Natural Language Processing - A Surveillance Stepping Stone to Identify Child Abuse

Area of Science:

• Medical Informatics
• Natural Language Processing
• Child Abuse Detection

Background:

• Child abuse injuries require timely and accurate identification in emergency settings.
• Existing methods for identifying abuse-related injuries may be inconsistent.
• Natural Language Processing (NLP) offers potential for automated analysis of clinical notes.

Purpose of the Study:

• To refine an NLP algorithm for identifying child abuse-related injuries.
• To identify opportunities for integrating NLP into real-time clinical decision support (CDS) tools.
• To improve the identification and evaluation of potential child abuse cases in emergency departments (EDs).

Main Methods:

• An NLP algorithm was applied in 'silent mode' to 353 ED provider notes across 9 EDs.
• Algorithm triggers were refined to enhance specificity and reduce false positives.
• Adherence to clinical guidelines and demographic disparities in evaluation/reporting were assessed.

Main Results:

• 73 cases falsely triggered the NLP due to interpretation errors; refinements improved specificity.
• Adherence to recommended evaluation standards was 63%.
• Significant demographic disparities in evaluation and reporting were observed based on ED type, insurance, and race/ethnicity.

Conclusions:

• NLP algorithm refinement in silent mode improved accuracy for detecting child abuse injuries.
• Real-time CDS integration can assist ED providers in identifying and evaluating injuries associated with child physical abuse.
• Addressing demographic disparities is crucial for equitable care.