AI Predicts Community-Based Opioid Deaths Using Social Media Posts

Researchers from Stony Brook University have developed an artificial intelligence (AI) model that uses social media posts to forecast opioid death rates more accurately than methods relying on traditional data measures, such as opioid-related death rates in previous years and socioeconomic factors, according to a study published this month in npj Digital Medicine.

The model, known as Transformer for Opioid Prediction (TrOP), leverages recent advances in sequence modeling and transformer networks, changes in yearly community-specific social media language on Twitter, and past opioid-related mortality data to predict future changes in opioid-related deaths in US counties.

To build TrOP, the researchers used language data from the County Tweet Lexical Bank, a dataset that contains word usage on Twitter gathered from over 2,000 US counties from 2011 onward. From these data, the research team created groupings, or topics, of words that appeared together for each county from 2011 to 2017.

The language data were combined with data on yearly opioid-related deaths per county sourced from the Centers for Disease Control and Prevention (CDC). Then, the dataset was limited to include only US counties that reported opioid-related deaths each year from 2011 to 2017. The final dataset included information from 212 million people across 357 counties.

Using this dataset, TrOP, a recurrent neural network, and a linear auto-regression model were tasked with forecasting opioid-related death rates for 2016 and 2017.

Overall, TrOP forecasted the actual yearly death rates within 1.15 deaths per 100,000 people, a 3 percent mean absolute percent error (MAPE), while the other two models with the same input achieved a 7 percent MAPE.

“The main result of this study was in fact a statistical evaluation of how well our AI model’s predictions lined up with what really happened, and doing this without assuming any particular language used in social posts should equate to mortality,” explained H. Andrew Schwartz, PhD, associate professor in computer science at Stony Brook University and senior author of the study, in the press release discussing the findings.

The study authors suggested that an AI model like TrOP has the potential to help public health stakeholders improve existing community-based assessments of opioid-related death risk and how that risk may evolve.

However, the researchers further noted that any model using social media posts should include empirical data findings and external validity checks, stating that “all results should always be taken in consultation with other evidence of community opioid-related death risks.”

The researchers also stated that such a tool can only be one part of addressing the opioid crisis in American communities.

“Community-specific predictions are just one line of attack on what is a very complex problem that certainly needs a multifaceted and evolving response,” said Schwartz. “What this work with TrOP provides is the potential for more accurate predictions for opioid-related deaths, which would give professionals and community leaders the ability to better prepare and target areas where there is likely to be an increase in mortality.”