Opioid Overdose Death Prediction with Graph Neural Networks

Area of Science:

• Public Health
• Data Science
• Epidemiology

Background:

• The opioid crisis significantly impacts Ohio, with overdose death rates exceeding national averages.
• Rural and Appalachian regions are disproportionately affected by opioid overdose deaths.
• Accurate county-level prediction of opioid overdose deaths is crucial for effective intervention.

Purpose of the Study:

• To develop and evaluate a novel Spatial-Temporal Graph Neural Network (ST-GNN) framework for predicting county-level opioid overdose deaths in Ohio.
• To integrate spatial relationships and temporal dynamics using graph neural networks (GNNs) and Long Short-Term Memory (LSTM) networks.
• To compare the ST-GNN framework's performance against traditional statistical models and other deep learning approaches.

Main Methods:

• Utilized quarterly opioid overdose death data for 88 Ohio counties from Q1 2017 to Q2 2023.
• Developed an ST-GNN framework combining GNNs for spatial dependencies and LSTMs for temporal patterns.
• Incorporated a nine-dimensional dynamic feature set (e.g., naloxone administrations, high-risk prescribing) and a static Social Determinants of Health (SDoH) index.

Main Results:

• The ST-GNN framework demonstrated superior predictive performance compared to baseline models.
• The model showed enhanced accuracy, particularly in predicting overdose deaths in larger counties.
• A supplementary classification-based strategy significantly improved prediction stability and reliability for smaller counties.

Conclusions:

• Spatial-temporal modeling is essential for accurately predicting opioid overdose deaths.
• Customized training strategies, including classification for smaller counties, enhance prediction reliability.
• The findings support improved public health decision-making and resource allocation for addressing the opioid crisis in Ohio.