Equitable Data-Driven Facility Location and Resource Allocation to Fight the Opioid Epidemic

TL;DR

A predictive dynamical model and a prescriptive optimization problem are integrated to compute high-quality opioid treatment facility and treatment budget allocations for each US state and show that policy-makers should target adding treatment facilities to counties that have fewer facilities than their population share and are more socially vulnerable.

Abstract

The opioid epidemic is a crisis that has plagued the United States (US) for decades. One central issue is inequitable access to treatment for opioid use disorder (OUD), which puts certain populations at a higher risk of opioid overdose. We integrate a predictive dynamical model and a prescriptive optimization problem to compute high-quality opioid treatment facility and treatment budget allocations for each US state. Our predictive model is a differential equation-based epidemiological model that captures opioid epidemic dynamics. We use a process inspired by neural ODEs to fit this model to opioid epidemic data for each state and obtain estimates for unknown parameters in the model. We then incorporate this epidemiological model into a mixed-integer optimization problem (MIP) that aims to minimize opioid overdose deaths and the number of people with OUD. We develop strong relaxations based on McCormick envelopes to efficiently compute approximate solutions to our MIPs with a mean optimality gap of 3.99%. Our method provides socioeconomically equitable solutions, as it incentivizes investments in areas with higher social vulnerability (from the US Centers for Disease Control's Social Vulnerability Index) and opioid prescribing rates. On average, our approach decreases the number of people with OUD by 9.03 $\pm$ 1.772%, increases the number of people in treatment by 88.75 $\pm$ 26.223%, and decreases opioid-related deaths by 0.58 $\pm$ 0.111% after 2 years compared to baseline epidemiological model predictions. Our solutions show that policy-makers should target adding treatment facilities to counties that have fewer facilities than their population share and are more socially vulnerable. We demonstrate that our optimization approach should help inform these decisions, as it yields population health benefits in comparison to benchmarks based solely on population and social vulnerability.

Figures (11)

• Figure 1: Flow diagram summarizing our approach. Blue circles indicate data sources, orange blocks our Neural ODE and Mixed-Integer Optimization models, and the green block the resulting decisions.
• Figure 2: Flow diagram of our state-level compartmental ODE model of the opioid epidemic.
• Figure 3: Estimated parameters for each US state.
• Figure 4: State MIP solutions consisting of additional facilities and treatment budget allocations.
• Figure 5: Average effect of our MIP solutions on compartments $A$, $R$, and $D$ for various solution groupings.