OBJECTIVES: Our study proposes to use two Machine Learning (ML) based methods to predict Opioid Use Disorder (OUD) in patients who are prescribed opioids and to identify the associated high impact predictors.

METHODS: The study design is a longitudinal retrospective cohort using IQVIA Pharmetrics data covering a 9-year period from 2007-2015. The index date was established as the date of the first opioid prescription, and OUD was identified using ICD-9 codes. We constructed predictive models using Random Forests (RF) and Support Vector Machines (SVM) to predict OUD, and performed feature selection to identify high impact parameters from a list a social, demographic, clinical and pharmacy variables. We validated our models using 5-fold cross-validation

RESULTS: The study population included 927,395 individual patients, each with at least one opioid prescription. The final analytical dataset included 97 variables. Both SVM and RF had an area under the curve metric of 1, however, the RF model had a better precision (SVM-0.93, RF-0.97), while the SVM model had a better recall (SVM-0.99, RF-0.97). These models were further used to identify the top 25 high impact parameters. Both models had a high mean accuracy of nearly 99% after cross-validation for classification. On comparing the top 25 variables identified by each method it was observed that there was a 50% concordance on the variables identified by the methods, however, their ranking differed significantly.

CONCLUSIONS: Using two ML based methods we have predicted OUD in patients exposed to prescription opioids with high sensitivity and accuracy. The models were also used to perform feature selection so as to identify the high impact variables to include in our subsequent statistical models.