Using natural language processing to identify opioid use disorder in electronic health record data

Area of Science:

• Public Health
• Health Informatics
• Epidemiology

Background:

• Rising opioid prescriptions correlate with increased opioid use disorder (OUD) and adverse outcomes.
• Accurate epidemiological surveillance of OUD is crucial for effective prevention strategies but faces challenges.
• Electronic health records (EHR) offer a potential data source for OUD surveillance.

Purpose of the Study:

• To compare the effectiveness of two methods for identifying OUD in EHR data: natural language processing (NLP) and ICD-10-CM diagnostic codes.
• To ascertain the prevalence of OUD using both NLP and ICD-10-CM codes.
• To evaluate the contribution of NLP in identifying OUD cases missed by diagnostic codes.

Main Methods:

• EHR data from hospital and emergency department visits between 2017-2019 were analyzed.
• A rule-based NLP algorithm was developed and refined using a stepwise process with EHR data.
• ICD-10-CM discharge codes were extracted, and NLP was applied to unstructured clinical notes.

Main Results:

• A combined approach using NLP and ICD-10-CM codes identified 2,332 unique OUD cases, a 29.5% increase over ICD-10-CM codes alone (6.1% vs. 7.9% prevalence).
• NLP identified 521 OUD cases (22.3%) not captured by ICD-10-CM codes, while ICD-10-CM codes identified 430 cases (18.4%) missed by NLP.
• The NLP algorithm demonstrated high accuracy with an estimated sensitivity of 81.8% and specificity of 97.5% compared to expert manual review.

Conclusions:

• NLP-based algorithms effectively automate data extraction and identify OUD from unstructured EHR data.
• Combining NLP with ICD-10-CM codes provides the most comprehensive ascertainment of OUD in EHR.
• NLP should be integrated into epidemiological studies utilizing EHR data for improved OUD surveillance.