Automating the Addiction Behaviors Checklist for Problematic Opioid Use Identification

Area of Science:

• Medical Informatics
• Natural Language Processing
• Pain Management

Background:

• Opioid use disorder (OUD) is a significant risk for individuals with chronic pain managed by opioids.
• Electronic health records (EHR) offer potential for large-scale studies on problematic opioid use.
• Diagnostic codes for OUD are often unreliable and underutilized, necessitating improved identification methods.

Purpose of the Study:

• To evaluate the efficacy of regular expressions, an interpretable natural language processing (NLP) technique, in automating the identification of problematic opioid use.
• To compare the performance of this automated method against a validated clinical tool (Addiction Behaviors Checklist) and traditional diagnostic codes.

Main Methods:

• A retrospective cohort study analyzed deidentified EHR data from 8063 individuals with chronic pain from 2021-2023.
• Free-text clinical notes, demographics, and diagnostic codes were extracted.
• The automated approach was validated against a manually reviewed holdout set and an independent external dataset of 100 patients.

Main Results:

• The automated regular expression method demonstrated superior performance compared to diagnostic codes.
• At the primary site, F1 scores were 0.73 vs 0.08 and AUCs were 0.82 vs 0.52.
• At the validation site, F1 scores were 0.70 vs 0.29 and AUCs were 0.86 vs 0.59.

Conclusions:

• Automated data extraction using regular expressions can effectively identify patients with problematic opioid use.
• This technique facilitates earlier identification of at-risk individuals and enables new research avenues.
• It offers a more reliable alternative to diagnostic codes for studying opioid use in chronic pain populations.