Primary Submission Category: Methodological approaches to studying public health
Machine learning for harm reduction: Translating predictive models for community-based prevention practice
Authors: Bennett Allen Daniel Neill Robert Schell Jennifer Ahern Benjamin Hallowell Maxwell Krieger Victoria Jent William Goedel Abigail Cartus Jesse Yedinak Claire Pratty Brandon Marshall Magdalena Cerdá
Presenting Author: Bennett Allen*
Overdose deaths continue to accelerate in the United States, with mortality trends shifting across geography, drug involvement, and sociodemographics. Health authorities now must distribute finite prevention resources across wider and more diverse catchment areas. Machine learning offers a novel strategy for practitioners to anticipate trends in harm and allocate sparse resources proactively to maximize impact. However, prior applications of machine learning to public health prevention have relied on conventional assessments, limiting the utility of models as decision supports for practitioners. To bridge the gap between predictive modeling and prevention practice, we developed four practice-based machine learning model evaluation criteria and applied them to the problem of overdose prevention in Rhode Island. Our criteria are: (1) implementation capacity, or the feasible overdose prevention resource dissemination in a jurisdiction; (2) preventive potential, or the maximum possible reduction in overdose death; (3) health equity, or the fair allocation of overdose prevention resources across geography, race/ethnicity, and socioeconomics; and (4) jurisdictional practicalities, or the site-specific constraints of a given health authority and provider network. To illustrate the use of our criteria, we developed two predictive models: random forest and Gaussian processes. We used Rhode Island overdose mortality records from January 2016 to June 2020 (N=1,408) and neighborhood-level Census data. Our models predicted 7.5-36.4% of overdose deaths during the test period, illustrating the preventive potential of overdose interventions assuming 5-20% statewide implementation capacities for neighborhood-level resource deployment. We described the health equity implications of our models to guide interventions along urbanicity, racial/ethnic composition, and poverty. Our criteria offer a tool for public health practitioners to integrate machine learning into prevention practice.