Citation

Baker-Robinson W, Bhakta Y, Krushnic D, Maxim L, DeFrancesco S, Carlson K. Inj. Prev. 2022; 28(Suppl 1): A58.

Copyright

(Copyright © 2022, BMJ Publishing Group)

DOI

10.1136/injuryprev-2022-SAVIR.149

PMID

unavailable

Abstract

SAVIR 2022 Conference Abstracts

Statement of Purpose CDC Firearm Injury Syndrome Definitions are used to detect firearm injury related emergency department visits in Oregon Electronic Surveillance System for the Early Notification of Community-based Epidemics (OR ESSENCE) data. This study examines the validity of the CDC Firearm Injury Syndrome Definitions, and methods for improving firearm injury detection using natural language processing.

Methods/Approach We manually reviewed all OR ESSENCE records identified as firearm injuries by the CDC Firearm Injury V2 Syndrome Definition between April 1, 2017 and June 30, 2021 (n=2,386). Two reviewers coded each record for occurrence of firearm injury and classification of injury intent. Using manual classifications as criterion standards, we measured relevant operating characteristics for each CDC Firearm Injury Syndrome Definition. Our team then created several supervised natural language processing (NLP) and rule-based models for the classification of firearm injury intent using triage notes.

Results Of reviewed records, most were deemed true firearm injuries (PPV=87.0%). The overall CDC Firearm Injury Definition identified 79.0% of records based on patients' discharge diagnoses. The intent specific definitions were found to have low sensitivity and low positive predictive value, with the exception of the Unintentional Syndrome Definition (Sensitivity=90.8%). This was due to the overuse of ICD-10 diagnosis codes for unintentional firearm injury. Use of supervised NLP and rule-based models significantly improved the sensitivity and positive predictive value for detection of firearm injury intent.

Conclusion The overall CDC Firearm Injury Syndrome Definition performed relatively well; however, the intent definitions performed poorly. Accuracy of discharge diagnosis coding had a large effect on the validity of the intent definitions. Supervised NLP and rule-based models, have shown promising results for the classification of firearm injury intent using triage notes data.

Significance This work helps to further the rapid detection and dissemination of firearm injury-related emergency department visit data from OR ESSENCE.

Language: en