INTRODUCTION

Maladaptive behaviors such as tobacco use, unhealthy diet, and physical inactivity have far reaching health and economic implications. Even with the best available interventions, many unhealthy behaviors remain difficult to modify. Emerging research suggests that impaired cognition, particularly executive cognitive function, plays a central role in failed attempts at behavior change (Filevich et al, 2012). Executive cognitive function refers to a core set of cognitive processes that support cognitive (self) control (Botvinick et al, 2001). Working memory, a central domain of cognitive control, is required to focus on goal-related information and to sustain goal-directed behavior (Baddeley, 2003). Indeed, objective measures of working memory performance predict a variety of lifestyle behaviors, including smoking and maintenance of a healthy body weight (Gonzales et al, 2010; Hege et al, 2013; Patterson et al, 2010).

Functional magnetic resonance imaging (fMRI) studies have explored the effects of brief abstinence from smoking (eg, 24–72 h) on working memory and associated neural activation, using within-subject contrasts of smokers in the abstinent vs smoking satiety states (Falcone et al, 2014; Jacobsen et al, 2007; Loughead et al, 2009). In addition to reducing working memory performance (Mendrek et al, 2006; Myers et al, 2008), brief abstinence is associated with reduced working memory-related activation in the dorsolateral prefrontal cortex (DLPFC) and medial frontal/cingulate gyrus (MF/CG; Falcone et al, 2014; Loughead et al, 2010), decreased working memory-related suppression of activity in regions within the default mode network (posterior cingulate cortex (PCC), ventromedial prefrontal cortex (vmPFC); Beaver et al, 2011; Falcone et al, 2014; Loughead et al, 2010), and dysregulated inter-network functional connectivity during task performance and at rest (Jacobsen et al, 2007; Lerman et al, 2014; Sutherland et al, 2013). FDA-approved medications, such as varenicline and bupropion, can reverse abstinence-induced working memory deficits and related neural activation in studies of abstinent smokers tested during medication vs placebo (Loughead et al, 2010; Perkins et al, 2013). To our knowledge, no study has examined whether working memory-related brain signal predicts relapse to smoking.

We tested the hypothesis that abstinence-induced changes in working memory-related neural activation would predict the likelihood of early smoking relapse (biochemically confirmed in the first 7 days of a quit attempt) above and beyond standard clinical and behavioral measures. Prior work suggests that regions within the brain’s executive control network (eg, DLPFC, MF/CG) are vital for cognitive control (Fassbender et al, 2004; Krain et al, 2006), whereas regions within the default mode network (eg, PCC, vmPFC) are involved in self-referential and stimulus-independent thought processes that can compete with executive control functions (Gusnard et al, 2001; Smith et al, 2009). Given the importance of outwardly focused attention and cognitive control in smoking behavior change (Ashare et al, 2013b; Kollins et al, 2013; Sutherland et al, 2012), we expected that short-term relapse would be predicted by reduced activation in DLPFC and MF/CG, and less suppression of activation in PCC and vmPFC, as assessed during a pre-quit abstinence challenge (vs smoking satiety).

MATERIALS AND METHODS

Participants

Treatment-seeking smokers aged 18–65 years who reported smoking ⩾10 cigarettes/day for ⩾6 months were recruited through mass media (N=80). Participants provided written informed consent and completed a physical examination including a breath alcohol test, urine drug screen, and urine pregnancy test (females). Persons with a history of DSM-IV Axis I psychiatric or substance disorders (except nicotine dependence) assessed with the Mini International Neuropsychiatric Interview (Sheehan et al, 1998) and those taking psychotropic medications were excluded. Exclusion criteria also included: current use of chewing tobacco, snuff, or smoking-cessation products; pregnancy, planned pregnancy or breastfeeding; history of brain injury; left handedness; presence of fMRI contraindicated material in the body; low or borderline intelligence (<90 score on Shipley Institute of Living Scale; Zachary, 1986); and any impairment that would prevent task performance. Eligible participants completed the six-item Fagerström Test for Nicotine Dependence (Heatherton et al, 1991). The University of Pennsylvania Institutional Review Board approved all procedures.

Functional Brain Imaging Sessions

Two blood oxygen level-dependent (BOLD) fMRI sessions were acquired in counterbalanced order: (1) smoking satiety and (2) 24 h abstinence challenge. The 24-h period was selected because the majority of relapses occur in the first day of a quit attempt (Piasecki, 2006). Imaging sessions were separated by ∼2 weeks. Subjects were to refrain from alcohol and other drugs 24 h before the imaging session and those with a positive drug screen, a breath alcohol test >0.01, or a breath carbon monoxide (CO) test >9 ppm (abstinence challenge condition only) were excluded. Before imaging, participants completed the Minnesota Nicotine Withdrawal Scale (MNWS; Hughes and Hatsukami, 1986), Questionnaire of Smoking Urges (QSU-Brief; Cox et al, 2001), and the Positive and Negative Affect Schedule (PANAS; Watson et al, 1988). For the smoking satiety condition, participants smoked immediately before scanning to standardize exposure (∼30 min prior to task performance).

Working memory was assessed during BOLD imaging with a visual N-back task (Ragland et al, 2002) used in our prior research (Loughead et al, 2010; Loughead et al, 2009). This task presents complex geometric figures (fractals) under four conditions: 0-back, 1-back, 2-back, and 3-back. In the 0-back condition, participants respond with a button press to a specified target fractal; for the 1-back condition, participants respond if the current fractal was identical to the previous one; for the 2-back condition, if the current fractal was identical to the item presented two trials back; etc. No response was required for nontargets. Images were presented for 500 ms with an interstimulus interval of 2500 ms. The task began with a 48 s baseline rest period (fixation point) of which the first 24 s was discarded to ensure the MRI signal reached steady state. Each condition was presented three times in 20-trial blocks (33% targets; 60 s). Blocks were presented in order of increasing memory load for one set, after which conditions were presented pseudo-randomly; visual instructions (9 s) preceded each block to indicate the upcoming condition. Equivalent forms with unique stimuli were used and task version order was counterbalanced.

fMRI Data Acquisition

BOLD fMRI was acquired with a Siemens Trio 3T (Erlangen, Germany) system using a whole-brain, single-shot gradient-echo echoplanar sequence with the following parameters: TR/TE=3000/30 ms, FOV=220 mm, matrix=64 × 64, slice thickness/gap=3.4/0 mm, 48 slices, effective voxel resolution of 3.4 × 3.4 × 3.4 mm. RF transmission utilized a quadrature body coil, and reception used a 32-channel head coil. Prior to BOLD fMRI, 5-min magnetization-prepared, rapid acquisition gradient echo T1-weighted image (MPRAGE, TR 1620 ms, TE 3.87 ms, FOV 50 mm, matrix 192 × 256, effective voxel resolution of 1 × 1 × 1 mm) was acquired for anatomic overlays of functional data and to aid spatial normalization to standard atlas space.

Smoking Cessation Procedures

Approximately 1–2 weeks after imaging, participants took part in a standardized counseling session during which they discussed strategies for quitting (including nicotine fading, identifying and managing triggers, problem solving, relapse prevention, and stress management) with a trained smoking cessation counselor (Lerman et al, 2004). During this counseling session, participants set a target quit date (TQD) to occur ∼1 week later and received a self-help booklet to take home (UDHHS, 2006). Participants completed a brief in-person visit on the TQD, which included a 20-min booster counseling session to reinforce and refine strategies discussed at the initial session. See ‘Outcome Measure’ below for the assessment of subsequent quit status.

Image Preprocessing

BOLD time series data were preprocessed and analyzed by standard procedures using fMRI Expert Analysis Tool (FEAT version 5.98) of FSL (FMRIB's Software Library, Oxford, UK). Single-subject preprocessing included skull stripping using the FSL Brain Extraction Tool (Smith, 2002), slice time correction, motion correction to the median image using MCFLIRT (Jenkinson and Smith, 2001), high pass temporal filtering (138 s), spatial smoothing using a Gaussian kernel (6 mm full-width at half-maximum, isotropic), and mean-based intensity normalization of all volumes with the same multiplicative factor. The median functional volume was coregistered to the anatomical T1-weighted structural volume and transformed into a standard anatomical space (T1 MNI template), using FLIRT (Jenkinson et al, 2002; Jenkinson and Smith, 2001). Transformation parameters were later applied to statistical maps for group-level analyses.

ROI Definition

As noted above, prior work with the visual N-back task by our group (Ashare et al, 2013b; Falcone et al, 2014; Lerman et al, 2014; Loughead et al, 2009, 2010) and others (Owen et al, 2005; Satterthwaite et al, 2013) has identified three task-active (right DLPFC, left DLPFC, MF/CG) and two task-negative (PCC, vmPFC) regions sensitive to abstinence challenge. ROI masks were functionally defined using the main effect of memory load (controlling for condition) in a whole-brain repeated measures condition (smoking satiety, abstinence challenge) by memory load (N-back level) ANOVA. Seven-day quit outcome was not included in this analysis. The resulting activation map was thresholded for family-wise error correction at p⩽0.001 (Z⩾5.87) to account for five a priori ROIs and then segmented by clusters⩾150 voxels using a watershed algorithm implemented in MATLAB (The Mathworks, Natick, MA). The five functionally defined a priori ROI masks (right DLPFC, left DLPFC, MF/CG, vmPFC, and PCC) were transformed to native subject space using parameters obtained in the time series analysis (Table 1). Each participant’s mean percent signal change was calculated for the memory load conditions in each ROI and exported for further statistical analysis.

Table 1 Functional Regions of Interest Identified by Main Effect of Memory Load in a Whole-brain Repeated Measures ANOVA
Full size table

Outcome Measure

The smoking cessation outcome measure was short-term abstinence; specifically, the ability to remain completely abstinent from smoking (not even a puff) for 7 days following the TQD. Participants completed a monitoring visit on day 7, during which smoking behavior was assessed using a timeline follow-back method (Brown et al, 1998) and biochemically confirmed using NicAlert urine test strips (Nymox Pharmaceutical Corporation, Hasbrouck Heights, NJ). NicAlert test strips utilize an immunochromatographic assay to provide a semiquantitative measure of the concentration of cotinine (the primary metabolite of nicotine) in urine. Results appear as categorical levels of usage (range: 0–6). Following manufacturer guidelines, NicAlert results of level 2 or below (corresponding to a urine cotinine concentration of ⩽100 ng/ml) were required to biochemically confirm abstinence; levels 3–6 were considered indicative of current smoking. The 7-day monitoring period was chosen because 50–75% of smokers who relapse do so within the first 7 days of a quit attempt (Hughes et al, 2004). Further, this measure is a well-validated indicator of long-term smoking status; continuous abstinence for the first week of a quit attempt is a highly significant predictor of abstinence at 6 months following the TQD (Ashare et al, 2013a). Sixty-one smokers relapsed and 19 quit successfully for this period. Of this sample, data from four relapsed and four quit participants were excluded from analysis owing to poor imaging data quality (root mean square mean relative motion >0.4, total non-responses >66%).

Statistical Analysis

Descriptive statistics were obtained for all variables. To examine expected abstinence challenge effects, mean percent signal change was modeled using regression with subject-level random effects, and estimated using maximum likelihood techniques (Stata xt-reg; Stata Corporation, College Station, TX) with mixed effects (Gaussian model). The models of abstinence challenge effects on BOLD signal included terms for the main effects of condition, categorical memory load and relevant covariates (age, sex, FTND score, Shipley IQ score, and condition order). Models of abstinence effects on smoking measures and task performance used similar regression models (subject-level random effects, maximum likelihood techniques with mixed effects), which utilized the measure of interest (CO reading, MNWS, QSU-B, N-back accuracy, or N-back correct response time) as the independent variable and included terms for main effects of condition, categorical memory load (for task performance) and relevant covariates (age, sex, FTND score, Shipley IQ score, and condition order). Multi-level mixed effects logistic regression (Stata xtmelogit) was used to examine whether BOLD signal in the a priori regions predicted performance outcomes (N-back accuracy, N-back correct response time), controlling for condition and categorical memory load.

Predictive models

Forward stepwise regression with a probability of 0.2 for removal and 0.1 for entry was used to select predictors for a logistic regression model (STATA logistic) of dichotomized 7-day relapse. Sex, age, and nicotine dependence level (Heatherton et al, 1991) were entered as baseline candidate predictors of relapse (Gourlay et al, 1994; Hymowitz et al, 1997; Japuntich et al, 2011; Sweitzer et al, 2013). Change scores (abstinence challenge minus smoking satiety) for paired data collected during abstinence challenge (MNWS, QSU-Brief, PANAS Positive, PANAS Negative, task performance (accuracy and response time), right DLPFC, left DLPFC, MF/CG, vmPFC, and PCC) were also included in the model. Prior work has demonstrated that nicotine abstinence effects are induced at the highest working memory load (3-back level; Loughead et al, 2009, 2010). Therefore, we examined task performance and BOLD signal acquired during performance of 3-back trials (minus BOLD signal acquired during performance of 0-back trials to account for task-related cognitive processes other than working memory). We required age and nicotine dependence to be entered based on clinical relevance (Hymowitz et al, 1997; Sweitzer et al, 2013); sex was nonsignificant, and allowed to dropout.

Following logistic regression, receiver operating characteristic (ROC) analysis was used to assess the predictive potential of the models (Steyerberg, 2008). Classification models, based on linear scores, selected a cut point for separating predicted classes. ROC curve analysis varied the cut point across the range of data, plotting true-positive rate against the false-positive rate for each increment. Estimates of predictive accuracy are expressed as area under curve (AUC). The ROC curve is a plot of the sensitivity vs 1-specificity of a classification system, represents the accuracy of the system, and provides a comparable metric across experiments (Bradley, 1997). Following literature standards (Fawcett, 2003; Harrell et al, 1996), we consider ROC curve AUC of 0.50–0.60 as indicating prediction at chance; 0.60–0.70 as indicating poor prediction; 0.70–0.80 fair prediction; 0.80–0.90 good prediction; and 0.90–1.0 excellent prediction.

Two resampling methods were used to validate predictive models. First, bootstrap procedures were used to create 1000 replicates of the data with records chosen randomly with replacement. Stepwise model regression (see above) was applied to each replicate to assess model uncertainty (Austin and Tu, 2004b). The frequency of a variable’s inclusion in a model can be summarized and variables selected most often are considered to be the most robust (ie, least influenced by outliers and noise). To examine the model’s potential for prediction in new cases, we performed leave-one-out cross-validation (LOOCV) using standard jackknife procedures (Steyerberg, 2008).

Exploratory Analysis

In a second validation analysis, a sub-sample of relapsers was selected and matched for age and sex with the group of quitters, resulting in a sample of 15 pairs. Matching was conducted without replacement, and there were no cases where more than one relapser was matched to a quitter. An exploratory whole-brain condition (abstinence challenge vs smoking satiety) by smoking cessation group (relapse vs quit) voxelwise 2 × 2 ANOVA at the 3-back level (minus 0-back) was conducted to identify novel regions sensitive to quit-group status. Resulting statistical maps were corrected for multiple comparison using random field theory with group maps cluster corrected at Z>1.96 and probability of spatial extent p<0.05.

RESULTS

Baseline Sample Characteristics and Abstinence Challenge Effects

Of 72 participants included in the analysis, 44.4% were female and 38.9% were Caucasian. The mean age was 41.9 years (SD 14.4), mean FTND score was 4.8 (SD 1.8), and mean Shipley IQ score was 103.5 (SD 8.4). Mean CO, MNWS, QSU-B, and N-back response times all showed significant condition effects, indicating compliance with the abstinence requirements (reduced CO readings during abstinence challenge), more severe withdrawal symptoms (MNWS) and craving (QSU-B), and slower response times on the N-back task during abstinence challenge compared with smoking satiety (ps⩽0.05, Table 2). There was no significant condition effect on N-back accuracy. Abstinence challenge effects were observed on BOLD signal for MF/CG (β=−0.17, 95% confidence interval (CI) −0.26 to −0.08, p<0.001), right DLPFC (β=−0.13, 95% CI −0.26 to −0.01, p=0.039), left DLPFC (β=-0.17, 95% CI −0.31 to −0.03, p=0.016), and PCC (β=−0.21, 95% CI 0.05 to 0.37, p=0.008), consistent with the effects we previously reported in an analysis of the abstinence challenge data for a subset of these participants (Falcone et al, 2014) and in prior studies (Loughead et al, 2009, 2010). Specifically, in the task-positive regions (MF/CG, left DLPFC, and right DLPFC), abstinence challenge (vs smoking satiety) was associated with reduced activation during task performance; whereas in the task-negative PCC, abstinence challenge was associated with less deactivation during task performance compared with smoking satiety. There was no significant effect of abstinence challenge on BOLD signal in the vmPFC. We examined associations between BOLD signal, accuracy, and correct response times, controlling for condition and memory load. BOLD signal in right and left DLPFC was positively associated with accuracy (left DLPFC: OR=1.13, 95% CI=1.04 to 1.23, p=0.006; right DLPFC: OR=1.09, 95% CI=1.01 to 1.18, p=0.04). There was no association between correct response time and BOLD signal in any region (ps>0.05).

Table 2 Smoking Characteristics and Performance Across Condition (Abstinence Challenge, Smoking Satiety) and Group (Quit vs Relapse) (N=72)
Full size table

Predictive Models for Quitting vs Relapse

Forward stepwise regression retained MNWS, left DLPFC, and PCC as predictors (Figure 1) with participant age also contributing to the model. Relapse (a reduced odds of quitting success) was predicted by older age (OR=1.07, 95% CI 1.01–1.13, p=0.02) and greater withdrawal change score (MNWS abstinence challenge minus smoking satiety, OR=1.22, 95% CI 1.01–1.37, p=0.04). BOLD percent signal change from two ROIs added significantly to the predictive model of relapse; reduced left DLPFC activation in abstinence challenge (vs smoking satiety) predicted greater odds of relapse (quitting failure) (OR=0.14 per standardized increase in percent signal change, 95% CI 0.03–0.74, p=0.02), and less deactivation in the PCC (task-negative region) during abstinence challenge (vs smoking satiety) predicted increased odds of relapse (OR=3.45 per standardized increase in percent signal change, 95% CI 1.05–11.32, p=0.04).

Figure 1
figure 1

(a) Visualization of functionally defined N-back ROIs in the left dorsolateral prefrontal cortex (DLPFC) and posterior cingulate cortex (PCC). (b) Forward stepwise logistic regression retained the left DLPFC, left PCC, and withdrawal scores. Greater abstinence-induced change in withdrawal (increase), left DLPFC percent signal change (reduced activation), and PCC percent signal change (less deactivation) were predictive of relapse.

PowerPoint slide

Full size image

The predictive value of the relapse model was examined using ROC curve analysis. The full model (age, FTND, MNWS, QSU-B, MF/CG, right DLPFC, left DLPFC, PCC) produced a relatively high AUC of 81% corresponding to 87.5% correct prediction rate at the optimal cut point (the point of optimal model performance; Figure 2). This model represents a significant improvement over clinical predictors (age, FTND) alone (AUC 67%, X2(1)=4.06, p<0.05). We also examined an intermediate model, using clinical and withdrawal scores (age, FTND, MNWS, QSU-B), which achieved an AUC of 73%, which did not differ significantly from the full or clinical models (AUC 67%, X2(1)=1.74, p<0.19).

Figure 2
figure 2

ROC curves for three predictive models of 7-day quit status. The full model (black) includes clinical, withdrawal and brain variables, yielding an AUC of 81%. Clinical predictors alone (red) achieved an AUC of 67% and the intermediate model, using only clinical and withdrawal scores, an AUC of 73% (blue).

PowerPoint slide

Full size image

The full model was validated using two resampling methods. Bootstrapping (Austin and Tu, 2004a) generated 1000 replicates of the data and conducted model selection on each replicate. Accounting for age and nicotine dependence, the most frequent variables selected were withdrawal, left DLPFC, and PCC—the same variables identified by stepwise procedures. LOOCV further evaluated the potential inflation of model performance. Compared with the 81% AUC for the full model, LOOCV procedures yielded 71% AUC, corresponding to 82.7% correct prediction rate at the optimal cut point. Descriptively, performance of the model under LOOCV fell in the ‘fair prediction’ range in contrast to ‘good prediction’, demonstrated by the training data. Shrinkage of coefficients obtained during the cross-validation of the ROC analyses is expected and falls within the typical range (Whelan and Garavan, 2014). The full model LOOCV AUC (71%) exceeds the rule-of-thumb cutoff for acceptable discrimination (an AUC of at least 70%; Hosmer and Lemeshow, 2000). Notably, when subjected to LOOCV procedures both the clinical and intermediate model AUC fell in the ‘chance’ range.

The exploratory whole-brain condition (abstinence challenge, smoking satiety) by group (relapse, quit) voxelwise 2 × 2 ANOVA in the matched sample (N=30) identified significant interaction effects in three clusters (Table 3). These clusters overlapped considerably with our functionally defined a priori ROIs in the left DLPFC, right DLPFC, and MF/CG (Figure 3a). An interaction effect in the PCC was seen at an uncorrected threshold (p<0.05). For the MF/CG and right DLPFC clusters, percent signal change did not differ between the quit and relapsed groups during the smoking satiety condition; however, when abstinent, MF/CG and right DLPFC signal decreased in relapsers and increased in quitters (Figure 3b). In the left DLPFC, the quit group (vs relapse) activated less during the smoking satiety condition and this pattern reversed under abstinence challenge (quit group showed increased signal). Consistent with the ROI-based analysis described above, abstinence challenge reduced signal in all clusters for the relapse group.

Table 3 Condition (Smoking Satiety, Abstinence Challenge) by Group (Quit, Relapse) Whole-brain Interaction Results (Z>1.96, p<0.05) for Matched Sample (N=30)
Full size table
Figure 3
figure 3

(a) Whole-brain condition (abstinence challenge, smoking satiety) by group (relapse, quit) voxelwise ANOVA interaction effect. Clusters (orange/yellow) are corrected for multiple comparison (Z>1.96 and probability of spatial extent p<0.05) and superimposed outlines (blue) delineate a priori ROIs. (b) MF/CG and right DLPFC clusters (see above) show comparable activation under the smoking satiety condition in both groups. There is greater activation in the left DLPFC in the smoking satiety condition for the relapse group (vs quit). Under abstinence, challenge signal in all clusters decreases for the relapse group and increases for the quit group.

PowerPoint slide

Full size image

DISCUSSION

These data suggest that disruptions in working memory-related neural activation during early abstinence may distinguish successful quitters from those who fail. Specifically, decreased left DLPFC and increased PCC percent BOLD signal change and increased withdrawal symptoms (in the abstinence challenge vs smoking satiety conditions) predicted subsequent relapse. Further, an exploratory whole-brain analysis revealed clusters overlapping with the functionally defined a priori ROIs in the DLPFC and MF/CG; however, the effect in the PCC was seen at an uncorrected threshold only (p<0.05). Both analyses are consistent in that the relapse group is characterized by a reduction in left DLPFC during abstinence challenge vs smoking satiety. In terms of model prediction, receiver operator curve analysis demonstrated 81% AUC for predicting smoking relapse using these predictors, a significant improvement over the model with clinical variables only. Cross-validation results revealed an expected degree of shrinkage; specifically, we observed a 10 percentage point change (from 0.81–0.71), which falls within an expected range for both model shrinkage (see Whelan and Garavan, 2014) and discrimination (an AUC of at least 0.70; Hosmer and Lemeshow, 2000).

The current findings supporting the predictive value of working memory-related DLPFC BOLD signal change during abstinence challenge are consistent with emerging data on cognitive neuroscience and behavior change. The DLPFC is at the core of the brain’s cognitive control network, supporting sustained attention to and memory of long-term goals, and goal-directed decision-making (MacDonald et al, 2000; Niendam et al, 2012). Indeed, when the DLPFC is engaged, activation in reward-sensitive regions (eg, nucleus accumbens, ventral striatum) can be shifted away from the drive to obtain immediate gratification in the form of smoking or eating unhealthy foods (Hare et al, 2009; Kouneiher et al, 2009; Nestor et al, 2011). Likewise, studies of resting state functional connectivity have shown that activity in the executive control network (which includes DLPFC) is inversely correlated with activation in the default mode network (Fox et al, 2005). The PCC is a central hub in the default mode network (Andrews-Hanna et al, 2010; Raichle et al, 2001), and suppression of activation (or deactivation) is integral to shifting attention away from stimuli irrelevant to a goal directed behavior (eg, quitting). Stimuli that are disruptive to an abstinent smoker include external environmental cues as well as internal processes such as craving and intrusive thoughts about wanting to smoke (Ashare et al, 2013b; Lerman et al, 2014).

Dopamine levels in the DLPFC are modulated, in part, by signaling at both α4β2 and α7 nicotinic acetylcholine receptors (nAChRs) in the ventral tegmental area (Arnsten and Jin, 2014; Nocente et al, 2013). Consistent with these molecular mechanisms and with the present data, we have shown that varenicline, a partial agonist at α4β2 and full agonist at α7 nAChRs, restores working memory performance and increases DLPFC activity during abstinence challenge compared with placebo during smoking abstinence (Loughead et al, 2010). In animal models, α7 nAChRs modulate both dopamine and NMDA receptor function in the DLPFC (Croxson et al, 2011) and working memory deficits caused by ketamine can be restored with an α7 agonist (Brozoski et al, 1979). These findings suggest the hypothesis that modulators of α4β2 and/or α7 nAChRs could be particularly beneficial for smoking cessation among smokers who exhibit larger abstinence-induced working memory deficits. Dopamine levels in DLPFC are also influenced by levels of the enzyme catechol-O-methyltransferase (COMT), and the COMT val158met variant has been shown to modulate working memory performance and associated BOLD signal change (Ashare et al, 2014; Loughead et al, 2009). Smokers with the val/val genotype (associated with a more active enzyme, resulting in decreased prefrontal dopamine levels) demonstrated impaired performance on the N-back task accompanied by reduced activation in the DLPFC during abstinence challenge compared to smoking (Loughead et al, 2009). However, pharmacologic inhibition of COMT with short-term tolcapone treatment had minimal effects on smoking behavior and no effects on working memory-related BOLD signal in either the DLPFC or PCC (Ashare et al, 2013b). Future pharmacoimaging trials for smoking cessation incorporating working memory testing and fMRI may be useful to evaluate whether increased DLPFC activity, and possibly reduced PCC activity, serves as an intermediate biomarker of smoking cessation treatment efficacy (Bough et al, 2013).

Nonpharmacologic interventions that augment DLPFC function, support neuronal plasticity, and improve executive cognitive control may also have utility for smoking cessation. Cognitive exercise training, real-time fMRI feedback, and transcranial direct current brain stimulation (tDCS) have shown some promising initial results, supporting the enhancement of both working memory performance and DLPFC function (Dresler et al, 2013; Snowball et al, 2013; Zhang et al, 2013). However, evidence for beneficial effects on smoking behavior from preliminary studies is mixed (Sheffer et al, 2013; Wing et al, 2013). Other recent data suggest that concurrent treatment with both cognitive training and DLPFC-targeted tDCS may be most profitable (Segrave et al, 2014).

Although the current study reveals the significant contribution of neural activity to prospectively measured smoking relapse, the relatively small sample of successful quitters is a limitation. The low success rate is a consequence of using brief behavioral counseling without medication to support the quit attempt. However, a majority of quit attempts take place under similar conditions, making our results representative of the natural environment. Given the low success rates for unaided quitting, our study focused on a short-term (7 day) relapse measure rather than long-term cessation. It should be noted, however, that this measure is highly predictive of long-term quitting success (Ashare et al, 2013a). A replication sample is required to validate the model; however the observed shrinkage in AUC with the LOOCV procedure suggests that our results are not likely to represent inflated predictive power.

Although the broad implementation of neuroimaging assessment is not currently clinically or economically feasible, the mechanisms identified are potential targets for improved assessment instruments to assist in personalized intervention strategies. In addition to guiding assessment, predictive models can identify therapeutic targets for pharmacotherapies or neuroscience-based nonpharmacologic interventions to promote smoking cessation. Regardless of therapeutic approach, models sensitive to relapse risk can guide the refinement of existing interventions and provide early signal regarding the efficacy of novel tobacco dependence treatments.

FUNDING AND DISCLOSURE

Dr Lerman has served as a consultant and/or received research funding from Pfizer. Dr Gur has received compensation for professional services from US Courts. The remaining authors declare no conflict of interest.