Introduction
Sensation seeking increases during adolescence (Kandel & Logan, Reference Kandel and Logan1984; Steinberg et al. Reference Steinberg, Albert, Cauffman, Banich, Graham and Woolard2008), often at the expense of safer choices. Some risk taking, for example, practising difficult sporting maneuvers or applying to highly ranked schools or jobs, is beneficial to growth and survival. Other risks taken by youth, however, are associated with deleterious behaviors, such as substance use and substance use disorders. The propensity for risky behaviors, such as substance use, in youth may be related to reward circuitry development, specifically, reduced ventral striatal function and volume (Schneider et al. Reference Schneider, Peters, Bromberg, Brassen, Miedl, Banaschewski, Barker, Conrod, Flor, Garavan, Heinz, Ittermann, Lathrop, Loth, Mann, Martinot, Nees, Paus, Rietschel, Robbins, Smolka, Spanagel, Strohle, Struve, Schumann and Buchel2012), and a delay in the development of prefrontal cortical regions implicated in cognitive control alongside the emergence of increased dopaminergic activity in subcortical regions during puberty (Steinberg et al. Reference Steinberg, Albert, Cauffman, Banich, Graham and Woolard2008). Reward circuitry comprises a widespread neural network, including the ventral striatum (VS), amygdala and insula, and specific prefrontal cortical regions: the ventrolateral prefrontal cortex [VLPFC; Brodmann area (BA) 47], the dorsal anterior cingulate cortex (dACC; BA24/32), and the medial and middle prefrontal cortex (mPFC; BA10). Reward circuitry-related activity, along with sensation-seeking personality traits and risk-taking behaviors, characterized early-onset drinking (Nees et al. Reference Nees, Tzschoppe, Patrick, Vollstadt-Klein, Steiner, Poustka, Banaschewski, Barker, Buchel, Conrod, Garavan, Heinz, Gallinat, Lathrop, Mann, Artiges, Paus, Poline, Robbins, Rietschel, Smolka, Spanagel, Struve, Loth, Schumann and Flor2012). In addition, on a naturalistic risk-taking task, activity in the bilateral insula, parietal, orbitofrontal, and motor cortices, as well as the left anterior cingulate cortex, together were able to discriminate between making a risky or safe choice on the next trial with 67% accuracy (Helfinstein et al. Reference Helfinstein, Schonberg, Congdon, Karlsgodt, Mumford, Sabb, Cannon, London, Bilder and Poldrack2014). Additionally, in adolescence, cortical maturation often corresponds with substance use onset (Shaw et al. Reference Shaw, Kabani, Lerch, Eckstrand, Lenroot, Gogtay, Greenstein, Clasen, Evans, Rapoport, Giedd and Wise2008). Animal studies reported differential changes in cortical thickness in adolescent animals exposed to substances (Vetreno et al. Reference Vetreno, Yaxley, Paniagua, Johnson and Crews2016), while adolescent marijuana users showed reduced cortical thicknesses relative to non-users (Lopez-Larson et al. Reference Lopez-Larson, Bogorodzki, Rogowska, McGlade, King, Terry and Yurgelun-Todd2011). The extent to which measures of reward circuitry function and structure in youth predict future substance use, however, remains to be determined. Identifying in youth such predictors, alongside clinical and demographic predictors, would not only provide objective neural markers to identify risk of future substance use disorders, but would also provide targets to ultimately guide early intervention, treatment choice, and novel treatment developments.
Predicting clinical outcome from neuroimaging measures is a burgeoning field of research (Berkman & Falk, Reference Berkman and Falk2013). Measures of neural structure and function predicted response to psychotherapy, cognitive–behavioral therapy, and psychotropic medications in adults and children with major depressive disorder (MDD) and anxiety disorder (AnxD) (McClure et al. Reference McClure, Adler, Monk, Cameron, Smith, Nelson, Leibenluft, Ernst and Pine2007; Forbes et al. Reference Forbes, Olino, Ryan, Birmaher, Axelson, Moyles and Dahl2010; Pizzagalli, Reference Pizzagalli2010; Masten et al. Reference Masten, Eisenberger, Borofsky, McNealy, Pfeifer and Dapretto2011; Fu et al. Reference Fu, Steiner and Costafreda2013; Hum et al. Reference Hum, Manassis and Lewis2013; Morgan et al. Reference Morgan, Olino, McMakin, Ryan and Forbes2013; Shin et al. Reference Shin, Davis, VanElzakker, Dahlgren and Dubois2013). Additionally, in youth, future positive mood and energy dysregulation was predicted by a combination of reward circuitry functional connectivity, white matter structure and clinical scores, together explaining 28% of the variance in clinical outcome (Bertocci et al. Reference Bertocci, Bebko, Versace, Fournier, Iyengar, Olino, Bonar, Almeida, Perlman, Schirda, Travis, Gill, Diwadkar, Forbes, Sunshine, Holland, Kowatch, Birmaher, Axelson, Horwitz, Frazier, Arnold, Fristad, Youngstrom, Findling and Phillips2016). The latter study in particular points to the feasibility of using a multimodal neuroimaging approach to identify markers of neural function that, in combination with clinical and demographic measures, can predict future behavioral outcomes in youth with psychiatric disorders. Large sample sizes, multimodal neuroimaging techniques, and statistical analyses that can evaluate large numbers of potential predictor variables are needed to fully examine the extent to which combinations of measures predict future outcomes in youth. LASSO (Least Absolute Shrinkage and Selection Operator) regression is one such statistical technique that has been adopted for use in genetic studies (Kohannim et al. Reference Kohannim, Hibar, Jahanshad, Stein, Hua, Toga, Jack, Weiner and Thompson2012a , Reference Kohannim, Hibar, Stein, Jahanshad, Hua, Rajagopalan, Toga, Jack, Weiner, de Zubicaray, McMahon, Hansell, Martin, Wright and Thompson b , Luo et al. Reference Luo, McShan, Kong, Schipper and Haken2015; Wang et al. Reference Wang, Xu and Liu2015; Zemmour et al. Reference Zemmour, Bertucci, Finetti, Chetrit, Birnbaum, Filleron and Boher2015), and is gaining favor in clinical research (Christensen et al. Reference Christensen, Zoetmulder, Koch, Frandsen, Arvastson, Christensen, Jennum and Sorensen2014; Yan et al. Reference Yan, Tsurumi, Que, Ryan, Bandyopadhaya, Morgan, Flaherty, Tompkins and Rahme2015; Bertocci et al. Reference Bertocci, Bebko, Versace, Fournier, Iyengar, Olino, Bonar, Almeida, Perlman, Schirda, Travis, Gill, Diwadkar, Forbes, Sunshine, Holland, Kowatch, Birmaher, Axelson, Horwitz, Frazier, Arnold, Fristad, Youngstrom, Findling and Phillips2016). This technique evaluates a large number of potential predictor variables, relative to the number of study participants, while minimizing model error and minimizing the risk of overfitting through cross-validation (CV).
The goal of the present study was to identify measures of reward circuitry function and cortical structural thickness that predicted future substance use in a large group of youth in the Longitudinal Assessment of Manic Symptoms (LAMS) study. LAMS is an ongoing multi-site study examining longitudinal relationships among the course of symptoms, outcomes, and neural mechanisms associated with different clinical trajectories in youth with symptoms characterized by behavioral and emotional dysregulation (Findling et al. Reference Findling, Youngstrom, Fristad, Birmaher, Kowatch, Arnold, Frazier, Axelson, Ryan, Demeter, Gill, Fields, Depew, Kennedy, Marsh, Rowles and Horwitz2010; Horwitz et al. Reference Horwitz, Demeter, Pagano, Youngstrom, Fristad, Arnold, Birmaher, Gill, Axelson and Kowatch2010). We hypothesized that in LAMS youth, future substance use would be predicted by increased prefrontal–cortical–striatal reward circuitry activity and reduced whole-brain cortical thickness. We also aimed to determine the proportion of future substance use predicted by neuroimaging measures, and to test the discriminatory power of identified predictors.
Method
Participants
A total of 130 youth, recruited from the LAMS1 cohort of 707 youth for whom parents were seeking psychiatric assessment and treatment, participated in the neuroimaging component of LAMS2. All 130 youth from LAMS1 entered LAMS2 with a variety of symptoms and diagnoses. Inclusion criteria for the LAMS1 cohort were: no out-patient treatment at a LAMS clinic in the last 12 months; 6–12 years of age; and without a sibling who was screened for LAMS1 (Findling et al. Reference Findling, Youngstrom, Fristad, Birmaher, Kowatch, Arnold, Frazier, Axelson, Ryan, Demeter, Gill, Fields, Depew, Kennedy, Marsh, Rowles and Horwitz2010). Families of eligible children completed the Parental General Behavior Inventory 10-Item Mania scale (PGBI-10M). Children who scored ⩾12 on this scale, and an age–sex-matched group of those who scored <12, were invited to participate in LAMS1. The 130 youth in the LAMS2 neuroimaging component were selected to include approximately equal numbers of youth: (1) with high (⩾12) v. low (<12) PGBI-10M scores; (2) who were older (⩾13 years) v. younger (⩽12 years); (3) who were male v. female (each site was age- and sex-matched for each PGBI-10M subgroup).
Exclusion criteria for participating in the LAMS2 neuroimaging component included systemic medical illnesses, neurological disorders, history of trauma with loss of consciousness, use of non-psychotropic central nervous system-affecting medications, intelligence quotient (IQ) < 70 assessed by the Wechsler Abbreviated Scale of Intelligence, positive drug and/or alcohol screen on scan day, significant visual disturbance, inability to communicate in English, autistic spectrum disorders/developmental delays, pregnancy, claustrophobia and metal in the body.
Parents/guardians and youth provided written informed consent and assent, respectively, after receiving a complete study description.
The final sample included 73 LAMS youth (age: mean = 13.91, s.d. = 2.00, range = 9.89–17.71 years; 30 females; Table 1). A total of 57 LAMS youth were excluded for behavioral data loss (n = 5), excessive movement during neuroimaging acquisition (n = 33), or cortical thickness processing problems (n = 19; inability to read the pixelated data, mislabeled parcellations, non-symmetric colors, or missing cortical regions). Included youth were older, had higher IQ, and higher socio-economic status (SES) relative to excluded youth (Table 1).
Demographic information, clinical variables, and current medication usage describing the total LAMS sample and comparing LAMS participants included v. excluded from neuroimaging
LAMS, Longitudinal Assessment of Manic Symptoms; s.d., standard deviation; IQ, intelligence quotient (Wechsler intelligence test); SES, socio-economic status; HS, high school; GED, general education development test; CALS, Child Affect Lability Scale (parent rating); PGBI-10M, Parent General Behavior Inventory 10-Item Mania Scale; KDRS, Kiddie Schedule for Affective Disorders and Schizophrenia for School-Age Children Present Episode Depression Rating Scale; KMRS, Kiddie Schedule for Affective Disorders and Schizophrenia for School-Age Children Mania Rating Scale; SCARED, Screen for Child Anxiety Related Emotional Disorders (child rating).
* Significant (p < 0.05); statistical comparison between included and excluded participants.
Reward task
Reward-related neural activity measures were acquired using a well-validated card-guessing task with a reward component (Forbes et al. Reference Forbes, Hariri, Martin, Silk, Moyles, Fisher, Brown, Ryan, Birmaher, Axelson and Dahl2009; Bebko et al. Reference Bebko, Bertocci, Fournier, Hinze, Bonar, Almeida, Perlman, Versace, Schirda, Travis, Gill, Demeter, Diwadkar, Ciuffetelli, Rodriguez, Olino, Forbes, Sunshine, Holland, Kowatch, Birmaher, Axelson, Horwitz, Arnold, Fristad, Youngstrom, Findling and Phillips2014; see online Supplementary material).
Neuroimaging data analysis
Functional magnetic resonance imaging (fMRI) data were collected on: (1) a 3 T Siemens Verio MRI scanner at Case Western Reserve University, (2) a 3 T Philips Achieva X-series MRI scanner at Cincinnati Children's Hospital, and (3) a 3 T Siemens Trio MRI scanner at the University of Pittsburgh. We preprocessed and analysed fMRI data using Statistical Parametric Mapping software (SPM8; http://www.fil.ion.ucl.ac.uk/spm). An axial three-dimensional magnetization prepared rapid gradient echo (MPRAGE) sequence [192 axial slices 1 mm thick; flip angle = 9°; field of view = 256 × 192 mm; repetition time (TR) = 2300 ms; echo time (TE) = 3.93 ms; matrix = 256 × 192] acquired T1-weighted volumetric anatomical images covering the whole brain. A reverse interleaved gradient echo planar imaging (EPI) sequence (38 axial slices 3.1 mm thick; flip angle = 90°; field of view = 205 mm; TR = 2000 ms; TE = 28 ms; matrix = 64 × 64) acquired T2-weighted blood oxygen level-dependent (BOLD) images covering the whole cerebrum and most of the cerebellum. Preprocessing involved realignment, co-registration, segmentation, normalization into a standard stereotactic space [Montreal Neurological Institute (MNI); http://www.bic.mni.mcgill.ca] and spatial smoothing using a Gaussian kernel (full width at half maximum: 8 mm). A two-level random-effects procedure was used to conduct region-of-interest (ROI) analyses. At the first level we constructed whole-brain statistical maps to evaluate the win > control and loss > control contrasts. Movement parameters obtained from the realignment stage of preprocessing served as covariates of no interest.
A single anatomically defined, bilateral ROI mask containing reward-related regions (Nusslock et al. Reference Nusslock, Almeida, Forbes, Versace, Frank, LaBarbara, Klein and Phillips2012; Caseras et al. Reference Caseras, Lawrence, Murphy, Wise and Phillips2013) from the WFU PickAtlas (Maldjian et al. Reference Maldjian, Laurienti, Kraft and Burdette2003) was used to avoid conducting multiple statistical tests over individual ROIs: dACC (BA24/32), mPFC (BA10), orbitofrontal cortex (BA11), VLPFC (BA47), amygdala, insula and VS [bilateral spheres centered on ±9, 9, −8; radius = 8 mm based on meta-analyses (Postuma & Dagher, Reference Postuma and Dagher2006; Di Martino et al. Reference Di Martino, Scheres, Margulies, Kelly, Uddin, Shehzad, Biswal, Walters, Castellanos and Milham2008)]. Using a one-sample t test, we extracted significant activity to win > control and loss > control (voxelwise p < 0.001), corrected with a three-dimensional cluster forming threshold of p < 0.05 (http://afni.nimh.nih.gov/pub/dist/doc/program_help/3dClustSim.html) over the entire ROI. Means of significant clusters were extracted using the MarsBaR (Brett et al. Reference Brett, Anton, Valabregue and Poline2002) toolbox in SPM.
Additionally, we examined gray matter structure across the whole brain as in other neuroimaging studies examining relationships between cortical thickness and risky behavior (Lopez-Larson et al. Reference Lopez-Larson, Bogorodzki, Rogowska, McGlade, King, Terry and Yurgelun-Todd2011; see online Supplementary material). Structural thicknesses were calculated using the freely available Freesurfer (Fischl, Reference Fischl2012) software. An unbiased within-subject template space and image were created. Next, skull stripping, Talairach transformation and atlas registration were completed. Finally, generation of spherical surface maps and parcellations with common information from the within-subject template was performed. The quality of surface reconstruction and segmentation was visually assessed. Each structure was extracted and adjusted for individual mean whole-brain thickness.
Clinical assessments
On or near scan day, parents/guardians completed the PGBI-10M to assess their child's behavioral and emotional dysregulation severity (Youngstrom et al. Reference Youngstrom, Meyers, Demeter, Youngstrom, Morello, Piiparinen, Feeny, Calabrese and Findling2005, Reference Youngstrom, Frazier, Demeter, Calabrese and Findling2008), and the Children's Affect Lability Scale (CALS) to assess their child's affective regulation (Gerson et al. Reference Gerson, Gerring, Freund, Joshi, Capozzoli, Brady and Denckla1996). On scan day, parents and LAMS youth completed the Kiddie Schedule for Affective Disorders and Schizophrenia for School-Age Children Mania Rating Scale (KMRS) (Axelson et al. Reference Axelson, Birmaher, Brent, Wassick, Hoover, Bridge and Ryan2003) and Depression Rating Scale (KDRS) (Kaufman et al. Reference Kaufman, Birmaher, Brent, Rao, Flynn, Moreci, Williamson and Ryan1997) to assess hypo/mania and depressive symptom severity, respectively. LAMS youth also completed the Screen for Child Anxiety Related Emotional Disorders (SCARED) on scan day to assess anxiety symptoms over the last 6 months (Birmaher et al. Reference Birmaher, Khetarpal, Brent, Cully, Balach, Kaufman and Neer1997).
Substance use measure
To assess substance use at scan day and post-fMRI scan (mean follow-up time: 741, s.d. = 181.41 days), questions concerning substance use from the Schedule for Affective Disorders and Schizophrenia for School-Age Children (KSADS) (Kaufman et al. Reference Kaufman, Birmaher, Brent, Rao, Flynn, Moreci, Williamson and Ryan1997), the Child and Adolescent Symptom Inventory (Lavigne et al. Reference Lavigne, Cromley, Sprafkin and Gadow2009), and age-appropriate versions of the Centers for Disease Control and Prevention's Youth Risk Behavior Survey (middle school: 10–12 years of age; 2005 version; high school: 13–17 years of age; 2003 version; adult: 18–22 years of age; 2010 version) (www.cdc.gov/yrbs) were used. A report of substance use (more than a few sips of alcohol and/or any illicit drug use) on any of these measures put the participant into the substance user group.
Data analytic plan
The outcome measure used in this analysis was yes/no lifetime substance use. Of the 73 youth, 36 reported substance use 24 months post-scan. Clinical predictor variables on or near scan day included positive mood and energy dysregulation (PGBI-10M score), depressive symptoms, manic symptoms, anxious symptoms, and affective lability, diagnoses [attention-deficit/hyperactivity disorder (ADHD), bipolar spectrum disorder, MDD, disruptive behavior disorder and AnxD], medication status (taking v. not taking each psychotropic medication class: stimulant, non-stimulant ADHD, mood stabilizer, antipsychotic, and antidepressant medications). Demographic variables included age, IQ and sex. Baseline measures of maternal education, parental life-stress (number of stressful events related to child's illness), and parental living arrangement (living with a new partner or alone) were also included as predictors (Kokkevi et al. Reference Kokkevi, Richardson, Florescu, Kuzman and Stergar2007a , Reference Kokkevi, Arapaki, Richardson, Florescu, Kuzman and Stergar b ). Neuroimaging predictor variables included the above BOLD measures to win > control and loss > control and the above whole-brain gray matter cortical thickness variables. We additionally included scan site, and days between scan and follow-up as predictor variables.
Given that our outcome variable was dichotomous and there were more predictor variables than observations, we used LASSO regression analysis with binomial family (logistic LASSO regression) for variable selection and reduction using the freely available GLMNET package in R (Friedman et al. Reference Friedman, Hastie, Simon and Tibshirani2014). LASSO is a modified form of least squares regression that penalizes complex models with a regularization parameter (λ) (Tibshirani, Reference Tibshirani1996). This penalization method shrinks coefficients toward zero, and eliminates unimportant terms entirely (Tibshirani, Reference Tibshirani1996; Friedman et al. Reference Friedman, Hastie and Tibshirani2010, Reference Friedman, Hastie, Simon and Tibshirani2014), thus minimizing prediction error, reducing the chances of overfitting through CV, and enforcing sparsity (Tibshirani, Reference Tibshirani1996).
GLMNET approximates the log-likelihood and then uses a coordinate descent algorithm (Wu & Lange, Reference Wu and Lange2008; Ricket, 2013) computed along a regularization path (an inner weighted least squares loop) to optimize the penalized log-likelihood. Coefficients are stabilized by coordinate descent (optimization of each parameter separately, holding all others fixed). Regularization adds constraints to a problem to avoid over-fitting. Regularization in GLMNET for a binomial regression is performed by producing the path of tuning parameter (λ) along the range of included variables, thus identifying the optimal λ (http://web.stanford.edu/~hastie/glmnet/glmnet_alpha.html). GLMNET then uses CV to compute the mean CV error for each penalty term to guard against type III errors (testing hypotheses suggested by the data). We used a k = 10-fold CV approach.
A test statistic or p value for LASSO that has a simple and exact asymptotic null distribution is still under development (Lockhart et al. Reference Lockhart, Taylor, Tibshirani and Tibshirani2014). We thus provide three other measures that are meaningful for data inference: (1) rate ratio (exponentiated coefficients) of the non-zero coefficients identified in the LASSO model; (2) Cox & Snell R 2 for variance in future substance use explained by the model; (3) classification table results (cut-off = 0.1) from a hierarchical logistic regression analysis in SPSS, using the eight predictor variables identified from the LASSO model.
Post-hoc sensitivity analysis
Of the 36 LAMS youth who at 24 months post-scan reported substance use, 15 also reported using substances at or prior to the scan. To test the importance of the combination of predictor variables derived from the LASSO, we examined the classification table from the logistic regression analysis after removing the 15 youth with substance use at scan. Additionally, to identify the non-zero variables related to future substance use only, we performed a new LASSO analysis, removing these 15 youth and including all of the original p = 108 predictor variables.
Scan site signal variability reduction
We reduced signal variability between scan sites in two ways. First, we monitored the signal:noise ratio monthly to ensure scanner stability over time with a Biomedical Informatics Research Network (fBIRN) phantom at each scan site (http://www.birncommunity.org). Second, we used scan site as a covariate in the LASSO models.
Results
Neuroimaging results
LAMS youth showed significant activation to the win > control contrast in the bilateral dACC (BA32) (MNI: −3, 20, 46 and 3, 20, 46), left mPFC (BA10) (MNI: −39, 47, 1 and −39, 50, 16) and the bilateral ventral anterior insula MNI: 33, 23, −5 and −48, 17, 1); and to the loss > control contrast, in the bilateral dACC (BA32) (MNI: −9, 8, 52; 3, 20, 46; and 9, 29, 31) and the ventral anterior insula (MNI: 30, 20, −8 and −33, 20, 7) (voxelwise p < 0.001, clusterwise corrected p < 0.05, Table 2).
Reward-related activity in 73 LAMS youtha
LAMS, Longitudinal Assessment of Manic Symptoms; MNI, Montreal Neurological Institute; BA, Brodmann area; k, cluster size; df, degrees of freedom; p, uncorrected voxelwise probability value; dACC, dorsal anterior cingulate cortex; t, t test statistical value; mPFC, middle prefrontal cortex.
a Region-of-interest analyses using voxelwise p < 0.001 and cluster-corrected p < 0.05. Table rows represent the peak voxel within the specified region.
LASSO results
Eight predictors together minimized mean squared error, enforced sparsity (Friedman et al. Reference Friedman, Hastie, Simon and Tibshirani2014) and optimized model fit (see Fig. 1 and online Supplementary material). These eight predictors and the direction of the relationships were as follows.
LASSO (Least Absolute Shrinkage and Selection Operator) plots generated in GLMNET. (a) Plot of variable fit. Each curve corresponds to an independent variable in the full model prior to optimization. Curves indicate the path of each variable coefficient as lambda (λ) varies. (b) Plot of non-zero variable fit after cross-validation. Representation of the 10-fold cross-validation performed in GLMNET using LASSO which evaluates the error associated with each lambda. Lambda.min corresponds to the λ which minimizes mean squared error. Lambda.1se corresponds to the λ that is 1 s.e. from the Lambda.min. The solid black line corresponds to the optimal lambda selected due to significantly improved model fit over the Lambda.min and Lamba.1se based on χ2 residual deviance comparisons (see online Supplementary material). Values are cross-validated means, with standard errors represented by vertical bars.
Substance use 24 months post-scan was predicted by greater left middle prefrontal cortical activity to win, lower left ventral anterior insula activity to loss, and thicker caudal anterior cingulate cortex. In addition, older youth, higher depression scores, lower mania (KMRS) scores, more parental stressful events and not being on an antipsychotic medication at scan predicted future substance use (Table 3).
Non-zero coefficients generated from GLMNET using a LASSO regression with a binomial family model
LASSO, Least Absolute Shrinkage and Selection Operator.
a The exponentiated coefficient is the rate ratio change in the dependent variable (future substance use) corresponding to a one-unit change in the predictor variable.
The full model explained 60.4% of the variance in future substance use. Hierarchical logistic regression showed that left middle prefrontal cortical and left ventral anterior insula activity, together with left caudal anterior cingulate cortical thickness, explained 14.4% of future substance use variance over and above the clinical and demographic variables (45.7%; depression and mania scores, parental stress, age, and antipsychotic medication use). Additionally, a cut-off ⩽0.1 from the logistic regression classification table correctly predicted 36/36 of future substance users and misidentified 12/37 of non-users as future substance users, correctly identifying 61/73 participants (83.6%).
Post-hoc sensitivity analysis
After removing the 15 youth who reported substance use at scan, the model remained significant and the Cox & Snell R 2 effect size increased from 0.6 to 0.63. The classification table using the eight non-zero predictor variables identified above (cut-off ⩽0.1) correctly predicted 21/21 future substance users and misidentified 6/37 non-users as future substance users (Cox & Snell = 0.631).
Additionally, in a new LASSO regression analysis including only the 58 youth who were not using substances at scan time, non-zero predictors of substance use were similar to the main analysis. Non-zero predictors were depression score, antipsychotic medication, parental stress at baseline, left middle prefrontal cortical activity to win, and right insula thickness. Notably absent variables in this post-hoc LASSO analysis that may be driven by substance use prior to scan but were predictive of eventual use (see post-hoc classification results above) included left caudal anterior cingulate thickness, left ventral anterior insula activity to loss, and mania scores.
Discussion
Our goal was to assess the ability of neuroimaging measures of reward circuitry activity and cortical thickness to predict future substance use in psychiatrically unwell youth. We used LASSO regression, along with CV, an approach that penalizes complex models with a regularization parameter and identifies the parameter that minimizes error, rendering unimportant coefficients as zero. Our LASSO analysis showed that engaging in substance use 24.3 months post-scan was predicted by a combination of neural activity to win and loss, cortical structure, and clinical and demographic characteristics. These findings explained 60.4% of the variance in substance use 24.3 months after neuroimaging assessment. Furthermore, neuroimaging measures incrementally predicted 14.7% of the variance, i.e. approximately a quarter of the explained variance, in this outcome measure. All eight predictor measures correctly classified 100% of youth who would use substances 24 months later, while misidentifying only 32% of non-users as future users. Including all identified non-zero variables in a logistic regression analysis, both with and without the 15 current users, successfully identified all future substance users 24 months post-scan.
In humans, the mPFC has been shown to be activated both by cognitively demanding tasks, e.g. working memory, and reward, and may subserve the higher cognitive aspects of reward value processing and related, goal-directed behaviors (Pochon et al. Reference Pochon, Levy, Fossati, Lehericy, Poline, Pillon, Le Bihan and Dubois2002). Our present finding of elevated left middle prefrontal cortical activity to reward in youth may thus reflect undue attention to, and higher-order processing of, reward obtained during the task, which, in turn, may predispose to risk-taking behaviors, such as substance use. The left lateralization of our finding may reflect the role of the left hemisphere in approach-related behaviors (Davidson et al. Reference Davidson, Ekman, Saron, Senulis and Friesen1990; Davidson, Reference Davidson1992) (Fig. 2).
Comparisons of neural measures of substance users and non-users 24.3 months post-scan and representation of the region on an average brain image. (a) Reward-related left middle prefrontal cortex (mPFC) and left ventral anterior insula activity. (b) Left caudal anterior cingulate thickness between the two groups (representative image). Thickness variables were adjusted for individual mean cortical thickness. Values are means, with standard errors represented by vertical bars. BOLD, Blood oxygen level-dependent.
We showed that lower ventral anterior left insula activity to loss > control predicted more substance use in the future, although this was no longer the case after excluding the 15 youth who were using substances at scan. Subdivisions of the insula have been shown to have distinct patterns of functional connectivity (Deen et al. Reference Deen, Pitskel and Pelphrey2011). The ventral anterior insula is functionally connected to the anterior cingulate cortex and may have role in the processing of emotion (Deen et al. Reference Deen, Pitskel and Pelphrey2011). Our finding that lower left ventral anterior insula activity to loss predicted future substance use may thus suggest that reduced perception of emotion during loss may have a role in the development of risky behavior in youth. In support of this, in abstinent drug users, insula activity was reported during decision-making (Stewart et al. Reference Stewart, Connolly, May, Tapert, Wittmann and Paulus2014a , Reference Stewart, May, Poppa, Davenport, Tapert and Paulus b ), while attenuation of bilateral insula activity was shown to predict relapse after 1 year among abstinent methamphetamine-dependent youth (Gowin et al. Reference Gowin, Harle, Stewart, Wittmann, Tapert and Paulus2014). Furthermore, individuals with insula lesions placed higher bets and showed less sensitivity to odds compared with controls (Clark et al. Reference Clark, Bechara, Damasio, Aitken, Sahakian and Robbins2008). In healthy individuals, however, greater insula activity was associated with the safer choice during performance of a risky stock market decision-making paradigm (Kuhnen & Knutson, Reference Kuhnen and Knutson2005). The above findings, taken together with our finding that lower left ventral anterior insula activity to loss may have been associated with substance use at scan, may thus suggest that LAMS youth who engaged in substance use may have perceived less emotion and, as a result, may have been less sensitive to the risks involved, and consequent losses sustained, when making decisions during the card number guessing task.
We also showed that greater right insula thickness predicted future substance use in the 58 youth who were not using substances at scan. Animal studies suggest normative thinning of subcortical and cingulate regions with age (Vetreno et al. Reference Vetreno, Yaxley, Paniagua, Johnson and Crews2016). Furthermore, the right insula is implicated in conscious awareness of interoception (Naqvi & Bechara, Reference Naqvi and Bechara2009). Our finding regarding right insula thickness may thus suggest that abnormal neurodevelopment of this region (i.e. reduced pruning) may predispose to abnormally heightened awareness of interoceptive processes that, in turn, may have a deleterious impact on decision-making, but this needs further study.
Other studies have shown that neuroimaging measures may predict future substance use (Becker et al. Reference Becker, Wagner, Koester, Tittgemeyer, Mercer-Chalmers-Bender, Hurlemann, Zhang, Gouzoulis-Mayfrank, Kendrick and Daumann2015), although, in contrast to our findings, a previous report indicated that measures of neural activity may be less important predictors of risky behaviors than other factors in youth. This study reported that a factor consisting of insula, putamen, caudate nucleus, amygdala, cerebellar vermis and prefrontal cortex activity, when combined with a personality factor and a genetic factor, was the least important factor in predicting drinking in adolescence (Heinrich et al. Reference Heinrich, Muller, Banaschewski, Barker, Bokde, Bromberg, Buchel, Conrod, Fauth-Buhler, Papadopoulos, Gallinat, Garavan, Gowland, Heinz, Ittermann, Mann, Martinot, Paus, Pausova, Smolka, Strohle, Rietschel, Flor, Schumann and Nees2016). The fact that a significant proportion of the variance in future substance use was predicted by neuroimaging measures in our study, however, highlights a need for future studies to further examine the role of neuroimaging measures as predictors of risky behaviors in youth.
We additionally showed that greater cortical thickness in the caudal anterior cingulate cortex predicted future substance use, but not after excluding the 15 youth who used substances at scan. In young adults, the left caudal anterior cingulate cortex was thicker in binge drinkers relative to light drinkers (Mashhoon et al. Reference Mashhoon, Czerkawski, Crowley, Cohen-Gilbert, Sneider and Silveri2014). Additionally, normative cingulate cortical thinning was not observed in animals exposed to ethanol (Vetreno et al. Reference Vetreno, Yaxley, Paniagua, Johnson and Crews2016). Thus, similar to the left insula activity to loss finding above, greater anterior cingulate cortical thickness may be a marker of current substance use. More studies are needed to better understand this structural finding.
Non-neuroimaging variables also predicted future substance use. Consistent with the literature, older participants (Kandel & Logan, Reference Kandel and Logan1984; Grant & Dawson, Reference Grant and Dawson1997) and youth with higher depression scores (Deykin et al. Reference Deykin, Levy and Wells1987; Grigsby et al. Reference Grigsby, Forster, Unger and Sussman2016) more often reported future substance use. Youth not prescribed an antipsychotic medication at time of the neuroimaging assessment were also more likely to use substances in the future, probably reflecting the moderating effect of these medications on psychotic and risk-taking behaviors. Intriguingly, youth with lower mania scores were also more likely to report future substance use. This may reflect the fact that youth with lower mania scores were less likely to be taking antipsychotic medication (p = 0.006), and thus did not benefit from the moderating effect of antipsychotic medications behaviors. While we do not suggest that youth be prescribed antipsychotic medication as a measure to reduce risk of future substance use, our findings do suggest that common patterns of neural activity may be associated with psychotic symptoms and substance use. This warrants further study. Finally, increased parental stress due to a child's illness predicted future substance use in youth. This accords with research showing that parental psychological distress is associated with emotional and conduct problems in children (Amrock & Weitzman, Reference Amrock and Weitzman2014; Reeb et al. Reference Reeb, Wu, Martin, Gelardi, Shirley Chan and Conger2015). Our findings thus add to present understanding of the role that parental stress and related behaviors may have on child behavior long term, and suggest that these factors may be used to identify those high-risk families most in need of intervention.
Limitations of the present study included the inability to assess the contribution of pubertal development and other psychosocial factors that show associations with substance use, such as sibling and peer substance use and parental monitoring (Kokkevi et al. Reference Kokkevi, Richardson, Florescu, Kuzman and Stergar2007a , Reference Kokkevi, Arapaki, Richardson, Florescu, Kuzman and Stergar b ), as they were not measured at scan time. Although the age of greatest risk for substance use was not yet reached by some youth in our sample, a larger proportion of the LAMS sample report substance use than is expected from the general population (Substance Abuse and Mental Health Services Administration, 2013). As the children in the LAMS sample are, and have been, behaviorally and emotionally dysregulated for at least 5 years and for as many as 10 years, and are at risk for a myriad of psychiatric disorders, it is, perhaps not unexpected that they engage in substance use at a higher rate than we see in healthy children. Finally, this analysis was designed post-hoc and we therefore were not able to control for substance use at the initial scan visit. Additionally we suspect that some of the misidentification as a substance user may, in fact, be due to the subjective account of substance use by participants. Although the statistical methods utilized here (LASSO with CV) do well at identifying predictors, the estimates may shrink, and error rates for classification of users may be higher, in new, independent samples.
We believe this is the first study to use functional and structural neuroimaging measures to predict future substance use in youth. Specifically, we show that approximately a quarter of the explained variance in future substance use was predicted by neuroimaging measures, especially measures of reward circuitry function. Furthermore, the high discriminative ability to identify future substance use in youth highlights the utility of using a combination of neuroimaging, clinical and demographic measures to help identify those youth most at risk of future substance use. This is an important step toward identifying neurobiological measures characterizing youth at risk of substance use, and provides promising neural targets for the development of novel future therapeutic interventions.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291716003147
Acknowledgements
The present study was supported by National Institute of Mental Health (NIMH) grants 2R01 MH73953-09A1 (B.B. and M.L.P., University of Pittsburgh), 2R01 MH73816-09A1 (S.K.H., Children's Hospital Medical Center), 2R01 MH73967-09A1 (R.L.F., Case Western Reserve University) and 2R01 MH73801-09A1 (M.A.F., Ohio State University), and the Pittsburgh Foundation (M.L.P.).
Declaration of Interest
L.E.A. has received research funding from Curemark, Forest, Lilly, Neuropharm, Novartis, Noven, Shire, Supernus and YoungLiving [as well as National Institutes of Health (NIH) and Autism Speaks] and has consulted with or been on advisory boards for Arbor, Gowlings, Ironshore, Neuropharm, Novartis, Noven, Organon, Otsuka, Pfizer, Roche, Seaside Therapeutics, Sigma Tau, Shire, Tris Pharma and Waypoint and received travel support from Noven.
B.B. has or will receive royalties for publications from Random House, Inc. (New Hope for Children and Teens with Bipolar Disorder) and Lippincott Williams & Wilkins (Treating Child and Adolescent Depression). He is employed by the University of Pittsburgh and the University of Pittsburgh Medical Center and receives research funding from the NIMH.
T.W.F. has received federal funding or research support from, acted as a consultant to, received travel support from, and/or received a speaker's honorarium from the Simons Foundation, Ingalls Foundation, Forest Laboratories, Ecoeos, IntegraGen, Shire Development, Bristol-Myers Squibb, NIH and the Brain and Behavior Research Foundation.
R.L.F. receives or has received research support, acted as a consultant and/or served on a speaker's bureau for Alcobra, American Academy of Child & Adolescent Psychiatry, American Physician Institute, American Psychiatric Press, Bracket, CogCubed, Cognition Group, Coronado Biosciences, Dana Foundation, Elsevier, Forest, Guilford Press, Ironshore, Johns Hopkins University Press, Jubilant Clinsys, KemPharm, Lundbeck, Medgenics, Merck, NIH, Neurim, Novartis, Otsuka, Oxford University Press, Pfizer, Physicians Postgraduate Press, Purdue, Rhodes Pharmaceuticals, Roche, Sage, Shire, Sunovion, Supernus Pharmaceuticals, Teva, Transcept Pharmaceuticals, Tris, Validus and WebMD.
M.A.F. receives royalties from Guilford Press, American Psychiatric Press and CFPSI. She receives research funding from Janssen and honoraria from Physicians’ Post-Graduate Press.
R.A.K. is a consultant and faculty for the REACH Institute. He is a consultant for a DSMP for Forest and Pfizer. He is employed by the Ohio State Wexner Medical Center.
M.L.P. is a consultant for Roche Pharmaceuticals.
J.L.S. receives research support from Siemens Healthcare.
E.A.Y. has consulted with Pearson, Western Psychological Services, Janssen, Lundbeck, Joe Startup Technologies and Otsuka about assessment, as well as having grant support from the NIH.
J.R.C.A., D.A.A., G.B., M.A.B, L.B., V.A.D., E.E.F., M.K.G., S.K.H., S.M.H., S.I., S.B.P., C.S., M.J.T. and A.V. report no competing interests.
