Use of prescription opioids 'beyond the bounds' of medical guidance can lead to opioid dependence. Yet recent efforts to predict extra-medical use of prescription pain relievers (EMPPR) have relied on electronic medical or pharmacy records. Because peak incidence of EMPPR occurs during adolescence- a time of relative health- administrative data may be inadequate. In this study, with data from a United States (US) population sample, we develop and internally validate an EMPPR prediction model. We analyzed data from 234,593 individuals aged 12-to-17-years, as sampled by the US National Survey of Drug Use and Health, 2004-2018, an annual cross-sectional survey. We encoded 14 predictors with onset prior to EMPPR initiation, including age, sex, and facets of drug and psychiatric history. We ranked these predictors by clinical utility before sequentially adding each to a regularized logistic regression model. On held-out test data (n = 23,685), the model performs well with 14 predictors, with an area under the precision recall curve (AUPRC) is 0.155. The area under the receiver operator curve (AUC) is 0.819, exceeding a recent benchmark on this dataset. Results are robust to survey redesign that occurred in 2015, and are not moderated by past-year use of medical services. In conclusion, while selection of predictors is limited to those with known timing prior to initiation of EMPPR rather than any cross-sectional variable, this model discriminates well. Good classification occurs even with a small set of clinically available predictors- age, a history of depression and alcohol, cigarette, and cannabis use.
Keywords: Drug dependence; Extra-medical; Machine learning; Nonmedical use.
Copyright © 2021 Elsevier Inc. All rights reserved.