Misuse of illicit and prescription opioids is a significant global problem that affects the health and economic welfare of individuals, families, and society. The U.S. opioid overdose rate has quadrupled since 1991 [1]. In 2017, more than 47,000 Americans died of an opioid overdose, and 36% of these deaths involved prescription opioids [2]. A major goal in treating opioid use disorder (OUD) is abstinence, or complete cessation, of opioid use, other than the use of prescribed opioid replacement therapy. There is not a single, clinically accepted definition of cessation that specifies the length of abstinence required before an individual is no longer considered to have OUD [3, 4]. Diagnostic and Statistical Manual of Mental Disorders, 5th Edition (DSM-5) considers sustained remission from OUD as a one-year period during which no criteria for the disorder (other than craving) are met [5].

Population differences affect multiple aspects of the current epidemic. Although opioid use nationally is higher among individuals of European ancestry (EAs) than African-Americans (AAs), the opioid death rate has increased more sharply among AAs than EAs [6]. AAs have less access to treatment for OUD [7], are less likely to obtain opioid prescriptions for pain management [8], and are incarcerated at a higher rate for illicit opioid use [9] than EAs. Previous research on OUD-related outcomes has been conducted primarily in combined ethnic groups or in EAs only [10], limiting the identification of key population differences in opioid use and treatment outcomes.

Although moderately correlated with opioid cessation, factors contributing to opioid treatment completion such as age, employment status, and age at first drug use have been identified from a mixed ethnicity sample [3]. Other factors are likely to influence cessation, such as pain experiences, general health, and the use of antidepressants [11–13]. Delineation of these factors could inform OUD treatment strategies that may differ across population groups; or could be useful for individuals with OUD who aim to reduce or stop their opioid use. However, studies thus far have tended to focus on a small number of clinically relevant factors such as the dosage, duration, and formulation of medication-assisted treatment of substance use disorders [14–16]. Large epidemiological studies of OUD [17–19] comprised of thousands of variables would allow a systemic, hypothesis-free query to identify factors predicting opioid cessation.

Statistical methods are generally limited in their ability to sort through large numbers of predictors [20]. Data mining using machine learning, which is particularly well suited for identifying predictive factors among thousands of variables [21, 22], has successfully identified predictor variables for a diverse set of outcomes [23–27]. Here, we applied multiple machine learning techniques to evaluate a large set of clinical, demographic, general health, and behavioral variables in a large, racially mixed cohort of individuals who were ascertained for cross-sectional genetic studies of substance use disorders, but not necessarily treated for OUD, to identify factors that are associated with opioid cessation (defined as self-reported last illicit opioid use and/or prescription opioids misuse > 1 year before the interview date). Our study identified additional factors associated with cessation, including several that are population-specific. These findings support an individualized approach to improve the outcome of cessation attempts.

Characteristics of the study samples are shown in Table 1. The sample included 1,069 unrelated AAs and 2,252 unrelated EAs, and 123 AA and 305 EA participants who were members of families containing a pair of siblings concordant for opioid or cocaine dependence. There is a higher proportion of females among individuals who ceased opioid use in both AAs (OR = 1.35, P = 6.7 × 10⁻³) and EAs (OR = 1.31, P = 1.1 × 10⁻³) compared to those who did not cease. Participants who ceased opioid use were also older by an average of 3.18 years in the AA group (P = 1.0 × 10⁻¹⁰) and 6.1 years in the EA group (P = 2.2 × 10⁻¹⁶) than those who did not cease use. The mean number of lifetime DSM-5 OUD criteria did not significantly differentiate individuals who ceased opioid use from those who did not.

We employed both regression and non-regression-based machine learning approaches to evaluate the association of more than 3,000 variables related to SUDs and other psychiatric disorders, other health-related behaviors, and demographic variables with opioid cessation among EAs and AAs assessed in a cross-sectional study of opioid, cocaine, and/or alcohol dependence. We observed moderate-to-high predictive accuracy across all methods; SVM, on average, marginally outperformed the other methods. Although the specific set of associated variables differed in EAs and AAs, a common profile emerged. People who ceased opioid use tended to be older, initiated drug use later in life, had used opioids for a shorter period, experienced fewer problems related to cocaine or alcohol use, were currently employed, and had recovered from other psychiatric disorders including depression and PTSD compared to those whose opioid use persisted.

Previous research using machine learning for addiction outcomes focused mainly on predictive accuracy, although a few studies attempted to identify and interpret specific variables that were associated with the outcomes [3, 4, 40, 41]. Acion et al.[3] reported that ensemble super learning was superior to other machine learning methods, and used penalized regression, SVM, and neural networks for predicting SUD treatment success indicated by treatment discharge status in a Hispanic cohort. In that study, less than 10% of participants had problems with cocaine or illicit opioids and fewer than 30 variables were assessed. In contrast, we evaluated several thousand variables, including detailed measures of drug-use activities and psychiatric disorders, and ranked the importance of the top-ranked variables with four distinct machine learning algorithms. Gowin et al.[40] identified regional brain activity changes predicting relapse from imaging data on fewer than 70 methamphetamine-dependent patients without including any lifestyle factors. Che et al.[4] applied deep learning to electronic health record data to identify people with short-term or long-term opioid use or dependence. Similar to our study, they identified associations with comorbid substance use and anxiety disorders [4]. Several other studies used only regression-based methods to identify variables associated with opioid and stimulant dependence [42], cocaine dependence [43] and alcohol dependence [44], which might not capture other relationships among variables. Several of the non-regression-based methods we employed have also been applied in other studies, which focused mainly on brain magnetic resonance imaging (MRI) traits as predictors of substance use disorder diagnoses [24, 25, 27, 40].

We identified association of opioid cessation with several variables that were previously associated with OUD or OUD-related conditions including co-morbid drug use, antisocial behavior, suicidal thoughts, HIV infection, and asthma [45–50]. Our finding that the majority of people who ceased opioids (60% in AAs and 66% in EAs) also ceased cocaine use is consistent with evidence of high rates of co-occurring OUD and cocaine use disorder (CUD) [51, 52]. This finding also supports the use of treatment strategies that target both disorders [51, 53] and suggests that ceasing use of one substance might influence the ability to cease use of the other. Alternatively, ceasing both opioid and cocaine use may reflect self-selection for inclusion in our genetic studies in which 43% of AA and 32% of EA participants were ascertained for CUD. Our findings are also consistent with observations that a failure to address tobacco use lowers the efficacy of opioid cessation treatment [54] and that a behavioral intervention in patients with antisocial personality disorder reduces substance use [55]. Unlike problems that are associated with other drug use and lower odds of opioid cessation, we found that cannabis use-related problems (e.g., two marijuana symptoms lasting a month, marijuana interfering with work) are associated with higher odds of opioid cessation. This finding is puzzling and not immediately explainable. Although our observation that cannabis users had better success quitting opioids is consistent with prior reports of association of cannabis use and reduced opioid withdrawal symptoms and pain [56, 57], recently we showed that cannabis as a replacement for opioids as treatment strategy for OUD could be harmful [58]. Previous findings of the co-occurrence of drug addiction, suicide attempts, depression, family conflicts, and PTSD, which may suggest bi-directional casual relationships [45–50], are consistent with our observation that better management of comorbid psychiatric problems (fewer recent suicide attempts and psychiatric symptoms) increases the likelihood of opioid cessation or vice versa.

Acion et al.[3] and we identified age, employment status, and age at first drug use as factors for treatment success. The protective effect of older age may be due to ascertainment bias because persons who survived severe dependence are more likely to have stopped using opioids. Full-time employment likely reduces the time or urge for persons dependent on opioids to seek and use the drug. In addition, drug screening associated with some jobs may reduce the likelihood of current opioid use [59]. Quitting opioids also make it easier to find/maintain a job. The inverse correlation of age at first drug use and opioid cessation may reflect the increased difficulty of reversing the effect of long-term opioid exposure on the brain reward system [60] or increased severity associated with earlier onset.

Several variables that were significantly associated with opioid cessation related to non-substance-related behavior were population specific. Although these findings may be due in part to differences between AAs and EAs in willingness to endorse these behaviors, previous studies showed that AAs were more likely than EAs to report prolonged gambling and problems associated with gambling [61, 62]. One explanation for our findings of significant associations of ceasing opioids with a self-reported HIV diagnosis in AAs is that OUD patients with severe or life-threatening illnesses are more likely to seek or adhere to treatment [63], an idea supported by evidence that HIV-infected patients have better treatment outcomes for OUD [64–66]. Alternatively, poorer general health may lead to reduced drug use [67] (the so-called “sick quitter”). In contrast, antisocial behavior, recovery from PTSD, and being an atheist were associated with opioid cessation in EAs only. Prior research may provide insight into these EA-specific patterns. One study reported antisocial behaviors in EA children were significantly associated with substance initiation while the association was less strong in AA children [68], although the impact on opioid use was not assessed. PTSD and being an atheist identified in EAs might be due to the racial difference in exposure to traumatic events and belief diversities [69, 70]. Previous evidence about the effect of religion on SUDs is contradictory. One study showed that loss of religiosity between childhood and adulthood was associated with increased substance use while recent religiosity increased the odds of illicit drug use in the past year [71]. Alternatively, the smaller sample of AAs might have limited our ability to detect these associations in that group.

The current study has several strengths. First, because the input dataset contains thousands of variables related to drug use activities, psychiatric disorders, medical history, and demographics obtained from several thousand individuals meeting DSM-5 criteria for OUD, we were able to explore many factors in addition to those included in other studies. Second, both linear and non-linear machine learning methods were employed to model the true underlying relationship between the variables and outcome, which increased the number of factors we identified. Third, we evaluated three models for each machine learning method in order to better understand the contribution of opioid and other drug use information. Fourth, we considered only independent variables in the association analyses to prevent over-representation of correlated factors. Finally, although there is no published “gold standard” predictive model against which to compare our results, the 80% predictive accuracy we achieved is similar to that seen in other machine learning studies [3, 4, 40, 41].

Limitations of this work should be noted. First, given the cross-sectional nature of our data and the over 90% relapse rate for OUD [72], many individuals classified as not using opioids may have subsequently relapsed to opioid use. However, it has been shown that prior abstinence is predictive of future abstinence, therefore people who ceased opioids are more likely to cease again even when relapse occurs [73]. Second, the machine learning analyses were based on samples that may have been underpowered to detect associations with some variables compared to other studies that included tens of thousands of individuals [74]. Third, most persons in our cohort were evaluated prior to the current opioid epidemic and may not reflect recent secular trends in the prevalence and associated features of OUD. Fourth, associations of cessation with some variables and overall prediction accuracy may have been inflated because our analysis did not fully account for familial correlations. Fifth, in spite of the large number of variables that were included in the machine learning analyses, potentially important variables such as the reasons for first use and details of treatment and support programs were unavailable. However, we identified attending an opioid addiction self-help group as associated with successful cessation, which is consistent with the reported benefit of self-help groups [75]. Sixth, the rate of response to many interview questions was substantially higher in EAs, while the sample size was twice that of AAs, which could account for some of the observed racial differences in predictive models. Related to this concern, our use of mean imputation for missing data may have been overly conservative. Other methods, including cold/hot deck (such as K nearest neighbors) and multiple imputation [76], may have provided additional information but were not appropriate for this dataset given its size, use of continuous, ordinal, and binary variable coding, non-linear relationships among variables, and lack of an appropriate external reference. Seventh, given the limited amount of temporal information with respect to many of the potential risk factors for opioid cessation in this cross-sectional sample, it is difficult to determine the effect direction for many of the observed associations. Because of these limitations, our findings require external validation in larger samples before they can be incorporated in prediction models for clinical purposes. Finally, while some of the factors we identified are plausible and consistent with prior studies, other factors such as atheism are not immediately interpretable. Thus, because our research is atheoretical, results should be interpreted with caution and be validated before implemented in clinical practice.

In conclusion, we analyzed a large number of variables including demographic, behavioral, health and drug use activities using machine learning techniques with feature selection and found variables in a wide range of domains that were associated with cessation. These included some that are consistent with prior literature, plausible but have not been well studied, and do not have readily apparent explanations for their associations. Our findings suggest hypotheses for future studies and could inform how one might increase the likelihood of cessation with and without treatment. These results also support several widely known treatment strategies for OUD, such as treating psychiatric comorbidity, adding wraparound services such as employment counseling, and simultaneously addressing polydrug use problems. Finally, in an era of increasing availability of digitized health-related records, our study provides a framework for disease outcome prediction using high dimensional phenotypic data collected via a research instrument.