthcare utilization, and neighborhood level socioeconomic factors, yielding sensitivity and specificity of the final model of 0.83 and 0.56, respectively. A very high-risk group (top 1% of predicted risk) yielded a positive predictive value of 0.14% and sensitivity of 13%. Their findings suggest that focusing on the very high-risk group would reduce screening burden by a factor of 11.7, compared to universal screening in this low prevalence setting.

Based on the premise that access to firearms is associated with increased suicide risk, Trujeque et al. compared the performance of 6 natural language processing approaches to characterize firearm access in Veterans Health Administration (VHA) clinical notes of 36 685 patients.⁵ They expanded pre-existing firearm term sets using subject-matter experts and generated 250-character snippets around each firearm term appearing in notes. Annotators labeled 3000 snippets into 3 classes, which were then used to compare 4 non-neural machine learning models (random forest, bagging, gradient boosting, logistic regression with ridge penalization) and 2 large language models (BioBERT and Bio-ClinicalBERT) for classifying firearm access as “definite access,” “definitely no access,” or “other.” Firearm terms were identified in 41.3% patient records; 33.7% of snippets were categorized as definite access, 9.0% as definitely no access, and 57.2% as “other.” Five of six models had acceptable performance, with BioBERT performing best (weighted F1 = 0.88). The authors conclude that given the high frequency of firearm-related terms in clinical notes of VHA patients, the ability to use text to identify and characterize patients’ firearm access could enhance suicide prevention efforts by identifying patients for clinical interventions.

Ancona et al. applied a supervised machine learning model, least absolute shrinkage and selection operator (LASSO) regression, to the St Louis region-wide hospital-based violence intervention program data repository (2010-2020) and compared the performance of the final 61-variable model to other models for distinguishing between new vs follow-up encounters for firearm injury.⁶ In a predominantly male and Black sample with a median age of 26 years, the machine learning model had excellent discrimination (0.92, 0.88-0.96) with high sensitivity (0.95, 0.90-0.98) and specificity (0.89, 0.81-0.95) and had significantly higher specificity than the comparative approaches (proxy measures of emergency department visits and time between firearm injury encounters, and diagnostic code encounter type designation [initial vs subsequent or sequela]). Accurate classification of firearm injury encounters provides an essential foundation for further analyses.

With a focus on interpersonal firearm violence, Kafka et al. developed a key term search and trained supervised machine learning classifiers (Naïve Bayes, decision tree, and random forest) to label court record documents (n = 1472) from alleged violent crimes as firearm-related or not firearm-related.⁷ The key term search (71 terms) had the highest recall with no false negatives but lowest precision. The decision tree performed best in terms of F1 score, with its logic equivalent to conducting a keyword search for “gun,” “shoot,” “handgun,” “bullet,” “revolver,” and “rifle.” The findings of this Case Report from a single state (Washington) support the efficiency of a simple automated method to label court records of alleged violent crimes.

Based on the premises that that (1) nutrition fact labels and drug fact labels have practical utility and (2) Model Cards⁸ and TRIPOD⁹ checklists are designed for developers and researchers and lack utility for others and rarely address biases related to race and ethnicity, the Perspective by Zhu et al. proposes criteria and template for a generic, extendable Model Facts label.¹⁰ They argue its relevance to firearm injury given the demographics reflected in data sources for firearm injury research and illustrate application of the Model Facts label to a violence risk identification model and a suicide risk prediction model. Comprising four sections: application, accuracy, demographics, and warnings, the Model Facts label is designed to enhance transparency as a foundation for trust. The authors suggest that further testing of the label and education about its use are required.

These research and applications papers, case report, and perspective offer a mere glimpse of the potential for health informatics and data science to address the important topic of firearm injury and I look forward to receiving future contributions.

Funding

None declared.

Conflict of interest

None to declare.

References