Area-specific autoencoder spatiotemporal graph neural networks for opioid overdose death prediction

Ohio has been severely impacted by the opioid crisis, with opioid overdose (OD) death rates exceeding national averages. Accurate OD death prediction supports proactive prevention and treatment allocation. Existing methods often focus on ZIP Code Tabulation Area (ZCTA)-level prediction for small-area resource allocation; however, performance at this resolution is poor due to substantial fluctuations in OD death counts, which introduce noise. This raises a critical methodological question: what is the optimal population threshold for OD death prediction that balances predictive accuracy with geographic resolution?

We perform a theoretical analysis of variance and error bounds to establish the minimal population required for robust prediction. Building on this analysis, we propose an Area-specific Autoencoder Spatiotemporal Graph Neural Network (AAE-STGNN) framework for OD death count prediction using urine drug test (UDT) data as dynamic features and Social Determinants of Health (SDoH) as static features. The framework consists of two key components: (1) an Area-specific Autoencoder (AAE), which learns latent spatial representations while incorporating the minimal population threshold, and (2) a Spatiotemporal Graph Neural Network (STGNN), which models geographic adjacency between areas and dynamic features across time.

Empirical evaluations demonstrate that AAE-STGNN outperforms state-of-the-art (SOTA) approaches, achieving improved accuracy and robustness. We also provide the OD death count trend estimation to support public health decision-making.

These findings underscore the importance of selecting an optimal spatial granularity and leveraging spatiotemporal modeling techniques for data-driven public health surveillance and targeted intervention in the opioid crisis.