Problem
About one in twelve adults in the United States has a substance use disorder. Research into genetic risk, early identification, and treatment needs large cohorts pooled across many health systems, and that requires a reliable, shareable way to say who has an SUD and who doesn’t.
Diagnostic codes don’t provide it. SUD is under-documented, partly because of stigma, and studies report sensitivity as low as 0.17 for opioid use disorder codes. Much of the evidence sits instead in free-text clinical notes.
Approach
- Probabilistic phenotypes from multiple sources. We combine billing codes, medications, and clinical notes into phenotypes that express uncertainty and can travel between institutions.
- Interpretable NLP. We favor methods clinicians can audit, such as regular expressions with negation and context rules, packaged as sudregex.
- Automating validated instruments. We translate clinician-validated tools, such as the Addiction Behaviors Checklist, into rules that run over notes at scale.
- Better reference labels. Phenotypes are only as good as the chart review behind them, so we also study how to make expert review faster without losing accuracy.
Selected results
Automating the Addiction Behaviors Checklist reached an F1 score of 0.73 for problematic opioid use among 8,063 patients with chronic pain, against 0.08 for ICD codes. At an external validation site the figures were 0.70 and 0.29. (JAMA Psychiatry 2025)
A systematic review screened 623 studies of NLP for SUD information extraction and included 35. Two thirds (65.7%) used rule-based approaches, and most reported acceptable-to-outstanding discrimination. (Current Addiction Reports 2026)