Association of Neural Responses to Drug Cues With Subsequent Relapse to Stimulant Use

Importance Although chronic relapse is a characteristic of addiction to stimulants, conventional measures (eg, clinical, demographic, and self-report) do not robustly identify which individuals are most vulnerable to relapse. Objectives To test whether drug cues are associated with increased mesolimbic neural activity in patients undergoing treatment for stimulant use disorder and whether this activity is associated with risk for subsequent relapse. Design, Setting, and Participants This prospective cohort study of 76 participants included a control group for baseline group comparisons. Veteran patients (n = 36) with stimulant use disorders were recruited from a 28-day residential treatment program at the Veterans Affairs Palo Alto Health Care System. Healthy controls (n = 40) were recruited from the surrounding community. Baseline data were collected between September 21, 2015, and January 26, 2018, from patients and healthy controls using functional magnetic resonance imaging during a performance of a reward cue task. Patients’ stimulant use was subsequently assessed after treatment discharge (at approximately 1, 3, and 6 months) to assess relapse outcomes. Main Outcomes and Measures Primary measures included neural responses to drug and food cues in estimated mesolimbic volumes of interest, including the medial prefrontal cortex, nucleus accumbens (NAcc), and ventral tegmental area. The primary outcome variable was relapse (defined as any stimulant use), assessed both dichotomously (3 months after discharge) and continuously (days to relapse). Brain activity measures were contrasted between groups to validate neural measures of drug cue reactivity, which were then used to estimate relapse outcomes of patients. Results Relative to controls (n = 40; 16 women and 24 men; mean [SD] age, 32.0 [11.6] years), patients (n = 36; 2 women and 34 men; mean [SD] age, 43.4 [13.3] years) showed increased mesolimbic activity in response to drug cues (medial prefrontal cortex, t74 = 2.90, P = .005, Cohen d = 0.66; NAcc, t74 = 2.39, P = .02, Cohen d = 0.54; and ventral tegmental area, t74 = 4.04, P < .001, Cohen d = 0.92). In patients, increased drug cue response in the NAcc (but not other volumes of interest) was associated with time to relapse months later (Cox proportional hazards regression hazard ratio, 2.30; 95% CI, 1.40-3.79). After controlling for age, NAcc response to drug cues classified relapsers (12 patients; 1 woman and 11 men; mean [SD] age, 49.3 [14.1] years) and abstainers (21 patients; 1 woman and 20 men; mean [SD] age, 39.3 [12.3] years) at 3 months with 75.8% classification accuracy. Model comparison further indicated that NAcc responses to drug cues were associated with relapse above and beyond estimations of relapse according to conventional measures. Conclusions and Relevance Responses in the NAcc to stimulant cues appear to be associated with relapse in humans. Identification of neural markers may eventually help target interventions to the most vulnerable individuals.

the research team worked closely with the clinical treatment team to ensure any issues that arose during participation in the study were rapidly and adequately addressed. A member of the research team escorted patient participants between the treatment facility and study location to ensure safe transit. The treatment team was aware of the study protocol, including the cuereactivity paradigm. Patients had access to social workers, case managers, psychologists and psychiatrists should they have needed supportive follow-up.

Follow-up assessments
Patients were notified that their participation would entail follow-up appointments one, three, and six months after the date of their treatment discharge. Follow-ups were not conducted with controls. At the end of the scanning session, patients completed a follow-up form requesting their contact information during this period, as well as information for two other close individuals (relatives, friends, or case worker) that could be contacted if the patient could not be reached.
Patients then received three Amazon.com gift cards and were told that one card would be activated with $50 after completing each follow-up appointment.
Patients were contacted by phone, text, and/or email for follow-up appointment scheduling and reminders. If a patient could not be reached, then collateral sources (relatives, friends, and/or VA records) were contacted for information about how to reach the patient as well as their potential stimulant use. At each follow up, stimulant use was assessed using the Time-Line Follow-Back method 3 . This self-report measure of relapse shows moderate-to-high consistency with urine toxicology screens 4 . Patients were asked to identify any dates since their last appointment that they used stimulants. Due to the sensitive nature of self-reporting use of illicit substances, we limited our interview questions to date(s) of use and did not systematically record quantity of use for a given use day. The Brief Addiction Monitor questionnaire was also administered at each follow-up to assess recent (past month) use of other illicit substances. Follow-ups were conducted either via phone (80.8%) or in-person (19.2%), and urine toxicology screens were collected in a subset of the in-person interviews (53.3%). In every instance where urine drug samples were collected (n=8), they were consistent with patients' self-report 5 .
Because treatment was abstinence-based, relapse was defined as any stimulant use in the time since treatment or the previous assessment. Relapse status was assigned based on patients' self-reported stimulant use in a follow-up interview or based on available medical records that explicitly noted a patient's use of stimulants (in no case did these sources conflict). Fourteen of the patients continued in some form of monitored program (either outpatient or residential) following treatment discharge, which provided an additional source of corroborating evidence of recovery outcomes.
Stimulant use data was successfully collected for all but one patient at the 1-month assessment.
Two additional patients were lost to follow-up prior to the 3-month assessment, and 3 additional patients were lost prior to the final follow-up targeted at 6 months posttreatment discharge.
Patients lost to follow-up (total n=6) did not significantly differ from other patients on any of the variables reported in eTable 2. In the interest of both minimizing the number of patients lost to follow-up and having a relatively equal number of relapsers and abstainers, we analyzed treatment outcome in the sample of 33 patients with either confirmed abstinence or confirmed relapse 3 months after treatment discharge. In this sample, the median follow-up duration was 194 days (SD=83.0; range, 90-463 days).

Neuroimaging scan acquisition
Scans were acquired with a 3 Tesla GE Discover MR750 scanner and a Nova Medical 32channel head coil. Functional (T2*-weighted) images were acquired using an echo-planar imaging pulse sequence with the following parameters: TR=2 s, TE=25 ms, flip angle=77°, FOV=232 × 232, 80 × 80 acquisition matrix with 46 axial slices (no slice gap). Images were acquired in an interleaved order and voxel dimensions were 2.9 mm 3 . Structural (T1-weighted) scans were acquired using GE's BRAVO sequence with the following parameters: TR=7.2 ms, TE=2.8 ms, flip angle=12°, FOV=256 × 230, 256 × 256 acquisition matrix with 186 slices, with slice thickness of 0.9 mm. Diffusion-weighted images were also acquired (described in a separate report).

Data processing and analysis
Pre-processing: the first six volumes of each functional scan were discarded to allow magnetization to reach steady-state. Images were then corrected for differences in slice acquisition times and head movement using Fourier interpolation, and smoothed using a 4 mm full-width at half maximum Gaussian kernel. Raw activity was then converted to percent signal change within each voxel and high-pass filtered to remove low frequency drift (admitting frequencies > 1 cycle per 90 s).
The first functional volume of each task was co-registered to a subject's T1-weighted anatomical volume in native space, and anatomical volumes were subsequently co-registered to an anatomical template (TT_N27) in standard Talairach space. These two transformations were combined and applied to functional data, aligning all participants' data in a common standardized space (i.e., Talairach space).
Regressors for whole-brain analyses: A general linear model was fit to each voxel time series that included task-related regressors (described below) as well as nuisance regressors.
Nuisance regressors included six rigid-body movement parameters estimated during motion correction, as well as averaged activity time series extracted from white matter and cerebrospinal fluid VOIs 6 . To model the task, regressors were defined to model activity during the cue, image, and rating periods of each trial. Reaction time for ratings was also modeled as a regressor to capture signal variability associated with motor responses. Regressors of interest were then created which indicated each trial type (drug, food, alcohol, and neutral). For these regressors, each trial was modeled as an 8 second boxcar function over the entire trial duration, beginning with cue onset and terminating with the offset of the rating period. These regressors were then convolved using a single gamma function to account for hemodynamic lags (Cohen, 1996). Food vs neutral trials, drug vs neutral trials, and drug vs food trials were then contrasted, producing three contrast maps for each subject.
VOI definition: VOI masks of MPFC, NAcc, and VTA were created. The MPFC VOI was defined with an 8 mm diameter sphere bilaterally centered on Talairach coordinates ±5, 47, 0. The NAcc VOI was anatomically defined based on "Left-Accumbens" and "Right-Accumbens" labels in the Desai atlas. The VTA VOI was created based on a previously-described structural landmark demarcation of midbrain dopamine nuclei 7 (substantia nigra and VTA collectively). The bilateral VTA VOI was defined as the medial aspects of this mask, which included voxels spanning left and right coordinates from x=-5 to x=6. Because we had no predictions regarding laterality, we averaged left and right VOIs to reduce the number of statistical tests.
Whole-brain classification analysis: Binary classifiers were trained to distinguish relapsers from abstainers based on patients' whole brain responses to drug cues. To maximize the number of instances for each class, we defined relapsers as the first half of participants in the patient sample to relapse, producing 15 relapsers and 15 abstainers at 215 days posttreatment discharge (only patients with at least six months of follow-up data were included in this analysis, leaving 30 patients). Features were selected using a support vector machine classifier with recursive feature elimination (SVM-RFE) in the python toolbox scikit-learn 8 . Features were defined as voxelwise regression beta coefficients modeling activity during drug cue trials.

Behavior
Mixed-level repeated measures analyses of variance (ANOVAs) confirmed that group (patient, control; between-subjects factor) and cue type (food, drug, neutral; within-subjects factor) influenced self-reported ratings collected both during and after the scan as predicted.

eFigure 2. Ratings of Cue Images in a Pilot Sample of Healthy Control Participants
Affective ratings were collected from a separate group of participants (n=24) online using Amazon's Mechanical Turk platform. Participants were located in the United States and were between 18-60 years of age. Participants rated each image on 7-point scales indicating valence (from "very negative" to "very positive"), arousal (from "not at all aroused" to "highly aroused"), wanting (from "strongly don't want" to "strongly want"), and familiarity (from "not at all familiar" to "very familiar"). Valence and arousal ratings were then mean-deviated within subject and rotated to index positive arousal and negative arousal 9 , consistent with the circumplex model of   The 5% most informative features were selected using an SVM-RFE classifier with C parameter of 10.00 and then back-projected into standardized brain space (Talairach warped) and clusterthresholded at 10 voxels. Positive weights indicate greater drug cue-induced activity for relapsers, while negative weights indicate greater drug cue-induced activity for abstainers.
Coordinates are in Talairach space.