Improving Early Detection of High-Risk Behavioral Crises Through Natural Language Processing

• In Event: Computational Methods in Criminal Justice Settings

Abstract

Extreme Risk Protection Orders (ERPOs), which prevent individuals at high risk of harming themselves or others from accessing firearms, may be petitioned by law enforcement agencies to prevent gun deaths and injuries. To ensure a timely response and a successful preventive intervention, early identification of individuals that pose an immediate significant danger is paramount. However, specialized units, such as crisis response teams, are often not the first point of contact between high-risk individuals and law enforcement, which can increase the time between the high-risk event, referral to the appropriate follow-up unit, and the preventive intervention. Through Natural Language Processing (NLP) techniques, we develop an early identification tool to process narratives from incident/offense and behavioral crisis reports, flagging potential high-risk events. This study presents the design and development process of the tool, including feature engineering, unstructured data processing, and classification strategies. In addition to discussing preliminary performance metrics, we especially consider the implications and risks associated with the deployment of machine learning methodologies for the given use case, as well as the corresponding preventive measures implemented.

Author

• Ana Ortiz Salazar, Seattle Police Department