AI can predict opioid overdoses from crime and socioeconomic data

AI can predict opioid overdoses from crime and socioeconomic data
Opioid abuse is on the rise nationwide. An estimated 1.7 million people in the United States suffered from substance use disorders related to prescription opioid pain relievers in 2017, and from July 2016 through September 2017 in 45 states, the U.S. Centers for Disease Control and Prevention recorded a 30% uptick in overdoses. Additionally, according to a recent study published in the journal Pain, roughly 21% to 29% of patients prescribed opioids for chronic pain misuse them.
It is, needless to say, imperative that the trend is reversed, and toward that end, researchers at the East Technical University in Turkey and the University of Pittsburgh say they’ve made encouraging progress. In a new paper (“CASTNet: Community-Attentive Spatio-Temporal Networks for Opioid Overdose Forecasting“) published on the preprint server, they describe an AI system capable of forecasting overdoses from socioeconomics and patterns of crime incidents.
“[Our] proposed model allows for interpreting what features, from what communities, have more contributions to predicting local incidents as well as how these communities are captured through forecasting,” explained the paper’s coauthors. “[S]tudies have identified relationships between opioid use and crime incidences, including cause (that opioid use leads to criminal activities), effect (that involvement in criminal behavior leads to drug use), and common causes (that crime and drug tend to co-occur).”
The researchers’ algorithm — CASTNet — learns numerical representation of the “dynamics” in communities that share similar behaviors in a “community-attentive” fashion. Overdose contributors (features) from several communities inform predictions for given locations within the AI model’s purview, and moreover enable the model to identify which local and global features are most predictive and isolate high-risk communities.
The team employed two types of features to inform their AI’s projections: static and dynamic. The former included 2010 census data about economic statuses, education level, vacant housing, median household income, high school graduation rates, and more, while the dynamic features captured per-neighborhood crime stats culled from public safety data portals, such as the number of total crimes and the number of total opioid overdose incidents.
To keep the scope manageable, the team focused on two regions — the City of Chicago (47 neighborhoods) and City of Cincinnati (50 neighborhoods) — for which they collected the geolocation, time, and category for each crime feature. For Chicago specifically, they collected opioid overdose death records from the open source Opioid Mapping Initiative Open Datasets, and for Cincinnati, they used the EMS response data.
The coauthors report that CASTNet achieved better performance than the baseline architecture against which it was tested, and that it selected crimes like “narcotics,” “assault,” “theft,” and “burglary” as the most important features for future opioid overdose deaths in the same locations (along with diversity and population density).
“Based on these results, the neighborhoods with higher population and lower or moderate gender diversity may require additional resources to prevent opioid overdose in both cities,” wrote the researchers. “Also, economic status is important for neighborhoods of both cities, which is consistent with the previous work that suggested communities with a higher concentration of economic stressors (e.g. low income, poverty) may be vulnerable to abuse of opioids as a way to manage chronic stress and mood disorders.”
They leave to future work investigating the link between opioid use and other social phenomena.