Online Program

Comparison of Statistical Modeling and Machine Learning Approaches for Exploring Factors Driving Opioid Prescribing in National Outpatient Health Care Data Using Complex Survey Design (302396)

Chaoran Hu, University of Connecticut 
*Junghi Kim, FDA 
Yong Ma, FDA/CDER 
Qing Pan, George Washington University 
Jaejoon Song, FDA/CDER 
Xiao Tan, George Mason University 

Keywords: Complex survey, machine learning, penalized logistic regression

The opioid crisis represents a large and growing public health burden in the United States. Understanding factors driving opioid prescribing in outpatient settings is vital to implementing targeted interventions to reduce preventable harm from potentially inappropriate opioid prescriptions. The National Ambulatory Medial Care Survey, which collects annual, cross-sectional data from outpatient physician office visits, can be used to undertake such investigation. We applied both machine learning methods and a penalized logistic regression for the analysis of NAMCS data. In the penalized logistic regression analysis, we identified number of medications other than opioid prescribed in the outpatient visit, reason for visit, injury within 72 hours, number of past year visits, and physician specialty as the main factors associated with opioid prescription outpatient settings. Using decision tree-based ensemble algorithms approach, we identified similar factors that were important in predicting opioid prescription. A cross-validation study was conducted to investigate the performance of the above methods.