Uncovering Patterns in the Evolving Drug Supply to Detect Emerging Overdose Trends Using Machine Learning - PROJECT SUMMARY/ABSTRACT The ongoing overdose crisis in the United States is exacerbated by an increasingly unpredictable illicit drug supply, including the emergence of novel substances and high-potency drug combinations. Such rapid and evolving shifts in the drug supply make it increasingly difficult for public health responses, such as resource allocation for overdose prevention and harm reduction programs. Therefore, the ability to detect change points in the drug supply, whether involving single substances or co-occurring combinations, is crucial for mitigating overdose risks. Given the complex spatial and temporal dynamics of drug supply shifts and overdose trends, advanced machine learning methods are necessary for granular analysis. We will leverage the highly restricted National Forensic Laboratory Information System (NFLIS-Drug) data, which provides detailed drug composition reports from law enforcement seizures, to track trends in the prevalence of illicit substances. We will further integrate this data with county-level overdose mortality data from the Centers for Disease Control and Prevention (CDC), allowing for a systematic assessment of how changes in drug supply influence overdose fatalities across both time and geography. Our study has two specific aims. Aim 1 is to identify spatiotemporal change points in the prevalence of key substances from 2013 to 2026, including fentanyl, heroin, stimulants, xylazine, and novel synthetic benzodiazepines, as well as their combinations. We will apply correlated anomaly detection models, such as Gaussian Process Subset Scanning and spatial-temporal outlier detection methods, and benchmark them against alternative machine learning approaches to ensure the robust and accurate detection of significant shifts in the drug supply. Aim 2 aims to evaluate the impact of these change points on overdose trends. Using spatiotemporal deep learning methods, specifically Convolutional Long Short-Term Memory networks, we will integrate change points from NFLIS-Drug data with county-level CDC overdose mortality data. This approach will account for temporal and spatial dependencies, time lags, and various county-level characteristics, enabling us to assess how shifts in the illicit drug supply influence fatal overdoses. Understanding drug supply dynamics is critical for optimizing resource allocation and informing policy decisions. By identifying regions and periods with significant changes in the drug supply, public health agencies can allocate resources more effectively to areas with heightened overdose risks. Furthermore, determining which substances and their combinations drive overdose surges and how early these shifts occur will provide insights into the optimal timing of public health responses. By leveraging our unique access to granular data and applying cutting-edge data science techniques, this research aims to advance the utility of NFLIS-Drug data, improve the timeliness and effectiveness of overdose prevention strategies, and inform policy and resource allocation to address emerging overdose crises.
