Predictive Modeling for Opioid-Related Overdose Prevention: Towards Safer Opioid Prescribing and Guiding Treatment Selection for Opioid Use Disorder

Abstract

Opioid-related overdoses have continued to cause record numbers of deaths in the United States. We aimed to inform overdose prevention by applying predictive modeling in two interrelated projects at the public health and individual levels respectively. At the public health level, interventions have traditionally been informed by observational studies. Predictive models using statewide prescription drug monitoring program (PDMP) data can inform more timely intervention. No predictive models using PDMP data for overdose prediction have been prospectively validated to our knowledge. We prospectively validated models predicting fatal and nonfatal opioid-related overdose using data from Tennessee. Models did not need to be retrained or updated to record similar discrimination on newer data, though models were miscalibrated. Prescription-level predictions from models can be aggregated to inform intervention at the county, state, or federal levels. At the individual level, opioid use disorder (OUD) confers additional overdose risk. Medication for OUD (MOUD) are effective but remain underused, particularly when specialists like addiction psychiatrists are not leading OUD care decisions. To help empower non-specialists, we developed a predictive model to guide treatment selection for OUD. For each individual, the model generates probabilities of responding to each MOUD. Different combinations of electronic health record and ZIP-level data were tested. While developing this model, we interviewed clinicians to understand their perceptions of the model. The model had low specificity and was limited by local prescribing practices that favored buprenorphine. ZIP-level information did not significantly improve the performance of electronic health record-based models. Interviews revealed that non-specialists would trust a tool that was endorsed by experts and appeared to follow clinical reasoning. However, clinicians also wanted to consider factors not well captured by the EHR, such as housing status and facility preferences. This dissertation applied predictive modeling to (1) predict overdose risk at the public health level and (2) guide treatment selection for OUD on the individual level, with the goal of informing overdose prevention.