By Shahzadi Adeena, Class of 2025

Figure 1:  Pill container laying on counter with cap off

The United States has been battling an opioid epidemic for over twenty years, with opioid-related deaths increasing by 350% in this period. A critical obstacle is that the driving force (such as usage of prescription drugs or synthetic opioids) differs within communities and over time. Current methods to follow drug epidemics focus on economic, broad healthcare, and survey outcomes, but do not adequately capture variations in abuse or shifts in epidemics. Inaccurate observations pose a challenge in providing resources to the affected areas on time. Mathew Matero and H. Andrew Schartz of Stony Brook University, along with Salvatore Girgio, a data scientist at the National Institute on Drug Abuse, designed an AI-based model to more efficiently predict rates of annual opioid deaths at the county level across the US. 

The model, Transformer for Opioid Prediction (TrOP), utilized data from the County Tweet Lexical Bank, a dataset containing word usage on Twitter from more than 2,000 U.S. counties starting in 2011. For each county, yearly topics that appeared together from 2011 to 2017 were sorted from this language data. The researchers combined this language data with data on opioid-related deaths per county queried from the Centers for Disease Control and Prevention, limiting the dataset to counties that reported opioid-related deaths for all years from 2011 to 2017 inclusive. The resulting data included a sample of approximately 212 million people, nearly two-thirds of the population of the U.S., across 357 counties. To evaluate the accuracy of TrOP, the researchers compared past opioid mortality and language use in communities to models utilizing recurrent deep learning methods, linear auto-regression, and heuristic baselines using previous years’ mortality estimates. A key finding of the researchers was the statistical evaluation of how well their AI model’s predictions matched actual opioid mortality rates; they found that the proposed model of predicting opioid-related deaths correctly estimated the actual yearly death rates to within 1.15 deaths per 100,000 people (3% mean absolute percent error).

The researchers will continue to evaluate TrOP’s accuracy considering changes in social media usage. They suggest improvements to using AI for opioid mortality forecasting by developing models that include time-dependent or specified socioeconomic variables, for which data is not currently available at county-year level. Inclusion of these variables would account for the distribution shifts in annual death rates that cannot be predicted from language features alone.

Works Cited:

[1] M. Matero, et al. Opioid death projections with AI-based forecasts using social media language. National Digital Medicine 6, (2023). doi: 10.1038/s41746-023-00776-0

Image retrieved from https://www.pexels.com/photo/orange-and-white-prescription-bottle-on-table-3602778/