Drug overdose remains a leading cause of death in the United States. In 2022, nearly 110,000 people died due to a drug overdose, representing an increase of almost 150 % in the age-adjusted rate of drug overdose deaths over the past decade (Spencer et al., 2025). These accelerations in drug-related harms have been attributed to an increasingly toxic illicit drug supply, with the introduction of highly potent adulterants such as fentanyl and xylazine (Jones et al., 2018; Alexander et al., 2022). The ongoing criminalization, stigmatization, and dehumanization of people who use drugs inhibits access to adequate resources and supports needed for healthy living, including naloxone and other harm reduction supplies (Bratberg et al., 2023). In addition, limited federal and state funding for harm reduction programs has historically resulted in low coverage of these services in the United States, and presents additional challenges for community-based organizations to reach the substantial number of people at risk for fatal overdose (Cohen et al., 2022; Skinner et al., 2024a). Finally, rapid changes in the unregulated drug supply in the US over the last two decades have increased the number of people at risk for drug overdose, placing increased demand on limited harm reduction services.

These factors contribute to an overdose crisis that reflects a combination of entrenched structural vulnerability with acute risk due to volatility in the drug supply or disruptions to local harm reduction service provision. As such, fatal overdose burden can vary substantially within jurisdictions. In Rhode Island, for example, rates of death due to drug overdose between 2016 and 2020 ranged from 15.5 deaths to 46.9 deaths per 100,000 residents across its 240 census tracts (Cartus et al., 2023). In the context of limited resources and substantial geographic variation in overdose deaths, harm reduction services ought to be directed to neighborhoods at highest risk. Despite several studies that have identified a range of community-level characteristics associated with higher rates of drug overdose deaths (Cartus et al., 2023; Skinner et al., 2024b), few studies have identified longer-term patterns in neighborhood-level overdose susceptibility over time. In addition, overdose surveillance information is often monitored at a municipality, county, or region level, and may be lagged by several months, limiting its ability to inform timely resource allocation. Importantly, with sparsity in overdose events at small geographic resolutions and significant reporting delays of drug overdose deaths in public health surveillance systems, novel forecasting tools, including those leveraging advances in machine learning for neighborhood-level prediction, may have utility in informing the rapid delivery of harm reduction resources to specific communities where they are most needed (Schell et al., 2022; Allen et al., 2023).

While several studies have reported on the potential application of machine learning for prediction of drug-related harms at the individual level to inform clinical practice (Lo-Ciganic et al., 2019, 2022; Gellad et al., 2023), studies that have applied and evaluated these approaches at the community level are rare. In November 2021, our team launched PROVIDENT (PReventing OVerdose using Information and Data from the EnvironmeNT), a randomized trial that seeks to assess the utility of a machine learning model to predict neighborhood-level rates of drug overdose deaths in guiding the allocation and delivery of harm reduction resources across Rhode Island neighborhoods (Marshall et al., 2022). The present study aims to apply techniques of sequence analysis to identify neighborhood-level trajectories in predicted overdose susceptibility over the trial period. In doing so, we hope to distinguish neighborhoods that were persistently versus sporadically (or never) predicted by the PROVIDENT model to be at high risk of elevated overdose mortality over time, as a means of informing tailored harm reduction interventions in these neighborhoods.

Among 809 CBGs, there were 60 observed unique trajectories (out of 64 possible combinations) in overdose risk designations over the six prediction periods. Among these unique trajectories, clustering of trajectories favored a solution with five trajectory groups. As presented by Fig. 1 and Supplementary Figure 1, these five trajectory groups correspond to CBGs that were designated as “high risk”:

Group 1: Rarely or never (73.4 %, n = 594)

Group 2: With decreasing likelihood (4.2 %, n = 34)

Group 3: With increasing likelihood (5.7 %, n = 46)

Group 4: Often (8.8 %, n = 71)

Group 5: Always (7.9 %, n = 64)

Neighborhoods in trajectory groups that were more frequently predicted to be at high risk of elevated overdose mortality over time also had higher rates of non-fatal overdose during the trial period, relative to those in trajectory groups that were more sporadically predicted to be at high future fatal overdose risk (Table 1). The trajectory groups also largely followed a racialized socioeconomic gradient, with CBGs in groups defined by more consistent designations as “high risk” throughout the trial period having fewer white residents and poorer socioeconomic circumstances than those defined by less consistent or no designation as “high risk” (Table 1). Relative to CBGs that were rarely or never designated as “high risk,” residents of neighborhoods that were always designated as “high risk” were 3.16 times more likely to be Black (11.1 % vs. 3.5 %) and 4.96 times more likely to be Hispanic or Latino (37.8 % vs. 7.6 %). Residents and households in neighborhoods that were always designated as “high risk” were also 3.65 times more likely to report incomes below the federal poverty line (26.0 % vs. 7.1 %), 2.61 times more likely to be without health insurance coverage (8.1 % vs. 3.1 %), 4.33 times more likely to receive cash public assistance (37.2 % vs. 8.6 %), and 2.76 times more likely to be paying more than 30 % of their incomes towards housing costs (32.2 % vs. 11.7 %), compared to neighborhoods that were rarely or never designated as “high risk.”

By identifying the trajectory groups that emerged from fatal overdose forecasting over six prediction periods, we found that a majority (81.3 %) of neighborhoods were consistently designated as either “high risk” or not “high risk” over the trial period. Specifically, the PROVIDENT machine learning model consistently designated 7.9 % of neighborhoods as “high risk” and 73.4 % of neighborhoods were consistently not designated to be at “high risk.” The three trajectory groups that comprised the remaining 18.7 % of neighborhoods were largely differentiated by non-fatal overdose burden, racial and ethnic composition, and socioeconomic need. Our results align with previous research that found that overdose hotspot neighborhoods were characterized by higher proportions of Black and Latino residents, unemployment, and cost-burdened households (Samuels et al., 2024). While structural vulnerability predisposes certain communities to elevated overdose risk, retrospective hotspot analyses only capture past events. Sequences of PROVIDENT model predictions add value to existing data sources by capturing the underlying trajectories of risk, providing foresight into emerging patterns. This information allows policymakers and public health practitioners to design effective, long-term solutions for neighborhoods that remained consistently at “high risk” over time, and to plan for scaling intervention capacity as areas are newly identified as “high risk”.

Our findings highlight a cluster of 7.9 % of neighborhoods in Rhode Island that were consistently designated as “high risk” for future overdose deaths and likely require the most focused allocation and delivery of harm reduction resources. Structural supports, including affordable housing, economic opportunity, and racial justice efforts, are additionally necessary to address the elevated fatal overdose burden in these neighborhoods that remain “high risk” over time (Bratberg et al., 2023; Skinner et al., 2024b; Park et al., 2020; Goedel et al., 2019; Wu and Evangelist, 2022; Milaney et al., 2021; Austin et al., 2021; Pamplin et al., 2023). Additionally, these neighborhoods demonstrate a need for fixed-site harm reduction resources, such as overdose prevention centers, syringe services programs, and access to voluntary substance use disorder treatment.

Although these high-risk neighborhoods may be familiar to local community-based harm reduction organizations, predictive tools like the PROVIDENT machine learning model could be most useful for identifying the 18.7 % of neighborhoods with variable prioritization status over time. Distinguishing between the three trajectory groups that were at moderately high risk over time (as identified by our data-driven approach to clustering) may be less meaningful from an intervention standpoint (i.e., for allocating harm reduction resources). Neighborhoods across each of these “middle” trajectory groups likely also require structural intervention, paired with dynamic direction of harm reduction resources and supports for people who use drugs. Notably, the emerging at-risk neighborhoods (i.e., the 5.7 % of neighborhoods with increasing likelihood of designation as “high risk” over the trial period) were often located in urban jurisdictions, and may benefit from city-level interventions.

Finally, the 73.4 % of neighborhoods that were consistently designated as not “high risk” still experience a non-zero risk of future overdose mortality and require baseline supports with close monitoring to identify increases in overdose burden. For instance, efforts to increase access to community-based and over-the-counter naloxone available in pharmacies may help to prevent fatal overdoses in all municipalities.

Although the PROVIDENT predictive model has been previously validated and exceeded its performance targets, the findings of the present study must be considered in the context of the model’s potential limitations. First, the model’s validity is not necessarily consistent for all CBGs. Additionally, uncertainty in ACS measures (a key source of model predictors) is greatest for small subgroups in small geographic units such as CBGs (Wei et al., 2023). It is possible that these ACS data-related issues with sampling error and measurement error could be better addressed using census tract estimates (Wei et al., 2023; Arribas-Bel et al., 2016), but CBGs were preferred by the intended audience of the PROVIDENT model results (i.e., community harm reduction organization staff) as a more actionable geographic unit for guiding resource allocation (Allen et al., 2024). Throughout the trial period, we held several trainings with our partner community organizations to make outreach staff aware of the limitations of the model and to help guide what they may and may not infer from its predictions. In particular, we noted that predicting overdose risk at small spatial units (e.g., census block groups) may be prone to error resulting from uncertainty in ACS measures, among other factors.

Nonetheless, future research should examine the relative validity of overdose prediction trajectories at the CBG versus census tract level. Methods that account for uncertainty in ACS measures in small subgroups remain under development (Wei et al., 2023), and innovative methods that can improve area-based prediction modeling to account for CBG-level error variance remain highly important. Alternatively, methods for retrospective analysis of data by space and time (e.g., local indicators of spatial associations, spatial scan statistics) could be used for identifying areas historically at high risk of overdose and for targeting interventions. Future research could apply a similar analytic approach to the present study to create trajectories based on whether CBGs were identified as “overdose hotspots” over time, which may produce different patterns than our findings that were based on predictions. Spatial machine learning approaches (Shaamala et al., 2025) for predicting overdose hotspots (such as those using long short-term memory networks) also offer a potential avenue for further investigation. In particular, deep learning methods (designed to handle long-term dependencies in sequential data) have been successfully used to predict flood risk dynamically over time (Zhou et al., 2023). Similar to flood risk, overdose is highly variable across space and time, requires many inputs to predict accurately, and, as a result, can be quite computationally intensive. Future research should investigate the feasibility of these methods for predicting neighborhood-level overdose risk.

Our results are limited to the Rhode Island context, but provide preliminary evidence that machine learning predictions may inform overdose prevention resource allocation by identifying neighborhoods that are both emerging hotspots of overdose risk and those consistently predicted to be at high risk of elevated overdose mortality.