JMIR Publications

ABSTRACT

Background:

Overuse and misuse of prescription opioids has become a significant public health burden in the United States [1]. About 11.5 million people are estimated to have misused prescription opioids for non-medical purposes in 2016 [1, 2]. This has led to a significant number of drug overdose deaths in the US, over 63,600 in 2016. Understanding spatio-temporal characteristics of opioid overuse is an important first step in developing interventions to mitigate it. Prior studies have examined spatio-temporal clusters of different kinds of metrics associated with opioid misuse, such as the number of patients, mortality, and narcotics-related emergency and non-emergency calls. These studies have been restricted to specific regions in the US because of the datasets used. Spatial scan statistics, restricted to circular shaped regions, has been the primary technique used in these studies.

Objective:

The goal of this study is to identify spatio-temporal hot-spots of opioid users and opioid prescription claims using Medicare data. For Virginia, West Virginia and North Carolina, we characterize the attributes of the counties and the providers who are part of anomalous clusters.

Methods:

We examine spatio-temporal clusters with significantly higher number of beneficiaries and rate of prescriptions for opioids using Medicare payment data from the Centers for Medicare & Medicaid Services (CMS). We use network scan statistics to detect significant clusters with arbitrary shapes. We use the Kulldorff scan statistic to examine the significant clusters for each year (2013, 2014, and 2015), and an Expectation-based version to examine the significant clusters relative to past years. Logistic regression is used to characterize the demographics of the counties that are a part of any significant cluster and data mining techniques are used to discover the specialties of the anomalous providers.

Results:

We examine significant spatial clusters with respect to prescription claims and beneficiary counts in three states: Virginia, North Carolina, and West Virginia, and we find some common patterns: the counties in the most significant clusters are fairly stable in 2014 and 2015, shrinking from 2013. The odds of a county being in a cluster are generally correlated with higher fractions of whites in the county, lower household income and employment level, and a higher percentage of population with Medicare and Medicaid access. These counties also tend to have a lower percentage of population with a direct purchase insurance plan. The Expectation-based scan statistic, which captures change over time, reveals different clusters than the Kulldorff statistic. Providers with an unusually high number of opioid beneficiaries and opioid claims surprisingly include specialties such as physician assistant, nurse practitioner and family practice.

Conclusions:

Network based scan statistics is a powerful approach to find anomalous spatio-temporal patterns of opioid usage. Our analysis of the Medicare claims data for Virginia, North Carolina and West Virginia reveals interesting spatial and temporal patterns of opioid usage, and provides characteristics of the counties and provider specialties that have higher odds of being anomalous.