Accurate spatiotemporal mapping of drug overdose deaths by machine learning of drug-related web-searches

Abstract

Persons who inject drugs (PWID) are at increased risk for overdose death (ODD), infections with HIV, hepatitis B (HBV) and hepatitis C virus (HCV), and noninfectious health conditions. Spatiotemporal identification of PWID communities is essential for developing efficient and cost-effective public health interventions for reducing morbidity and mortality associated with injection-drug use (IDU). Reported ODDs are a strong indicator of the extent of IDU in different geographic regions. However, ODD quantification can take time, with delays in ODD reporting occurring due to a range of factors including death investigation and drug testing. This delayed ODD reporting may affect efficient early interventions for infectious diseases. We present a novel model, Dynamic Overdose Vulnerability Estimator (DOVE), for assessment and spatiotemporal mapping of ODDs in different U.S. jurisdictions. Using Google® Web-search volumes (i.e., the fraction of all searches that include certain words), we identified a strong association between the reported ODD rates and drug-related search terms for 2004–2017. A machine learning model (Extremely Random Forest) was developed to produce yearly ODD estimates at state and county levels, as well as monthly estimates at state level. Regarding the total number of ODDs per year, DOVE’s error was only 3.52% (Median Absolute Error, MAE) in the United States for 2005–2017. DOVE estimated 66,463 ODDs out of the reported 70,237 (94.48%) during 2017. For that year, the MAE of the individual ODD rates was 4.43%, 7.34%, and 12.75% among yearly estimates for states, yearly estimates for counties, and monthly estimates for states, respectively. These results indicate suitability of the DOVE ODD estimates for dynamic IDU assessment in most states, which may alert for possible increased morbidity and mortality associated with IDU. ODD estimates produced by DOVE offer an opportunity for a spatiotemporal ODD mapping. Timely identification of potential mortality trends among PWID might assist in developing efficient ODD prevention and HBV, HCV, and HIV infection elimination programs by targeting public health interventions to the most vulnerable PWID communities.

Identifying counties at risk of high overdose mortality burden during the emerging fentanyl epidemic in the USA: a predictive statistical modelling study

“…other wise have been considered to be at risk. Although it was not possible to do so in this study, comparing the performance of our model with that of other models introduced in the literature (such as that by Sumetsky and colleagues, 18 Campo and colleagues, 23 and Lyle Cooper and colleagues 24 ) might advance the broader effort to develop better performing models. Further, given that this is a nascent line of research, we highlight the impor tance of evaluating the performance of various modelling strategies.…”

“…Although similar models have been used to inform funding allocation, such as the CDC's drug-related HIV outbreak risk assessment model, 12,22 these models have not been validated against data and have not been designed to provide yearly predictions-studies by Sumetsky and colleagues 18 and Campo and colleagues 23 are two exceptions. Model validation is key to both ensuring that the tools used for policy guidance are providing accurate information and to improving our understanding of the epidemic processes.…”

“…Previous works have applied Bayesian spatial-temporal models, polynomial func tions, and a variation of the random forest algo rithm. 18,23,24 Future research should aim to replicate and compare these methods to identify strengths of each approach, which can inform future model development.…”

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

“…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.…”

“…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.…”