Development and Validation of a Natural Language Processing Tool to Identify Injuries in Infants Associated With...

Area of Science:

• Medical Informatics
• Pediatric Emergency Medicine
• Child Abuse Detection

Background:

• Clinicians may overlook subtle injuries in infants that are indicative of physical child abuse.
• Failure to recognize these injuries can allow abuse to continue or escalate, posing significant risks to child welfare.

Purpose of the Study:

• To develop a natural language processing (NLP) algorithm for identifying infant injuries associated with physical abuse.
• To evaluate the accuracy of the developed NLP algorithm in detecting these high-risk injuries.

Main Methods:

• An NLP algorithm was created to detect ten specific injuries linked to infant abuse.
• The algorithm was refined through iterative review and applied to 1344 emergency department (ED) provider notes.
• Results were validated against expert review of the same clinical notes.

Main Results:

• The NLP algorithm demonstrated high accuracy in identifying high-risk injuries in infants.
• Sensitivity was 92.7% and specificity was 98.1%.
• Positive predictive value was 60.3% and negative predictive value was 99.8%.

Conclusions:

• An NLP algorithm can accurately identify infants with injuries suggestive of child abuse in ED settings.
• This tool shows promise in assisting clinicians to better recognize and address potential child abuse cases.