Domain-Enhanced Attention Enabled Deep Network for Behavioral Health Identification from 911 Narratives

Appropriately addressing behavioral health challenges is a problem in the United States. Behavioral health challenges can manifest in the form of crimes committed by individuals with mental illness, social and domestic violence, etc., which may result in emergency calls to first responders. For example, in domestic violence situations, involved parties may need various supports and resources if no criminal charges are filed. In these instances, when a 911 call is made, it mobilizes the entire first responder system (e.g., police, fire, EMT), each of whom write their respective incident reports. A majority of the calls that first responders are addressing are social and behavioral in nature, which they are often ill-equipped to handle. The recurring nature of these calls and limited available resources, first responders encounter challenges in providing appropriate responses to community residents. Therefore, early identification of these cases through incident reports could guide service providers and first responders to take appropriate actions on behalf of the community resident. However, the presence of noise and contextual variability within these unstructured public narrative reports makes it challenging for identifying social and behavioral health based cases. To address this challenge, in this study, we propose a novel model to analyze the first responders’ public narratives: the model learns contextual features automatically from the public narratives by employing self-attention techniques and incorporates automatically extracted domain knowledge by comparing expert-given keywords to represent document features to identify behavioral health cases. Unlike conventional approaches, integrating domain expert knowledge with the contextual representation of the public narrative report enriches the model’s understanding of social and behavioral health cues and enhances its overall performance. Extensive evaluation showcases the efficacy and effectiveness of our proposed novel model. Our evaluation shows that our model identifies behavioral health with an accuracy of 82% and F1-score of 86%, outperforms the baseline models.