Methodological approaches for the prediction of opioid use-related epidemics in the United States: a narrative review and cross-disciplinary call to action

The opioid crisis in the United States (US) has been defined by waves of drug- and locality-specific Opioid use-Related Epidemics (OREs) of overdose and bloodborne infections, among a range of health harms. The ability to identify localities at risk of such OREs, and better yet, to predict which ones will experience them, holds the potential to mitigate further morbidity and mortality. This narrative review was conducted to identify and describe quantitative approaches aimed at the “risk assessment”, “detection” or “prediction” of OREs in the US. We implemented a PubMed search composed of the: 1) objective (e.g. prediction), 2) epidemiologic outcome (e.g. outbreak), 3) underlying cause (i.e. opioid use), 4) health outcome (e.g. overdose, HIV), 5) location (i.e. U.S.). In total, 46 studies were included, and the following information extracted: discipline, objective, health outcome, drug/substance type, geographic region/unit of analysis, and data sources. Studies identified relied on clinical, epidemiological, behavioral and drug markets surveillance and applied a range of methods including statistical regression, geospatial analyses, dynamic modeling, phylogenetic analyses and machine learning. Studies for the prediction of overdose mortality at national/state/county successfully anticipate outbreak risk and respond preemptively. We present a multi-disciplinary framework for the prediction of OREs in the US and reflect on challenges research teams will face in implementing such strategies along with good practices.


Introduction
The United States (US) is experiencing one of the most devastating public health crises of modern times as a result of serial and intertwining epidemic waves of opioid use and associated health harms, opioid use related epidemics (OREs) hereafter, defined as increased number of cases of overdoses, and/or infections such as HIV, associated with opioid use for a given locality than would normally be expected. [1][2][3][4][5] Over 750,000 lives have been lost to overdose since 2000 in the US, of which 70,630 occurred in 2019 alone. 6,7 Further, preliminary data indicates that the rate of drug overdose deaths has likely accelerated throughout the COVID-19 epidemic. 8 In addition, a number of opioid use associated HIV outbreaks [9][10][11][12][13] have been identified across the country over the past few years; acute HCV incidence has increased by 350% between 2010 and 2016 [14][15][16] and over 15,000 Hepatitis A virus (HAV) cases have been identified across several states since 2016, mostly among homeless and opioid using communities. 17 As novel and synthetic opioids of high potency enter the market, the magnitude and speed of these epidemics increases and the capacity of the public health system to effectively respond is further reduced. The failure to curb this crisis has in great part been a consequence of lagging public health responses, which have so far been reactive rather than preemptive. 18 To effectively respond to geographically diverse emerging OREs, affecting different populations and involving increasingly potent substances, we argue that a shift in perspective, founded on epidemic preparedness, is needed. Preparedness is an integral component in the control of infectious disease epidemics, 19,20 as well as in the management of natural or man-made disasters, 21-23 but it has not been systematically integrated in the context of OREs despite their exceptional impact on public health. Anticipating such threats requires a strong surveillance infrastructure and the development of forecasting tools, which we categorize into three broad groups responding to the following objectives: early detection, risk assessment and prediction.
Early detection refers to the identification of OREs at the initial stages of the epidemic, leading to the implementation of control measures (such as contact tracing for infectious diseases 24 or distribution of drug testing strips in the context of drug adulteration as seen with fentanyl 25,26 ), thereby preventing further disease transmission or mortality associated with the epidemic.
Risk assessments determine and rank the susceptibility of specific demographics (e.g., by ethnicity) and geographical areas (e.g., counties) to experiencing an ORE based on assessing the prevalence of risk factors observed in past epidemics or identified in observational studies. Risk assessment generally result in a score or ranking which indicates how likely it is an ORE occurs in each given locality or among specific populations. Risk assessments can inform the allocation of resources and ensure that local public health stakeholders are prepared for such eventualities.
Predictions estimate the future incidence of opioid use disorders (OUD), overdose, and associated health harms. Models (i.e., statistical, mathematical, heuristic) can be developed and trained on data-related to the mechanisms of disease and drug use spread, progression and severity over time, with the capacity to take into account dynamic factors such as population demographics or drug market changes. 27 These models can then be applied to project the dynamics of future ORE occurrence, temporally and geospatially. Whereas risk assessment is concerned with the likelihood that an ORE may occur (for example, the relative likelihood of a county-level fentanyl overdose outbreak over the next three years, as compared to other counties), prediction extends this by attempting to predict the trajectory of the ORE (for example, the specific annual county-level fentanyl overdose death rate over the next three years for all counties). As such, they provide more specific information about the magnitude and likely course of an epidemic before its emergence or along its development. The ability to predict OREs and their magnitude and severity allows for the assessment of potential interventions to mitigate the harms. If we can make well-informed predictions, we are best equipped to assess the best course of action to prevent future harm. Figure 1 illustrates the use of these types of tools at different stages of the epidemic, as well as their relative impact on reducing incidence. This narrative review aimed to identify the different quantitative methods that have been applied towards the early detection, risk assessment and prediction of OREs in the US. Having a birds-eye view of existing approaches, the type of findings/outcomes that they enable, the data they require and their strengths and limitations, is a necessary step towards maximizing scientific knowledge and promoting coordination. We describe and characterize the main approaches used to date in the context of OREs in the US and share an open repository of studies on our project website (https://www.emergens-project.com/repository), which we continue to update, and invite contributions from the research community.

Methods
We implemented a PubMed search to identify studies aiming to detect or forecast outbreaks of opioid use and related health harms in the US. This included five main components; 1) the key health outcomes of interest: based on previous research, we included OUD, overdose, HIV, hepatitis, tuberculosis, as well as opioid-related admissions; 2) opioid use: to restrict the search to outbreaks that were related to opioid use while making sure we captured all relevant studies, we included terms such as "substance use" and "people who inject drugs"; 3) the epidemiological outcome of interest, corresponding to outbreaks, hotspots, clusters and related terms; 4) the study objective, using terms related to "detection", "prediction" and "risk assessment"; and 5) the location, corresponding to the "United States". The complete list of search terms is provided in appendix. The search was last updated on January 29 th 2021, CM and AB scanned the abstracts and selected relevant articles for full-text consideration by the research team. A snowball approach was used to identify additional articles by scanning the references and citations of each selected article. a quantitative nature. Determination of inclusion in the study was done through discussion with all authors, who had expertise in a range of fields (i.e. geospatial analyses, phylogenetics, statistics, dynamic modeling, HIV and substance use epidemiology, and computer sciences) and who evaluated the contribution of each study to the detection/ forecasting of OREs. Authors extracted the following information from the articles within their field of expertise: field of study, objective, health outcome, substance/drug, geographic region, geographic unit of analysis, methods and data sources. While our objective was to describe the landscape of methods that can be applied in future ORE prediction efforts, rather than evaluate studies' quality, we provided an assessment of their limitations when appropriate. Given the breadth of the research question, a narrative review format was chosen.
This study did not involve human subjects as it relied on secondary reporting of aggregated data from multiple studies with no identifiable information.

Results
Our initial search returned 1,265 articles. After full text screening and dialog with full research team, 46 articles were determined to meet the inclusion criteria for this study. Selected studies are presented in Table 1 alongside their corresponding objective and other extracted information. These studies were implemented at the national, state, county, and local (ZIP Code) levels. Health outcomes included OUD, overdose, HCV, HIV, and TB. Substances included: "any substance", opioids, and novel substances. We describe their potential contribution to policy guidance at different stages of the epidemic and further disaggregate by type of health outcome (i.e., overdose, bloodborne infections and tuberculosis). Studies are organized first by overarching objective (Early Detection, Risk Assessment, Prediction) and then by type of outcome (Overdose, Bloodborne-Infections).

Early detection
The ability to detect OREs is dependent upon the clinical, epidemiological, behavioral and drug market surveillance infrastructure in place as data can be examined to identify unusual patterns. We describe selected studies below and provide further information in Table 1.
Overdose-In late 2005 to early 2006, it was discovered that heroin was being adulterated with fentanyl and an increase in heroin-and opioid-related deaths in Illinois was detected and reported to the CDC in April of 2006. 29 Friedman found that if public health officials were prospectively analyzing poison call center data for heroin-related exposures, the use of time series analysis -specifically autoregressive integrated moving average (ARIMA) modeling -would have detected this ORE one month earlier, in March of 2006. 29 Similarly, Li et al. found that the location of 311 municipal service calls in Columbus, OH was associated with opioid overdose geographic hotspots. 33 A study conducted in Texas used a combination of data from emergency room visits, the Youth Risk Behavior Survey, poison control cases, and qualitative interviews to characterize a "cheese" heroin use outbreak. 31 Another study implemented in Kentucky used data from a combination of death certificates, overdose deaths from the State Medical Examiners' Office, emergency room visits, and prescription drug monitoring programs to detect potential opioid use clusters at state level. 30 Marks et al.  34 They used a natural language processing (NLP) approach to identify opioid-related tweets. Their findings indicate that the annual number of heroin-and synthetic opioid-related Twitter posts were associated with annual heroin-and synthetic opioid-overdose death rates, respectively. 34 Chary et al used NLP (semantic distance specifically) to identify Tweets related to the non-medical use of prescription opioids (NMPO) from 2012 to 2014 and showed these correlated with 2013-2015 NSDUH state level data on NMPO across time, with the strongest correlation found among those aged 18-25. 35 In a recent study, Sarker et al used machine learning and NLP to identify Twitter chatter related to "self-reported opioid abuse or misuse" in Pennsylvania from 2012 to 2014. 36 They showed the latter was significantly correlated with yearly opioid related overdose deaths at county level across the three years and that it was also correlated with four pertinent measures of drug use from the NSDUH at substate level, although significance was not attained due to small sample sizes. This holds promise for the real time monitoring of opioid use and associated overdose and could therefore contribute to the early detection of overdose outbreaks. Another type of internet data explored has been online news. Hswen et al investigated the correlation between opioid related (Google) news and opioid related overdose deaths across the US and found that geographical variations in these two outcomes were not consistent, suggesting news will unlikely prove useful for the realtime monitoring and detection of opioid related deaths, but instead might reflect or affect public opinion and policy decision making regarding OUD. 37 Bloodborne-infections-As mentioned, collecting information on substance use characteristics is key to enabling ORE detection and this is particularly true in the context of infectious disease surveillance. For example, Fitzmaurice et al. utilized national HIV surveillance data to identify injection drug use (IDU) related HIV outbreaks using a heuristic approach. 28 They examined three years (2013-5) of HIV diagnosis data, generating an average annual HIV diagnosis rate and corresponding standard error at the state-and countylevels. Then, they examined HIV diagnosis rates for the subsequent year (2016) and generated "alerts" for all localities whose HIV diagnosis rate was at least two standard deviations greater than that localities average rate for the prior three years. Another study showed that enhanced HCV surveillance implemented in New York state and prioritizing follow-up of positive laboratory markers for HCV infection among persons aged <30 years allowed researchers to identify a cluster of HCV transmission related to IDU among people from one particular county who had attended the same high school. 38 While effective, enhanced surveillance and comprehensive outbreak investigations involving contact tracing and data linkage generally rely on additional funding and are labor intensive.
genotypes) to help identify persons involved in the same chain of disease transmission and add value to conventional contact investigation. They may help identifying unknown or unusual transmission settings or factors, uncover inter-jurisdictional transmission and identify additional persons involved in an outbreak. 39 9 Similarly, routine HIV screening at the first legal syringe services program (SSP) in Florida led to the identification of ten anonymous HIV seroconversions. 43 Through phylogenetic analysis, they were able to identify both drug use and sexual HIV transmission networks linking SSP clients with individuals outside of this group. 43 Falade-Nwulia et al. found that, among HCV-positive PWID in Baltimore, women and HIV-positive individuals were more likely to be associated with an HCV transmission cluster, 44 showing that phylogenetic approaches can help identify individuals at higher risk of infection. Further Zhou et al present a novel molecular analysis to identify recent infections and drug resistance mutations among new HIV diagnoses. 45 Systematically applying this method to new diagnoses could help identify and interrupt new transmission clusters. While the cost of sequencing has decreased over time, it is still not routinely implemented across settings due, in part, to high costs. Interestingly, routine molecular analyses could lead to more cost-effective screening and contact tracing strategies through prioritizing specific transmission clusters.
Tuberculosis-In a CDC outbreak investigation, the review of medical records and interviews identified substance-use as a factor associated with a growing cluster of TB transmission in Florida. 46 In a study of genotype clusters of tuberculosis, Althomsons et al. suggested that "routinely reported [genotype] data may identify small clusters that are likely to become outbreaks and which are therefore candidates for intensified contact investigations". 47 The early identification of these growing clusters is fundamental to targeting prevention campaigns.

Risk assessment
Risk assessments aim to identify risk factors for specific health outcomes and to characterize their distribution in order to identify communities at higher risk of experiencing outbreaks of the health outcome of interest. As could be expected, many of the studies identified under this objective employed geospatial methods.
Overdose-Geospatial approaches have been applied to substance use surveillance data to identify localities at high-risk (i.e. "hotspots") of increased overdose burden. Marotta et al. applied exploratory spatial data analysis (ESDA) to identify hotspot clusters of opioid overdose in the state of New York. 48 ESDA involves first visualizing the data (i.e. mapping overdose death rates by county), performing a test of global spatial auto-correlation (via Moran's I), and then identifying clusters of counties applying Local Indicators of Spatial Association (LISA) analysis. 48 Stopka applied a similar approach to identify clusters of opioid overdose and high rates of prescription opioid administration among localities (by ZIP code) across Massachusetts. 49 Albright et al also followed a similar approach to identify hotspots of OUD among veterans in Alabama. 50 Hernandez et al. and Brownstein et al. both applied the spatial scan statistic (SaTScan) approach to analyze the spatial distribution of opioid overdose across Ohio and of opioid use across New Mexico, respectively. 51, 52 Basak et al also applied this method to identify hotspots for opioid use and opioid prescription claims in Virginia. 53 Pesarsick et al investigated local clustering of nonfatal overdose in southern Pennsylvania using EMS data from cases in which naloxone was administered and an improvement in the Glasgow coma score was recorded. They identified local clusters through applying the Kulldorf scan statistic. 54 These EMS data allow for near real time risk assessments and service adjustment.
Geospatially driven regression strategies can also be applied to identify locality-specific factors associated with a higher risk of overdose mortality occurring. Cerda et al. applied a spatio-temporal Bayesian Poisson model to examine the relationship between zip code level factors and prescription opioid overdose-related hospitalizations. 55 They found that localities in California with greater pharmacy density, greater prevalence of manual labor jobs, and lower income were all associated with higher rates of prescription opioid hospitalizations. 55 Cao et al also applied geospatial regression strategies to characterize the ZIP code-level relationship in Maryland between drug-related emergency departments visits and series of predictors including geolocated Twitter data on sentiment related to crime, drug use and depression, ACS data on poverty, unemployment, housing prices, and education level. 56 Their findings suggest that this approach can be used to identify areas at higher risk of experiencing drug use related health harms.
The consistent measurement of outcomes of interest (and factors associated with them) across geographic settings and over time represents a primary limitation to the effective implementation of geospatial analyses. Geospatial analyses often rely on data collected by multiple disconnected organizations with wide variations in data collection and reporting practices. For example, it would certainly be of interest to incorporate the geospatial distribution of synthetic opioid overdose throughout the US into a risk assessment modelunfortunately, due to inconsistent reporting of the involvement of specific drugs in mortality records by local coroner's offices, 57 such an endeavor is not feasible. A CDC report of overdose deaths found that only 20 of 50 US states from 2013-2017 reported consistent and trustworthy synthetic opioid overdose counts. 58 In addition, it is important to acknowledge that geospatial methods have so far predominantly been used for explanatory rather than predictive purposes, and therefore further validation of predictive adaptations is needed.
Wastewater epidemiology is emerging as a potential useful tool to monitor and characterize geographic differences in substance use patterns. Endo et al present results from a study involving robotic wastewater sampling across residential manholes in an urban municipality in North Carolina and targeted mass spectrometry to detect opioids, including illicit and prescription opioids as well as naloxone and buprenorphine. 59 No correlation was found between total opioid exposure in each catchment area and the number of overdoses, but the latter did correlate with both naloxone and buprenorphine exposure. Both Duvallet et al. and Gushgari et al. applied a similar approach to identify the presence of opioids in urban wastewater. 60,61 Gushgari et al. found that opioid presence in wastewater in two US cities was associated with future opioid overdose deaths, indicating it may be a valuable risk assessment tool. 61 Such technology can be used to quantify levels of different opioids in a community, thereby identifying areas at higher risk for overdose, as well as the presence of new or more potent opioids.
Bloodborne infections-One of the most cited studies, published by Van Handel et al 62 as part of a CDC initiative, applied a multilevel Poisson regression to rank US counties at highest risk of an IDU-related HIV outbreak, using HCV incidence as a proxy measure for IDU prevalence. They calculated a risk score for future potential HIV transmission risk in each county based on the model's coefficients and on the distance to high HIV prevalence counties. 62 They identified 220 counties at highest risk of IDU-related HIV outbreaks and these findings have been used to inform funding allocation for substance use treatment and harm reduction programs. Rickles et al replicated and expanded on the Van Handel risk assessment approach, focusing solely on counties in Tennessee. 63 They incorporated a wider variety of potential indicators and applied both factor analysis and principal component analysis in order to select variables for their final model.  64 They complemented the study with an optimized outlier analysis, a geographic information system method, to identify HCV hotspots when accounting for HCV rates in neighboring small areas. Their results differed from those of Van Handel, which they argue is likely due to their analysis being more recent (2019 vs. 2016) and using different methods for handling regions with limited data. 64 While its impact has been influential and has driven resource allocation, the approach employed by Van Handel et al. is subject to important limitations that need be discussed. First the vulnerability score calculated for each county was defined cyclically: they trained their model with HCV as the outcome and then they applied the model on the same data it was trained on to generate vulnerability scores. As such, is unclear if their approach was optimal compared to simply using HCV incidence as the vulnerability score. It is also unclear if HCV incidence represents an accurate proxy of IDU (as partly dependent on chronic prevalence and its measurement is uneven) and, therefore, if their approach effectively captured the at-risk population in each county. Next, they applied no validation strategy, internal or external, to their findings. Without such validation the accuracy of their risk assessment is unclear, which is important both in the context of resource allocation based on the results and in determining whether stakeholders should implement this strategy again in the future. Importantly Van Handel et al. 62 and related studies use the proportion of people who are white as a key model predictor. The opioid crisis has long been characterized primarily affecting white people in the US. 65 However, this was related to preferential pain management treatment (and over-prescription of opioids) among white people. 65 Robinson et al. 66 note that using race as a predictive factor, when it is systemic racism driving the observed pattern, risks reinforcing the impact of systemic racism (in this instance, by overestimating the harm of future OREs in localities with a higher proportion of white people and potentially depriving ethnically diverse urban areas from needed resources). Finally, it is unclear if applying a regression strategy without the ability to capture geospatial correlations resulted in biased findings. Given the proliferation of geospatial modeling strategies, identifying the appropriate method to rigorously incorporate geospatial effects is challenging. This concern also further highlights the importance of including a strategy to validate model performance.
At a smaller geographic scale, Des Jarlais et al assessed location of substance use and injection risk behaviors among PWID accessing substance use treatment centers in New York City to determine zip codes considered "hotspots" of HIV and HCV risk. 32 This type of local level study can inform the prioritization of outreach efforts such as mobile HIV/HCV testing or SSP, as well as naloxone distribution efforts. However, they are difficult to standardize as they rely on treatment (or harm reduction) centers' capacity. Coordination efforts across these centers through a central organization and through dedicated funding mechanisms could address this issue.

Prediction
Prediction goes one step further than risk assessment by quantifying the future incidence of a health outcome of interest among specific populations or geographical locations. This is challenging in general and particularly in the context of OREs given the rapidly changing dynamics of drug markets and drug use itself. However, estimating future outcomes requires the systematic and rigorous analysis of observed patterns and their underlying mechanisms, thereby leading to a better understanding of OREs.
Overdose-An interesting study led by Sumetsky et al. tested the performance of two statistical methods (standard log-linear vs. log-logistic Bayesian hierarchical Poisson conditionally autoregressive (CAR) spatial models) in predicting overdose deaths by county in two states from 2001-2014. 67 They generated a county-level "carrying capacity" (i.e., the number of people at risk of overdose in given county) based on a set of factors, a wellfounded decision given that a majority of individuals are not at risk of overdosing. However, one factor informing this carrying capacity was the proportion of county population that was white, replicating the issue highlighted by Robinson above and potentially limiting its validity to the prescription opioid wave of the opioid crisis. It is important to capture risk factors across different classes of opioids and other drugs as their impact is driven by different mechanisms and affects different populations. 68 A recent study by Cooper et al. used three-degree polynomial functions to investigate fatal overdose dynamics from 2012 to 2016 by state, disaggregating rates by heroin, semisynthetic and synthetic opioids. 69 They identified states with highest elasticity (i.e. rate of change over time) for each of the opioid sub-epidemics. Such an approach is useful for characterizing changes in overdose rates over time (i.e., elastic versus inelastic) and displays the potential benefit of applying non-linear functions for predictive purposes. Further, calculating overdose rate elasticity may be applied within a broader predictive framework, as counties with greater elasticity may be subject to greater fluctuations in overdose death. Based on the write-up, it is not readily clear how the model coefficients were reached and accuracy of predictive models are not reflected on -making it challenging to evaluate how this study should be adapted for future use. Further, while it is of interest to consider nonlinear regression, it is worth investigating whether the equation they used, or a single equation structure generally, could explain changes in overdose death rates across all localities.
Two studies utilized Google Trends data to predict opioid use trends and related outcomes. 70,71 Google Trends analyzes a portion of Google web searches over a period of time and normalizes the data to compare trends of different search terms from the same region during the same period. Young et al. built a linear mixed model regression using opioid related search terms as the independent variable and opioid related emergency department admissions as the dependent variable between 2005 and 2011 in 10 metropolitan statistical areas to then predict opioid related hospital admissions for the next two 3-month periods. 70 Perdue et al also used Google Trends to investigate whether searches for novel drugs were associated with prevalence of use of these novel drugs, using the "Monitoring the Future" survey as their validation dataset, finding that internet searches for novel drugs were associated with novel drug use in teenagers. 71 Young et al. display that Google Trends may be effectively used in a predictive analysis. One limitation of their analysis is that it is limited to metropolitan statistical areas, as opposed to smaller localities such as counties in rural areas that may need to mobilize a preventative response -unfortunately, Google Trends are not available at county level. Recently, Campo et al presented a novel machine learning algorithm informed with Google search terms related to drug use to predict next year's overdose death rates at county level across the US from 2005 to 2017, as well as monthly estimates at state level across this period. 72 They showed that the model has good accuracy based on mean average error (with variation across states) and that for 2017 (which data was not used for model training), it identified 75 of the top 100 counties based on overdose death rates. Both the Young et al and Campo et al studies display that incorporating Google Trend data to inform predictive models holds the potential to improve overall model performance.
Pitt et al used a compartmental dynamic model of opioid use to estimate future overdose mortality rates at national level, finding that in the absence of further interventions, 510,000 opioid-related deaths would occur between 2016 and 2025. 73 Unlike statistical models, which relate a set of predictors with the outcome of interest through capturing their association, dynamic models explicitly represent the mechanisms generating a specific outcome, such as progression through different stages of opioid use associated with differential risk of fatal overdose, or contact patterns leading to infectious disease transmission. Dynamic models can be split into two main categories: compartmental models, which represent groups of individuals and their "average behaviors" and agent-based models, which explicitly represent each individual and their behavior. Chen et al also developed a compartmental model differentiating non-medical use opioid prescription (with and without disorder) and illicit opioid use, and estimated that in the absence of interventions 700,400 opioid-related deaths would occur between 2016 and 2025. 74 This is higher than estimated by Pitt et al in part due to the assumption that both the incidence of illicit opioid use and its lethality would follow an upward trend based on observed increases in synthetic opioid related deaths. They found that eliminating the non-medical use of prescription opioids would have a marginal effect on the overdose epidemic, indicating multi-pronged interventions are needed. Ballreich et al used a detailed dynamic model including 32 compartments which disaggregate the population by type of opioid used and whether it was initiated through prescription opioid use or not, as well as seven different medication assisted treatment (MAT) states, to simulate the opioid epidemic in the US between 2020 and 2029. 75 They estimated 484,429 opioid related deaths across this time period in the absence of further interventions. Similar to Pitt et al and Chen et al, they found that reducing opioid prescribing (40% reduction) would have a small effect on the epidemic, while scaling up naloxone access (leading to a 19% reduction in the overdose fatality rate) would prevent about 15% of deaths and increasing MAT access (tripling uptake) would have the greatest impact with a 25% reduction in mortality. Each of these compartmental modeling studies have been implemented at the national level, and therefore fail to account for heterogeneity in opioid prescriptions, substance use behaviors and OUD treatment access across the country. As such they can only provide general insight and guidance about the likely course of the epidemic and appropriate responses. In addition, assumptions regarding future trends in overdose deaths between the different studies lead to significantly different estimates and therefore to different implications for interventions (such as for future treatment need). That said, comparing outputs across modeling studies and identifying the reasons behind discrepancies is an effective way to better understand epidemic dynamics.

Bloodborne infections-While
Campo et al used overdose deaths as their outcome of interest, they present this as a proxy for IDU and state their model could be used to predict or raise alert about the risk of IDU related outbreaks of blood borne infections. Indeed, they highlight that Scott County was identified by their model as the top Indiana County for overdose deaths in 2011 (and 12 th highest US county), which coincided with the HCV and HIV outbreaks, suggesting that these could have been prevented or mitigated to a large extent had these signals been attended to. Gonsalves et al applied a compartmental model to retrospectively evaluate the response to the Scott County HIV outbreak. 11 They found that implementing the response in April 2011 (instead of in early 2015), just after HCV outbreaks were identified in several counties in the state, could have limited the outbreak to 10 cases. Such a model could be used prospectively in the event of an outbreak to inform the response. Goedel et al. applied an agent-based modeling approach, projecting that the proactive implementation of SSP in Scott County would have decreased HIV incidence by 90%. 76 Fraser et al also used data from the Scott County outbreak to predict future HCV incidence and estimate the directly acting antivirals (DAA) treatment coverage needed to achieve a 90% reduction in chronic HCV prevalence and incidence by 2025 and 2029, either in isolation or in combination with MAT and SSP. 77 They found that 159 per 1,000 PWID would need to be treated every year and this would be halved if MAT and SSP were scaled up to 50% of PWID. While these studies rely on detailed data collected during a specific infectious disease outbreak, which is not available across settings, they are useful case studies to demonstrate the importance of scaling up harm reduction and treatment services early in the context of OREs and also provide concrete estimates of treatment coverage needed to achieve HCV elimination goals and will be useful to similar settings. Bobashev et al. employed an agent-based modeling approach to explore the impact of different types of heroin (powder versus black tar), different types of needles (high versus low dead space), and various injection practices (such as syringe sharing) on HIV incidence. 78 This approach highlights both the feasibility and necessity of incorporating variations in drug type (driven by drug market availability) and route of administration when attempting to predict OREs. In a more experimental fashion, substance use has been modeled as a communicable behavior. Marks et al. estimated the future incidence of IDU in a generic North American setting, assuming different coverages of MAT, which has been associated with a lower rate of assisting others to initiate IDU. 79 They found that a 60% MAT coverage could translate in a 23% reduction in the annual rate of IDU initiation.

Implications
Proposing a Framework for ORE Surveillance-Based on these findings we propose a framework for the enhanced surveillance and forecasting of OREs which integrates the different methodological approaches identified to enable the early detection, risk assessment and prediction throughout the epidemic's course as shown in Figure 2. To develop a multidisciplinary ORE surveillance and analytic infrastructure, it is essential to consider the multiple health outcomes arising from opioid use, including overdose, infectious diseases and mental health disorders (which we attempted to include in our search but failed to identify relevant studies) and to maximize the different types of surveillance systems associated with each.

Types of Surveillance Data:
Clinical surveillance remains the cornerstone of OREs' early detection and local authorities have merged a variety of data sources to increase the sensitivity of their surveillance efforts. Indeed, the ability to disaggregate routine infectious disease, hospital admissions, EMS records and mortality data by substance use characteristics is crucial to detecting OREs. Enhancing this information with epidemiological surveillance data, including poison call center data, as well as data from substance use treatment and harm reduction organizations further increases sensitivity. Substance use behavioral surveillance had until recently been grounded on routine population surveys such as the NSDUH, as well as opioid prescription and OUD treatment records.
Recently, internet-based surveillance of opioid use related searches and social media posts has been proposed as an alternative to traditional surveillance approaches, potentially enabling the real time monitoring of substance use trends. It has been shown to correlate well with NSDUH estimates as well as with overdose outcomes. While representativeness and validity remain difficult to assess, especially considering lower internet access among some people suffering from OUD such as those experiencing homelessness, this is a promising field.
Drug market surveillance, through drug seizures, drug testing services and wastewater testing, can provide the often-missing part of the equation. Ideally, these different sources of data would be easily accessible in real-time to allow for continuous "merged-monitoring" analyses. 27 How these data are collected or processed, will determine key attributes, such as the availability and granularity of geospatial, socio-demographic, social network or genetic information.

Types of Methods That Can Be Employed:
Depending on the health outcome of interest, the surveillance data available and their characteristics, different analytic methods can be used to enable the early detection, risk assessment and prediction of OREs. The studies identified have implemented a variety of statistical regression, geospatial, machine learning, dynamic modeling, and phylogenetic/molecular analyses. Some of these methods could in theory be applied towards any of the three objectives, as long as the analysis is designed accordingly.
For example, geospatial methods can be used to detect clusters/hotspots for both opioid use and associated harms, or to characterize the geographical heterogeneity in these outcomes thereby contributing to early detection and risk assessment efforts. Furthermore, they can estimate the likely future geographical distribution of opioid use related outcomes, thereby providing predictive capacity.
As illustrated through the identified studies, statistical regression methods have been used in the context of early detection, risk assessment and prediction and are often the method of choice given the limited availability of granular geospatial data. The emergence of big data has led to its replacement by machine learning, using other classification and regression algorithms such as random forests, artificial neural networks, and support vector machines.
Other methods such as phylogenetic/molecular analyses are mostly relevant in the context of early detection (or retrospectively to better characterize an outbreak). They have enhanced traditional infectious disease outbreak investigations and they can also allow the identification of rapidly growing clusters of transmission among people who use drugs.
Dynamic modeling, on the other hand, is mostly used for prediction purposes and can provide guidance on effective intervention strategies. Additionally, modeling substance use as a communicable behavior could provide early information about the potential spread of a new drug in a specific community and at larger scales. Different methods should be combined to enhance the sensitivity and accuracy of early detection or forecasting analyses and to guide control measures through the epidemic's course.
Further, in Table 2, we provide further insight on the different types of methods identified in the review as well as details on when to employ them and their potential limitations.

Objective and scope
This narrative review described studies illustrating methodologies which can be employed for the early detection, risk assessment or prediction of future OREs. Further, we have developed a multidisciplinary framework for enhancing ORE surveillance and forecasting endeavors. Few efforts have been made to anticipate OREs and appropriate methods for evaluating their predictive performance have largely been absent. However, recent contributions show promising results and these methodologies represent an interdisciplinary toolkit that can be utilized in the efforts to prevent, identify, and mitigate substance use associated health harms.
Here, we highlight the potential of these different approaches, as well as challenges that must be overcome to undertake this research and we propose good practices (see Table 3 for a summary). Importantly, while we attempted to capture all key terms through our search strategy, we note that we might have missed some relevant papers but expect that the selected sample provides a comprehensive overview of the current tools available across different fields. While we attempted to incorporate mental health outcomes in our search, we regret that no eligible studies were identified, likely illustrating the lesser focus on population level prediction in the field of mental health and the increased complexity in identifying particular conditions, aside from suicide. One study by Yao et al did apply machine learning to Reddit data to investigate suicidality among people who use opioids. 80 It was not included because its main objective was to validate their language processing methods, however, these hold potential for suicide prediction or risk assessment efforts.

Methodological limitations
Given the common application of explanatory models within substance use-related research, it is important to note that predictive and explanatory modeling endeavors are distinct pursuits with their own considerations. Statistical and geospatial regression methods in the context of OREs have primarily been used for explanatory purposes, to identify factors independently associated with overdose mortality or infectious disease transmission. Only a few studies have extended this for predictive purposes. Indeed, it is sometimes the language, rather than the methods, which determines whether these models are used for explanatory or predictive purposes. Explanation and prediction, however, are distinct endeavors requiring distinct methodological considerations. 81 Introducing rigorous predictive methodologies into ORE prediction studies, such as evaluation of predictive performance, cross-validation, as well as steps to avoid common pitfalls, such as overfitting, is a necessary step. Collaboration with researchers in machine learning and bioinformatics may aid with this introduction.
Innovation in the field of dynamic modeling could enhance the use of these methods to predict the spread of substance use as communicable behavior -theoretical model development by Behrens, Caulkins, and colleagues display that dynamic modeling is wellequipped to address the mechanisms underlying social communication of substance use behaviors. 82 This would require more complete data on social and substance use networks; fortunately, this information is becoming increasingly available. As suggested by preliminary studies, the use of Google searches, as well as other internet data including social media data, could provide complementary information to anticipate OREs. However, evidence is limited and experience using these data in the context of flu epidemics' predictions has shown that differentiating (health seeking) behaviors from interest is challenging, requiring more complex methods to implement corrections. [83][84][85][86]

Data constraints
Implementing these methods is dependent on having timely access to data. The epidemiological surveillance infrastructure to monitor substance use and associated harms must be in place, and structural barriers to access requisite data and to communicate findings across local, state, and federal public health agencies must be removed. In the case of Scott County, no advanced analytic techniques were required to identify the IDU-related HIV outbreak, but a failure to communicate information around the outbreak pushed its detection back by months. 87 Additionally, the non-participation of the state of Indiana in the molecular surveillance program limited the efficiency of the response early in the outbreak.
Timeliness of data is crucial in rapidly evolving OREs. Typically, publicly available epidemiological data related to substance use is limited and outdated. For example, as of February 2021, the Centers for Disease Control and Prevention (CDC) was only reporting county-level overdose death rates through the end of 2019. 88 Assuming no other sources of this data are available, our ability to predict future county-level overdose death rates is limited by the fact that the data from the previous two years is unavailable. This is reflected in the historical timeframes of the predictive studies presented here: Sumetsky, Cooper and Campo (each published in 2020) only predicted overdose deaths up to 2017. 67,69,72 Importantly, drug market data collected by law enforcement agencies is typically unavailable or restricted to public health research teams when it provides valuable information to monitor both changes in drug availability and properties. 68,89 Alternative strategies to monitor drug markets, such as wastewater testing have been investigated and hold promise. 90

Epistemological considerations
Detecting and predicting OREs is also dependent on identifying predictor data which can be used to train and inform analytic approaches. As touched on in the results, though, it is important to consider unintended consequences of certain predictor choice. Robinson et al. highlight how predictive endeavors which use identity-based variables such as race may inadvertently reinforce dynamics of systemic prejudice. 66 This is especially important given that the opioid crisis has long been characterized as primarily affecting non-Hispanic white people in the US. 65 Quite reasonably, many studies we identified used the white proportion of the population in each locality as a key predictor in their analytic approaches. The question that is important to ask is whether it appears that race or systemic racism is driving the disparity. It is argued that the opioid crisis initially impacted white populations more severely because they received preferential pain management care and were essentially overprescribed opioid painkillers. 65 In the past decade, excessive opioid prescribing practices have been curtailed and it has become clear that the opioid crisis is being felt across all racial and ethnic populations in the US. 91,92 Looking to historically available data, white race will continue to be a significant predictor despite likely changes in the dynamics of the opioid crisis -it is, thus, important that researchers and reviewers engage critically with the appropriateness of its inclusion. Broadly, we recommend that researchers explicitly justify the use of immutable, identity-based characteristics (such as race) in predictive modeling endeavors to protect from the risk of replicating systemic disparities. Further and more generally, we argue that the use of qualitative methods is informative in driving variable selection, as they can provide rationale for the importance of given factors. 93

Knowledge dissemination
Given the variety of methods, data sources, and potential outcomes to explore, the importance of transparency and knowledge-sharing is crucial to the timely development of OREs' prediction tools. As noted by Woelfle et al., an open science approach can accelerate the time to breakthroughs and findings. 94 Given the multitude of disciplines that need to be coordinated to best address OREs and their harms, precious time can be saved by ensuring that research teams are not unnecessarily replicating one another's work. On top of traditional academic publishing, it will be important to leverage non-peer reviewed publication platforms such as arXiv which can allow for the quick dissemination of novel approaches while still protecting individual researchers' rights to their creative works. Furthermore, predictive modeling efforts are only as useful as they are applicable. Opensource solutions can improve the ability of public health department to implement predictive solutions for themselves.
Finally, it is important to note that the potential for developing effective technologies for predicting ORE and for understanding their expected utility is dependent on transparent platforms which allow research teams to make future-oriented predictions. For example, during the COVID-19 pandemic, modeling teams have attempted to project future incidence and mortality, which organizations such as the CDC have continued to share. 95 As such, we have the ability to examine the accuracy of a wide range of COVID-19 predictive models over space (i.e. various geographic regions) and time and we can make evaluations of their predictive utility. It is crucial that future-oriented predictions of OREs be transparently aggregated and evaluated. We believe that the introduction of such a platform can work to promote inclusion, collaboration and innovation and will be the most effective way to both ascertain the utility of ORE prediction endeavors and to provide timely evidence-based guidance.

Conclusion
This is an interdisciplinary field that has not yet coalesced. It is our intention that this review and accompanying resources can begin the process of bringing together the urgent multidisciplinary action needed to stem the harms of ORE in the US and globally. We have launched a website (https://www.emergens-project.com/repository) where research on this topic will be aggregated and invite readers to share their publications. The intention of this site is to create a space where stakeholders can build on one another's work and disseminate approaches and findings. This review and subsequent resources can act as a first step to coalescing this field and accelerating advancements that may be readily implemented.

Supplementary Material
Refer to Web version on PubMed Central for supplementary material.

Background
Localized opioid use epidemics of overdose and infectious diseases in the United States are a significant cause of morbidity and mortality. Anticipating these epidemics to plan an appropriate response, as done in the infectious diseases and disaster management fields, is urgently needed to reduce health harms.

Translational Significance
This narrative review identifies quantitative methodological approaches that have been employed for detecting and predicting opioid use-related epidemics. We synthesize these approaches and provide a multi-disciplinary framework outlining how methods from various sub-disciplines may be used in coordination to improve opioid use-related epidemic response.  The black curve corresponds to the epidemic at baseline, while the dashed orange, blue and green curves correspond to the epidemic in the presence of interventions resulting from early detection, risk assessment and prediction, respectively. The time periods for each of these analyses are also colored in blue, green, and orange, respectively. We hypothesize that accurate prediction would have the strongest prevention impact because it would confer time to plan and implement an appropriate response, followed by risk assessment, which is less specific and therefore less informative, and by early detection, which is highly specific but occurs once the epidemic has started spreading. However, this impact will depend on how this evidence is used by decision makers.  The diagram should be read from the bottom upwards, with each layer corresponding to a different component determining the choice of method in a step by step manner: 1) health outcome of interest, 2) type(s) of surveillance data available, 3) characteristics of the collected data, 4) objective, 5) analytical method. First, the bottom layer refers to the surveillance and analytical infrastructure -these are pieces that must be in place to collect data and to analyze it. The role of surveillance, in the context of these studies, is often undertaken by public health agencies and institutions such as the CDC. As such, the first step is identifying available ORE-driven outcomes. These generally represent the outcome of focus for a given research project. Then, for identifying available measures of the outcome and potential predictors, researchers should ask which types of surveillance data are available to them. Clinical data (i.e., EMS, hospital records, death records), epidemiologic data (i.e., poison call centers, harm reduction services, 311 calls), behavioral data (i.e., observational studies, internet data), and drug market data (i.e. DEA, drug sample testing, wastewater sampling) represent four types of data of importance to consider. At this stage, depending on study purpose, we recommend that researchers aim to identify sources for each type of data. Next, after identifying potential data sources, data should be extracted. We have identified five types of data (traditional epi, internet data, genetic data, geospatial data, and social network data) -identifying which types of data are available can inform study objective. Prior to selecting the method to employ, we then recommend choosing an overarching objective. Failing to do so can lead to confusion amongst the research team about the underlying purpose of a study. For example, certain approaches may be well-suited for risk assessment but not prediction, and a failure to explicitly identify study objective prior to choosing a method may result in choosing an inappropriate analytic approach. Finally, once the research team has identified their health outcome, the data that is available, and the overarching objective of their study, they can select the method(s) best suited to their data and objective. In addition, multiple methods can be used in parallel to increase the sensitivity and accuracy of findings. Papers that met study inclusion criteria. They are first organized by over-arching objective of the study (early detection, risk assessment, or prediction) and then by outcome type (overdose, bloodborne-illness, tuberculosis). The following information was extracted and presented in the table for each study: methodology (i.e., statistical regression, geospatial analyses, etc), overarching study objective, substances measured, health outcome of interest, region in which study took place, unit of analysis (i.e., state, county, locality), specific methods employed, and data sources. Presentation of the overarching methodologies identified in the review, the objectives (early detection, risk assessment, prediction) that they can be employed for, and the strengths and limitations inherent to each methodological approach. • Internet data is easily accessible.
• Internet data can be collected in realtime (i.e., social media posts, Google Trends) • Huge quantities of data can be collected.
• Quality and applicability of data for study purposes will generally be poor (involves many steps of data cleaning) • Ability to identify data specific communities at high risk is challenging (i.e., difficult to identify posts and content corresponding to vulnerable groups such as PWID) • Not representative of populations at highest risk which do not have internet access • Without qualitative and culturally-specific expertise (i.e., vernacular and social media norms), identifying relevant data points may be challenging or even impossible.

Dynamic Modeling Prediction
Can be used to: • make short and long-term projections of ORF impact • simulate population dynamics at the individual level (i.e. agent-based modeling) predict the efficacy of interventions aimed at attenuating harms of ORFs.
• predict the impact of changes to drug markets • Can incorporate qualitative data in model design • Over-simplification of real-world phenomena. • Data thirsty, limiting its application to specific localities or, on the contrary, to large regions, ignoring heterogeneity • Often rely on published research to inform model parameters, thus are subject to bias introduced by other studies.
• Data used to inform models is not always specific to the locality the model is being applied to.

Phylogenetics
Early Detection Can be used to identify: • locality-specific outbreaks of ORE driven infections • interpersonal spread of ORE driven infections drug use social networks • Can be used to confirm the connection between cases arising from a given ORE.
• Genetic data is challenging and expensive to collect • Generally, only available for specific localities undertaking phylogenetic driven research and prevention initiatives.
• Results, generally, cannot be generalized to inform ORE detection, assessment, or prediction in other localities Transl Res. Author manuscript; available in PMC 2021 August 01.  Challenges and good practices when undertaking research aimed at the early detection, risk assessment, and/or prediction of opioid use related-epidemics. Note: Challenges are organized into three categories: challenges related to data; challenges related to methods; and challenges related to dissemination of results.

Data
Lack of standardized opioid related epidemic clinical, epidemiologic, behavioral and drug market surveillance • Invest in the systematic monitoring of clinical and mortality outcomes, including the collection of substance use indicators • Implement enhanced surveillance protocols for substance use related infectious disease outbreaks, including strategic phylogenetic analyses • Invest in the monitoring of poison center call data and similar resources and make these available to researchers • Enable the systematic collection and sharing of data at substance use treatment and harm reduction services • Explore alternative methods to collect substance use behavioral surveys, including participatory surveillance through online platforms • Implement and continuously refine internet-based substance use surveillance • Invest in drug market surveillance, implement wastewater sampling, free drug testing services for people who use drugs and standardized testing and reporting of drug seizures by DEA • Invest in estimating the prevalence of locality-specific high-risk drug use indicators like injection drug use to better estimate locality "carrying capacity" Expertise Across a Range of Disciplines is Required • Focus on building multi-disciplinary workforce • Develop and promote funding mechanisms for multi-disciplinary, multi-method research • Utilize qualitative methods to aid in the design of modeling strategies, which are well-equipped to identify fast moving changes in the landscape of the opioid crisis Limited and delayed access to valuable data sources collected at national level • Enable timely access to local NSDUH data (while the sampling is designed to provide data representative at state level, the local estimates can provide key insight and uncertainty can be handled rigorously) • Enable timely access to local NFLIS data • Enable timely access to local NEMSIS data • Enable timely access to routine national/state infectious disease surveillance data as well as all other surveillance data available Disorganized management of available data Create a log of available data sources, including what geographic divisions and localities they are available and for what time divisions (i.e. month/year) and periods • Focusing on smaller geographic regions (such as counties or states, instead of the entire US) will likely allow for richer set of available data to be used • Develop and share coding pipelines such that when new data is available, analytic data sets can be updated efficiently

Explanatory and Predictive
Modeling are Distinct Methodological Tasks • Place emphasis on predictive performance, not on explanatory findings • Include evaluation of internal validity (such as cross-validation) to ensure model is not overfit. • Include evaluation of external validity to ensure model performance will meaningfully address research question at hand Validity of Internet based substance use surveillance is difficult to ascertain • Internet data represents a valuable resource to complement logistically complex household surveys and research efforts should be dedicated to refining current methods to ensure sensitivity and specificity of their findings in this new field Results from dynamic models making long term predictions (5+ years) are very sensitive to assumptions on trends in behaviors and sizes of at-risk populations • Be very transparent about assumptions for future behavioral/mortality trends • Carry out sensitivity analyses • Compare to findings with similar studies and encourage model comparison exercises Invest in estimation of population sizes of people who use drugs Different methods will provide complementary insights about OREs • Use multiple methods providing short and long term as well as small and large geographical scale predictions to inform decision making, including qualitative research Predictive Models Can Replicate Systemic Racism (and otherisms) Depending on Variable Selection • Explicitly justify or avoid using immutable, identity-based characteristics (such as race) as factors in predictive modeling endeavors • Exception to this is if the modeling approach employed can effectively account for the mediating and moderating pathways by which such identity-based characteristics impact the outcome of interest.

Dissemination
Transl Res. Author manuscript; available in PMC 2021 August 01.