Substance Use
Evaluating the predictive performance of administrative data sources to forecast overdose deaths at the neighborhood level with machine learning in Rhode Island John Halifax* John Halifax Bennett Allen William Goedel Benjamin Hallowell Maxwell Krieger Alex Skinner Magdalena Cerdá Brandon Marshall Daniel Neill Jennifer Ahern
Street-based harm reduction efforts (including mobile outreach) to people who use drugs is a practical and cost-effective delivery model for overdose prevention. The PROVIDENT trial in Rhode Island evaluates the effectiveness of machine learning (ML) predictive analytics to forecast overdoses and prioritize neighborhoods for mobile outreach. Rhode Island is uniquely positioned to adopt this approach due to its robust overdose surveillance system. In other jurisdictions, fewer data sources may be readily available.
To provide a template for other settings where data availability may be limited, we evaluate the predictive performance of seven combinations of data sources across two ML models. Public American Community Survey (ACS) data was used as the starting set, with other sources across five domains (built environment, emergency medical services [EMS] non-fatal overdose response, prescription drug monitoring program [PDMP], carceral release, and historical fatal overdose data) appended separately. A seventh dataset merged all six data sources. The two prediction approaches were linear regressions and random forests embedded in a nested cross-validation design. The primary outcome was continuous fatal overdose at the census block group (CBG) level, which was used to calculate the proportion of statewide overdoses captured by CBGs in top percentiles by predicted fatal overdose.
Preliminary results indicate that ACS and EMS data together frequently outperformed or at least neared models trained on all data sources. Both modeling approaches performed well, with linear regressions generally outperforming random forests. Findings suggest that neighborhood-level fatal overdose prediction is feasible using ACS data combined with only one other public health data source, establishing an accessible template. Initial results demonstrate that jurisdictions may be able to leverage existing data to accurately predict area-level overdose to guide targeted overdose prevention.
