Fatal Overdose: Predicting To Prevent

In their article describing the overdose epidemic in the US since 1979(Jalal et al., 2018), Jalal and colleagues investigate its long term trends to “understand the epidemic dynamics and perhaps predict its future course”. Indeed, predicting an epidemic could provide public health benefit through warning and informing communities at risk, scaling up prevention methods and training healthcare providers. Longer term predictions can inform which geographical areas should be prioritized for investments in healthcare system strengthening to mitigate health harms associated with substance use. However, in their thorough dissection of the overdose epidemic through time and space, disaggregating deaths by substance, age, sex, race and urbanicity, only the overall growth could be described as responding to natural laws, following a surprisingly smooth exponential curve for the past 38 years.

This finding is even more striking given the chaotic manifestation of multiple heterogeneous sub-epidemics. Indeed, through their heat maps, the authors effectively illustrate how each substance left its own mark and depict the constantly changing landscape of overdose mortality. Based on these observations, they state that “mortality curves from individual drugs do not show regular or predictable growth patterns” and conclude that: “the recent historical variability with which some specific drugs have waxed and waned makes predictions about the future role of specific drugs far more uncertain”. Instead, Jalal et al tantalisingly propose that a “fundamental longer-term process” may be driving the apparently predictable long-term exponential dynamics of the national epidemic. They posit that economic and technological factors have increased supply, along with increased demand resulting from erosion of community cohesion and the deepening of sentiments of despair in the face of social, political and economic hardship. They challenge us to think and seek further, in an effort to identify the larger forces behind the epidemic.

We agree it is a priority to identify longer-term processes fuelling the overdose epidemic. The persistent failure of legal and policy interventions such as the misguided war on drugs continues, and a change in approach is long overdue. However, the United States is large and heterogeneous, as the COVID-19 pandemic illustrated. Socio-structural circumstances driving substance use are particular to each community’s historical, geographical and cultural background and they interact differently with different substances. We believe that in order to identify the processes behind this exponential curve, we need to go back to the unit level and understand sub-epidemics. The apparent differences between national and county-level overdose epidemics could be reconciled through a better understanding of the latter. This situation is akin to disagreements between micro and macro-economic models and a lack of unifying theory bridging the two. However, both the macro and the micro-level overdose epidemics are experienced by the same individuals, whereas micro- and macro-economic phenomena occur at different levels; there is for example a disconnect between individuals’ financial experiences and trade agreements between countries. In addition, a subtler understanding of local epidemics, through methods development and collection of surveillance data, should also enable their prediction at the level which can most usefully inform actionable public health intervention, which brings us back to Jalal’s et al original question concerned with forecasting sub-epidemics.

In this commentary, we set out to address the substantial challenges in predicting the trajectories of local overdose sub-epidemics rightly pointed out by Jalal et al. We present methodological advances in the field of substance use disorders (SUD) and overdose epidemic prediction which highlight the feasibility and utility of prediction efforts. We then detail key data (individual, socio-structural, and drug market related) needed to inform prediction models and better simulate the determinants shaping SUD and local overdose sub-epidemics to inform policy. We expect that this endeavor may also help us make sense of patterns which on the surface appear irregular and in finding answers to Jalal’s fundamental question.

Analytical tools for the prediction of overdose epidemics

Several studies illustrate the potential benefits of applying analytical tools to leverage surveillance data, incorporating knowledge on risk factors at the individual, socio-structural and drug market levels. We recently performed a literature review(Marks, Carrasco-Escobar, et al., 2021) of studies implementing quantitative methods for the early detection, risk assessment, and prediction of opioid related epidemics of overdose and infectious diseases (specifically HIV, HCV and tuberculosis) in the US; a publicly available living repository of relevant studies is available (https://www.emergens-project.com/copy-of-repository). While we found few prediction studies investigating the opioid epidemic at population level, the field is rapidly growing and leveraging methods from a range of disciplines, including geospatial analyses and machine learning.

Early identification or prediction of SUD or fatal overdose outbreaks in the short term

While Jalal et al describe a long-term exponential process spanning over 40 years, the overdose epidemic is formed by innumerable smaller overdose outbreaks, and identifying predictors could provide clues about key drivers behind this curve. Pragmatically, short term prediction (i.e. within the year preceding the outbreak) could also allow us to mount a pre-emptive response and prevent deaths. For example, geospatial analyses identified clusters of higher than expected prescription rates(Basak, Cadena, Marathe, & Vullikanti, 2019; Brownstein, Green, Cassidy, & Butler, 2010), and some showed associations with opioid-overdose death rates(Marotta et al., 2019; Stopka et al., 2019). These studies identify communities potentially at higher risk of overdose outbreaks and inform interventions to address inappropriate prescription practices, if present. Studies merging multiple surveillance sources have shown it was feasible to detect changes in the drug supply (i.e. presence of fentanyl) several weeks in advance of increases in overdose presentations at the ED(L. S. Friedman, 2009), demonstrating the importance of improving surveillance infrastructure to provide warning to communities and deploy emergency interventions (e.g. fentanyl test strips, deployment of temporary low threshold supervised consumption spaces). Modeling from British Columbia has shown that despite overdose deaths rising steeply during the fentanyl epidemic, the toll would have been substantially greater in the absence of emergency harm reduction interventions(Irvine et al., 2018; Irvine et al., 2019). Of course, their deployment relies on resource availability, healthcare and harm reduction workforce capacity and, most importantly, political will.

When focusing on overdose as the outcome, several studies, including those by Sumetsky et al(Sumetsky et al., 2021), Campo et al(Campo, Gussler, Sue, Skums, & Khudyakov, 2020), and Marks et al(Marks, Abramovitz, et al., 2021), predict overdose death rates at county level in the next year. All showed that using statistical models informed with (mostly publicly available) data on overdose death rates in previous years, as well as with socio-demographic, healthcare access, drug market and internet search data, enhanced predictive ability of which counties would experience overdose outbreaks. Importantly, these studies all retrospectively predicted overdose death rates. One of the main barriers to predicting future overdose rates is that the models rely on past year overdose death data, yet the CDC data has so far had a two-year lag. Fortunately, the CDC has started releasing data sooner (i.e. suppressed data for 2020 and provisional data for 2021 were made available end 2021). In the first week of January 2022, our team released predictions of overdose death rates at county level for 2022, identifying counties at highest risk of experiencing outbreaks(Charles Marks, 2022). To our knowledge, this is the first attempt to provide early warning about the likely evolution of the overdose epidemic for policy guidance and we hope it will encourage multiple research groups to generate equivalent predictions using a variety of prediction models. Such models could also use data from state opioid dashboards which can be timelier, but these are only available in a limited number of states. In our model, we found the overdose rate in nearby counties in the previous year to be the strongest predictor for next year overdose rates in a given county, suggesting a geographical spread of the epidemic may be operationalized through the drug supply. Unfortunately, this is difficult to confirm given drug seizure data is only publicly available at state level in an aggregated form. Retrospective access to drug market data would yield important insights, thereby contributing to answering Jalal’s question.

Predicting SUD and overdose in the long-term to prevent next generations from perpetuating the cycle of SUD

Most prediction studies examine timescales of one year or less, which is justified given the urgent need for guidance to prevent further overdose and infectious disease outbreaks. However, Jalal illustrates pointedly that this is a sustained and accelerating crisis, and it will most likely have implications for future generations, potentially supporting its continued exponential growth. Over 2 million children (or 3% of children in the US) live with at least one parent with SUD in the past year(Lipari & Van Horn, 2013). The foster care system has seen a steep increase in the proportion of children entering as a result of parental substance use (from 14.5% in 2000 to 36.3% in 2017(Meinhofer & Anglero-Diaz, 2019)) and it is uncertain how many have been orphaned as a consequence of the opioid epidemic. Children of parents with SUD are at higher risk of developing SUD themselves due to trauma and structural conditions that often go hand in hand with SUD, including economic deprivation and limited access to healthcare(Biederman, Faraone, Monuteaux, & Feighner, 2000; Hoffmann & Cerbone, 2002; Yule, Wilens, Martelon, Simon, & Biederman, 2013). In addition, many more children are growing up in communities deeply affected by substance use epidemics, potentially exposing them to substance use from a younger age and to challenging environments characterized by higher crime rates and low social capital. There is evidence suggesting a relationship between current and past overdose epidemics. Segel and Winkelman(Segel & Winkelman, 2021) found an association between state level opioid-related overdose death rates from 1999–2004 and opioid-related (but also sedatives- and stimulants-related) overdose death rates over 10 years later when controlling for basic socio-economic factors.

Modeling is warranted to yield insights as to how intergenerational dynamics contribute to long-term overdose population dynamics. To inform these models, we need to better understand the driving factors which determine the development of SUD from early life stages at the individual, family and community level through retrospective and prospective longitudinal studies(Moska et al., 2021). Predictive approaches are emerging(Wagner et al., 2021) that could help identify families and communities at greatest need for interventions. They could also contribute to our understanding of the observed exponential trend, as we may find that the number of families and communities exposed to those specific set of circumstances, or that the severity of these circumstances, has grown over time.

Predicting intervention impact at population level to effectively change overdose epidemic trajectories

While predicting overdose epidemic trajectories is needed to warn communities and deploy timely prevention, it is also crucial for models to investigate intervention impact. In other words, how do we “flatten the curve” observed by Jalal and colleagues? How do we best allocate resources for interventions such as increasing access to medications for opioid use disorders, overdose education and naloxone distribution programs or providing peer support upon release from incarceration, as well as for structural interventions such as increases in minimum wage, housing access, and employment opportunities? As mentioned by Jalal, reproducing the natural history of SUD, as well as the effects of treatment is needed to adequately understand the mechanisms shaping fatal overdose epidemic and the impact of interventions. However, these dynamic models require understanding of the interplay between overdose risk and individual and socio-structural determinants (including drug market changes) – thus necessitating strengthening of data collection in these areas.

Strengthening our surveillance infrastructure to predict overdose epidemics

The strengthening of surveillance data is required to make sense of the seemingly unpredictable growth patterns of sub-epidemics as observed by Jalal, and to form meaningful predictions. Here we describe key areas requiring consolidation.

Individual-level determinants

Our ability to use existing knowledge on individual level determinants of SUD and overdose to inform predictive models is limited by the structure and accessibility of the data. Sources of information describing key risk factors such as previous overdose, access to treatment for SUD, history of mental health issues and treatment, exposure to incarceration, homelessness, unemployment, and availability of health insurance are collected and held by separate institutions or agencies and generally only available at aggregate level. A few exceptions include data from the Veteran Health Administration (VHA) or Medicare, and from isolated examples of administrative data linkage efforts(Keen, Young, Borschmann, & Kinner, 2020; Krawczyk et al., 2020), which are not always feasible. The standardization of data-linkage processes as implemented in the State of Massachusetts(Bearnot, Pearson, & Rodriguez, 2018) would take us a long way in terms of assessing individual risk in different communities and clarifying associations with social and structural determinants.

Social and structural determinants

Several studies(Dasgupta, Beletsky, & Ciccarone, 2018; Galea, Ahern, & Vlahov, 2003; Monnat, 2018; Peters, Monnat, Hochstetler, & Berg, 2020; Rigg, Monnat, & Chavez, 2018) have shed light on social and economic factors driving community susceptibility to overdose, but greater granularity of data is needed. A study of US counties stratified by drug epidemic type(Peters et al., 2020) suggests poverty is mostly a driver in rural areas which suffered from the decline of specific industries, leading to high rates of unemployment and underemployment. In contrast, social disorganization is found to play a more important role than poverty in urban epidemics driven by heroin. The study highlights how further granularity in the geographical scale, temporality, and phenomenon investigated is needed. For example, county-level poverty will not tell us about particular communities exposed to very high poverty levels and therefore its effect on overdose risk will get diluted when included in a predictive model. The American Community Survey now provides zip code level data which will help with future predictive endeavors and geospatial data is being collected more systematically across agencies.

The socio-economic indicators measured also need to be clearly defined. For example, while rates of unemployment might seem like an obvious indicator to measure, Peters et al(Peters et al., 2020) argue it is high rates of unemployment among low-skilled occupations that will have a particularly harmful impact, and therefore surveys must enable this distinction. Finally, further diversity in the measured indicators is needed to capture the complexity of how different factors operate to influence overdose risk. In particular, accounting for resilience is important when informing models, as it might offset the effect of risk drivers. In their mapping tool of fatal opioid overdose and associated factors, NORC incorporated a prosperity index score, reflecting both social and economic risk and resilience(NORC at the University of Chicago, 2021). Further, unprecedented events such as the pandemic, harrowing incidents such as the killing of George Floyd, or common occurrences such as closures of automotive manufacturing factories, create new vulnerabilities(Bor, Venkataramani, Williams, & Tsai, 2018; J. Friedman & Akre, 2021; Venkataramani, Bair, O’Brien, & Tsai, 2020). Capturing their effect can require developing new measures (e.g. COVID stress scales(Taylor et al., 2020), protest aggregator(Neyman & Dalsey, 2021)) or the application of existing instruments (e.g. perceived racism/racial discrimination(Atkins, 2014)). Paying careful attention to political and socio-economic change at community level and expanding both data collection and extraction accordingly to capture their effect will enable a better understanding of the mechanisms behind overdose epidemics as well as more accurate predictions. Coming back to Jalal’s question, the structural forces driving the exponential curve have likely materialized in multiple forms over time and space, but through better characterizing the impact of these different manifestations on the overdose epidemic, we will be in a better position to find a common denominator and to advocate for appropriate solutions.

Drug markets

Different substances have distinct patterns of availability and characteristics, reaching different groups by urbanicity, age, ethnicity or socio-economic status as shown in Jalal’s dissection of the epidemic by drug. Consequently, they have different impacts on health, requiring specific tools to detect, prevent, and treat them. However, monitoring drug markets (both licit and illicit) at population level is challenging.

Information on illicit markets is mainly available through data on drug seizures shared by the DEA, but, as mentioned, this is only publicly available at state level and it only provides estimates of counts, rather than volume (or morphine milligram equivalents (MME) for opioids, allowing control for the wide range in opioid potency). Increased fentanyl presence in drug seizures has been associated with increased overdose deaths(Rosenblum, Unick, & Ciccarone, 2020). Similarly, increased fentanyl presence in the stimulants’ supply(Park et al., 2021) has been observed in areas where stimulant-fentanyl related overdose death rates have increased(Fleming, Barker, Ivsins, Vakharia, & McNeil, 2020), showing that these data can be used to alert communities at risk. More timely and granular drug seizure data in terms of geographical level and seizure characteristics would represent a valuable surveillance tool to leverage for prediction.

Innovative data sources for drug trends are increasingly available. Harm reduction free drug checking services represent a valuable source of data to better characterize illicit substances and provide direct prevention benefits(Laing et al., 2021). Participatory-based sentinel surveillance to monitor specific markets, such as the “Ecstasy and Related Drugs Reporting System” in Australia(Price et al., 2021), have been successfully implemented. Routine testing among patients accessing primary care or emergency departments can be used to assess trends in use over time.(Wakeman, Flood, & Ciccarone, 2021) Wastewater testing for multiple drugs, as is done in Europe(EMCDDA, 2021), can measure population level consumption of substances(Gushgari, Venkatesan, Chen, Steele, & Halden, 2019), consumption trends, and potentially identify the presence of new drugs. Using indirect data collection methods, such as monitoring google searches or mining drug forums on social media for data on specific substances, SUD symptoms or help-seeking can also provide complementary information. The latter can also have real time public health benefits for those accessing these forums, for example when they provide information on fentanyl contaminated batches in specific locations(F, 2017). Rapid ethnographic assessments(Ciccarone, Ondocsin, & Mars, 2017) can quickly identify substance characteristics, such as powder color and taste, that are associated with increased risk, as well as behaviors and coping mechanisms in response to changes in drug markets(Mars, Ondocsin, & Ciccarone, 2018).

Monitoring licit drug markets is in theory simpler but often requires purchasing access to data from companies, meaning it is not available to all researchers or local public health departments. However, it provides key insight on potential emerging waves of substance use. For example, rates of prescribing stimulants increased steadily between 2008 and 2018 among US adults, aligning with increases in stimulant related fatal overdoses over the past few years(Ciccarone, 2021). Access to these data at local level and in real time could provide an early alert to changing patterns in prescriptions that could lead to increases in SUD.

Again, the capacity to act upon these near real-time data and implement successful overdose prevention interventions under a relatively short time period is yet uncertain, but we expect to see increasing evidence supporting drug checking and drug market surveillance efforts in the near future.

For longer-term predictions, higher level analyses of drug market trends at global level (including from dark web sales), accounting for geo-political circumstances and manufacturing capacity across countries as carried out by the DEA and other security agencies could provide insight about the substances that will dominate the market(Broseus et al., 2016). This is illustrated by Jalal in the context of “the rise and fall of cocaine-related overdose deaths”, which likely mirrored production and supply trends in Colombia, in turn driven by events during the civil war.

Obtaining more detailed data on drugs’ purity, potency, availability and price over time, which has unfortunately been so limited over the years due to their illicit nature, would allow us to rigorously examine the role of the drug supply in supporting the exponential growth of the overdose epidemic(Caulkins, 2007; Scott, Caulkins, Ritter, Quinn, & Dietze, 2015). Similarly, investigating global level dynamics of the world market might shed light on specific patterns that contribute to an amplification of drug availability or toxicity driving mortality.

Closing Comments

The simplicity of the exponential overdose curve observed by Jalal and colleagues raises important fundamental questions about core drivers of overdose and other “deaths of despair”. Yet, like the COVID-19 pandemic, the smooth overall U.S. epidemic trajectory masks numerous sub-epidemic trajectories across states and regions with substantial heterogeneity. We believe that understanding and predicting these sub-epidemics is hard, but necessary to inform local public health response in partnership with affected communities. Investing in prediction and in creating the datasets needed to enable it will help us manage the overdose crisis through pre-emptive action and will likely also contribute to understanding its course at a macro-level and to identifying key structural forces underpinning it. We see much potential in mass-research collaborations using the common-task approach in which multiple groups would attempt to predict county level overdose rates, as well as in model comparison initiatives to assess robustness of policy conclusions, as has been proposed for opioid use disorder(OUD Modeling writing group, 2021) and done in the fields of HIV(Eaton et al., 2012) or COVID-19(Aguas et al., 2020). We hope this commentary will encourage and facilitate such initiatives.