The United States continues to experience an unprecedented overdose crisis. Between 2002 and 2022, the drug overdose mortality rates in the United States increased four-fold, with drug overdose deaths surpassing 100,000 in both 2021 and 2022 (Spencer et al., 2024). The lack of accessible, stable, and preventative care for people who use drugs (PWUD) in the United States too often results in healthcare that is reactive, unnecessarily expensive, and inadequate (Buresh et al., 2021; Adeniran et al., 2023; Vohra et al., 2022). In response, community-based overdose prevention services to meet PWUD where they are is a practical and cost-effective care delivery model for wound care, infectious disease management, substance use disorder treatment access, and other overdose prevention services (e.g. naloxone distribution) (Hill et al., 2023; Springer, 2023; Robinowitz et al., 2014; Chan et al., 2021; Rosecrans et al., 2022; Pepin et al., 2023).

While funding for overdose prevention and response has increased in the United States, public health resources are still subject to finite limits. Practitioners need strategies to ensure that they deploy available resources efficiently to maximally reduce overdose mortality. Data science methods can be used to prioritize communities at highest risk of future overdose deaths, thereby improving the efficiency of care delivery for PWUD, particularly in the context of limited resources (Saloner et al., 2018; Wilt et al., 2019). Machine learning techniques have advanced accurate prediction of public health outcomes at both the individual and neighborhood level, with performance typically enhanced by utilizing a variety of data sources (Mhasawade et al., 2021). Prior scholarship has established the capacity of multiple machine learning approaches to forecast overdoses spatially, and the use of such forecasts to improve public health intervention deployment is currently under study statewide in Rhode Island and elsewhere (Neill and Herlands, 2018; Schell et al., 2022; Marshall et al., 2022; Allen et al., 2023; Allen et al., 2024).

The use of neighborhood-level predictive modelling to guide overdose crisis response has potential for applications beyond Rhode Island, but portability considerations have not been well explored. Rhode Island is an excellent environment to trial predictive modelling approaches due to its centralized, statewide public health infrastructure. Public health data and funding at all levels in the state are overseen by a single health authority that coordinates action in partnership with a network of community-based harm reduction organizations, which in turn have strong ties to technical experts at local academic institutions (Yedinak et al., 2024).

However, such “ideal” conditions are not necessarily generalizable to other jurisdictions. Where public health activities are governed by multiple state and local agencies, access to multiple data sources may be infeasible. Overdose prediction efforts by public health departments may also be hindered by a lack of technical expertise. A previous machine learning effort in Rhode Island did not prioritize usability or accessibility of the model for non-technical stakeholders, solely focusing on maximizing predictive performance (Allen et al., 2024). We consider this model, coded in two languages and resource-intensive to train, inaccessible as a template for non-technical public health stakeholders, highlighting the need for an accessible template (Allen et al., 2024). To inform the feasibility of using predictive analytics to guide overdose prevention intervention by local authorities across the United States and beyond, this study used Rhode Island data from several commonly used overdose and public health surveillance data sources to A) assess which commonly collected sociodemographic, public health, and overdose monitoring data sources have the best predictive performance in the Rhode Island context; and B) apply simple-to-implement machine learning and univariate techniques to assess performance and to provide a template (including statistical code in R) for application in new settings. We hypothesized that different combinations of data sources would have varying degrees of predictive performance in forecasting overdose deaths.

From January 1st, 2019 to December 31st, 2021, the period of analysis, Rhode Island experienced 1072 fatal overdoses, averaging 179 fatal overdoses per six-month period with fatal overdoses increasing moderately with time. Of the 352 covariates available when combining all six data sources, variable selection included 86 unique covariates across the 28 linear models and random forests. Supplementary Tables 1 and 2 rank these included covariates by variable importance across the linear models and random forests respectively. For linear models, the most influential variables were derived from across the Built Environment, RIDOC, ACS, PDMP, and EMS data sources. In contrast, the random forests relied more narrowly on variables from the PDMP, ACS, and RIDOC sources to predict future fatal overdose. Variable importance measures resulted in thirteen individual variables being selected for use in univariate rankings, with some variables being the most important variable across multiple dataset-model combinations.

Assessing overall model performance, MSE was lower (better) for linear models than random forest models trained on the same data sources and window in all cases. Shown in Fig. 2, the models trained on All Data had the lowest (i.e. best) MSE for both model classes.

Assessing approaches by the percentage of statewide fatal overdoses captured metrics, the models trained on All Data provided an upper benchmark of performance. Shown in Fig. 3, linear models slightly outperformed their random forest counterparts across the four prioritization scenarios on average across the two windows. Both predictive approaches performed well across all data combinations in the first three prioritization scenarios (top 5th-top 15th percentiles of CBGs), generally achieving prioritization goals on average. Scaling efficient CBG prioritization to the top 20 % proved more challenging. Four linear models trained on All Data, ACS + EMS, ACS + PDMP, and ACS + RIDOC were the only models to achieve the a priori goal on average, with no random forest achieving the a priori goal and only the random forest trained on All Data coming close. Performance of both approaches was generally better in window two, with the three linear models trained on All Data, ACS + PDMP, and ACS + RIDOC most consistent across the two windows for all four prioritization scenarios. Despite meeting the a priori goal on average in the top 20 % scenario, the linear model trained on ACS + EMS data did not do so in window one.

In terms of proportion of overdoses captured averaged over the two windows, univariate rankings generally underperformed compared to models, but often the performances were quite similar. For clarity, only univariate results for variables selected from the All Data models are shown in Figs. 2 and 3, with full results in Supplementary Figs. 2 and 3. In limited instances, univariate rankings in a single window outperformed one or both models trained on the corresponding dataset, with best performing variables derived from the ACS, EMS, and PDMP sources. These instances were generally in window one, where both model classes generally performed worse than window two. Better performance by the multivariable models is most evident in the top 15th and 20th percentile prioritization scenarios.

This analysis establishes a template for utilizing and evaluating simple-to-implement univariate or machine learning methods to predict area-level fatal overdose, generating six-month forecasts that can guide overdose prevention efforts even when only limited data sources are available. In the Rhode Island context, linear regression approaches trained on only the ACS + EMS, ACS + PDMP, and ACS + RIDOC data sources on average performed well, and comparably to both the linear and random forest models trained on All Data. This finding suggests that, in the context of limited data availability, practitioners might consider acquisition of at least one of emergency medical service responses for nonfatal opioid-involved overdose, prescription drug monitoring program, or carceral release data in conjunction with socio-economic census data to train simple prediction models to forecast fatal overdoses at the neighborhood level. This recommendation relies on the assumption that both the processes that drive overdose and the predictor data available in new jurisdictions are similar to those in Rhode Island. Future work in new jurisdictions should be mindful of the limits of this assumption and consider the differences between Rhode Island and other jurisdictions to adapt to new contexts. Wherever there are a variety of predictors readily available, it would be advisable to incorporate all of them into a predictive model that includes a variable selection step.

Results indicate that on average in Rhode Island the linear regression approach outperformed the more complex random forest algorithm and the simpler univariate rankings, providing a method to prioritize neighborhoods for overdose prevention efforts that is quick to train and accessible to interpret for non-technical stakeholders. Due to limitations in generalizing these specific findings, we recommend that future work in new jurisdictions still evaluate multiple prediction approaches ranging in complexity. The code provided in Supplementary Materials offers a starting place for such evaluation.

The standout predictive performances of emergency medical service responses for suspected, nonfatal, opioid-involved overdoses, prescription drug monitoring programs, and carceral release data align well with established knowledge. For example, recent release from incarceration is strongly associated with increased risk of fatal overdose in Rhode Island, at both the individual and neighborhood levels (Brinkley-Rubinstein et al., 2018; Cartus et al., 2023). However, unlike many other jurisdictions, carceral release data in Rhode Island comes from one agency as the state has a single, combined jail and prison system and no federal prisons. While carceral release data may improve predictive ability efficiently (i.e., the RIDOC dataset consisted of only three variables), barriers to comparable data access from carceral institutions may inhibit the portability of this approach beyond Rhode Island.

The first wave of the North American overdose crisis was characterized by prescription opioid involved deaths, with PDMPs established in response to prescription opioid misuse (Smith et al., 2023). Recent trends in overdose mortality have been driven by illicitly manufactured synthetic opioids, primarily fentanyl, and design and utilization heterogeneity among PDMPs may hinder PDMP data portability beyond Rhode Island (Ciccarone, 2021). Prescription opioid related variables were less included in and less important for models considering PDMP data compared to buprenorphine related variables. Of the 39 times PDMP variables were included in models in this analysis, 22 of these instances involved a buprenorphine-related variable, which in linear models were always positively associated with future fatal overdose. An over-reliance on PDMP data may skew models to prioritize areas with existing treatment services, such as buprenorphine access.

Despite less stable window to window performance, emergency medical services data may be more natural to integrate into overdose forecasting efforts by local stakeholders. Emergency medical services data for responses to suspected overdoses have been used previously to identify geographic units to target overdose prevention resources in eastern United States (Dworkis et al., 2018; Pesarsick et al., 2019), and non-fatal overdose is a known predictor of fatal overdose at the individual level (Guo et al., 2021). A prediction approach that incorporates EMS responses for nonfatal opioid-involved overdose data likely benefits dually from improved prediction performance and more direct prioritization of areas to target preventative care where the wider health system is already responding. While in the window one top 15th and 20th percentiles scenarios ACS + EMS models did not achieve prioritization goals, the univariate ranking by Total EMS Responses still performed well, indicating the utility of EMS data in the face of poor individual model performance. A nationally standardized case definition for an EMS response for non-fatal overdose might improve the generalizability of these data regarding their predictive performance in other settings (Hallowell et al., 2021). The case definition for nonfatal opioid-involved overdose used in this analysis could form such a basis, or could be broadened to capture nonfatal stimulant-involved overdoses (Hallowell et al., 2021).

The stronger predictive performance of the linear regression models than the more complex random forest models was surprising given the size and assumed complexity of possible predictor interactions. However, traditional regressions outperforming machine learning methods is not without precedent (Christodoulou et al., 2019). The random forests may have suffered from their inability to extrapolate as statewide fatal overdose counts increased longitudinally from training to test set periods, and predictor relationships may lack significant complexity (Hastie et al., 2009). The strong performance of the univariate rankings (though typically underperforming the multivariable models) supports this lack of significant complexity, and the linear models may have found a sweet spot in terms of complexity.

While previous work has explored the utility of using predictive analytics to address the overdose crisis, the majority of these modelling efforts have focused on targeting clinical interventions based on individual-level overdose risk (Bharat et al., 2021). Efforts to predict fatal overdoses at the neighborhood level have used a variety of sophisticated methods (Schell et al., 2022; Allen et al., 2024; Lo-Ciganic et al., 2019; Yedinak et al., 2021; Bozorgi et al., 2021), but often without specific regard for accessibility and interpretability for non-technical stakeholders in whose hands the models may make the most difference. Machine learning approaches to guide community response to the overdose crisis rely on engagement with public health professionals and community stakeholders for effective implementation (Yedinak et al., 2021). Our aim is to establish simple-to-implement methods evaluated across different data sources as a template for the uptake of predictive analytics to inform preventative outreach by overdose crisis stakeholders who may have less technical expertise or have access to data from fewer domains.

This analysis is subject to several limitations. The high quality and standardized nature of data used in this analysis, due to diligent management by Rhode Island’s centralized state health department, may not be reproducible in other jurisdictions with more fragmented public health authorities. All data used to train and evaluate this study’s models either precede or are concurrent with the COVID-19 pandemic, limiting generalizability to future time periods. Data sources’ predictive performances may not be generalizable to other jurisdictions as associations making certain variables predictive for overdose in Rhode Island may be different in other regions. Practitioners working in new jurisdictions or outside the United States should consider incorporation of further data sources that fit conceptually as predictive of fatal overdose. The application of these methods in other jurisdictions is a crucial next step in our research to evaluate this approach. Using only fatal overdose as the outcome variable may not capture the greater burden of nonfatal overdoses and substance use disorders, which preventative programs should also target. Some model predictors may reflect areas already with services (e.g. count of freestanding ambulatory clinics or total buprenorphine prescriptions), introducing a bias in model predictions. Mitigation of the overdose crisis at the neighborhood level should be complemented by similar endeavors at the individual level. All predictive models that inform the delivery of healthcare will only ever be as effective as the care they facilitate. Further work is required to address the upstream causes of overdose mortality (Dasgupta et al., 2018).

This analysis presents a template for straightforward modelling approaches to forecast fatal overdoses at the neighborhood level with data from administrative sources commonly used as part of overdose surveillance. A screened linear regression model trained on only either ACS and emergency medical service response to non-fatal overdoses, ACS and carceral release data, or ACS and prescription drug monitoring program data performed comparably on average to the same model trained on six diverse administrative data sources. This template has potential to enable jurisdictions in the United States and beyond to predict overdose at the neighborhood level and coordinate deployment of overdose prevention services to where they are needed most.