Edited by: Stella Dracheva, Icahn School of Medicine at Mount Sinai, United States
Reviewed by: Graeme Henderson, University of Bristol, United Kingdom; Eugene A. Kiyatkin, National Institute on Drug Abuse (NIDA), United States
This article was submitted to Neuropharmacology, a section of the journal Frontiers in Neuroscience
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Opioid abuse is now the primary cause of accidental deaths in the United States. Studies over several decades established the cyclical nature of abused drugs of choice, with a current resurgence of heroin abuse and, more recently, fentanyl’s emergence as a major precipitant of drug-related deaths. To better understand abuse trends and to explore the potential lethality of specific drug–drug interactions, we conducted statistical analyses of forensic toxicological data from the Wayne County Medical Examiner’s Office from 2012–2016. We observed clear changes in opioid abuse over this period, including the rapid emergence of fentanyl and its analogs as highly significant causes of lethality starting in 2014. We then used Chi-square Automatic Interaction Detector (CHAID)-based decision tree analyses to obtain insights regarding specific drugs, drug combinations, and biomarkers in blood most predictive of cause of death or circumstances surrounding death. The presence of the non-opioid drug acetaminophen was highly predictive of drug-related deaths, likely reflecting the abuse of various combined acetaminophen-opioid formulations. The short-lived cocaine adulterant levamisole was highly predictive of a short post-cocaine survival time preceding sudden non-drug-related death. The combination of the opioid methadone and the antidepressant citalopram was uniformly linked to drug death, suggesting a potential drug–drug interaction at the level of a pathophysiological effect on the heart and/or drug metabolism. The presence of fentanyl plus the benzodiazepine midazolam was diagnostic for in-hospital deaths following serious medical illness and interventions that included these drugs. These data highlight the power of decision tree analyses not only in the determination of cause of death, but also as a key surveillance tool to inform drug abuse treatment and public health policies for combating the opioid crisis.
Drug-related deaths now exceed all other causes of accidental death; the majority of these deaths involves opioid abuse (
In the mid-2000s, the detection of a multistate pattern of fentanyl fatalities led to concerted public health and law enforcement responses believed to have cut short that outbreak (
The WCMEO conducts medico-legal investigations to determine the cause and manner of death in fatalities resulting from violence, in persons without recent medical attendance, and under other circumstances as outlined in the laws of the State of Michigan. The WCMEO makes such determinations on the basis of medical, police and other records, scene investigations, and autopsies, including pathology and toxicology findings. Death investigations by the WCMEO are mandated by law and not subject to review by an Institutional Review Board.
Comprehensive toxicological analyses were performed for the WCMEO by NMS labs
Statistical analyses were conducted using IBM SPSS Statistics version 24. Quantitative toxicological results for each drug were transformed into binary variables of 0 (for a negative finding) or 1 (for a positive finding, signifying the presence of a drug and/or its metabolites). These values were organized into a matrix that included de-identified case classification as drug death or non-drug deaths, which was then used for statistical analysis with an initial focus on changes in the prevalence of different drugs of abuse over time and the extent of co-abuse in drug deaths.
To evaluate the predictive power of drug-positive toxicology in the classification of cases as drug deaths vs. non-drug deaths, IBM SPSS Decision Tree Analysis was performed using the Pearson CHAID method. CHAID allows for the explorations of relationships between multiple predictor binomial variables (e.g., the presence or absence of a drug) and a categorical response variable (e.g., cause of death); it does so by building decision trees using chi-squared statistics that model these relationships hierarchically and provide an organized and interpretable visual representation of interactions within the data. In the present study, this method of analysis was used to identify individual drugs and drug combinations that were most significant in classifying drug-related deaths. Bonferroni-adjusted
To first ascertain the drugs most commonly associated with drug-related deaths for 2012–2016, and to understand how stable the patterns of drug abuse were over time, we examined toxicological data from all drug-related deaths in Wayne County, Michigan during this time period, as described in the Materials and Methods. Figure
Opioid drugs associated with drug-related deaths in Wayne County from 2012–2016. In each panel, the number of drug deaths associated with a specific opioid drug
The total number of drug deaths with a codeine-positive toxicology was approximately one-half that seen for morphine (Figure
A large number of drug deaths involved hydrocodone, which, unlike morphine and codeine, was markedly consistent on a year-to-year basis from 2012 through 2016 (Figure
The illicit stimulant cocaine and the prescription benzodiazepine alprazolam are two of the drugs most commonly co-abused with opioids and associated with opioid deaths (
Undoubtedly, the most striking change during this 5-year period pertained to detection in blood of the synthetic opioid fentanyl and its analogs. Fentanyl-positive cases (Figure
We next used decision tree (CHAID) analysis to predict the relationship between the presence of any individual drug and the likelihood of drug death. Every 2012–2016 WCMEO case with a quantifiable blood level of any drug or other chemical (i.e., not solely drugs of abuse;
Figure
Chi-square Automatic Interaction Detector (CHAID) analyses of drug-related deaths and non-drug-related deaths in Wayne County from 2012–2014. Determination of the drugs and drug combinations most predictive of drug deaths (DD, red) vs. non-drug deaths (NDD, green). In each node, the total number of cases and their partitioning between DD and NDD (n, %) is shown. For ease of visualization, the major branches have been divided after the first parent node into morphine-positive cases
Unexpectedly, in cases with morphine-negative toxicology (Figure
In the tree branch negative for both morphine and acetaminophen (Figure
Surprisingly, we observed that cases negative for morphine (a proxy for heroin), acetaminophen (a proxy for opioid/acetaminophen combination formulations), or methadone, but positive for cocaine (Figure
Figure
CHAID analyses of drug-related deaths and non-drug-related deaths in Wayne County from 2015–2016. Determination of the drugs and drug combinations most predictive of drug deaths (DD, red) vs. non-drug deaths (NDD, green). In each node, the total number of cases and their partitioning between DD and NDD (n, %) is shown. For ease of visualization, the major branches have been divided after the first parent node into morphine-positive cases
Undoubtedly, the most importance difference between the 2012–2014 and 2015–2016 CHAID analyses was the prominent appearance of fentanyl in the latter decision tree, as fentanyl now predicts drug death in both morphine-positive cases (2.353E-15; Figure
Interestingly, a small subset of morphine-negative, fentanyl-positive cases was also positive for midazolam, a short-acting benzodiazepine used primarily for in-hospital procedures, often in combination with fentanyl. These cases were overwhelmingly (81%) not drug deaths (Figure
In 2015–2016, in cases negative for both morphine and fentanyl (Figure
Examining the more branched 2015–2016 CHAID tree presented in Supplementary Figure
The data presented here resulted from systematic collection and unbiased statistical analyses of forensic toxicological results. They highlight the fact that continued abuse of prescription opioids, growing abuse of heroin, and a dangerously lethal infusion of fentanyl and its analogs account for most drug-based deaths in Wayne County, Michigan. This general trend mirrors that observed elsewhere in the United States, as does a renewed tendency for poly-drug abuse involving concurrent use of cocaine, benzodiazepines, or other drugs of abuse. The increasing prominence of fentanyl and its analogs is particularly alarming, as these opioids are harbingers of persistently increasing drug-related mortality. Left unabated, fentanyl and its more potent analogs could well displace heroin as the primary cause of drug-related deaths in Wayne County and nationwide.
The use of CHAID-based decision tree analyses provided new insights regarding specific drugs, drug combinations, and biomarkers in blood most predictive of cause of death or circumstances surrounding death. The presence of the non-opioid drug acetaminophen was highly predictive of drug-related deaths, likely reflecting the use of various combined acetaminophen-opioid formulations on the market. Detection of the short-lived cocaine adulterant levamisole was predictive of a short post-cocaine survival time preceding a sudden, often violent, non-drug death. The combination of the opioid methadone and the SSRI citalopram was universally associated with drug death, indicating the possibility of a drug–drug interaction with an effect on the heart and/or drug metabolism. Lastly, fentanyl, alongside the short-acting benzodiazepine midazolam, was diagnostic for death following serious illness and medical interventions that included administration of these drugs in a hospital setting. The extent to which the specific drugs, drug combinations, and blood biomarkers most predictive of drug death in Wayne County generalize to other areas of the United States remains to be determined, but is readily amenable to CHAID analysis.
Collectively, our results provide strong support for the use of forensic toxicology coupled with CHAID and similar analyses, not only in medico-legal death investigations, but also in a broader context as predictive surveillance tools that can inform public health and medical intervention strategies for substance abuse disorders and associated diseases (e.g., hepatitis A, B, and C, human immunodeficiency virus) that have far-reaching societal effects.
All pertinent data for this work are included in the manuscript and Supplementary Material. Data not included but referred to in text or numerically summarized in graphs and tables can be requested by contacting the corresponding author.
MS, CLS, CJS, and MB designed the study. CJS collected the samples. MS, CLS, MR, and ST performed the statistical analyses. MS, ST, CJS, and MB interpreted the results, prepared the figures, and wrote the manuscript. All authors read, critically revised, and approved the manuscript.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
The Supplementary Material for this article can be found online at: