Comprehensive Summary
Cartwright and colleagues developed a natural language processing (NLP) model to identify patients with firearm violence exposure using electronic health records (EHRs). Clinical notes from OCHIN, a nonprofit healthcare network spanning 40 U.S. states, were collected between January 2012 and December 2022. An NLP text classifier built in Python was trained to detect direct and indirect exposure to firearm violence. While promising, the model was subject to potential biases, including counterfactual information, adversarial inputs, and embedding-related bias. To mitigate risks, a Stakeholder Advisory Committee was established to oversee lexicon development, monitor model performance, and guide clinical application. The authors note the need for future work including younger patients (<5 years old), more recent datasets, and broader population samples.
Outcomes and Implications
This research study is particularly poignant because despite the high incidence of firearm violence, structured data is largely limited to acute primary injury. This results in a lack of support and research regarding those who faced exposure to firearms without injury. Through developing a model which can identify patients with exposure to firearm violence, patients can be provided with the appropriate mental health care and monitored for both short and long term impacts of exposure. Clinically, this is relevant when determining how to proceed with longitudinal healthcare, and to tailor healthcare plans to patients’ specific needs.