Substance Use
A Machine Learning-Based Forecasting Tool to Predict Future Overdose Deaths: Results of the PROVIDENT Randomized Community Intervention Trial Yu Li* Brandon Marshall Marshall Marshall Marshall Marshall Marshall Marshall Marshall Marshall Marshall Marshall Brown University
Overdoses are an ongoing public health problem, necessitating timely, proactive responses directed towards communities at highest risk. We developed, validated, and disseminated a machine learning-based forecasting tool that predicts future overdose deaths at the neighborhood level in Rhode Island. We then conducted a randomized, population-based, community intervention trial to determine whether municipalities randomized to receive predictions about which neighborhoods are most at risk of future overdose would experience greater reductions in overdoses compared to municipalities receiving traditional surveillance data. To evaluate the effect of the intervention on the primary endpoint, municipal-level fatal and non-fatal overdose rates, we used a generalized linear mixed effects model, including fixed effects for pre-/post-time, trial arm, their interaction, and municipal-level covariates. Randomization occurred in November 2021; the analysis included 2 years of post-intervention data (see figure). We randomized 20 and 19 municipalities to the intervention and standard-of-care, respectively; neighborhood characteristics and historical overdose rates were balanced. Over the study period, 36 service providers registered to use the PROVIDENT decision tool and to view the neighborhood forecasts in the treatment arm, logging in 562 times; however, engagement declined over time. We observed no effect of the intervention on the study endpoint (adjusted incidence rate ratio=1.06, bootstrapped 95%CI: 0.96-1.20, p=0.358). Results were robust to sensitivity analyses under different model specifications. While machine learning-based approaches can successfully predict neighborhoods at highest risk of overdose, dissemination of these forecasts to policymakers and community leaders had no significant effect on overall overdose burden. Future research is needed to improve integration of predictive analytic approaches into statewide governance and service provider networks.
