Predicting Opioid Use Disorder (OUD) Using A Random Forest

Abstract

Opioid Use Disorder (OUD), defined as physical or psychological reliance on opioids, is quickly becoming a public health epidemic. This research demonstrates how supervised machine learning can be used to predict adults at risk for OUD by considering interactions between various demographic, socioeconomic, physical, and psychological features in an integrated manner. A labeled data set was built from the responses to the 2016 edition of the National Survey on Drug Use and Health (NSDUH). This labeled data set was used to train and test a random forest classifier while accounting for class imbalance. The classifier can predict adults at risk for OUD accurately (sensitivity = 0.81, specificity = 0.76, AUC = 0.86), although the prevalence of OUD is only about 1%. Early initiation of marijuana (prior to 18 years of age) emerges as the dominant predictor for developing OUD in adult life. This is surprising because it ranks higher than both mental illness and disability; two conditions that are often comorbid with substance use disorders. Thus, curbing early initiation of marijuana may be the best prevention strategy. This highlights the crucial role that educators, counselors, and parents can play in alleviating the United States' opioid overdose crisis.