Methodological Advances in Leveraging Neuroimaging Datasets in Adolescent Substance Use Research

Abstract

Purpose of Review

Recent innovations in the statistical analysis of neuroimaging data related to adolescent substance use are highlighted. Going beyond assumptions of homogeneity in small studies of regional localization, the focus is on novel approaches that integrate across regions of the brain and levels of analysis in order to detect individual differences in use along with antecedents and consequences.

Recent Findings

Three analysis approaches are considered. Multimodal approaches like the construct-network framework combine neural, behavioral (including cognitive), and self-report indicators to create comprehensive representations of risk factors for adolescent substance use. Machine learning approaches link adolescent substance use to complex patterns of brain activity detected using prediction-focused algorithms. Person-specific approaches reflect heterogeneity in functional brain connectivity associated with adolescent substance use.

Summary

When applied to specialized datasets, multimodal, machine learning, and person-specific approaches have significant potential to provide unique insights into the neural processes underlying adolescent substance use.

Introduction

Adolescent substance use impacts personal, familial, and societal health. Despite increased knowledge about the dangers and long-term negative consequences of use as well as public prevention campaigns (e.g., Drug Abuse Resistance Education or D.A.R.E.) [1, 2], adolescent substance use rates have remained relatively steady in recent years, according to data from the nationally representative Monitoring the Future study [3]. This is true with respect to illicit substances, including marijuana (about 24% of adolescents use each year) and alcohol (about 59% of adolescents have had an alcoholic beverage by the end of high school). Patterns vary for nicotine, however; there have been continuing decreases in cigarette smoking, but dramatic increases in vaping, over the past couple of years [3]. Although there is historical variability in the types of substances used and the ways in which they are used, the consistency of adolescent substance use across cohorts is likely tied to the interplay among biological (e.g., genes, pubertal hormones), psychological (e.g., childhood mental illness), social (e.g., parental and peer substance use), and contextual (e.g., neighborhood) processes unique to this developmental period [4–8]. As all of these processes are mediated, at some level, by the brain, research on the neural correlates of adolescent substance use— and the psychophysiological mechanisms linked to it—is imperative. This paper outlines recent methodological developments that have the potential to propel this field forward: After synthesizing the state of the extant science, knowledge gaps due to current research limitations are identified, and three methodological advances that partly overcome those limitations are presented.

Neural Basis of Adolescent Substance Use

Neural structure and function is prominent in the biopsychosocial interplay underlying adolescent substance use. Not only do normative patterns of adolescent brain development, with divergent trajectories for limbic and frontal regions, bias adolescents’ motivational behaviors (e.g., in ways that predispose them to risk taking and social status seeking) [9, 10], but the adolescent brain is also particularly sensitive to the positive influences of substances, such as dopamine-linked stimulation of reward-related brain regions in response to alcohol, while at the same time being resistant to some negative side effects, such as alcohol-induced psychomotor impairment (reviewed in [11, 12]).

The resistance to negative side effects is short-lived, however, in that there are long-term decrements in cognition and aberrations in brain development and function associated with adolescent substance use (reviewed in [13, 14]). For instance, functional magnetic resonance imaging (fMRI) studies revealed that adolescent marijuana use is associated with thickened cortices, especially in frontal regions, and with less white matter integrity, likely indicative of failed or delayed gray matter pruning and white matter proliferation characteristic of typical adolescent brain development; marijuana users also show altered frontal and parietal regional activation during cognitive tasks (reviewed in [15•, 16]). Moreover, adolescent alcohol use is associated with smaller prefrontal cortical volumes, potential cortical thickening, reduced white matter integrity, altered frontal and parietal regional activation during cognitive tasks, and sensitivity in reward-related regions, such as the nucleus accumbens, during reward processing tasks (reviewed in [15•, 17]). Although consistent findings have emerged, particularly with respect to frontal lobe structure and function, more work is needed. For example, it is unclear to what extent neural effects reflect risk for versus consequences of substance use, the role of gender in brain-behavior relations is rarely considered, and marijuana use is often confounded with alcohol use.

Knowledge Gaps Linked to Methodological Limitations

Significant knowledge also remain concerning the translation of neuroimaging findings into concrete recommendations for clinicians, policy makers, and the families of struggling adolescents. These knowledge gaps stem partially from the limitations of relatively small, cross-sectional studies that only have the statistical power to detect large, homogeneous effects. To ameliorate these methodological limitations and strengthen inferences, large, diverse, longitudinal studies have emerged; some have particular substance use or related foci and involve the collection of neuroimaging data (e.g., fMRI and electroencephalography or event-related potentials) and data from other modalities (e.g., behavioral, genetic). For instance, the IMAGEN study is following 2000 14-year-old adolescents every 2 years to 3 years [18], and the Adolescent Brain Cognitive Development (ABCD) study is following over 11,800 9-year-old and 10-year-old early adolescents for 10 years [4]. However, emerging findings from these studies suggest that neural and behavioral (including cognitive) indicators only explain a small portion of the variance in substance use and associated outcomes, such as externalizing psychopathology [19–22].

The disappointing predictive power of neural and behavioral indicators in large-scale, diverse studies can be attributed to several factors. First, substance use–linked constructs (e.g., reward sensitivity and inhibitory control) are typically operationalized with indicators from a single level of analysis, such as regional blood oxygen level–dependent (BOLD) fMRI responses to reward anticipation or the proportion of inhibitory errors on a go/no-go task, but substance use is typically self-reported. Therefore, as long-noted in psychometrics [23], method-related bias may artificially limit estimates of predictive power: Simply because substance use is self-reported, the predictive utility of self-report indicators may be favored, while the associations between neural and behavioral (including cognitive) indicators and substance use may be restricted. Second, substance use is highly multidetermined, in that it results from the complex interplay among biological, psychological, social, and contextual processes [4–8, 24], so it is not likely that one or even a handful of factors explains a large portion of the variance in adolescent use. Third, the vast majority of research on adolescent substance use assumes, implicitly or explicitly, that neural processes operate similarly across individuals, but given the multidetermined nature of substance use and individual differences in brain development and function, significant heterogeneity is likely present [25, 26]. Thus, it is not surprising that mean-level results concerning the average adolescent substance user have small or non-replicable effects with unclear applicability to individual adolescents.

Methodological Advances for Detecting Neural Contributions to Adolescent Substance Use

Limited predictive power has led some to suggest that small effect sizes are “the new normal” in clinical neuroscience [27••], implying that neural indicators—even from large, representative and longitudinal studies—may ultimately be unable to inform meaningful policy recommendations or clinical interventions relevant to adolescent substance use. The challenge to establish robust and translatable findings can at least be partially overcome, however, by leveraging recent methodological advancements related to the analysis of neuroimaging data. Thus, the aim of this paper is to showcase how three such advancements can be utilized to address factors thought to contribute to the meager ability of neural indicators to predict adolescent substance use. First, limitations inherent in the use of single neural indicators to predict self-reported use can be assuaged with multimodal approaches that combine across neural, behavioral (including cognitive), and self-report indicators. Second, difficulties linked to the multidetermined nature of substance use can be countered with the use of machine learning approaches that optimize prediction by identifying complex patterns among large sets of variables. Third, inaccuracies stemming from violated assumptions of homogeneity among substance-using adolescents can be corrected by employing person-specific approaches that actually exploit heterogeneity in order to map patterns of neural connectivity at the individual level.

In what follows, each approach (i.e., multimodal, machine learning, and person-specific) is described in the context of neuroimaging research on adolescent substance use, illustrated through the review of select empirical studies and considered in light of its potential for future research. The similarities and differences among the approaches as well as challenges in their collective implementation are discussed in the final section.

Multimodal Approaches

Multimodal approaches integrate neural correlates of adolescent substance use with indicators from other levels of analysis, including behavior (e.g., cognitive tasks) and self-reports. One exemplar of this approach is the construct-network framework [28]. The motivation for this framework was rooted in foundational work on the concept of method variance [23], which is systematic error that is present in measurements from the same modality. For example, self-reported indicators of externalizing psychopathology, including substance use problems, may be systematically biased by individuals’ general tendency to endorse versus deny personal difficulties, while neuroimaging indicators, such as BOLD fMRI responses from different brain regions, may be systematically biased by factors such as motion during data collection. Method variance, therefore, has the potential to either reduce (if measures are from different modalities) or artificially inflate (if measures are from the same modality) observed relations between psychological indicators and adolescent substance use, and this may partially explain the modest predictive performance of brain-based indicators in past work.

Rather than viewing indicators from different levels of analysis (e.g., neural versus self-report) as distinct, the construct-network approach [28] combines relevant indicators across modalities in a statistical model of a single core construct. This is based on the assumption that the shared variance of indicators from different modalities will more precisely reflect variation across individuals in the underlying biopsychosocial construct than combinations of same-modality measures, which are biased by method variance. Drawing on previous work [29] and on an extension to a new community-recruited adult sample, Patrick and colleagues [28] demonstrated how this approach can be used to create a psychoneurometric factor of disinhibition. This latent factor was made up of both self-reported externalizing behavior and individuals’ amplitudes of P3 event–related potentials (i.e., an electrophysiological measure that has been previously linked to externalizing behaviors and disinhibition). The resulting psychoneurometric factor was strongly linked to substance use, even though individual electrophysiological measures had only weak and variable relations with use. More recent work using this approach with adult samples has incorporated additional indicators, including cognitive tasks, into the disinhibition factor and provided evidence that links between psychoneurometric factors of disinhibition and substance use largely reflect genetic influences [30, 31•]. Hence, a multimodal approach operationalized within the construct-network framework provides a valuable way to bridge the gap between neural correlates of substance use and indicators from other modalities; it reduces the impact of method variance and will likely increase predictive utility.

The construct-network approach has been developed in adult samples and primarily applied to electrophysiological indicators of brain activity, so it has significant, unrealized potential for applications in large longitudinal datasets of adolescent development, such as IMAGEN and ABCD. Given the abundance of indicators from neural (e.g., multiple structural and functional MRI–based measures), behavioral (e.g., cognitive), and self-report modalities in these datasets, work focused on identifying novel multimodal factors may both provide much-needed dimensionality reduction and foster the discovery of biopsychosocial constructs that explain more substantial variance in adolescent substance use than neural indicators alone. Toward this end, Brislin and colleagues [32•] have already developed and validated a measure of disinhibition (the same underlying construct extensively investigated in previous work using the construct-network approach [28••, 29–31]) from self-report items available in IMAGEN and found that it displayed a robust negative relationship (r = − .34) with P3 potentials in an independent sample of undergraduate students, suggesting that it is comparable to other psychoneurometric disinhibition factors. This provides researchers with an exciting opportunity to identify multimodal neural correlates of disinhibition (e.g., utilizing MRI-based measures) in the large-scale, longitudinal IMAGEN study and to link them to adolescent substance use in a construct-network framework. Similar work in other large datasets aimed at uncovering novel multimodal measures of psychological constructs relevant to adolescent substance use will likely maximize information from neuroimaging indicators and highlight their importance for clinical prediction of substance use to benefit struggling youth and the parents and practitioners who support them.

Machine Learning Approaches

Machine learning approaches are powerful not only for characterizing the many correlates of adolescent substance use but also for translating neural indicators of brain structure and function into robust, individual-level predictions. These approaches include a large and growing array of methods—from simple logistic regression models to support vector machines and other advanced algorithms—that share several distinctive features [33•]. First, machine learning methods work by identifying the subset of dimensions from a larger set of complex, high-dimensional data (e.g., whole-brain neuroimaging data or data from many modalities) that are most relevant for predicting a given outcome. Second, these methods are typically a-theoretical, in that they are optimized for predictive accuracy regardless of whether the patterns identified correspond to hypotheses or mechanisms from explanatory theories. Finally, the standard of success is a method’s ability to make robust predictions. To evaluate this, these methods are developed in a training dataset (e.g., a sample or subset of available data) and are subsequently used to predict outcomes in a novel test dataset (e.g., a separate sample or hold-out portion of available data); training datasets are usually significantly larger than test datasets.

Two recent studies used machine learning methods to prospectively predict adolescent alcohol use, highlighting both the promise and challenges of this approach for this line of research. In one study, regularized logistic regression models were used in the IMAGEN dataset [34••]. The goal was to predict binge drinking at age 14 (baseline) and again at age 16 (2-year follow-up) using multimodal indicators collected at age 14, such as fMRI activation maps and structural MRI regional gray matter volumes, genetics, cognitive performance, personality self reports, and demographics and life history (e.g., previous substance use, family risk factors, and romantic relationships). The models produced robust predictions of adolescent binge drinking at both ages that were significantly better than chance, and impressively, they generalized both to hold-out portions of the sample and to a novel sub-sample from IMAGEN. Follow-up analyses indicated that the life history, personality, and brain-based variables contributed most to the models’ predictive power. Although life history variables, such as prior substance use and cigarette smoking, were more strongly related to binge drinking than brain-based indicators, the predictive accuracy of the full model exceeded the accuracy of models that only used information from a single modality, emphasizing the multidetermined nature of adolescent substance use and suggesting that neural indicators contribute incrementally to predictive models utilizing data from other modalities.

In another study, random forest classification models were used to identify the most important predictors of moderate-to-heavy alcohol use by age 18 from an array of multimodal indicators from ages 12 through 14, when youth were naïve to substances [35]. The model suggested that reduced cortical thickness and neural activation across multiple brain regions during a visual working memory task, poor cognitive performance, and demographic and psychosocial factors (e.g., male gender, socioeconomic status, externalizing behaviors) contributed to model predictions. The sample was too small (N = 137), however, to assess predictive power in a test dataset, so the accuracy and generalizability of this model are currently unknown. Even though this work provides preliminary indications that brain-based developmental risk factors contribute to predictive models of adolescent substance use, it also highlights the need for a paradigm shift in future research, as collecting neuroimaging data on a scale large enough to both develop and validate machine learning models will require large collaborative efforts, such as IMAGEN and ABCD.

As intimated in these empirical examples, prediction-focused machine learning methods have considerable potential for future basic and applied work on adolescent substance use. For instance, the extension of these methods beyond neuroimaging data to include multimodal indicators (e.g., demographic and life history risk factors) facilitates the precise characterization of neural contributions to adolescent substance use. Moreover, through their a-theoretical application, they may even inform explanatory developmental or clinical neuroscience theories by identifying novel multivariate patterns in data that may have been overlooked in hypothesis-driven research [36]. Finally, machine learning methods can improve the translational utility of clinical neuroscience research by generating personalized predictions from neuroimaging data [37]; for example, findings reviewed above suggest that neuroimaging data contribute to predictive models that generate risk scores for individual youth, which, in turn, may inform future targeted prevention efforts. Thus, machine learning approaches are poised to capitalize on the value inherent in large-scale longitudinal neuroimaging datasets in order to advance understanding of adolescent substance use risk as well as to define the nature of use and to inform treatment.

Person-Specific Approaches

Each adolescent is unique. This is exemplified by individual differences in substance use behaviors and heterogeneity in substance use disorders; it is also apparent in the brain development and function underlying use [25, 38]. Yet, in related research, there is an overwhelming assumption of homogeneity accompanied by the widespread utilization of analyses that average across individuals (e.g., general linear models). These analyses are based on the assumption that variability around the mean is random and will cancel out in averages, which then serve as predictions when generalizing results to other samples. But, if variability around the mean is systematic, reflecting meaningful heterogeneity, the mean-level results will actually fail to represent any single person in the sample, let alone serve as reasonable predictors for other individuals [26, 39]. Therefore, it is not surprising that therapeutic efforts based on mean-level models have unsatisfying results when employed with unique adolescents [40, 41]. Fortunately, the intensive time series nature of functional neuroimaging data (e.g., fMRI data are collected from each voxel in the brain at hundreds of volumes) permit person-specific analyses; instead of analyzing variability across individuals, person-specific models analyze variability within a person across the time series to create results unique to each adolescent in a sample. Indeed, mathematical theory proves that person-specific analyses are the only accurate way to model heterogeneous psychological phenomena [39].

Group Iterative Multiple Model Estimation (GIMME) [42] is a person-specific directed functional connectivity approach that estimates connections among a priori brain regions of interest (ROIs), and it has particular potential for research on the neural contributions to adolescent substance use. GIMME is a state-of-the-art program that uses a data-driven procedure to create sparse connectivity maps for each individual in a sample. It does this by implementing unified structural equation models (uSEMs) [43, 44] to explain the variation in resting state or task-based functional neuroimaging data; uSEMs combine SEMs and vector autoregressive models in order to estimate contemporaneous (volume-locked) and lagged (predicting from one volume to the next) connections among ROIs. Although all connections are estimated at the individual level, GIMME will identify group-level connections if there is statistical evidence for sample homogeneity. Due in large part to this integrated group- and individual-level mapping, GIMME has outperformed other undirected functional (e.g., principal component analysis), directed functional (e.g., SEMs), and effective (e.g., dynamic causal modeling) connectivity mapping approaches in large-scale simulation studies (reviewed in [45]). Moreover, findings based on data simulated to reflect heterogeneity across adolescent neural networks highlight how GIMME uniquely bridges the gap between the theory and methods used to understand the neural basis of adolescent risk taking behaviors, such as substance use, because it permits directional inferences (e.g., from limbic-to-cortical ROIs) and captures variability across adolescents and development [46].

GIMME has been applied in neuroimaging studies of substance use (e.g. [47]), including one relevant to adolescent alcohol use [48•]. Late adolescents were thrice scanned while completing a go/no-go task in which images of alcoholic or non-alcoholic beverages were the go or no-go cues; they were scanned the summer before college and during their first and second semesters. Lagged and contemporaneous connections among ROIs in cognitive control and emotion processing networks were estimated at the group and individual levels at each scan, and then connections within and between the networks were tabulated for each individual. Findings revealed heightened connectivity within a cognitive control network during students’ first semester—and this neural increase coincided with their alcohol cue exposure and use. Figure 1 shows person-specific networks generated from GIMME for two illustrative participants during this transitional first semester of college. Notice that all connections are directional (i.e., predict from one ROI to another) and have person-specific magnitudes (i.e., beta weights that differ between participants). Moreover, some connections (thick lines) are estimated for all participants and others (thin lines) are estimated only for individuals. Finally, some connections (solid lines) are contemporaneous and others (dashed lines) are lagged. Connectivity among cognitive control ROIs (i.e., rostral and dorsal anterior cingulate cortex and bilateral dorsolateral prefrontal cortex) distinguished the first semester of college; one participant (Fig. 1A) had 8 such connections, while the other (Fig. 1B) had 9. As this figure showcases, person-specific directed functional connectivity maps generated by GIMME characterize the expected neural heterogeneity and are well-suited to subsequent analyses (e.g., longitudinal change in networks and links between networks and substance use through the extraction of graph theoretic metrics, such as network density).

Fig. 1.

Two illustrative person-specific networks generated from GIMME while participants completed an fMRI-based go/no-go task in which images of alcoholic beverages were the response cues; participants were late adolescents in their first semester of college. Group-level connections were estimated for all participants, and individual-level connections were estimated for unique participants. Contemporaneous connections represent prediction at the same time point (i.e., functional volume), and lagged connections represent prediction to the next time point. All connections have an associated beta weight. Person-specific networks fit the data well. a χ²(67) = 70.52, p = .36, CFI = 1.00, NNFI = .99, RMSEA = .022, SRMR = .051; b χ²(68) = 81.62, p = .12, CFI = .99, NNFI = .99, RMSEA = .043, SRMR = .039. OFC, orbitofrontal cortex; ACC, anterior cingulate cortex; DLPFC, dorsolateral prefrontal cortex. Data from Beltz et al. [48•]

There is significant potential in applying person-oriented models to substance use research (see examples in [46]), particularly research on the neural correlates of adolescent substance use available in large, longitudinal datasets, such as IMAGEN and ABCD. GIMME is just one option, but within it, there are several analytic extensions that facilitate rigorous and innovative investigations. For instance, GIMME has mechanisms for verifying the directionality of connections within networks and for identifying connections within a priori or data-driven subgroups of participants [49, 50]. Beyond GIMME, person-specific approaches provide a valuable first step in overcoming limitations associated with one-size-fits-all prevention and intervention efforts by accurately characterizing brain function at the individual level. It will be paramount for future work to extend person-specific descriptions to clinically relevant predictions in line with calls for personalized, precise mental healthcare, especially for the treatment of adolescent substance use disorders [51, 52].

Conclusions

Three underutilized statistical analysis approaches—multi-modal, machine learning, and person-specific—are uniquely positioned to leverage data available in large-scale, longitudinal datasets in order to increase understanding of the neural basis of adolescent substance use and inform prevention and intervention strategies. Although other approaches are available, these three are unique in utilizing multimodal indicators, optimizing predictive power and focusing on individuals. Regarding multimodal indicators, both the construct-network framework and machine learning methods can integrate neural indicators (of structure and function from electrophysiology or MRI) with indicators from other levels of analysis—be they behavior (including cognition), self reports, or individual difference markers. Some person-specific methods, such as GIMME, can also integrate across levels as long as time series from the multimodal measures can be aligned.

Regarding predictive power, machine learning methods are ideal for predicting adolescent substance use. In fact, a common criticism of multimodal (e.g., construct-network) and person-specific (among many other) approaches is that they focus on forming and testing explanations of adolescent substance use, rather than making robust predictions about risk for use. Although the goals of prediction and explanation are not mutually exclusive, a strong emphasis on explanation may have led to the proliferation of theories that elegantly describe data from idiosyncratic samples and well-controlled experiments, but that have poor or unclear predictive power in the real world [33•]. Nevertheless, it is also important to note that predictive models in developmental science still must hedge interpretations based on the longitudinal nature of data (e.g., risk factors must be measured before substance use outcomes [35•]). Ultimately, predictive and explanatory research programs can be mutually beneficial because theoretical frameworks can guide research questions and measurement choices in predictive modeling, and they can facilitate interpretations of complex outputs from predictive analyses.

Regarding individuals, person-specific models impeccably characterize heterogeneity in neural processes underlying adolescent substance use. Because most multimodal and machine learning approaches base results on metrics of between-person variation or patterns, they can generate average findings that do not equally apply to all individuals. Although machine learning approaches facilitate individual-level predictions, predictions for some individuals may be grossly inaccurate, whereas results for individuals from person-specific models are accurate, but are only predictive under certain conditions (e.g., temporal stability), reflecting trade-offs between robustness of prediction and descriptive accuracy at the individual level.

Despite the advantages of multimodal, machine learning, and person-specific approaches for studying adolescent substance use, there are inherent limitations. Each approach requires unique data (e.g., from large, multimodal datasets or intensive time series from an individual) and technical skills (e.g., in data analysis and coding). If these are indeed the approaches of the future, then they beg the question of whether current research practices (e.g., based on single investigator labs) and training programs (e.g., of statistics and data science) are optimally situated for innovative science (see also [33•]). Nonetheless, resources are currently available to begin to close the knowledge gap between the conceptual and statistical science. They include open datasets, such as ABCD, and models from collaborative, interdisciplinary research teams (e.g. [32, 34, 48],) that pave the way for highly impactful future research on the neural underpinnings of adolescent substance use with translational potential.