Toward biomarkers of the addicted human brain: Using neuroimaging to predict relapse and sustained abstinence in substance use disorder

Abstract

The ability to predict relapse is a major goal of drug addiction research. Clinical and diagnostic measures are useful in this regard, but these measures do not fully and consistently identify who will relapse and who will remain abstinent. Neuroimaging approaches have the potential to complement these standard clinical measures to optimize relapse prediction. The goal of this review was to survey the existing drug addiction literature that either used a baseline functional or structural neuroimaging phenotype to longitudinally predict a clinical outcome, or that examined test-retest of a neuroimaging phenotype during a course of abstinence or treatment. Results broadly suggested that, relative to individuals who sustained abstinence, individuals who relapsed had (1) enhanced activation to drug-related cues and rewards, but reduced activation to non-drug-related cues and rewards, in multiple corticolimbic and corticostriatal brain regions; (2) weakened functional connectivity of these same corticolimbic and corticostriatal regions; and (3) reduced gray and white matter volume and connectivity in prefrontal regions. Thus, beyond these regions showing baseline group differences, reviewed evidence indicates that function and structure of these regions can prospectively predict – and normalization of these regions can longitudinally track – important clinical outcomes including relapse and adherence to treatment. Future clinical studies can leverage this information to develop novel treatment strategies, and to tailor scarce therapeutic resources toward individuals most susceptible to relapse.

1. Introduction

Drug addiction is a chronic disorder marked by high rates of relapse even after months or years of abstinence, long after self-reported craving and withdrawal have abated (Dennis et al., 2007). Accordingly, clinical self-report measures only modestly predict relapse and future drug use in the laboratory (Paliwal et al., 2008) and the clinic (Miller and Gold, 1994), are subject to a range of problems and biases that may reduce their reliability and validity (Crowne and Marlowe, 1960; Moeller and Goldstein, 2014; Williamson, 2007), and present challenges of translation between human and preclinical studies (Moeller and Stoops, 2015; Sinha et al., 2011). For these reasons, and in recognition that the persistence of drug-seeking and relapse may at least partly stem from underlying neurobiological alterations associated with chronic consumption of the drug, more recent studies of relapse prediction have incorporated neuroimaging approaches. These neuroimaging procedures, which include magnetic resonance imaging (MRI), electroencephalography (EEG), and positron emission tomography (PET) (among others), allow researchers to examine noninvasively how structural and/or functional brain abnormalities may contribute to relapse and other important clinical outcomes.

To date, studies correlating brain phenotypes with abstinence mainly have been cross-sectional, for example testing whether active/recent users and abstinent users differ on neuroimaging markers associated with inhibitory control, cue-reactivity, or gray matter volume (GMV) (Bell et al., 2014; Castelluccio et al., 2014; Connolly et al., 2013; Ersche et al., 2005; Li et al., 2013; Parvaz et al., 2016b). However, a growing number of studies have begun employing longitudinal designs, examining whether a particular neuroimaging phenotype predicts future clinical outcomes. Importantly, longitudinal studies can both inform the direction of association and can account for at least some of the extraneous variables (beyond abstinence) that may differ between active and former users. For example, active users and former users studied cross-sectionally may additionally differ on treatment motivation (Prisciandaro et al., 2014), self-regulation (Heatherton and Wagner, 2011), recent drug use that may prime further use (Donny et al., 2004), the expectation of receiving an imminent drug reward (Wilson et al., 2012), and potentially many other factors. In longitudinal studies, addicted individuals ostensibly begin the study with equivalent motivation and underlying neurobiology, and only later diverge on abstinence. For this reason, longitudinal studies are putatively less likely than cross-sectional studies to yield epiphenomenal neural signatures of abstinence.

These longitudinal prediction designs were the focus of the current review. The goal was to examine the extent to which neuroimaging phenotypes [i.e., structural and functional MRI (fMRI), EEG, and PET] can prospectively predict clinical relapse in human drug addiction. Animal studies were not included in this review because, although animal models are vital for cause-and-effect understanding of addiction pathophysiology, animal abstinence can be easily and externally enforced compared with human abstinence (Garavan and Weierstall, 2012). PubMed was searched on August 1, 2016 for the following keywords: “addiction and (brain imaging or biomarker) and (relapse or abstinence)”; relevant citations were also gleaned from prior reviews on broadly similar topics [e.g., (Courtney et al., 2016; Garavan and Weierstall, 2012; Garrison and Potenza, 2014; Hanlon et al., 2013; Marhe et al., 2014)]. To be included in the current review, studies must have incorporated at least one baseline neuroimaging assessment of brain structure or function that was then used to predict a ≥ 1 month relapse-relevant outcome variable (e.g., relapse, abstinence length, drug use frequency, or adherence to a particular clinical approach or course of treatment) and/or a ≥ 1 month follow-up neuroimaging assessment (during a time span where treatment was sought and/or drug use was absent or reduced); we did not include studies examining short-term abstinence or withdrawal that occurred over several hours or days [e.g., (Lerman et al., 2014; Moeller et al., 2013)]. As this field has evolved over time to use larger sample sizes with better statistical power, studies included in this review were required to have ≥ 15 study participants per group (i.e., ≥ 15 total if conducting within-group/correlational analyses; ≥ 15 per group if conducting between-group analyses of relapsers and abstainers). Our goal was not to list all regions and activations for each study, but rather to identify a manageable number of relevant regions that appeared across multiple studies of a particular task or modality (see Table 1 for summaries of the longitudinal studies reviewed here). The neuroimaging phenotypes identified in this review, then, potentially could serve as targets for future therapeutic interventions, and could help identify which individuals might be most susceptible to relapse and most in need of additional resources to sustain abstinence.

Table 1.

Summaries of main findings in longitudinal neuroimaging prediction studies.

Neuroimaging Procedure	Relevant Region	Reference	Addiction	Longitudinal Time Span	Longitudinal Outcome	Analysis Type	Imaging Tool	Contrast or Measure	Result
Drug Cue-Reactivity Tasks	DMPFC	(Beck et al., 2012)	Alcohol	3 months	Relapser / Abstainer	ROI	1.5T fMRI	Drug > neutral	Relapser ↑ activation
		(Chua et al., 2011)	Smoking	4 months	Relapser / Abstainer	ROI	3T fMRI	Tailored > untailored quit message	Relapser ↓ activation
		(Wang et al., 2015)	Heroin	3 months	Treatment adherence	Whole-brain	3T fMRI	Drug > neutral	Treatment adherence ↑ activation
	Ventral striatum	(Beck et al., 2012)	Alcohol	3 months	Relapser / Abstainer	ROI	1.5T fMRI	Drug > neutral	Relapser ↓ activation
		(Li et al., 2015)	Heroin	3 months	Relapser / Abstainer	Whole-brain	3T fMRI	Drug > neutral	Relapser ↑ activation
		(Mann et al., 2014)	Alcohol	12 weeks	Time to relapse	ROI	1.5T fMRI	Drug > neutral	Shorter time to relapse ↓ activation (Naltrexone condition)
		(Reinhard et al., 2015)	Alcohol	11 weeks	Time to relapse	ROI	1.5T fMRI	Drug > neutral	Shorter time to relapse ↑ activation
	Posterior Regions	(Kosten et al., 2006)	Cocaine	10 weeks	Negative urines	Whole-brain	1.5T fMRI	Drug > neutral	Negative urines ↑ activation
		(McClernon et al., 2007)	Smoking	10 weeks	Test-retest (x3)	ROI amygdala	1.5T fMRI	Drug > neutral	Abstinence ↓ activation
		(Parvaz et al., 2016c)	Cocaine	6 months	Test-retest (x2)	ROI	ERP: LPP	Drug > pleasant	Abstinence ↑ pleasant LPP; ↓ drug LPP
		(Versace et al., 2014)	Smoking	24 weeks	Relapser / Abstainer	ROI	3T fMRI	Cluster analysis (drug, pleasant)	Relapser ↑ activation drug; ↓ activation pleasant
Inhibitory Control Tasks	DLPFC	(Brewer et al., 2008)	Cocaine	8 weeks	Treatment retention	Whole-brain	3T fMRI	Incongruent > Congruent	Treatment retention ↓ activation
		(Worhunsky et al., 2013)	Cocaine	8 weeks	Treatment retention	Whole-brain ICA	3T fMRI	Incongruent > Congruent	Treatment retention ↓ activation
	ACC	(Luo et al., 2013)	Cocaine	3 months	Relapser / Abstainer	Whole-brain	3T fMRI	Stop error > stop success	Relapser ↓ activation
		(Marhe et al., 2013a)	Cocaine	3 months	Days of use	ROI	3T fMRI	Drug > neutral	More cocaine use days ↓ activation
		(Marhe et al., 2013b)	Cocaine	3 months	Days of use	ROI	ERP: ERN	Incorrect > correct	More cocaine use days ↓ ERN
	Striatum	(Brewer et al., 2008)	Cocaine	8 weeks	Negative urines	Whole-brain	3T fMRI	Incongruent > Congruent	Negatives urines ↑ activation
		(Kober et al., 2014)	Cannabis	1 year	Days of abstinence	Whole-brain	3T fMRI	Incongruent > Congruent	More days abstinence ↑ activation
		(Worhunsky et al., 2013)	Cocaine	8 weeks	Negative urines	Whole-brain ICA	3T fMRI	Incongruent > Congruent	Negatives urines ↑ activation
Monetary Reward Tasks	Striatum	(Jia et al., 2011)	Cocaine	8 weeks	Negative urines	ROI	3T fMRI	Reward > neutral	Negatives urines ↓ activation
		(Stewart et al., 2014)	Meth	1 year	Relapser / Abstainer	Whole-brain	3T fMRI	[win, loss, tie] > baseline	Relapser ↓ activation
	Midbrain	(Balodis et al., 2016)	Cocaine	4–8 months	Test-Retest	Whole-brain	3T fMRI	Anticipation > baseline	Abstinence ↑ activation
		(Moeller et al., 2012)	Cocaine	6 months	Test-Retest	Whole-brain	4T fMRI	Reward > baseline	Abstinence ↑ activation
Decision-Making Tasks	Insula	(Gowin et al., 2014)	Meth	1 year	Relapser / Abstainer	ROI	3T fMRI	Risky > safe decision	Abstinence ↑ activation
		(Gowin et al., 2015)	Meth	1 year	Relapser / Abstainer	ROI	3T fMRI	Riskiest > safe decision	Abstinence ↑ activation
		(Paulus et al., 2005)	Meth	1 year	Relapser / Abstainer	ROI	1.5T fMRI	Prediction > non-prediction	Relapser ↓ activation
Resting-State	Striatum	(Camchong et al., 2013)	Alcohol	6 months	Relapser / Abstainer	Seed-based (putamen)	3T fMRI	Whole-brain connectivity	Relapser ↓ connectivity (e.g., with ACC, DLPFC, insula)
		(McHugh et al., 2013)	Cocaine	1 month	Relapser / Abstainer	Seed-based (putamen)	3T fMRI	Whole-brain connectivity	Relapser ↓ connectivity (e.g., with insula)
	Amygdala	(McHugh et al., 2014)	Cocaine	1 month	Relapser / Abstainer	Seed-based (partitioned amygdala)	3T fMRI	Whole-brain connectivity	Relapser ↓ corticomedial connectivity (e.g., with vmPFC)
	Anterior PFC	(Bauer, 2001)	Mixed	6 months	Relapser / Abstainer	Pre-selected frequency bands	EEG	High frequency beta activity	Relapser ↑ beta activity
	Whole-brain	(Jacobus et al., 2012)	Cannabis	4 weeks	Test-retest (x2)	Whole-brain	ASL	Cerebral blood flow	Abstinence ↑ blood flow
		(Janu et al., 2012)	Alcohol	1 year	Duration of abstinence	Whole-brain	PET 18FD G	18FDG uptake	Longer abstinence ↑ brain metabolism
Gray Matter Structure	ACC and/or OFC	(Beck et al., 2012)	Alcohol	3 months	Relapser / Abstainer	ROI	1.5T MRI	GMV	Relapser ↓ GMV
		(Cardenas et al., 2011)	Alcohol	8 months	Relapser / Abstainer	Whole-brain	1.5T MRI	GMV	Relapser ↓ GMV
		(Durazzo et al., 2011)	Alcohol	1 year	Relapser / Abstainer	ROI	1.5T MRI	GMV	Relapser ↓ GMV
		(Parvaz et al., 2016a)	Cocaine	6 months	Test-Retest (x2)	Whole-brain	4T fMRI	GMV	Abstinence ↑ GMV
		(Rando et al., 2011)	Alcohol	90 days	Time to relapse	ROI	3T MRI	GMV	Faster relapse ↓ GMV
	Larger PFC	(Durazzo et al., 2010)	Alcohol	12 months	Relapser / Abstainer	ROI	1.5T MRI	Perfusion MRI of GMV	Abstainer ↑ perfusion
		(Mon et al., 2011)	Alcohol	7–8 months	Test-retest (x3)	ROI	1.5T MRI	GMV	Abstinence ↑ GMV
		(Morales et al., 2012)	Meth	1 month	Test-retest (x2)	Whole-brain	1.5T MRI	GMV	Abstinence ↑ GMV
	Limbic Regions	(Van Dam et al., 2014)	Mixed	90 days	Days of use	ROI	3T MRI	GMV	More days abstinence ↓ GMV
		(Wrase et al., 2008)	Alcohol	6 months	Relapser / Abstainer	ROI	1.5T MRI	GMV	Relapser ↓ GMV
		(Xu et al., 2014)	Cocaine	12 weeks	Negative urines	ROI	3T MRI	GMV	Negatives urines ↓ GMV
White Matter Structure	Frontal Regions	(Mon et al., 2011)	Alcohol	7–8 months	Test-retest (x3)	ROI	1.5T MRI	White matter volume	Abstinence ↑ white matter volume
		(Pfefferbaum et al., 2014)	Alcohol	1–8 years	Relapser / Abstainer	ROI	1.5T MRI	DTI FA	Relapser ↓ volume (steeper age-related decline)
		(Xu et al., 2010)	Cocaine	8 weeks	Negative urines	Whole-brain	3T MRI	DTI FA	Negatives urines ↑ FA
		(Wang et al., 2013)	Heroin	1 month	Test-retest (x2)	Whole-brain	1.5T MRI	DTI FA	Abstinence ↑ FA

Abbreviations. ACC = anterior cingulate cortex, ASL = arterial spin labeling, DTI = diffusion tensor imaging, DLPFC = dorsolateral prefrontal cortex, DMPFC = dorsomedial prefrontal cortex, EEG = electroencephalogram, ERP = event-related potentials, FA = fractional anisotropy, fMRI = functional magnetic resonance imaging, GMV = gray matter volume, ICA = independent components analysis, LPP = late positive potential, Meth = methamphetamine, OFC = orbitofrontal cortex, PET = positron emission tomography, ¹⁸FDG = fluorodeoxyglucose, ROI = region of interest

2. Functional Imaging Phenotypes

2.1 Task-Based Activation

Neural activations elicited during a range of cognitive and emotional tasks have been used to predict clinical outcomes. Tasks have principally included cue-reactivity, response inhibition, monetary reward, and decision-making.

2.1.1 Drug Cue-Reactivity

Cue-reactivity tasks assess the degree to which drug-related cues, including actual drug paraphernalia or more abstract stimuli such as words and images, capture attention and evoke a craving response in drug-addicted individuals (Jasinska et al., 2014). These tasks are meant to tap into the excessive motivational significance carried by drugs and their associated cues (Goldstein and Volkow, 2011; Robinson and Berridge, 2008). Most studies have contrasted brain activation to drug stimuli against activation to neutral, non-drug-related stimuli; a subset of studies has contrasted activation to drug stimuli against activation to other appetitive reinforcers (e.g., sexual images, serene beaches, etc.). Most tasks have entailed passive cue exposure; fewer have incorporated an additional executive function such as inhibitory control [e.g., whether addicted individuals have the ability to halt a (presumably) prepotent tendency to respond to drug stimuli].

One region consistently engaged by cue-reactivity is the medial prefrontal cortex (PFC), extending into the rostral anterior cingulate cortex (ACC) (Engelmann et al., 2012; Kuhn and Gallinat, 2011; Schacht et al., 2013). This region, among other functions, participates in tagging stimuli as being salient and self-relevant (Abraham, 2013; van der Meer et al., 2010). In one study, compared with healthy controls and alcohol-dependent individuals who achieved 3-month abstinence, alcohol-dependent individuals who relapsed after 3 months had increased baseline fMRI activation to alcohol cues in the dorsomedial PFC (Beck et al., 2012). Dorsomedial PFC activation also predicted therapeutic adherence to extended-release naltrexone in heroin dependence (i.e., the number of monthly injections received) (Wang et al., 2015). A final study reported that increased fMRI activation in the dorsomedial PFC to tailored quit-smoking messages predicted actual quitting at 4-month follow-up (Chua et al., 2011). Taken together, these studies clearly implicate regions in and adjacent to the dorsomedial PFC in the prediction of clinical outcome, but the direction of association remains to be clarified. One speculative explanation could be that non-specific engagement of this region to drug cues could result in poor clinical outcomes, whereas engagement of this region to drug cues in a therapeutic or change-related context could result in positive clinical outcomes.

Another prominent region in these studies is the ventral striatum (nucleus accumbens) (Balodis and Potenza, 2015), a region prominently involved in reward anticipation and consumption (Diekhof et al., 2012). In a prospective 3-month study, compared with a non-relapsing group, a relapsing group of methadone-maintained heroin-addicted individuals had increased baseline fMRI activation to heroin pictures (versus neutral pictures) in the bilateral ventral striatum (and cerebellum) (Li et al., 2015). In a corroborating study of alcohol-dependent individuals, elevated ventral striatal (and ventral ACC and lateral OFC) fMRI activation to alcohol cues predicted a faster time to relapse over the next 11 weeks (Reinhard et al., 2015) (Figure 1A). In another study, a medication (naltrexone) reduced the risk of relapse specifically among alcohol-dependent individuals who exhibited higher baseline ventral striatal activity to alcohol cues (Mann et al., 2014). In a contradictory study, however, relapsers (versus abstainers and controls) had reduced activation to alcohol-related pictures in the ventral striatum (and midbrain) (Beck et al., 2012). The reasons for this difference of association in ventral striatum between these respective studies are not entirely clear, but could involve the differential treatments and medications received by the participants.

Figure 1.

Representative functional task studies. (A) In alcohol-addicted individuals who completed a cue-reactivity task, fMRI activation in ventral striatum during viewing of alcohol versus neutral pictures predicted a shorter time to relapse. (B) In cocaine-addicted individuals who completed a color-word Stroop task, fMRI activation during interference in the dorsolateral prefrontal cortex predicted the number of weeks in treatment. (C) In cocaine-addicted individuals who completed a monetary incentive delay task, higher fMRI activation in the midbrain was seen post-treatment compared with pre-treatment, and predicted the days abstinent from cocaine. (D) In methamphetamine-addicted individuals who completed a risky-decision making task, fMRI activation in the insula during risky (versus safe) decisions predicted relapse with greater sensitivity and specificity than standard clinical variables. (A) and (B) are adapted from (Reinhard et al., 2015) and (Brewer et al., 2008), respectively, both with permission from Elsevier; (C) and (D) are adapted from (Balodis et al., 2016) and (Gowin et al., 2014), respectively, both with permission from Nature Publishing Group.

Beyond these regions relevant to reward and salience, neuroimaging cue-reactivity studies have identified more posterior brain regions as being predictive of clinical outcome. In one of the earliest studies, activation in the left precentral and right superior temporal gyri during cocaine versus neutral movie clips predicted the total number of negative urine toxicologies in cocaine-dependent participants (assessed over the next 10 weeks) (Kosten et al., 2006). In a subsequent study, cluster analysis of fMRI data during exposure to pleasant, unpleasant, neutral, and smoking images identified two groups of smokers, who differed on their ability to maintain smoking abstinence over 24 weeks. The relapsing smoker group had higher activation to smoking images in multiple regions including the occipital, parietal, and temporal cortices, and precuneus; the abstinent smoking group had higher activations in these regions to the pleasant images (Versace et al., 2014). This finding is consistent with a subsequent study using EEG-defined event-related potentials (ERPs) (Parvaz et al., 2016c) – specifically the late positive potential (LPP), an ERP component that tracks stimulus arousal (Hajcak et al., 2010) and correlates with drug craving (Franken et al., 2008). LPPs in this study were defined along central and parietal cortices, elicited during the presentation of drug and pleasant images. After six months of treatment, abstinence, and/or significant drug use reduction, drug-elicited LPPs were diminished, whereas pleasant-elicited LPPs were enhanced (Parvaz et al., 2016c). A final study examined a region of interest (ROI) in the bilateral amygdala. Smokers completed three fMRI scans (first at baseline, second while enrolled in a treatment program consisting of reduced nicotine content cigarettes supplemented by nicotine patch, and third after treatment). Baseline amygdala activation was higher during the viewing of smoking versus neutral pictures at baseline, but these differences were eliminated during and after treatment (McClernon et al., 2007). Thus, higher activation to drug cues in these diffuse brain regions generally predicted worse outcome.

2.1.2 Inhibitory Control

Inhibitory control tasks assess the capacity to control ongoing behavior. By contrasting behavior and associated neural activity occurring during a condition of response inhibition against a condition of prepotent response, inhibitory control tasks have consistently revealed that drug-addicted individuals exhibit self-control impairments (Luijten et al., 2014; Smith et al., 2014), which may contribute to or exacerbate compulsive drug-taking (Goldstein and Volkow, 2011). Three of the most commonly used inhibitory control tasks, in order from simplest to most cognitively complex, include: (1) go/no-go tasks, where individuals must respond as quickly as possible to frequent go stimuli and inhibit responses to infrequent no-go stimuli) (Chambers et al., 2009); (2) stop-signal tasks, where individuals must successfully inhibit an action that has already begun in response to an initial stimulus, following the presentation of a second, stop stimulus (Aron et al., 2014; Verbruggen and Logan, 2008); and (3) Stroop tasks, where individuals must override a more automatic response tendency [e.g., reading the semantic content of a particular color word (e.g., “blue”) or an emotionally-charged word (e.g., “cocaine”)] and instead respond with an alternate task-specific demand (e.g., responding to the ink color of the word) (MacLeod, 1991; Smith and Ersche, 2014).

Dorsolateral prefrontal cortex (DLPFC) activation has emerged as an important predictor of clinical outcomes in these paradigms, with studies generally finding that reduced activation is associated with better outcomes. A focus on the DLPFC is not unsurprising, given that this region plays an important role in cognitive control across a range of paradigms (Niendam et al., 2012). In a key early study, reduced DLPFC fMRI activation during color-word Stroop interference predicted higher treatment retention in cocaine dependence (Brewer et al., 2008) (Figure 1B). Also in cocaine-dependent individuals, independent component analysis (ICA) identified five distinct functional networks that were engaged during color-word Stroop interference. Greater engagement of a “fronto-cingular” network, which included the DLPFC among other regions, predicted fewer completed weeks of treatment (Worhunsky et al., 2013).

ACC activation has been another important predictor of future drug-related outcomes, although the direction of association is less clear. Again, however, a focus on the ACC is not unexpected given the important involvement of this region in cognitive control (Niendam et al., 2012). In one study, treatment-seeking cocaine-dependent individuals performed a baseline stop-signal task during fMRI, and were followed over the next three months to assess relapse. In all participants, relapse was predicted by decreased error-related (i.e., the fMRI contrast ‘stop error greater than stop success trials’) activations in the dorsal ACC (Luo et al., 2013). Similarly, another study examined the error-related negativity (ERN) (Marhe et al., 2013b), an ERP component that appears approximately 100 ms following an incorrect response and that can be traced to ACC activity (Miltner et al., 2003). Reduced ERN during a flanker task (measured during the first week of treatment) predicted more cocaine use days (measured at 3-month follow-up) independently of baseline craving (Marhe et al., 2013b). In contrast, higher dorsal ACC activation during drug-Stroop interference (in combination with more self-reported craving) predicted more cocaine use days over the next three months in recently detoxified cocaine-dependent patients (Marhe et al., 2013a).

Striatal activation also appears relevant for predicting relapse in these response inhibition studies. For example, baseline fMRI activation during color-word Stroop interference in ventral striatum (positively) and insula/putamen (negatively) correlated with percent days of marijuana abstinence at 1-year follow-up in treatment-seeking cannabis-dependent individuals (Kober et al., 2014). In addition, longer duration of self-reported abstinence positively correlated with activation of right putamen (also with ventromedial PFC and left posterior cingulate cortex), and percent drug-free urine screens positively correlated with activation of right putamen (Brewer et al., 2008). Finally, greater engagement of a “subcortical” network (i.e., thalamus, striatum, amygdala, and hippocampus) and a “ventral fronto-striatal” network, determined by ICA, predicted a better treatment outcome (more cocaine-negative urine toxicologies) (Worhunsky et al., 2013). Thus, on balance, greater striatal activity during inhibitory control was associated with better outcomes.

2.1.3 Monetary Reward

Beyond cue-reactivity to drug-related stimuli, drug-addicted individuals have shown abnormal responsiveness to non-drug reinforcers (Aguilar de Arcos et al., 2008; Asensio et al., 2010; Lubman et al., 2009). The most common non-drug reinforcer examined in drug addiction has been money (Buhler et al., 2010; de Ruiter et al., 2009; Goldstein et al., 2007).

Similarly to the drug cue-reactivity studies, core ROIs in these monetary reward studies have typically included the ventral and dorsal striatum, with seemingly different directions of association. For the dorsal striatum [a region importantly implicated in drug habit (Everitt and Robbins, 2016)], methamphetamine-dependent participants completed an fMRI reinforcement learning task (i.e., a paper-scissors-rock task, where certain responses were reinforced (won) more than other responses on given blocks) (Stewart et al., 2014). Compared with 1-year methamphetamine abstainers, methamphetamine relapsers had lower bilateral caudate activation (and lower activation in multiple other regions including the ACC, bilateral insula, left inferior frontal gyrus) in response to all task outcomes (wins, losses, and ties) (Stewart et al., 2014). For the ventral striatum, cocaine-dependent individuals completed a monetary incentive delay (MID) task (Jia et al., 2011), which has been used to examine how addicted individuals respond to the anticipation and receipt of various monetary rewards (Balodis and Potenza, 2015). The addicted individuals in this study had greater activation during reward receipt than healthy controls in ventral striatum (and right insula). Over an 8-week follow-up period within the cocaine-dependent participants, higher ventral striatal (and thalamus) activation during reward outcome negatively correlated with abstinence length and percent urine-negative toxicology (the higher the activation, the worse the clinical outcome) (Jia et al., 2011).

The dopaminergic midbrain has also been identified in neuroimaging clinical outcome studies examining monetary reward, with higher activation predicting better outcomes. In an earlier study, cocaine-addicted individuals completed a monetarily rewarded fMRI drug-word task (similar to a drug-Stroop task), but that also provided varying payments for correct performance. Participants engaged the midbrain during the task more after 6 months of treatment/abstinence than at baseline (Moeller et al., 2012). A more recent study using a MID task corroborated these effects in the midbrain (as well as the thalamus and precuneus) while studying a larger sample of cocaine-addicted individuals (Balodis et al., 2016). In particular, fMRI midbrain activation to reward anticipation was higher at posttreatment than pretreatment; the higher this midbrain activation, the longer was the abstinence from cocaine (Balodis et al., 2016) (Figure 1C).

2.1.4 Decision-Making

Decision-making tasks model how individuals choose from a set of available options, with the goal of producing an optimal outcome given current internal and external demands (Paulus, 2007). Numerous studies have reported decision-making deficits and biases in drug addiction [for reviews, see (Gowin et al., 2013; Redish et al., 2008; Verdejo-Garcia, 2016)], and a subset of such studies have used neuroimaging procedures to predict prospective relapse.

More so than in other task domains, decision-making tasks have prominently featured the insula as a region that predicts clinical outcomes. The insula, along with multiple other cortical regions, has been consistently implicated in decision-making deficits in addiction (Gowin et al., 2013). In an early and influential study, methamphetamine-dependent individuals completed a basic fMRI decision-making task (i.e., participants were asked to predict, and choose accordingly, the location of a stimulus that would soon appear on a computer screen). Baseline task-related insula activation (in combination with activation in the posterior cingulate and temporal cortex) correctly predicted the relapse status of >90% of the methamphetamine-dependent individuals included in the study (Paulus et al., 2005). In another study also in methamphetamine dependence (Gowin et al., 2014), participants completed a risky fMRI decision-making task during which they chose between safe options (which yielded certain, but marginal monetary wins) and riskier options (which yielded more lucrative wins, but could also result in monetary losses). Attenuated insula activation during risky versus safe decisions predicted a faster time to relapse; insula activation provided more sensitivity and specificity in prediction than clinical variables, such as current abstinence or total lifetime uses of methamphetamine (Gowin et al., 2014) (Figure 1D). In a subsequent study using the same sample and task (but this time while further partitioning the risky options), participants who remained abstinent one year later showed greater fMRI activation in the insula and dorsal striatum (caudate and putamen) during the riskiest task condition (versus the safe option), with an opposite pattern of results for those who relapsed (Gowin et al., 2015). Thus, attenuated risk processing in both insula and subcortical brain areas increases the risk for early relapse.

2.1.5 Summary

Task-related fMRI activation during a range of cognitive and emotional contexts has predicted important clinical outcomes measured at least one month later. The different task paradigms reviewed – cue reactivity, inhibition, monetary reward, and decision-making – yielded activations predicting clinical outcome in both distinctive regions (e.g., dorsomedial PFC, insula, DLPFC) and overlapping regions (e.g., striatum). Although the emergence of ventral and dorsal striatal regions may have been aided by the use of ROI analyses in some studies, the consistent ability of these regions to predict outcomes was impressive nonetheless. A predictive effect in striatum emerged even for neutral, response inhibition tasks, which could be expected to activate this region less than emotionally salient reward tasks.

Taken together, results generally revealed the following patterns: (1) higher activation to drug cues generally predicted worse clinical outcomes, whereas (2) higher activation to non-drug cues (e.g., money) generally predicted better clinical outcomes; and (3) higher activation of PFC, but lower activation of striatal regions, during cognitive control seemed to predict worse outcomes, although (4) higher PFC activation to the outcome of cognitive control (e.g., error processing) appeared to predict better outcomes. One could speculatively interpret these collective effects as suggesting that abstinence may be sustained insofar as drug-addicted individuals are able to reduce reactivity/engagement to drug reinforcement while increasing reactivity/engagement to non-drug reinforcement and having a high capacity to exert self-control and respond to mistakes.

2.2 Resting State

2.2.1 fMRI Functional Connectivity

Beyond task activation, fMRI studies have examined resting-state functional connectivity (RSFC). RSFC captures the synchronicity of low-frequency, spontaneous fluctuations in fMRI signals that reflect fluctuations in neuronal activity (Shmuel and Leopold, 2008) between brain regions in the absence of external stimulation (Fox and Raichle, 2007). To date, fMRI studies have used a seed-based connectivity approach, examining functional covariation (at a specified threshold) between a priori ROIs and all other voxels in the brain.

Not surprisingly, research examining prediction of clinical outcomes has often focused on similar regions as the task-based studies, including the ventral and dorsal striatum. In perhaps the first study of this kind (Camchong et al., 2013), alcohol-dependent individuals were scanned at baseline and evaluated again at six months; RSFC was examined using seeds placed in the ventral striatum (and subgenual ACC). Relapsers showed decreased RSFC between ventral striatum with dorsal striatum (putamen), thalamus, dorsal ACC, DLPFC, insula, precuneus, and temporal and fusiform gyri; relapsers also showed a similar pattern of decreased connectivity between the subgenual ACC seed and many of these same regions (Camchong et al., 2013). In another study of cocaine-addicted individuals, although no differences emerged between 30-day relapsers and abstainers on RSFC between a bilateral putamen seed with a cluster that encompassed the posterior insula and right postcentral gyrus, posthoc tests showed that group differences between all addicted individuals and controls were driven by reduced RSFC in relapsers relative to controls (McHugh et al., 2013).

Another region that has been examined as an a priori seed region is the amygdala. Compared with cocaine-addicted individuals who remained abstinent after 30 days, cocaine-addicted individuals who relapsed during this time period had reduced RSFC between the left corticomedial amygdala (as a seed region) and ventromedial PFC/rostral ACC. This pattern of effects suggested impairment because it was inconsistent with that of healthy controls. Relapsers additionally showed increased RSFC between the bilateral basolateral amygdala (as another seed region) and visual processing regions (lingual gyrus/cuneus); this pattern of effects, however, did not necessarily indicate impairment because it was consistent with that of healthy controls (McHugh et al., 2014) (Figure 2A).

Figure 2.

Representative functional resting-state studies. (A) In cocaine-addicted individuals, relapse was associated with reduced fMRI connectivity between the left corticomedial amygdala and ventromedial PFC/rostral ACC, but increased fMRI connectivity between the basolateral amygdala and visual processing regions (lingual gyrus/cuneus). (B) In a study using arterial spin labeling, marijuana users at baseline had reduced cerebral blood in the dorsomedial PFC at baseline but not at 4-week follow-up. (A) is adapted from (McHugh et al., 2014), under the Creative Commons Attribution License; (B) is adapted from (Jacobus et al., 2012), with permission from Springer.

2.2.2 Other Methodologies Examining Rest

Non-fMRI BOLD approaches have also been employed [e.g., EEG, arterial spin labeling (ASL), and PET)], examining whether baseline activation predicts clinical outcomes. An early resting-state EEG study (5-minute continuous recording) examined 6-month relapse prediction in a heterogeneous group of treatment-seeking substance-dependent individuals. The dependent variable was high frequency (19.5–39.8 Hz) beta activity, which can be traced to anterior frontal brain regions [which are known to be functionally impaired in addiction (Goldstein and Volkow, 2011)]. Results revealed increased beta activity in the relapsers compared with the abstainers and healthy controls; beta activity was a superior predictor of relapse than severity of illness (Bauer, 2001).

A second study used ASL to examine cerebral blood flow (CBF) in heavy adolescent marijuana users at baseline and subsequently after 4 weeks of urine-verified abstinence. Compared with healthy controls, the marijuana users at baseline had reduced CBF in the dorsomedial PFC, left insula, left superior and middle temporal gyri, and left supramarginal gyrus; and they had increased CBF in the right precuneus. These group differences were not observed at 4-week follow-up, suggesting normalization with abstinence (Jacobus et al., 2012) (Figure 2B). However, it should be noted that not all cannabis users had a cannabis use disorder, and participants were also not excluded for alcohol use disorder (and alcohol use was higher in the cannabis group).

Finally, a third study used PET with ¹⁸FDG in 7-week detoxified patients with alcohol dependence following short-term treatment. Baseline brain metabolism, measured by ¹⁸FDG uptake, in the cerebellar vermis was positively correlated with duration of abstinence over the next year (Janu et al., 2012).

2.2.3 Summary

Relapse generally has been predicted by weakened fMRI RSFC among corticostriatal and corticolimbic brain regions, for example using seeds placed in striatum or amygdala, and such connections may be linked to impulsivity, reward processing, and motivation (Pariyadath et al., 2016). Studies using ASL and PET suggested that increased blood flow and brain metabolism predicted better clinical outcome, although these activations were in diffuse regions. In contrast, an early study using resting-state EEG reported that relapsers had increased activation presumably in the anterior PFC. These results collectively suggest that increased brain activation and connectivity during rest generally seems to predict better clinical outcomes. However, more studies are clearly needed in this area before firm conclusions can be drawn.

3. Structural Imaging Biomarkers

In addition to functional differences, drug-addicted individuals have consistently exhibited morphological differences in brain structure from healthy controls [for meta-analyses, see (Ersche et al., 2013; Hall et al., 2015; Sutherland et al., 2016; Wollman et al., 2015)]. These structural abnormalities have been primarily observed using voxel-based morphometry (VBM) to image GMV and diffusion tensor imaging (DTI) to image white matter tracts, although related techniques have also been employed (e.g., white matter volume). These structural studies have begun to examine longitudinal prediction of clinical outcome.

3.1 Gray Matter

GMV in the ACC and OFC have consistently predicted clinical outcome, generally showing that increased GMV in these regions is associated with better outcomes. For example, compared with alcohol-dependent abstainers and healthy controls, alcohol-dependent relapsers (as defined at a 3-month follow-up) had reduced GMV in the dorsomedial PFC, ACC, and OFC (Beck et al., 2012). These results were consistent with three other studies in alcohol dependence. In the first, lower GMV in the ACC/dorsomedial PFC (and medial occipital cortex) predicted faster relapse to alcohol use (while controlling for age, IQ, and years and recent frequency of alcohol use) (Rando et al., 2011). In the second, relapsers (defined one year later) had lower baseline GMV than the abstainers and controls in the lateral OFC; furthermore, within the relapsers, lower lateral OFC GMV was associated with more severe relapse (i.e., total number of drinks consumed) (Durazzo et al., 2011). In the third, relapsers had lower GMV than abstainers (each defined after approximately eight months) also in the lateral OFC (Cardenas et al., 2011). However, in this third study, lateral OFC did not differ between alcohol-dependent individuals and light drinkers. Instead, alcohol-dependent participants had smaller GMV than controls in a large cluster spanning the temporal cortex and insula (Cardenas et al., 2011). Finally, in a recent test-retest GMV study, cocaine-addicted individuals showed longitudinally increased GMV in the ventromedial PFC (and left IFG) from baseline to 6-months follow-up (during which drug use was reduced or eliminated) (Parvaz et al., 2016a).

GMV in other cortical regions has also been identified to predict clinical outcome. Two studies in alcohol-dependent participants reported that GMV in large frontal and parietal cortical ROIs increased with abstinence. In the first study, perfusion MRI of gray matter data was conducted at baseline and at 1-month follow-up, with relapse status defined at a third time point (12-month follow-up). In cross-sectional analyses, perfusion in both ROIs was lower in the relapsers as compared with the abstainers and controls at both baseline and 1-month follow-up. However, in longitudinal analyses, abstainers had higher perfusion in the prefrontal ROI, but not in the parietal ROI, across time (Durazzo et al., 2010). In the second study, alcohol-dependent participants were studied with VBM approximately 1 week (Time 1), 1 month (Time 2), and 7–8 months (Time 3) into abstinence. Consistent with the previous study, GMV in the frontal ROI increased from Time 1 to Time 2, and from Time 2 to Time 3, with these results indicating stepwise improvement in these regions over time. Here, however, the parietal ROI also linearly improved over these time points (as did a third ROI located in the temporal cortex, although this latter ROI only showed improvement from Time 2 to Time 3) (Mon et al., 2011). Finally, an interesting study of methamphetamine-dependent individuals showed longitudinal GMV increases from baseline to 1-month follow-up in a large portion of cortex (e.g., IFG, angular and temporal gyri, precuneus, insula, and occipital pole), and showed longitudinal GMV decreases in the cerebellum over this time period (Morales et al., 2012) (Figure 3A); GMV of healthy controls did not significantly change (Morales et al., 2012). Of note, this study also included a group of healthy control smokers, enabling examination of GMV differences that might be attributable to smoking (since many methamphetamine-dependent individuals also smoke cigarettes). These analyses indicated that GMV decreases in the OFC, a major region identified above, and caudate in part may be attributable to cigarette smoking; GMV deficits in frontal, parietal, and temporal cortices, were specific to methamphetamine-dependent individuals (Morales et al., 2012). These findings highlight the importance of appropriately controlling for cigarette smoking in future studies of this kind.

Figure 3.

Representative structural studies. (A) In methamphetamine-addicted individuals, gray matter volume increased in multiple cortical regions, but decreased in the cerebellum, after one month of abstinence compared with baseline. (B) In cocaine-addicted individuals, abstinence positively correlated with fractional anisotropy (FA) values in the frontal and parietal lobes, and in the body of corpus collosum. (A) is adapted from (Morales et al., 2012), with permission from Elsevier; (B) is adapted from (Xu et al., 2010), with permission from Nature Publishing Group.

Finally, GMV effects have been reported in limbic regions, including hippocampus, hippocampal gyrus, and amygdala. In treatment-seeking cocaine-addicted individuals, lower hippocampal volumes were correlated with more cocaine-negative urines during subsequent 12-week outpatient treatment (Xu et al., 2014). It should be noted, however, that the cocaine participants did not differ from controls on hippocampal volume at baseline (Xu et al., 2014). In contrast, a different study that included a large cohort of individuals with heterogeneous substance use disorders reported that lower limbic GMV (i.e., of hippocampus, parahippocampus, and fusiform gyri) predicted greater severity of relapse (days of use over 90 days) (Van Dam et al., 2014); similar to the preceding study (Xu et al., 2014), this limbic GMV cluster was not associated with substance use disorder per se, but rather was associated with experiencing childhood trauma (Van Dam et al., 2014). In an earlier study, 6-month alcohol relapsers, compared with abstainers and healthy controls, had reduced GMV in the amygdala; here, volume of the hippocampus (and ventral striatum) was lower in alcohol users than controls, but did not correlate with relapse (Wrase et al., 2008). Thus, further studies are needed to evaluate whether and in what direction hippocampal GMV predicts clinical outcome; the amygdala is an interesting candidate, but also requires additional investigation.

3.2 White Matter

Comparably fewer studies have investigated longitudinal prediction of clinical outcome using DTI in addiction, but frontal regions again appear to be important, showing positive associations with better clinical outcomes. In one study of treatment-seeking cocaine-addicted individuals, percent of cocaine-free urines over 8 weeks positively correlated with fractional anisotropy (FA) values, a DTI marker of white matter integrity, in the frontal and parietal lobes, and in the body of corpus collosum (Xu et al., 2010) (Figure 3B). In addition, heroin-dependent individuals had less FA than controls at baseline in white matter areas adjacent to the middle frontal cortex and dorsal ACC, but these group differences were eliminated at 1-month follow-up, suggesting improvement (Wang et al., 2013). An interesting recent study compared age-related white matter trajectories of alcohol relapsers and abstainers, measured over multiple follow-up periods throughout a 1–8 year period (Pfefferbaum et al., 2014). Relapsers, but not abstainers, had accelerated white matter declines in multiple brain areas (including PFC regions) (Pfefferbaum et al., 2014). Finally, in alcohol-dependent individuals, white matter volume in a large frontal ROI increased from one week to one month, and from one month to 7–8 months (Mon et al., 2011).

One study showed a white matter prediction effect in subcortical regions. Heroin relapsers (assessed at six months) had decreased FA in, among other white matter structures, the left anterior and posterior limb of internal capsule (adjacent to basal ganglia regions) (Li et al., 2016).

3.3. Summary

Studies have consistently linked addiction to lower GMV volume in multiple, but especially PFC, brain regions. It follows that increased GMV in these regions, or recovery of GMV in these regions over time, predicts and tracks with better clinical outcomes. Subcortical GMV effects, however, are less clear. For some brain regions (e.g., hippocampus), there have been different directions of association; for other brain regions (e.g., striatum, amygdala), there have not been enough studies to render definitive conclusions. It will be interesting for future studies of this kind to spotlight morphology of these subcortical regions in light of reliable evidence that substances of abuse alter microscopic neuronal structure in these regions (e.g., in ventral striatum) (Villalba and Smith, 2013). White matter/DTI studies have been far less frequent than GMV studies, but those that have been conducted similarly appear to conclude that greater white matter integrity (e.g., indexed by FA) is associated with better longitudinal clinical outcomes.

4. Future Directions

4.1 Underlying Neurochemistry

To facilitate the development of new treatments and prevent future relapse, it will be important to study the neurochemical predictors of clinical outcome (e.g., via PET with select radioligands). For example, in studies of cocaine addiction, baseline elevated mu opioid receptor binding, measured with [¹¹C]carfentanil, in frontal and temporal cortices predicted fewer weeks of cocaine abstinence during a 12-week treatment program and a shorter time to resuming cocaine use (Ghitza et al., 2010; Gorelick et al., 2008). Also in cocaine addiction, baseline dopamine D2-type receptor availability, measured with [¹¹C]raclopride, in ventral striatum predicted with 82% accuracy greater cumulative treatment attendance (Luo et al., 2014). Finally, alcohol-dependent participants had decreased type 1 cannabinoid receptor binding, measured using PET with [¹⁸F]MK-9470 in posterior brain regions, compared with healthy social drinkers, and these deficits failed to normalize with one month of abstinence (Ceccarini et al., 2014).

Select neurotransmitters and brain metabolites could also be examined with magnetic resonance spectroscopy (MRS). This technique has been used to examine differences in concentrations of select neurotransmitters and metabolites in addicted individuals versus healthy controls (Licata and Renshaw, 2010; Moeller et al., 2016), and differences between active and abstinent addicted individuals (Mon et al., 2012). No studies were identified, however, that used a longitudinal prediction design.

4.2 Genetic Modulation

Single nucleotide polymorphisms (SNPs) have been shown to modulate the effects of brain imaging phenotypes in select regions highlighted in this review (e.g., dorsal striatum, OFC) on relapse prediction. This modulation has been either direct (i.e., where a gene × brain phenotype interaction was detected) or indirect (i.e., where a gene modulated brain functional activation during a particular task, and this functional activation in turn predicted relapse). Relevant effects have been reported with SNPs of: GATA binding protein 4 gene (GATA4) that is relevant to atrial natriuretic peptide (ANP) receptors (Zois et al., 2016), kainate receptor gene (GRIK1) that is relevant to glutamatergic neurotransmission (Bach et al., 2015a), mu opioid receptor gene (OPRM1) (Bach et al., 2015b) and kappa opioid receptor gene (OPRK1) that are both relevant to opioid neurotransmission (Xu et al., 2013), and brain-derived neurotrophic factor (BDNF) gene (Hoefer et al., 2014).

To date, these imaging genetics studies have been limited to select SNPs. Gene discovery approaches that take into account imaging phenotypes, and that are validated via clinical outcome measures including relapse, could present a highly promising future direction. However, to achieve adequate statistical power, large-scale collaborative efforts from imaging and genetics consortia will be required. Despite these challenges, examination of genetics offers an opportunity both to further explore underlying neurochemistry of relapse prediction (i.e., in additional to PET approaches described above) and potentially to illuminate novel pharmacogenomic therapies (i.e., where specific individuals are targeted with specific treatments/medications based on their genetic profiles) (Haile et al., 2009).

4.3 New Task Paradigms

Additional paradigms, which relate to constructs that have not been explored, could also be examined in the prediction of relapse. These paradigms could include the induction of stress, which is highly relevant to opponent process/anti-reward theories of addiction (Koob and Volkow, 2016) and indeed has been examined in the context of imaging drug relapse in one study (Seo et al., 2013); or could include the demands of working memory, for which relapse prediction has been examined in cannabis users (Cousijn et al., 2014). Interesting future constructs not yet examined might include interoception, which involves the ability to perceive and interpret one’s own emotions and bodily sensations including those related to drug use (Paulus and Stewart, 2014; Verdejo-Garcia et al., 2012); and drug-related choice behavior, which involves making choices specifically under a drug-related context (Moeller and Stoops, 2015).

5. Conclusion

Drug addiction is a chronic disease with a high propensity for relapse, but few reliable and valid biomarkers are available that accurately predict relapse. The aim of this review was to identify brain functional and structural phenotypes that can inform which addicted individuals, after beginning treatment or abstinence, are more likely to relapse versus remain abstinent. Although the longitudinal studies reviewed here cannot speak to causal relationships between variables, they constitute an important advance over cross-sectional studies that compare groups of active and abstinent individuals, who might differ on many (potentially unmeasured and unaccounted for) variables upon entry to the study.

Results were broadly consistent with the hypothesis that relapse becomes more likely when addicted individuals show increased task activation to drug-related rewards, reduced task activation to non-drug rewards (e.g., money), and weakened functional connectivity in corticolimbic and corticostriatal circuitry; lower prefrontal GMV, white matter volume, and white matter integrity also seem generally less beneficial (Figure 4). However, it is important to acknowledge the presence of opposing findings between studies. That is, approximately 15% of studies reviewed here deviated from these broad conclusions, and it is likely that this percentage would be higher if all the subtleties and brain regions of each study were introduced and discussed at length. Beyond the obvious differences between studies in paradigms and approaches, it is important to acknowledge the considerable heterogeneity in research participant characteristics, including drug of choice, substance and mental health comorbidities, treatment modalities, and medication usage (among other factors). Large-scale studies are needed to address these sources of variability, which can potentially build toward a precision addiction medicine approach (van der Stel, 2015). Within such a framework, not unlike other specialties (Relling and Evans, 2015; Vargo-Gogola and Rosen, 2007), it might be possible to uncover the existence of different types of addictions, with each particularly suited to a certain type of treatment.

Figure 4.

Conclusions derived from this review that can inform hypotheses in future work. Relapse was generally associated with enhanced activation to drug-related cues and rewards, reduced activation to non-drug-related cues and rewards, and weakened functional connectivity in corticolimbic and corticostriatal brain regions. Relapse was also generally associated with reduced prefrontal gray and white matter volume and integrity.

We are optimistic that conclusions from our review ultimately can be applied in clinical practice. Although the field is not yet at the stage of broad implementation, promising neuroscience-based treatments indeed are beginning to emerge that show salubrious effects on future abstinence in addiction (Salling and Martinez, 2016). These include, among others, transcranial magnetic stimulation (Dinur-Klein et al., 2014) or transcranial direct current stimulation (Klauss et al., 2014) of the lateral PFC, deep brain stimulation of the ventral striatum (Kuhn et al., 2014; Voges et al., 2013), or real-time neurofeedback to augment the cortical mechanisms of attention (Horrell et al., 2010). Nevertheless, the future clinical viability of these neuroimaging approaches will depend importantly on their ability to outperform (or at least independently complement) behavioral, diagnostic, and self-report measures, which are cheaper, more easily administered, and require less technical expertise. Importantly, however, this review largely supported the clinical utility of these neuroimaging phenotypes, which often explained unique variance in the relevant clinical outcome after drug use variables and other clinical indicators (e.g., craving, frequency of use) were statistically covaried. Thus, neuroimaging approaches have the potential to identify, above and beyond currently available behavioral and diagnostic measures, the individuals who may require additional therapeutic resources for achieving sustained abstinence and better longer-term clinical outcomes.

Highlights.

• We review neuroimaging studies examining longitudinal prediction of relapse.

• We discuss MRI, fMRI, EEG, and PET studies meeting defined inclusion criteria.

• We reveal predictive effects in prefrontal and subcortical brain regions.

• These imaging phenotypes may provide biomarkers that have future clinical utility.