Date of Award

May 2021

Degree Type


Degree Name

Doctor of Philosophy


Biomedical and Health Informatics

First Advisor

Jake Luo

Committee Members

Michelle Maloney, Han Joo Lee, Hobart Davies, Jennifer Fink


IntroductionOpioid use disorder has continued to rise in prevalence across the United States, with an estimated 2.5 million Americans ailing from the condition (NIDA, 2020). Medically managed detoxification incurs substantial costs and, when used independently, may not be effective in preventing relapse (Kosten & Baxter, 2019). While numerous studies have focused on predicting the factors of developing opioid use disorder, few have identified predictors of readmission to medically managed withdrawal at an inpatient level of care. Utilizing a high-fidelity dataset from a large multi-site behavioral health hospital, these predictors are explored.

MethodsPatients diagnosed with Opioid Use Disorder and hospitalized in the inpatient level of care were analyzed to identify readmission predictors. Factors including patient demographics, patient-reported outcome measures, and post-discharge treatment interventions were included. Patients re-hospitalized to the inpatient level of care were binary labeled in the dataset, and various machine learning algorithms were tested, including machine learning techniques. Methods include random forest, gradient boosting, and deep learning techniques. Evaluation statistics include specificity, accuracy, precision, and Matthew's Coefficient.

ResultsOverall, there was a wide variation if correctly predicting the class of patients that would readmit to a medically managed level of inpatient detoxification. Out of the six models evaluated, three of the six did not converge, thus not producing a viable feature ranking. However, of the other three models that did converge, the deep learning model produced almost perfect classification, producing an accuracy of .98. AdaBoost and the logistic regression model produced an accuracy of .97 and .61, respectively. Each of these models produced a similar set of features that were important to predicting which patient profile would readmit to medically managed inpatient detoxification.

ConclusionsThe results indicate that overall reduction in the Quick Inventory of Depressive Symptomology, discharge disposition, age, length of stay, and a patient's total number of diagnoses were important features at predicting readmission. Additionally, deep learning algorithms vastly outperformed other machine learning algorithms.