Background: Prescription tranquilizer misuse is a risky behavior associated with fatal drug poisonings. Although various predictors have been examined, there is no published prediction model for tranquilizer misuse. This study develops and internally validates a tranquilizer misuse prediction model based primarily on drug histories of participants in a national cross-sectional survey. Predictors also include psychiatric, behavioral and demographic variables.
Methods: We analyzed data from 471,097 individuals aged 14-to-29-years in the United States, as sampled by the National Survey of Drug Use and Health, 2004-2018, an annual cross-sectional survey. We encoded 21 predictors with known or likely onset prior to tranquilizer misuse initiation, (e.g., early onset of cannabis use). With this dataset, we trained a neural network and regularized logistic regression model. While the assessment for tranquilizer misuse changed slightly in 2015, by pooling all years of survey data, predictions are robust to this source of variation.
Results: 1.44% of the pooled sample, 2004-2018, recently initiated tranquilizer misuse (unweighted estimate). On held-out test data (n = 43,714), logistic regression and the neural network performed equally well, with an area under the receiver operating characteristic curve (AUC) of ∼0.83 on the primary model, containing 12 variables known to occur before tranquilizer misuse.
Conclusion: Built for case prediction rather than case detection, this model restricted predictors to those with known timing prior to initiation of tranquilizer misuse. Yet its performance supersedes commonly accepted criteria for clinical prediction models (AUC > 0.80). Future work should incorporate survey analysis weights into the prediction model to minimize possible bias.
Keywords: Benzodiazepines; Drug Dependence; Extra-Medical; Machine Learning; Muscle Relaxants; Nonmedical Use; Xanax.
Copyright © 2020 Elsevier B.V. All rights reserved.