Optimizing opioid use disorder treatment with naltrexone or buprenorphine

Abstract

Background:

Relapse rates during opioid use disorder (OUD) treatment remain unacceptably high. It is possible that optimally matching patients with medication type would reduce risk of relapse. Our objective was to learn a rule by which to assign type of medication for OUD to reduce risk of relapse, and to estimate the extent to which risk of relapse would be reduced if such a rule were used.

Methods:

This was a secondary analysis of an open-label randomized controlled, 24-week comparative effectiveness trial of injection extended-release naltrexone (XR-NTX), delivered approximately every 28 days, or daily sublingual buprenorphine-naloxone (BUP-NX) for treating OUD, 2014-2017 (N=570). Outcome was a binary indicator of relapse to regular opioid use during the 24 weeks of outpatient treatment.

Results:

We found that applying an estimated individualized treatment rule—i.e., a rule that assigns patients with OUD to either XR-NTX or BUP-NX based on their individual characteristics in such a way that risk of relapse is minimized—would reduce risk of relapse by 24 weeks by 12% compared to randomly assignment treatment.

Conclusions:

The number-needed-to-treat with the estimated treatment rule to prevent a single relapse is 14. A simpler, alternative estimated rule in which homeless participants would be treated with XR-NTX and stably housed participants would be treated with BUP-NX performed similarly. These results provide an estimate of the amount by which a relatively simple change in clinical practice could be expected to improve prevention of OUD relapse.

1. Introduction

Once initiated, extended-release naltrexone (XR-NTX) was recently found to perform equivalently to buprenorphine-naloxone (BUP-NX) in terms of preventing relapse to opioid use disorder (OUD) in a large, randomized controlled trial (X:BOT).¹ However, this does not mean that the two medications would perform equivalently for each individual initiating medication. The population of individuals with OUD is diverse, particularly in terms of externalizing symptoms and psychiatric and substance use co-morbidities,² motivating efforts to personalize treatment.³ It is possible that XR-NTX may be more effective for some individuals and BUP-NX may be more effective for others. It was an aim of the X:BOT trial to identify such subgroup heterogeneities, and evidence suggests important differences by homeless status.⁴ Beyond this, it is plausible that identifying subgroups for whom one medication works better than another involves the intersection of multiple characteristics as opposed to considering each characteristic in isolation.⁵

Such goals have motivated research into the identification of optimal individualized treatment rules, which are rules that, when applied, maximize treatment success across the population.⁶ Rules may be as simple as “treat everyone with BUP-NX” or “treat homeless individuals with XR-NTX and stably housed individuals with BUP-NX”. However, the optimal rule may be significantly more complicated.

Motivated by the practical goal of better understanding heterogeneities in the performance of XR-NTX vs. BUP-NX in treating OUD, we apply a flexible and robust method from the optimal individualized treatment rule literature⁷ to 1) learn a rule to assign treatment with either XR-NTX or BUP-NX to individuals to reduce risk of relapse to OUD (the intent-to-treat average treatment effect), and 2) estimate the extent to which risk of relapse to OUD could be reduced if we would have applied the rule to assign X:BOT participants to medication type as compared to if the treatments were randomly assigned, as they were in the trial, or if everyone were treated with XR-NTX, or if everyone were treated with BUP-NX.

2. Materials and Methods

2.1. Data and Sample

We used data from the N=570 participants who were randomized to treatment with either XR-NTX or BUP-NX in the X:BOT comparative effectiveness trial (ClinicalTrials.gov: NCT02032433), described previously.^1,8,9 The estimates and inferences that follow apply to these X:BOT trial participants. The goal of the parent trial was to compare effectiveness of XR-NTX (N=283) to BUP-NX (N=287) in preventing opioid relapse, the trial’s endpoint. X:BOT was conducted across eight sites affiliated with the National Drug Abuse Treatment Clinical Trials Network from January 30, 2014–January 31, 2017. These sites were community-based addiction treatment programs with inpatient and outpatient medical management capabilities and frequent admissions with OUD. Participants were English-speaking adults with Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) OUD who had used non-prescribed opioids in the previous 30 days and who had been admitted to inpatient (or short-term inpatient) treatment for OUD. Participants were excluded if they had a chronic pain disorder requiring opioids, current methadone maintenance treatment, or had conditions or circumstances that were contraindications for participation or treatment. The Institutional Review Board at the New York State Psychiatric Institute determined this secondary analysis of de-identified data to be non-human subject research.

2.2. Measures

2.2.1. Randomization to medication

Participants were randomized to either XR-NTX or BUP-NX in a 1:1 ratio, stratified by site and severity of opioid use (≥ 6 bags of IV heroin per day in the 7 days prior to admission, or equivalent). The trial was open label. XR-NTX treatment consisted of intramuscular injections given approximately every 28 days. BUP-NX was dispensed at weeks 0, 1, 2, 3, 4, 6, 8, 10, 12, 14, 16, and 20 for participants to self-administer sublingually daily. Dose was based on clinical indication.

2.2.2. Outcome

The primary outcome for this trial was relapse, defined as using non-study opioids at least once per week for four consecutive weeks or using every day for 7 consecutive days, and occurring between 20 days post-randomization (prior to that positive urine tests could be due to opioid use during medically managed withdrawal) and just prior to the end of the 24-week treatment phase. As in the primary outcome paper, missed visits or refused urine samples were considered as positive for opioid use,¹ which research suggests is a reasonable assumption.^10–13 While the primary outcome paper examined time-to-relapse with survival analysis,¹ the present analysis used relapse at any time of the 24 week trial as a simple binary indicator (1=any relapse, 0=absence of relapse), absence of relapse being reflective of a good clinical outcome. The individualized treatment rule we aim identify minimizes the mean outcome (i.e., reduce risk of relapse).

2.2.3. Covariates/Effect modifiers

We considered numerous covariates that could potentially act as modifiers of comparative treatment effectiveness, based on theory and previous analyses:^1,4 study site; severe opioid use (yes/no); randomization timing (early, randomized within 72 hours of most recent opioid v. late); age; gender; race/ethnicity (black, white, Hispanic/Latino); level of education (<high school, high school or GED, >high school); employment status; marital status (married v. not); living with a person who has an alcohol use problem; living with a drug user; age at heroin initiation; current IV drug use; duration of opioid use; history of: i) amphetamine use, ii) sedative use, iii) cannabis use; DSM-5 alcohol use disorder; cocaine use disorder; history of a psychiatric illness; previous treatment for an opioid use disorder; average cost of primary opioid consumption (dollars/day); baseline depression score using the Hamilton Depression Scale¹⁴; baseline pain using the EuroQOL (none v. moderate/extreme) ¹⁵; and past opioid withdrawal discomfort level (10-point scale). Homeless status (“currently homeless or living in a shelter”) at baseline was considered based on a previous analysis of the X:BOT trial that identified it as the only significant moderator of medication responsivity. ⁴

2.3. Statistical analysis

Assignment to XR-NTX versus BUP-NX and outcome data had no missingness. Covariate missingness was minimal (< 1%), however to preserve use of the full sample, we imputed the few missing observations using chained equations.¹⁶

We first estimated an individualized optimal treatment rule to assign medication type—XR-NTX or BUP-NX—to each participant in a way that would minimize their expected risk of relapse to OUD. We denoted this rule as d(v), which was a function that took observed covariate values (using all covariates listed in Section 2.2.3), v, as input and output a treatment decision {0,1}, where 0 represented BUP-NX and 1 represented XR-NTX. The optimal treatment rule,¹⁷ may be written as $\text{d}(v)=\mathbb{1}(\text{E}(Y_1-Y_0|V=v)<0)$, where Y_a is the counterfactual outcome in a hypothetical world where treatment is set to A = a. Using A = 1 to represent XR-NTX and A = 0 to represent BUP-NX, we can understand the intuition as follows. A negative conditional average treatment effect, $\mathbb{1}(\text{E}(Y_1-Y_0|V=v)<0)$, means that the risk of relapse on XR-NTX is lower than that on BUP-NX among those with V = v, and so XR-NTX would be recommended for these individuals. Similarly, $\mathbb{1}(\text{E}(Y_1-Y_0|V=v)>0)$ means that risk of relapse on BUP-NX is lower than on XR-NTX among those with V = v, so BUP-NX would be recommended for those individuals. The rule d(v) can be estimated using a number of different approaches.^6,18–20 One type of approach focuses on modeling conditional (i.e., strata-specific) average treatment effects (CATE), the so-called “blip” function.¹⁷ Our estimators of d(v) are simple classifiers that estimate the CATE and decide to assign treatment according to the above rule. We used the efficient influence function of the average treatment effect^21,22 regressed on covariates to estimate the CATE, which resulted in doubly robust estimation (meaning that the estimator was consistent even under misspecification of either the treatment or outcome model). ²²

Our sample size of N=570 was modest relative to the large number of potential effect modifying variables reflecting the complex landscape of social factors that may influence treatment success or failure. With this in mind, and in order to obtain an interpretable rule, we used two different algorithms that both entailed dimension-reduction in modeling the CATE by regressing the efficient influence function of the average treatment effect (pseudo-outcome) on a subset of effect-modifying covariates.

The first algorithm was a recently developed “adaptive lasso” approach that used a twostage regularization algorithm²³ to data-adaptively prune the set of covariates down to only those that were true modifiers of the average treatment effect.⁷ Specifically, the approach supplements the typical lasso algorithm ²⁴ with covariate-specific weights ²³ in such a way that it consistently identifies the correct effect modifiers asymptotically.⁷ We used a cross-fitted estimator^25,26 of this adaptive lasso algorithm (with 10 folds) to estimate the optimal treatment rule, which we denote d*(v). This adaptive lasso approach resulted in an interpretable (as opposed to a “black box”) treatment rule. In order to assess the extent to which the linearity assumption made by the adaptive lasso results in a reduction of performance of the rule, we performed a sensitivity analysis that used more flexible algorithms for estimating the CATE. This did not result in significantly different results (Figure A2 in the appendix.)

We also considered a simple, single-variable algorithm to estimate the optimal treatment rule, which we denote d′(u). This rule, designed to be similar to previous research that examined single baseline variables as modifiers of the comparative treatment effect in X:BOT, ⁴ identified the single baseline covariate, denoted u, most correlated with the efficient influence function of the average treatment effect (the pseudo-outcome), and fit a regression of the pseudo-outcome as a function of the single covariate. The fitted values resulted in the estimated rule. This algorithm was also implemented as a cross-fitted estimator with 10 folds.

We next estimated the expected risk of relapse had the treatment rule estimated via adaptive lasso been applied to everyone in the trial, denoted $\mathcal{E}(Y_{{\widehat{\text{d}}}^{*}})$, and had the treatment rule estimated via the single-variable algorithm been applied to everyone in the trial, denoted $\mathcal{E}(Y_{{\widehat{\text{d}}}^{\prime }})$, where $Y_{{\widehat{\text{d}}}^{*}}$ is the counterfactual outcome under the rule estimated with the adaptive lasso algorithm, $\mathcal{E}(Y_{{\widehat{\text{d}}}^{\prime }})$, is the counterfactual outcome under rule estimated with the single-variable algorithm, and where $\mathcal{E}$ denotes the expectation averaged across the cross-fitted folds in the sample. Please see Section A1 in the appendix for the mathematical definition of these effects. We also estimated the contrasts comparing: 1) the expected difference in risks of relapse had the rule estimated via adaptive lasso been applied versus: a) the observed outcome under random treatment assignment, $\mathcal{E}(Y_{{\widehat{\text{d}}}^{*}}-Y)$, b) had everyone been assigned naltrexone, $\mathcal{E}(Y_{{\widehat{\text{d}}}^{*}}-Y_1)$, c) had everyone been assigned buprenorphine, $\mathcal{E}(Y_{{\widehat{\text{d}}}^{*}}-Y_0)$; and 2) the expected difference in risks of relapse had the rule estimated via the single-variable algorithm been applied versus: a) the observed outcome under random treatment assignment, $\mathcal{E}(Y_{{\widehat{\text{d}}}^{\prime }}-Y)$, b) had everyone been assigned naltrexone, $\mathcal{E}(Y_{{\widehat{\text{d}}}^{\prime }}-Y_1)$, c) had everyone been assigned buprenorphine, $\mathcal{E}(Y_{{\widehat{\text{d}}}^{\prime }}-Y_0)$For estimating the expected risks of relapse and their contrasts, we used a cross-fitted²⁶ doubly robust estimator that solved the efficient influence function estimating equation in one step,²⁷ substituting our estimates of the treatment rule for A. We estimated the variance as the sample variance of the efficient influence function.

R (version 3.6.1) was used for all analyses.²⁸ Code to replicate the analyses is available: : https://github.com/kararudolph/code-for-papers/blob/master/XBOT_txrule.

3. Results

Figure 1a shows the estimates of the following causal parameters: the expected risk of relapse under random assignment, $\mathcal{E}(Y)$; assignment to buprenorphine only, $\mathcal{E}(Y_0)$; assignment to naltrexone only, $\mathcal{E}(Y_1)$; assignment based on the rule estimated with the adaptive lasso algorithm (and denoted “lasso”), $\mathcal{E}(Y_{{\widehat{\text{d}}}^{*}})$; and assignment based on the rule estimated with the single-variable algorithm (denoted “single-variable”), $\mathcal{E}(Y_{{\widehat{\text{d}}}^{\prime }})$. Figure 1b shows estimates of the contrasts between the expected risk of relapse under each estimated rule as compared to the expected risk under each reference group (random assignment, buprenorphine only, and naltrexone only) on the additive scale, e.g., for the causal effects contrasting each estimated rule with observed random assignment, $\mathcal{E}(Y_{{\widehat{\text{d}}}^{*}}-Y)$ and $\mathcal{E}(Y_{\widehat{\text{d}}},-Y)$.

Figure 1:

A) Estimated risk of relapse by 6 months under the observed randomized medication assignment (“rand”), assigning buprenorphine to everyone (“bup”), assigning naltrexone to everyone (“ntx”), assigning based on the individualized treatment rule estimated using adaptive lasso (“lasso, rule”), and estimated using the single-variable algorithm based on homeless status (“single-variable, rule”). B) Estimated differences in risk of relapse by 6 months comparing the lasso-estimated and single-variable-estimated individualized treatment rule to i) observed randomized assignment (“rand”), ii) assigning buprenorphine (“bup”) to everyone, and iii) assigning naltrexone (“ntx”) to everyone.

The treatment rule estimated via adaptive lasso selected the following covariates as modifiers of the comparative treatment effect: homeless status (the strongest effect modifier); having cocaine use disorder in combination with being non-White, non-Black, and non-Hispanic race/ethnicity (incorporated as an interaction); being White; being non-Hispanic/Latino; having used opioids for fewer years; not having previous pharmacologic treatment for OUD; and having more than a high school education in combination with being non-White, non-Black, and non-Hispanic race/ethnicity (incorporated as an interaction) all predicted a lower risk of relapse on XR-NTX versus BUP-NX. The regression model is described further in Section A2 in the appendix, the coefficient associated with each covariate is provided in Table A1 in the appendix, and the estimated CATEs resulting from this model (i.e., estimated blips) are shown in Figure A1 in the appendix.

Assigning type of medication—XR-NTX vs. BUP-NX—according to the treatment rule estimated by adaptive lasso vs. random assignment reduces absolute risk of relapse by 7.1 percentage points (risk difference (RD): −7.1 (95% CI: −11.1, −3.2), and reduces the relative risk of relapse by 12% (relative risk (RR) 0.88 95% CI: 0.82, 0.95). Under the rule estimated with adaptive lasso, 107 participants (19%) would be assigned to naltrexone and the remaining 463 would be assigned to buprenorphine. The contrast between the adaptive lasso-estimated treatment rule versus assigning naltrexone to everyone is even larger (RD: −11.6, 95% CI: −18.6, −4.6). Applying the adaptive lasso-estimated rule still reduces relapse risk as compared to assigning buprenorphine to everyone, but the contrast is slightly less than the contrast with the other two reference groups and is not statistically significant (RD: −2.9, 95% CI: −6.7, 0.1).

In the single-variable algorithm, the baseline covariate of homeless status was consistently chosen as the most important effect modifier across cross-fitted folds. Thus, the treatment rule estimated with the single-variable algorithm translates to assigning XR-NTX if homeless and assigning BUP-NX if stably housed. Assigning treatment based solely on homeless status (resulting in 143 participants (25%) being assigned to naltrexone and the remaining 427 being assigned to buprenorphine) performs similarly as the rule estimated by adaptive lasso, reducing risk of relapse by 8.4 percentage points on the additive scale as compared to random medication assignment (−8.4, 95% CI:−12.3, −4.5) and 14% on the relative scale (0.86, 95% CI: 0.80, 0.93). Contrasts with the other two reference groups follow the same pattern as for the rule estimated by adaptive lasso.

Although the two rule estimates performed very similarly, they did not result in the same medication decision for all participants. They made the same decision for 532 participants (93%), assigning BUP-NX to 426 and XR-NTX for 106. The two rule estimates differed for the remaining 38 participants. Under the single-variable-estimated rule, 37 would be treated with XR-NTX, but would be treated with BUP-NX under the adaptive lasso-estimated rule. Under the single-variable-estimated rule, 1 would be treated with BUP-NX, but would be treated with XR-NTX under the adaptive lasso-estimated rule.

4. Discussion

We found that applying an individualized treatment rule estimated with an adaptive lasso algorithm to hypothetically assign patients with OUD to pharmacologic treatment with either XR-NTX or with BUP-NX would significantly reduce risk of relapsing to OUD by 12%, translating to a number-needed-to-treat. (NNT) of 14. This estimated treatment rule was primarily a function of homeless status as well as having cocaine use disorder in combination with being non-White, non-Black, and non-Hispanic/Latino race/ethnicity and having used opioids for fewer years (Table A1). Applying an alternative, simpler estimate of the individualized treatment rule in which homeless patients would be assigned treatment with XR-NTX and stably housed patients would be assigned BUP-NX performed similarly to the lasso-estimated rule, reducing risk of relapse by 14%) relative to random assignment. These results lend convergent support for an earlier finding that homeless status was an important modifier of XR-NTX vs. BUP-NX treatment effectiveness in X:BOT,⁴ and provide an estimate for the amount by which a relatively simple change in clinical practice could be expected to improve prevention of OUD relapse.

Homeless individuals with drug use disorders may experience multiple structural and contextual challenges to treatment. Medication for OUD (MOUD) typically involves long-term treatment. Medication must be accessed with frequent visits during this time-frame, requiring adherence to a consistent schedule.²⁹ For example, buprenorphine is typically dispensed every 2-4 weeks for individuals to take daily, requiring the individual to keep track of all doses dispensed. The requirements of buprenorphine treatment may be particularly onerous for homeless individuals, who may lack the consistency and safe and secure storage options required for treatment success. In contrast, XR-NTX is administered via monthly injection, thereby eliminating the need to keep track of medication, making it potentially more successful with patients who lack stable housing.

Relapse is one of several outcomes we want to optimize in the treatment of OUD, others include illicit opioid use, overdose, and death. It may also be of interest to minimize side effects of the treatments, which may also affect some subgroups more than others. The methods we used here can be extended to consider such competing risks by incorporating all relevant risks/benefits into the development of a single, optimal, more nuanced treatment rule, or by developing multiple optimal treatment rules—one for each risk/outcome—and creating a weighted summary.³⁰ Applying such extensions to match people with OUD with the treatment most likely to optimize a set of multiple, desirable outcomes while minimizing side effects simultaneously represents an area for future work.

It is possible that we were under-powered to learn and estimate an optimal treatment rule with the sample size of 570 (N=287 in BUP-NX arm, N=283 in XR-NTX arm). However, the X:BOT trial and research question we consider align well (both in terms of data structure and estimation approach) with several of the simulation scenarios considered by Luedtke et al. 2019, which suggested that the sample size of 300 participants per arm may be sufficient to learn an optimal treatment rule (i.e., reject the null hypothesis of no treatment effect modification).³¹

This relatively small sample size could also explain why the treatment rule estimated with the single-variable algorithm results in a slightly lower average risk of relapse than the lasso-estimated rule. In this particular case, we have a large number of covariates (n=35, including site indicators) relative to the number of participants. We speculate that performance of our finitedimensional parameters of interest could be negatively affected by the curse-of-dimensionality. ³² In Figure A2 in the appendix, we show that estimators we used that reduce dimensionality (adaptivelasso and single-variable) result in better performance in terms of larger reductions in predicted risk of relapse as compared to two implementations of flexible ensembles of machine learning algorithms, ²² weighted to optimize predictive accuracy (SuperLearner^33,34, implementation details provided in Section A3 of the appendix, results provided in Figure A2). However, an alternative explanation is that the true, unknown optimal treatment rule is simple, depending only on homeless status.

Another limitation was that we did not include any contextual-level covariates as potential treatment effect modifiers with the exception of indicators for each of the eight study sites. Such covariates, like the availability of harm-reduction programs in the local area, the availability and accessibility of homeless services, including housing, etc., may plausibly influence whether BUP-NX vs. XR-NTX works better for certain subgroups, and incorporating them would be another area for future work.

We could not consider methadone—which is highly effective and one of the most common medications for OUD³⁵—in this analysis, because it was not included in the trial.³⁵ Thus, we are limited to commenting only on XR-NTX vs. BUP-NX assignment. It would be of future interest to use a trial in which methadone is directly compared to both XR-NTX and BUP-NX to learn an optimal treatment rule for deciding between all three possible medications. If all three medications are not simultaneously compared in one trial, two or more trials, similarly conducted, with a common outcome, may be harmonized to learn an optimal rule for assigning medication type among the three options. We are currently pursuing such an option.

Despite the aforementioned limitations, this analysis benefited from several strengths. First, we used randomized trial data to estimate an optimal treatment rule, which is considered the gold standard,³⁶ because estimation of the conditional average treatment effect in such a setting (which forms the basis for estimating the individual treatment rule) should be approximately unbiased in expectation. Furthermore, the X:BOT clinical trial is relatively large in size with the number of participants in each trial arm approaching the number that prior simulation studies found to be sufficient for similar research questions.³¹ In addition, trial participants are well-characterized by numerous baseline variables thought to be predictive of relapse, which we incorporated in the analysis. In doing so, we used a doubly robust estimation approach⁷ that incorporates machine learning in model fitting via the adaptive lasso,²³ chosen with the sample size and the extensive set of covariates in mind.

Finally, and perhaps most importantly in a practical sense, the single-variable-estimated treatment rule based only on homeless status performed similarly to the lasso-estimated rule. Singlevariable treatment rules have been identified in other contexts in substance use disorder research.³⁷ Conceptually, it could be that one or few clinically measured predictors represent a “common final pathway” where relevant predictors converge. Identifying these “flag” covariates may inform clinically tractable tailoring algorithms. Although the NNT of 14 associated with this single-variable treatment rule is a small effect, it is within the range of other useful interventions for substance and mental health problems.³⁸ In this case, our results suggest a simple, easy adjustment to clinical practice could meaningfully reduce OUD relapse rates.