RADOR: Rapid Actionable Data for Opioid Response in KY

Opioid use disorder (OUD) remains a persistent public health crisis and epidemic. Kentucky had the fifth-largest drug overdose fatality rate in the United States in 2023, with around 79% of those deaths involving opioid substances. Frequently, efforts to reduce opioid overdoses and support opioid overdose control and prevention are limited by delays in data availability, fragmented data systems, and differences between community-level needs. In the context of the dynamically changing opioid epidemic, agencies and organizations responsible for monitoring and improving the health of the population need timely (state and local) data to make critical decisions on resource allocation and targeted responses.

The Rapid Actionable Data for Opioid Response in Kentucky (RADOR-KY) team at the University of Kentucky is working to address these problems. This team is building a statewide surveillance system, utilizing timely data from a variety of sources around Kentucky, to monitor and respond to the opioid crisis. At the CAAI, we work with this team to provide research support through machine learning and forecasting methods. A key aspect of the RADOR-KY project is to produce timely, relevant, and accurate forecasts of opioid overdose incidents around Kentucky. By aggregating Emergency Medical Services (EMS) response data at different geographical levels, we can determine where overdose events are expected to increase or decrease in the coming months. With this information, adequate support could be prepared and provided to those areas with the hope to treat victims in time and reduce the number of deaths associated with opioid-related incidents.

Time series forecasting is the process of using machine learning or statistical methods to analyze the trends and patterns of time-dependent data and create future predictions of that data. Past values of the target series are typically used to understand the trends and forecast them into the future, but additional data sources can also be used to aid in the predictions. We have worked with a variety of different data sources from the RADOR-KY project, many of which are helpful to use as covariates, or additional variables, with forecasting models. These additional data sources include:

• Temperature and precipitation monthly averages
• Social determinants of health for each county, such as unemployment rate, vehicle access, and age distributions
• Aggregated Medicaid claims containing counts of individuals diagnosed with or receiving treatment for Opioid Use Disorder, among other measurements
• Kentucky State Police drug seizures for opioid substances
• Kentucky Department of Corrections substance use risk measures for inmate intakes and releases
• Naloxone distribution counts for each county

We have also experimented with different machine learning models to perform forecasting, such as gradient boosting, the N-Linear model, and the Temporal Fusion Transformer (TFT). These models vary in their generalizability and complexity, but all are capable of utilizing the different covariates and creating forecasts on a variety of different series at the same time. This is important because we group the EMS data based on different geographical regions. Initially we focused on the county level, but due to data sparsity in less populated areas, larger Area Development District (ADD) regions have also been used.

We have found that when grouped by these larger regions, we can create forecasts for each region with low error (0.1158 RMSE with the N-Linear model). The covariates that have improved the model’s performance the most have been the drug seizures and Medicaid measures. The graphs below show what these predictions can look like. The predicted line in red shows a close adherence to the true values used in the test set. They share a similar shape, showing how the model learns the overall trends that different regions have in common, while also making adjustments to the exact shape and scale of the forecasts.

We will continue to test new models and external data sources to improve forecasting accuracy. Still, the current results show that forecasting opioid overdoses around Kentucky is possible with limited error and will prove useful to state agencies for determining when and where opioid overdoses can be expected to increase or decrease.