Opioid death projections with AI-based forecasts using social media language

Area of Science:

• Artificial Intelligence
• Public Health
• Computational Social Science

Background:

• Accurate prediction of opioid mortality is crucial for targeted interventions in the U.S. opioid epidemic.
• Existing methods struggle with the heterogeneity of communities and predicting longitudinal changes in mortality.
• AI-based language analysis shows potential for assessing community well-being and predicting trends.

Purpose of the Study:

• To develop and evaluate TROP (Transformer for Opioid Prediction), an AI model for community-specific opioid mortality trend projection.
• To leverage social media language and historical mortality data for accurate future overdose death predictions.
• To improve the precision of forecasting county-level opioid-related deaths.

Main Methods:

• Developed TROP, a transformer network model utilizing yearly Twitter language and past opioid mortality data.
• Trained the model over five years and evaluated its predictive accuracy over the subsequent two years.
• Compared TROP's performance against a baseline model using linear auto-regression and socioeconomic data.

Main Results:

• TROP demonstrated state-of-the-art accuracy in predicting county-specific opioid trends.
• TROP achieved a Mean Absolute Percentage Error (MAPE) of 3%, significantly outperforming the baseline model's 7% MAPE.
• TROP's predictions were within 1.15 deaths per 100,000 people, compared to the baseline's 2.93 deaths per 100,000.

Conclusions:

• AI-driven language analysis, specifically using transformer networks, can accurately predict future opioid mortality trends at the community level.
• TROP offers a powerful tool for identifying high-risk areas and informing location-specific aid strategies for the opioid epidemic.
• The model's superior accuracy suggests a promising approach for public health interventions and resource allocation.