Mega-Analysis of Gray Matter Volume in Substance Dependence: General and Substance-Specific Regional Effects
Abstract
Objective:
Although lower brain volume has been routinely observed in individuals with substance dependence compared with nondependent control subjects, the brain regions exhibiting lower volume have not been consistent across studies. In addition, it is not clear whether a common set of regions are involved in substance dependence regardless of the substance used or whether some brain volume effects are substance specific. Resolution of these issues may contribute to the identification of clinically relevant imaging biomarkers. Using pooled data from 14 countries, the authors sought to identify general and substance-specific associations between dependence and regional brain volumes.
Method:
Brain structure was examined in a mega-analysis of previously published data pooled from 23 laboratories, including 3,240 individuals, 2,140 of whom had substance dependence on one of five substances: alcohol, nicotine, cocaine, methamphetamine, or cannabis. Subcortical volume and cortical thickness in regions defined by FreeSurfer were compared with nondependent control subjects when all sampled substance categories were combined, as well as separately, while controlling for age, sex, imaging site, and total intracranial volume. Because of extensive associations with alcohol dependence, a secondary contrast was also performed for dependence on all substances except alcohol. An optimized split-half strategy was used to assess the reliability of the findings.
Results:
Lower volume or thickness was observed in many brain regions in individuals with substance dependence. The greatest effects were associated with alcohol use disorder. A set of affected regions related to dependence in general, regardless of the substance, included the insula and the medial orbitofrontal cortex. Furthermore, a support vector machine multivariate classification of regional brain volumes successfully classified individuals with substance dependence on alcohol or nicotine relative to nondependent control subjects.
Conclusions:
The results indicate that dependence on a range of different substances shares a common neural substrate and that differential patterns of regional volume could serve as useful biomarkers of dependence on alcohol and nicotine.
The social and economic costs associated with problematic use of drugs and alcohol place an enormous burden on the individual and society (1–5). In the United States alone, the National Institute on Drug Abuse estimates that the costs associated with problematic substance use—including medical care, law enforcement, and lost productivity—exceed $700 billion per year (6). Substance dependence is characterized by a loss of control over drug and alcohol taking behavior, which contributes to high relapse rates (7–10). The therapeutic landscape would be radically altered by the identification of a set of biomarkers that could be used to estimate risk at various stages of the disorder—for example, the risk of transition from occasional to problematic patterns of use or risk of relapse after treatment—and to prescribe the most appropriate treatments on the basis of the individual patient’s specific functional vulnerabilities (11, 12).
It remains to be determined whether regional differences in brain volume measured by MRI can provide clinically useful biomarkers of substance dependence. Although brain volumetric studies have routinely observed lower gray matter volume in individuals with substance dependence compared with healthy control subjects who do not have a substance dependence, the brain regions associated with dependence on a specific substance have not been consistent across studies (13–15). Since volumetric studies have tended to focus on one substance at a time, it is also not clear from this literature whether a shared set of brain areas will exhibit altered volume in all individuals with substance dependence regardless of the substance used. Human twin studies suggest that genetic vulnerability to substance dependence is accounted for principally by a shared set of variations regardless of the substance used, with proportionately smaller substance-specific effects (16). On the basis of preclinical research and data from other imaging modalities, several candidate brain regions have been proposed as playing a central role in substance dependence, including the striatum, the insula, and parts of the frontal cortex (reviewed in references 17–19).
The authors of the present study joined to form an international working group within the framework of the Enhancing Neuro-Imaging Genetics Through Meta-Analysis (ENIGMA) project (20, 21) to overcome issues related to low statistical power in individual neuroimaging studies. This first project of the Addiction Working Group has pooled data from 23 laboratories in 14 countries and represents the largest study of brain volumetric data in substance dependence research to date. The objective was to identify general and substance-specific associations between dependence and regional brain volumes. The large sample size facilitated the adoption of a rigorous cross-validation method to address the widespread failure to replicate neuroimaging results, which has been noted in several recent influential reports (22, 23). In addition, a support vector machine classifier was used to explore patterns of regional brain volume that could potentially serve as disease biomarkers.
Method
Behavioral Phenotyping
All procedures were performed in accordance with the Declaration of Helsinki. Data sets from the working group were selected that assessed individuals for dependence on one of five substances: alcohol, nicotine, cocaine, methamphetamine, and cannabis. A variety of diagnostic instruments were used to assess substance dependence (see Table S1 in the online supplement). Case and control data were gathered from 23 laboratories on 3,240 individuals, of whom 2,140 were diagnosed with current dependence on at least one of the five substances of interest. Individuals were excluded if they had a lifetime history of neurological disease, a current DSM-IV axis I diagnosis other than depressive and anxiety disorders, or any contraindication for MRI. Control subjects may have used addictive substances recreationally but were not diagnosed as dependent. Summary demographic statistics (sex distribution and mean age) on participants whose data passed the quality control steps described below are provided in Table 1. Site-specific summaries are provided in Table S1 in the online supplement.
Group or Dependence Subgroup | Total N | Female | Age (years) | ||
---|---|---|---|---|---|
N | % | Mean | SD | ||
All Groups | |||||
Control | 1,100 | 449* | 40.8 | 28.5* | 9.9 |
Case | 2,140 | 731 | 34.2 | 33.3 | 10.6 |
Alcohol | |||||
Control | 292 | 99 | 33.9 | 31.3* | 10.2 |
Case | 898 | 291 | 32.4 | 34.7 | 10.7 |
Nicotine | |||||
Control | 290 | 155* | 53.4 | 26.1* | 8.0 |
Case | 602 | 250 | 41.5 | 30.8 | 9.8 |
Cocaine | |||||
Control | 99 | 39* | 39.4 | 36.0* | 10.3 |
Case | 227 | 54 | 23.8 | 40.2 | 7.7 |
Methamphetamine | |||||
Control | 173 | 71 | 41.0 | 31.7 | 9.3 |
Case | 228 | 78 | 34.2 | 32.9 | 10.0 |
Cannabis | |||||
Control | 246 | 85 | 34.6 | 22.7* | 7.5 |
Case | 185 | 58 | 31.4 | 26.5 | 10.0 |
Sex Distribution and Mean Age of Case and Control Subjects, by Dependence Subgroup, in a Mega-Analysis of Gray Matter Volume in Substance Dependence
Preparation of Structural MRI Data
Structural T1-weighted MRI brain scans were acquired from all participants. Scanner and acquisition details at each site are provided in Table S1 in the online supplement. Data were prepared in FreeSurfer (version 5.3), a fully automated MRI processing pipeline that identifies seven bilateral subcortical and 34 bilateral cortical regions of interest (24, 25). A majority of the data sets were prepared using CBRAIN, a network of high-performance computing facilities in Canada (26). The volume of subcortical regions of interest and mean thickness of cortical regions of interest served as the dependent measure in all analyses. The use of FreeSurfer in multisite analyses has been validated in previous ENIGMA studies (27–30) that established a standardized protocol of quality control procedures performed at each site (http://enigma.ini.usc.edu/protocols/imaging-protocols/). This includes detection of outliers and visual inspection of all data in a series of standard planes (for more details, see the Supplemental Methods section in the online supplement). An additional level of visual inspection was performed centrally at the University of Vermont on a randomly selected subsample of participants to ensure uniformity of quality control across sites.
Linear Mixed-Effects Models With Cross-Validation
Differences in region-of-interest thickness or volume between substance-dependent participants and nondependent control subjects were assessed in each region of interest with two linear mixed-effects models, using SPSS Statistics for Windows, version 21.0 (IBM, Armonk, N.Y.). The linear mixed-effects model effectively accounts for site effects, including sites that did not collect data on nondependent control subjects (31). In model 1, substance-dependent individuals were treated as one group regardless of the substance used; individuals dependent on any of the five substances of interest were coded as “dependent” and control subjects as “nondependent.” Model 1 permitted inclusion of individuals who were dependent on more than one substance. In model 2, dependence on the five substances was coded as individual categories in a single fixed factor: individuals were coded as belonging to one and only one of six categories: “nondependent” or dependent on “alcohol,” “nicotine,” “cocaine,” “methamphetamine,” or “cannabis.” Model 2 did not permit inclusion of individuals who were dependent on more than one substance. In both models, MRI site was entered as a random factor, and sex, age, and total intracranial volume were included as covariates. Further analyses were performed to disconfirm the existence of a site-by-diagnosis interaction (see the Supplemental Methods section in the online supplement).
The replicability of neuroimaging results has recently been brought into question (22, 23). The large sample size of the present study facilitated the adoption of an optimized split-half strategy to verify the reliability of effects. The data were split into two halves (a discovery data set and a replication data set) with statistically matched stratification for age, sex, and intracranial volume within each site and dependence status. Since each region of interest was analyzed separately, a false discovery rate method (the Benjamini-Hochberg procedure) was used to control for multiple comparisons on the first half of the data (the discovery data set). Associations discovered in the first half of the data are reported here as significant only if they were replicated in the second half of the data (the replication data set), that is, if the sign of the difference in means was the same and the null hypothesis had a probability <0.05.
General Versus Substance-Specific Dependence Effects
Model 2 permitted a comparison of the estimated marginal mean region-of-interest volume or thickness between nondependent control subjects and participants dependent on each substance. Significance was defined as in model 1. The large impact of alcohol dependence on the data (see the Results section) influenced the decision to examine whether dependence on any of the substances other than alcohol was related to differences in region-of-interest volume or thickness compared with nondependent control subjects. This was assessed with a secondary linear contrast within model 2 that grouped dependence on nicotine, cocaine, methamphetamine, and cannabis (but not alcohol) in a comparison with nondependent control subjects.
Past-30-Day Use
Linear mixed-effects models were used to determine whether past-30-day nicotine or alcohol use was related to the volume or thickness of regions of interest identified by model 1 or 2 (i.e., those brain regions listed in Tables 2 and 3). (See the online supplement for more details.)
Left | Right | |||||
---|---|---|---|---|---|---|
p | Cohen’s d | p | Cohen’s d | |||
Region and Comparison | 1st Half of Data | 2nd Half of Data | Both Halves of Data | 1st Half of Data | 2nd Half of Data | Both Halves of Data |
Subcortical volume | ||||||
Substance-dependent versus nondependent controls | ||||||
Amygdala | 0.0002 | 0.0039 | –0.055 | 0.0011 | 0.0037 | –0.041 |
Hippocampus | <0.0001 | <0.0001 | –0.087 | <0.0001 | <0.0001 | –0.081 |
Nucleus accumbens | 0.0068 | 0.0214 | –0.025 | |||
Cortical thickness | ||||||
Substance-dependent versus nondependent controls | ||||||
Caudal middle frontal gyrus | <0.0001 | 0.0370 | –0.038 | |||
Fusiform gyrus | <0.0001 | 0.0231 | –0.036 | |||
Inferior parietal cortex | <0.0001 | 0.0298 | –0.026 | |||
Insula | <0.0001 | 0.0002 | –0.056 | 0.0007 | 0.0003 | –0.042 |
Isthmus of cingulate gyrus | <0.0001 | 0.0447 | -0.035 | |||
Medial orbitofrontal cortex | 0.0095 | 0.0491 | –0.029 | |||
Middle temporal gyrus | 0.0910 | 0.0065 | –0.030 | 0.0040 | 0.0474 | –0.026 |
Paracentral lobule | 0.0019 | 0.0015 | –0.031 | 0.0421 | 0.0056 | –0.024 |
Precentral gyrus | <0.0001 | 0.0025 | –0.039 | <0.0001 | 0.0042 | –0.042 |
Precuneus | <0.0016 | 0.0425 | –0.023 | |||
Superior parietal cortex | 0.0082 | 0.0472 | –0.022 | |||
Supramarginal gyrus | 0.0049 | 0.0131 | –0.027 | 0.0046 | 0.0319 | –0.026 |
Left | Right | |||||
---|---|---|---|---|---|---|
p | Cohen’s d | p | Cohen’s d | |||
Region and Comparison | 1st Half of Data | 2nd Half of Data | Both Halves of Data | 1st Half of Data | 2nd Half of Data | Both Halves of Data |
Subcortical volume | ||||||
Alcohol-dependent versus nondependent controls | ||||||
Amygdala | <0.0001 | 0.0021 | –0.107 | <0.0001 | 0.0003 | –0.111 |
Globus pallidus | 0.0274 | 0.0005 | –0.075 | |||
Hippocampus | <0.0001 | <0.0001 | –0.196 | <0.0001 | <0.0001 | –0.180 |
Nucleus accumbens | 0.0159 | 0.0013 | –0.048 | <0.0001 | <0.0001 | –0.088 |
Putamen | 0.0006 | <0.0001 | –0.098 | 0.0001 | 0.0014 | –0.080 |
Thalamus | 0.0149 | 0.0002 | –0.098 | |||
Cortical thickness | ||||||
Alcohol-dependent versus nondependent controls | ||||||
Caudal middle frontal gyrus | 0.0006 | 0.0219 | –0.062 | 0.0264 | 0.0298 | –0.054 |
Fusiform gyrus | 0.0017 | 0.0002 | –0.072 | <0.0001 | <0.0001 | –0.094 |
Inferior temporal gyrus | 0.0021 | 0.0146 | –0.056 | 0.0148 | 0.0214 | –0.047 |
Insula | 0.0023 | <0.0001 | –0.087 | |||
Isthmus of cingulate gyrus | 0.0009 | 0.0005 | –0.078 | |||
Lateral occipital cortex | 0.0013 | 0.0211 | –0.042 | |||
Lateral orbitofrontal cortex | 0.0322 | 0.0021 | –0.061 | |||
Medial orbitofrontal cortex | 0.0432 | 0.0197 | –0.060 | |||
Parahippocampal gyrus | 0.0281 | 0.0265 | –0.076 | |||
Paracentral lobule | 0.0002 | 0.0001 | –0.074 | 0.0053 | 0.0003 | –0.062 |
Posterior cingulate gyrus | <0.0001 | <0.0001 | –0.091 | 0.0004 | <0.0001 | –0.085 |
Precentral gyrus | 0.0091 | 0.0007 | –0.063 | 0.0008 | 0.0003 | –0.079 |
Precuneus | 0.0008 | 0.0003 | –0.062 | 0.0039 | 0.0002 | –0.061 |
Rostral anterior cingulate cortex | 0.0381 | 0.0090 | –0.082 | |||
Superior frontal gyrus | <0.0001 | 0.0030 | –0.071 | 0.0003 | 0.0060 | –0.071 |
Superior parietal cortex | 0.0198 | 0.0272 | –0.040 | |||
Superior temporal gyrus | 0.0239 | 0.0353 | –0.062 | |||
Supramarginal gyrus | 0.0493 | 0.0168 | –0.044 | |||
Temporal pole | 0.0518 | 0.0464 | –0.063 | |||
Cocaine-dependent versus nondependent controls | ||||||
Supramarginal gyrus | 0.0177 | 0.0491 | –0.047 | |||
Nicotine-, cocaine-, methamphetamine-, and cannabis-dependent versus nondependent controls | ||||||
Inferior parietal cortex | 0.0011 | 0.4630 | –0.029 | |||
Insula | 0.0057 | 0.0300 | –0.041 | 0.0303 | 0.0274 | –0.033 |
Support Vector Machine Classification
Support vector machine classification was implemented in MATLAB (MathWorks, Natick, Mass.) with a radial basis function kernel, tuned by parameter sweep in a 10-fold inner loop nested within an optimized split-half cross-validation (32) (for details, see the Supplemental Methods section in the online supplement). The radial basis function kernel facilitates the inclusion of nonlinear relationships in the classifier. In other words, the support vector machine can detect informative patterns in the data that may not be identified by traditional linear analyses such as models 1 and 2. To mitigate site, sex, age, and intracranial volume effects, region-of-interest data were residualized prior to classification. Five studies without control participants were excluded. Area under the receiver operating characteristic curve and corresponding p values based on equivalence with the Mann-Whitney U test were calculated to estimate generalizable classifier performance on the independent half of the data for each of two train–test scenarios (i.e., train on the first half, test on the second, and vice versa). A greater area under the receiver operating characteristic curve, which plots true positive rate against false positive rate, indicates a better separation of the substance-dependent and nondependent groups. The significance threshold for area under the curve was defined as a p value of 0.05 in both classification scenarios. The top 20 features of each classification were determined by the greatest change in cost function resulting from their individual removal from the classification (33).
Results
Basic demographic information (sex distribution and mean age) is provided in Table 1 and by site in Table S1 in the online supplement.
Model 1: Dependent Versus Nondependent Subjects
Subcortical volume in dependent individuals was significantly lower in the left and right hippocampus, the left and right amygdala, and the right nucleus accumbens (Table 2). Lower cortical thickness was observed in several areas, including the left and right insula, precentral gyrus, and supramarginal gyrus and the right medial orbitofrontal cortex. See Table 2 for a complete list and Supplemental Table S2 in the online supplement for an at-a-glance summary.
Model 2: Substance Dependence Groups Compared Separately to Nondependent Control Subjects
All subcortical regions of interest identified in model 1 plus the right thalamus, the left and right putamen, the right globus pallidus, and the left nucleus accumbens had significantly lower volume in model 2 when alcohol-dependent participants were compared with nondependent control subjects. In addition, alcohol-dependent participants exhibited lower average thickness in 27 cortical regions of interest (Table 3, Figure 1). Cocaine dependence was associated with lower cortical thickness in only one brain region (see Table 3, Figure 1). No cross-validated differences in regional volume or thickness were significant for dependence on nicotine, methamphetamine, or cannabis on their own. Since most effects were related to alcohol dependence, a secondary linear contrast was performed to explore the effect of removing alcohol from the analysis. The contrast compared participants dependent on any substance except alcohol against nondependent control subjects. It revealed that the left inferior parietal cortex and the insula bilaterally were significantly thinner in dependent individuals (see Table 3).
Substance-Specific Versus Shared Substance-General Effects
Three distinct patterns of results emerged, which are illustrated in Figure 2.
Pattern 1 (substance specific).
In most regions of interest where a significant difference was observed, the effect was demonstrated in model 2 to be related specifically to dependence on alcohol alone (27 regions of interest)—for example, the right nucleus accumbens (Figure 2)—or to both alcohol and cocaine—the right supramarginal gyrus (one region of interest) (see Figure 1, Tables 2 and 3).
Pattern 2 (substance general).
Six cortical regions of interest (e.g., the left supramarginal gyrus and the right medial orbitofrontal cortex) were associated with dependence in model 1 but were not significantly thinner in any one particular substance group relative to nondependent control subjects in model 2 (see Tables 2 and 3, Figures 1 and 2).
Pattern 3 (substance general).
Three cortical regions of interest (the left inferior parietal cortex and the right and left insula) were significantly thinner when all dependent groups were compared with control subjects (model 1) and when all dependent groups except alcohol were contrasted against control subjects (model 2). In addition, the left insula was significantly thinner in the alcohol-dependent group alone relative to control subjects (Tables 2 and 3, Figures 1 and 2).
Past-30-Day Use
The volume of several subcortical regions of interest were negatively associated with past-30-day use of alcohol after a false discovery rate correction for multiple comparisons: the left and right amygdala and nucleus accumbens, the right hippocampus, and the left globus pallidus. No brain regions were related to past-30-day nicotine use.
Support Vector Machine
The support vector machine produced a significant classification of alcohol- and nicotine-dependent individuals relative to nondependent control subjects (Figure 3) in both halves of the data (p<0.05). The classification of cocaine-dependent individuals approached significance. The top 20 structural predictors distinguishing dependence on each substance from nondependent control subjects in each classification are listed in Table S3 in the online supplement.
Discussion
Subcortical volume or cortical thickness was significantly lower on average in substance-dependent individuals compared with nondependent control subjects across widespread parts of the brain (i.e., 22 distinct regions of interest out of a total of 82) (see Table 2; see also Table S2 in the online supplement). Some of these differences were substance specific, and others appear to constitute a shared neural substrate associated with dependence regardless of the substance used (see Table 3 and Figure 1). A majority of the identified regions of interest were smaller or thinner specifically in the brains of alcohol-dependent individuals (e.g., the left and right posterior cingulate and superior frontal cortex). A more limited set of seven regions with lower cortical thickness across substance dependence groups included the left and right insula, the left inferior parietal cortex, the right medial orbitofrontal cortex, the left and right middle temporal gyrus, and the left supramarginal gyrus. No region of interest was significantly larger or thicker in substance-dependent individuals relative to control subjects. An unexpected finding of the study was the absence of substance-specific linear effects on brain volume related to nicotine, methamphetamine, or cannabis dependence despite the collection of large pooled samples. Also, the successful classification of individuals dependent on nicotine, alcohol, or cocaine using the support vector machine approach suggests that the development of clinically useful neuroimaging biomarkers of substance dependence may be more productive if based on broader patterns of brain function or structure rather than differences in unique brain regions considered alone.
The set of brain regions identified with substance dependence in general is supported by previous evidence. The insula performs a central role in the perception of the internal state of the body (34). Disruption of the insula could alter regulation of the intense positive and negative bodily states associated with drug taking and withdrawal, biasing the individual toward relapse as a maladaptive response to anticipated challenges to physiological homeostasis (35). It has been reported that smokers who have suffered brain damage involving the insula have subsequently lost the urge to smoke (36). The parietal cortex has been associated with attention and working memory (37, 38). Disruption of these processes could interfere with self-awareness about a substance use problem and the management of stressful situations. The medial orbitofrontal region of interest defined by FreeSurfer (also known as the ventromedial prefrontal cortex) encodes the subjective value of future rewards during decision making (39). Lesions of this region produce disadvantageous choices on gambling tasks that model real-life decisions (40). Altered neural activity in the insula and the medial orbital and parietal cortex has frequently been linked to substance dependence and may predict greater craving and risk of relapse (41–44). The present results support the idea that substance dependence is mediated by a shared set of mechanisms across substance groups. Indeed, twin studies suggest that vulnerability to substance dependence is accounted for principally by a shared set of genetic variations regardless of the substance used, with proportionately smaller substance-specific effects (16).
Although subtle in magnitude, the wide spatial distribution of alcohol-specific effects is a striking finding of the study. Alcohol consumption enjoys greater cultural acceptance in the countries from which the data for this study were sampled relative to the other substances examined (45). Alcohol is legal to buy and consume, and widely publicized government-sanctioned guidelines exist for “safe” low-dose use of alcohol. This tolerance of alcohol-related health risks is unlike the cultural views toward any of the other substances investigated here, whose use even in small amounts is discouraged (45). It should be noted that lifetime exposure to each substance could not be uniformly assessed in the data sets used here. As a consequence, the scope of the alcohol dependence effects may in part be related to greater absolute consumption of alcohol relative to the other substances. It was possible to assess past-30-day use of nicotine and alcohol, a limited proxy of level of exposure, in a sizable minority of the data sets. Several subcortical regions of interest, such as the amygdala and the nucleus accumbens, were significantly smaller in individuals who reported the highest numbers of alcoholic drinks consumed in the past 30 days, consistent with the notion that greater exposure could be responsible for the magnitude of the observed alcohol effects. Further studies will be required to clarify whether the greater number of observed alcohol-specific effects relative to the other substances is related to differences in toxicity or total exposure.
It is also notable that, besides the seven brain regions associated with dependence in general, there were no drug-specific effects for dependence on nicotine, methamphetamine, and cannabis. Although cross-validation demonstrated that the volumetric differences observed were reliable, the effect sizes were uniformly small (see Tables 2 and 3). This suggests that the lack of consistency in the literature (13–15) may be related to the insufficient power of most studies to detect true effects. Other imaging modalities, such as task-based functional MRI (41–44) and higher-resolution structural imaging, may be required to detect reliable substance-specific nicotine, methamphetamine, or cannabis effects if they exist. It is also possible that substance dependence has multiple, heterogeneous interactions with brain volume that are not well assessed by simple linear analyses. Evidence for this is provided by the support vector machine classification.
The support vector machine classification found that the pattern of regional volume differences could be used successfully to distinguish between nondependent control subjects and individuals dependent on alcohol and nicotine. The transformation of the data with a radial basis function kernel prior to classification facilitated the detection of nonlinear patterns that cannot be detected by models 1 and 2. Additionally, the support vector machine can identify a multivariate pattern of effects across numerous regions of interest, each of which, in isolation, may not pass statistical threshold. Thus, the support vector machine detected useful information in the pattern of results that was not apparent from the linear analysis. The significant classifications suggest that the overall pattern of volumetric effects may contain useful clinical information that would not be apparent if only traditional univariate linear analyses were performed. While influential features in the classification partly overlapped with the regions of interest identified by the univariate analyses—for example, brain regions associated with alcohol dependence, such as the hippocampus and amygdala—additional regions not identified by the linear mixed effects analyses (i.e., model 1 and model 2) were also involved (see Table S3 in the online supplement). Future efforts of the Addiction Working Group will include the incorporation of other imaging modalities with which it may be possible to distinguish individuals with dependence on additional substances, such as methamphetamine and cannabis, from nondependent control subjects. It would also be clinically useful to examine whether the support vector machine classifications developed in this study offer an index of the strength of substance dependence in individuals who go on to recover or relapse. It is worth noting that current blood and urine tests do not identify dependence, as the machine learning classifier in the present study does, but rather detect, and to an extent quantify, recent substance use. While the present findings are preliminary and the support vector machine classifications should be tested on other independent samples, if brain volume is confirmed as a viable biomarker of dependence, or of biological risk of dependence, it could be used to plan how prevention and treatment resources are allocated to individual patients as well as, potentially, to track intervention success. A structural MRI scan in combination with other factors known to be related to substance use problems (e.g., change in employment or marital status, health issues) could be used to assess risk of transition to problematic patterns of use or to quantify the current degree of dependence, which would influence the intervention strategy.
Several factors limit the interpretation of the study findings. Different diagnostic instruments were used to assess substance dependence (see Table S1 in the online supplement). Although the validity of each of these instruments has been well established, variation between instruments could add noise to the measured behavioral phenotype. This, however, could be an advantage because the extrapolation of significant findings to the general population is also likely to be more robust by virtue of generalizing across different methods of assessment. The absence of nutrition and education information, which are potential confounders, also limits the interpretability of the results. A perennial concern with multisite studies is variation attributable to different scanners and acquisition protocols. This issue was mitigated by using a standard data extraction protocol developed by the ENIGMA project that has been validated in previous multisite reports (20, 28–30) and by the formal consideration of potential site differences in all statistical analyses. As discussed above, the degree of exposure to the various substances was not characterized uniformly across studies, which limits, for instance, the interpretation of the widespread alcohol effects and whether alcohol represents a greater source of toxicity than the other substances examined. It should be emphasized, however, that this study examined brain volumetric associations with dependence and not with total lifetime substance use. A beneficial outcome of this first study of the Addiction Working Group will be to raise awareness of the data needed to estimate the relation between brain volume and total exposure and, more generally, of the utility of uniform phenotypic data for data pooling. Greater consideration of how data may be used in international collaborations may influence the collection of data in future studies, which will increase their impact beyond their primary research focus. The PhenX Toolkit (https://www.phenxtoolkit.org/), for example, provides an extensive catalog of standardized measures expressly intended to facilitate secondary cross-study comparisons. Finally, co-occurring substance use limits the interpretation of the findings. Pervasive recreational substance use is a general issue for all studies of human substance dependence. For example, it is likely epidemiologically that a methamphetamine user will be exposed to alcohol. Methamphetamine users who do not use any other addictive substance would be an unusual group who, in practice, would be difficult to identify but, more importantly, would not be characteristic of the real-world population of methamphetamine users—that is, there would be a selection bias. Unlike studies in animal models, it is not possible to randomly assign humans to groups with restricted exposure to one substance alone. The typical strategy, which was used in the data sets included in this study, is to screen subjects for dependence on other substances but not to exclude for nondependent use of other substances.
The field of neuroimaging faces a crisis of relevance if published studies cannot be replicated, as noted in a series of reviews (22, 23). The authors of the present study joined to form a working group within the preexisting framework of the ENIGMA project to assemble a sufficiently large sample to overcome issues related to low statistical power that affect most individual neuroimaging studies. Using a rigorous cross-validation method, several brain regions were found to have a reliable association with substance dependence, including a shared set of regions across substances, such as the insula and the medial orbitofrontal cortex. Although the univariate analyses failed to identify linear effects in relation to dependence on nicotine, methamphetamine, and cannabis specifically, a machine learning algorithm, which was also able to detect nonlinear patterns in the data, successfully classified individuals dependent on alcohol or nicotine relative to nondependent control subjects. This suggests that the overall pattern of volumetric effects may contain more useful information with regard to the development of a neuroimaging biomarker of substance dependence than is revealed by the magnitude of single brain regions examined in isolation.
1 : Global burden of disease attributable to mental and substance use disorders: findings from the Global Burden of Disease Study 2010. Lancet 2013; 382:1575–1586Crossref, Medline, Google Scholar
2 : The Health Consequences of Smoking: 50 Years of Progress: A Report of the Surgeon General. Atlanta, Centers for Disease Control and Prevention, National Center for Chronic Disease Prevention and Health Promotion, Office on Smoking and Health, 2014Google Scholar
3 : WHO Report on the Global Tobacco Epidemic 2011. Geneva, WHO, 2011Google Scholar
4 : 2010 national and state costs of excessive alcohol consumption. Am J Prev Med 2015; 49:e73–e79Crossref, Medline, Google Scholar
5 : National Drug Threat Assessment. Washington, DC, US Department of Justice, 2011Google Scholar
6 National Institute on Drug Abuse: Trends and Statistics. 2016. https://www.drugabuse.gov/related-topics/trends-statisticsGoogle Scholar
7 : The effectiveness of NHS smoking cessation services: a systematic review. J Public Health (Oxf) 2010; 32:71–82Crossref, Medline, Google Scholar
8 : Nicotine dependence: biology, behavior, and treatment. Annu Rev Med 2009; 60:247–260Crossref, Medline, Google Scholar
9 : Drug dependence, a chronic medical illness: implications for treatment, insurance, and outcomes evaluation. JAMA 2000; 284:1689–1695Crossref, Medline, Google Scholar
10 : Rates and predictors of relapse after natural and treated remission from alcohol use disorders. Addiction 2006; 101:212–222Crossref, Medline, Google Scholar
11 : Neuroimaging and biomarkers in addiction treatment. Curr Psychiatry Rep 2014; 16:513Crossref, Medline, Google Scholar
12 : Pragmatism instead of mechanism: a call for impactful biological psychiatry. JAMA Psychiatry 2015; 72:631–632Crossref, Medline, Google Scholar
13 : Are there volumetric brain differences associated with the use of cocaine and amphetamine-type stimulants? Neurosci Biobehav Rev 2013; 37:300–316Crossref, Medline, Google Scholar
14 : Alcohol and the human brain: a systematic review of different neuroimaging methods. Alcohol Clin Exp Res 2011; 35:1771–1793Crossref, Medline, Google Scholar
15 : Neuroimaging in cannabis use: a systematic review of the literature. Psychol Med 2010; 40:383–398Crossref, Medline, Google Scholar
16 : Co-occurrence of abuse of different drugs in men: the role of drug-specific and shared vulnerabilities. Arch Gen Psychiatry 1998; 55:967–972Crossref, Medline, Google Scholar
17 : Neurocircuitry of addiction. Neuropsychopharmacology 2010; 35:217–238Crossref, Medline, Google Scholar
18 : The neurocircuitry of impaired insight in drug addiction. Trends Cogn Sci 2009; 13:372–380Crossref, Medline, Google Scholar
19 : Neural systems of reinforcement for drug addiction: from actions to habits to compulsion. Nat Neurosci 2005; 8:1481–1489Crossref, Medline, Google Scholar
20 : The ENIGMA Consortium: large-scale collaborative analyses of neuroimaging and genetic data. Brain Imaging Behav 2014; 8:153–182Crossref, Medline, Google Scholar
21 : Genetic imaging consortium for addiction medicine: from neuroimaging to genes. Prog Brain Res 2016; 224:203–223Crossref, Medline, Google Scholar
22 : Excess significance bias in the literature on brain volume abnormalities. Arch Gen Psychiatry 2011; 68:773–780Crossref, Medline, Google Scholar
23 : Power failure: why small sample size undermines the reliability of neuroscience. Nat Rev Neurosci 2013; 14:365–376Crossref, Medline, Google Scholar
24 : Cortical surface-based analysis, I: segmentation and surface reconstruction. Neuroimage 1999; 9:179–194Crossref, Medline, Google Scholar
25 : An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. Neuroimage 2006; 31:968–980Crossref, Medline, Google Scholar
26 : CBRAIN: a web-based distributed computing platform for collaborative neuroimaging research. Front Neuroinform 2014; 8:54Crossref, Medline, Google Scholar
27 : Identification of common variants associated with human hippocampal and intracranial volumes. Nat Genet 2012; 44:552–561Crossref, Medline, Google Scholar
28 : Distinct subcortical volume alterations in pediatric and adult OCD: a worldwide meta- and mega-analysis. Am J Psychiatry 2017; 174:60–69Link, Google Scholar
29 : Subcortical brain volume abnormalities in 2028 individuals with schizophrenia and 2540 healthy controls via the ENIGMA consortium. Mol Psychiatry 2016; 21:585Crossref, Medline, Google Scholar
30 : Common genetic variants influence human subcortical brain structures. Nature 2015; 520:224–229Crossref, Medline, Google Scholar
31 : A study of clustered data and approaches to its analysis. J Neurosci 2010; 30:10601–10608Crossref, Medline, Google Scholar
32 : Advanced support vector machines and kernel methods. Neurocomputing 2003; 55:5–20Crossref, Google Scholar
33 : Gene selection for cancer classification using support vector machines. Mach Learn 2002; 46:389–422Crossref, Google Scholar
34 : How do you feel? Interoception: the sense of the physiological condition of the body. Nat Rev Neurosci 2002; 3:655–666Crossref, Medline, Google Scholar
35 : The role of interoception and alliesthesia in addiction. Pharmacol Biochem Behav 2009; 94:1–7Crossref, Medline, Google Scholar
36 : Damage to the insula disrupts addiction to cigarette smoking. Science 2007; 315:531–534Crossref, Medline, Google Scholar
37 : Superior parietal cortex is critical for the manipulation of information in working memory. J Neurosci 2009; 29:14980–14986Crossref, Medline, Google Scholar
38 : The frontoparietal attention network of the human brain: action, saliency, and a priority map of the environment. Neuroscientist 2012; 18:502–515Crossref, Medline, Google Scholar
39 : Greater preference consistency during the willingness-to-pay task is related to higher resting state connectivity between the ventromedial prefrontal cortex and the ventral striatum. Brain Imaging Behav 2016; 10:730–738Crossref, Medline, Google Scholar
40 : Characterization of the decision-making deficit of patients with ventromedial prefrontal cortex lesions. Brain 2000; 123:2189–2202Crossref, Medline, Google Scholar
41 : Disrupted ventromedial prefrontal function, alcohol craving, and subsequent relapse risk. JAMA Psychiatry 2013; 70:727–739Crossref, Medline, Google Scholar
42 : Pretreatment brain activation during Stroop task is associated with outcomes in cocaine-dependent patients. Biol Psychiatry 2008; 64:998–1004Crossref, Medline, Google Scholar
43 : An exploratory pilot study of the relationship between neural correlates of cognitive control and reduction in cigarette use among treatment-seeking adolescent smokers. Psychol Addict Behav 2013; 27:526–532Crossref, Medline, Google Scholar
44 : Neurobiological impact of nicotinic acetylcholine receptor agonists: an activation likelihood estimation meta-analysis of pharmacologic neuroimaging studies. Biol Psychiatry 2015; 78:711–720Crossref, Medline, Google Scholar
45 : Why does society accept a higher risk for alcohol than for other voluntary or involuntary risks? BMC Med 2014; 12:189Crossref, Medline, Google Scholar