Using natural language processing to identify child maltreatment in health systems

Area of Science:

• Medical Informatics
• Public Health Surveillance
• Child Health Research

Background:

• Child maltreatment (CM) rates in electronic health records (EHRs) are significantly underestimated compared to national prevalence.
• Improved documentation and surveillance methods are crucial for accurate CM identification.

Purpose of the Study:

• To evaluate natural language processing (NLP) for identifying CM in outpatient notes, beyond ICD codes.
• To compare CM detection rates between ICD coding and NLP.
• To analyze demographic differences in CM identification by age, gender, and race/ethnicity.

Main Methods:

• Utilized outpatient chart notes (2018-2020) for children aged 0-18 years from Kaiser Permanente Washington.
• Applied NLP to identify maltreatment-related terms categorized by concept unique identifiers (CUIs).
• Manually reviewed text snippets for validation and NLP algorithm retraining.

Main Results:

• NLP identified CM in 1.55%-2.36% of notes, a substantially higher rate than ICD codes (3.32 per 1000 children).
• NLP detected CM at a rate of 37.38 per 1000 children.
• Adolescents, females, Native American children, and those on Medicaid showed the most significant increase in CM identification with NLP.

Conclusions:

• NLP substantially enhances the estimated number of children affected by CM.
• Improved identification of vulnerable youth through NLP can aid in early intervention for mental health risks.