A Machine Learning Approach to Identify Socio-Economic Factors Responsible for Patients Dropping out of Substance Abuse Treatment

Abstract

In recent years, the subject of substance abuse has drawn considerable attention from researchers and policymakers alike. Researchers have been utilizing the wealth of patient level data available from various agencies to develop prediction models for the relationship between socioeconomic factors and substance abuse issues. According to the Substance Abuse and Mental Health Services Administration (SAMHSA), in 2017, 26% of patients admitted to treatment facilities drop out prematurely, which is significant when considering that roughly 1.5 million people are admitted to these treatment facilities every year, thereby revealing the need for an analysis to identify variables associated with such a large number of people not completing treatment. This study applies Multiple Logistic Regression (MLRM) as well as Random Forest Classification (RF) model to determine significant socioeconomic factors responsible for patients prematurely dropping out of substance abuse treatment for opioid misuse. A MLRM has its limitations when the dataset has a large number of categorical variables; machine learning methods such as RF have proven more effective and accurate when dealing with such data. Patient level data from the Treatment Episode Dataset-Discharge (TEDS-D 2017) was analyzed and the models were compared using the Area Under the Curve (AUC) operating characteristic. The MLRM was found to have an AUC of .68 while the RF model had an AUC of .89, thereby demonstrating the advantage of machine learning methods. The factors deemed significant from the RF model can provide healthcare professionals as well as administrative officials with the necessary information to help address the issue of patients prematurely dropping out of opioid misuse treatment.