Addressing diagnostic code variability in intimate partner violence surveillance through natural language processing:...

Intimate partner violence (IPV) and substance use disorders (SUDs) represent major public health challenges, yet identifying IPV cases in electronic health records (EHR) remains difficult due to inconsistent documentation practices and variable billing procedures.

We constructed three patient cohorts based on SUDs or overdoses involving stimulants, opioids, or both from the EHR. Using the Open Health Natural Language Processing (OHNLP) toolkit, we developed a rule-based classifier to identify IPV in clinical notes. We validated classifier performance through manual chart review and linked cohorts to Kentucky's fatal overdose surveillance system to examine the relationship between IPV and fatal overdose risk.

We analyzed 15,557,678 clinical notes from 29,447 patients with SUDs (2017-2023). Our classifier demonstrated strong performance: recall 0.91, precision 0.85, F1-score 0.89. IPV was detected in 5.6% of patients via NLP compared to 0.2-2.5% using various ICD-10-CM diagnostic code definitions; this variation reflected different levels of code specificity across published definitions. Notably, 90% of NLP-detected cases lacked corresponding diagnostic codes. The cohort with both stimulant and opioid disorders showed the highest IPV prevalence (7.5%), followed by opioid-only (5.2%) and stimulant-only (4.8%). Among fatal overdose cases, IPV documentation rates (4.4% via NLP, 5.1% via diagnostic codes) were similar to the overall cohort, suggesting missed intervention opportunities.

NLP-based analysis of EHR clinical notes identified substantially more IPV cases than diagnostic codes alone. The elevated IPV prevalence among people with polysubstance use highlights a particularly vulnerable population requiring integrated screening and intervention strategies. Routine NLP surveillance could significantly improve IPV case identification in populations with SUDs.