Natural Language Processing for Identification of Hospitalized People Who Use Drugs: Cohort Study

Area of Science:

• Health Informatics
• Clinical Research
• Public Health

Background:

• People who use drugs (PWUD) face higher risks of severe infections.
• Current identification methods using billing codes are inaccurate, potentially underrepresenting diverse patient groups.
• This limits understanding of economic, racial, and ethnic disparities among hospitalized PWUD.

Purpose of the Study:

• To evaluate Natural Language Processing (NLP) for improved identification of PWUD in electronic medical records.
• To specifically identify underrepresented populations, including low-income and minoritized racial/ethnic groups.
• To compare NLP's effectiveness against traditional identification methods.

Main Methods:

• A cohort of hospital admissions (2020-2022) was identified using ICD-10 codes, toxicology results, OUD medication prescriptions, and NLP keyword detection.
• Admissions were categorized as highly documented (all criteria) or minimally documented (NLP-only).
• Chart review served as the gold standard to calculate positive predictive value and assess racial, ethnic, and social vulnerability index impacts.

Main Results:

• NLP achieved a 54% positive predictive value, surpassing traditional methods in identifying PWUD hospitalizations.
• NLP significantly enhanced identification when integrated with other criteria.
• Racial/ethnic minorities and those with lower social vulnerability index had less PWUD-related documentation.

Conclusions:

• NLP is an effective tool for identifying hospitalizations of people who use drugs, exceeding traditional methods.
• NLP demonstrates potential for reducing healthcare disparities among PWUD.
• Further NLP refinement is recommended for comprehensive patient identification.