Using Predictive Models and Linked Datasets to Understand Risk of Fatal Opioid Overdose
Ferris, Lindsey M
MetadataShow full item record
Problem Statement The opioid crisis has had a devastating impact on the United States that will span generations. Public health agencies are increasingly looking toward data-driven solutions to understand risk factors, identify high-risk individuals, and direct interventions. Leveraging data captured by public health, healthcare, and other social and human service agencies will be increasingly common, as will applying sophisticated risk modeling to predict outcomes. This dissertation examines risk factors and models in the literature, compares multivariate predictive models with existing threshold-based risk identification, and measures the impact patient matching algorithms have on understanding risk when linking disparate patient-level datasets together. Methods A comprehensive review of the literature from 2008-2018 examined predictive model variables and performance related to opioid overdose using prescription history and other data sources. Using 2015 Maryland Prescription Drug Monitoring Program (PDMP) data and 2015-2016 death data, multiple risk identification methods for fatal opioid overdose were quantified and compared, including a multivariate risk model and common prescription-based thresholds. Finally, criminal justice data from 2013-2015 were matched with PDMP data at the patient-level using three matching algorithms to understand the impact on risk indicator prevalence and performance of a risk model. Results Risk models are increasingly being explored in the literature in recent years, although most use a payer-specific cohort and risk factor and measure definitions were inconsistent. Generally, risk models identified more individuals at risk of a fatal opioid overdose than simple risk thresholds, however, there may be value in combining the risk model with simple thresholds to identify high-risk individuals. Finally, the probabilistically matched population resulted in the highest degree of matching with arrest and death data, although risk model performance was comparable across all algorithms. Conclusions These results illustrate the ways predictive models based on PDMP data can assist with identifying high-risk individuals as a standalone tool or in combination with other risk stratification methods. The matching technique used to link person-level data across disparate data together affects the risk prevalence and factors, although model performance indicates a basic deterministic matching algorithm may be a suitable approach depending on resource constraints and scope of analysis.