Prediction complements explanation in understanding the developing brain

Monica Rosenberg; B. Casey; Avram Holmes

doi:10.1038/s41467-018-02887-9

Prediction complements explanation in understanding the developing brain

Rosenberg, Monica; Casey, B.; Holmes, Avram 2018-02-21 00:00:00 REVIEW ARTICLE DOI: 10.1038/s41467-018-02887-9 OPEN Prediction complements explanation in understanding the developing brain 1 1 1,2 Monica D. Rosenberg , B.J. Casey & Avram J. Holmes A central aim of human neuroscience is understanding the neurobiology of cognition and behavior. Although we have made signiﬁcant progress towards this goal, reliance on group- level studies of the developed adult brain has limited our ability to explain population variability and developmental changes in neural circuitry and behavior. In this review, we suggest that predictive modeling, a method for predicting individual differences in behavior from brain features, can complement descriptive approaches and provide new ways to account for this variability. Highlighting the outsized scientiﬁc and clinical beneﬁts of pre- diction in developmental populations including adolescence, we show that predictive brain- based models are already providing new insights on adolescent-speciﬁc risk-related beha- viors. Together with large-scale developmental neuroimaging datasets and complementary analytic approaches, predictive modeling affords us the opportunity and obligation to identify novel treatment targets and individually tailor the course of interventions for developmental psychopathologies that impact so many young people today. nderstanding how the brain gives rise to cognition and behavior is a fundamental goal of human neuroscience. Scientists, philosophers, and statisticians have long debated the Unature of understanding, but tend to agree that there are two routes to achieving it: 1–4 explanation and prediction . Despite the historical dominance of explanation as a route to understanding, scientists and philosophers of science have emphasized the importance of both 5–8 these approaches . As noted by the philosopher Heather Douglas, “explanation and prediction are best understood in light of each other and thus … should not be viewed as competing goals but rather as two goals wherein the achievement of one should facilitate the achievement of the other” . Foundational cross-species research has made signiﬁcant progress on the path towards neu- roscientiﬁc explanation. Researchers have described neural bases of cognition, characterizing how patterns of brain organization from neural circuits to functional networks relate to behavior 9–14 and psychopathology . Although this work has traditionally taken a cross-sectional, group- level approach to studying the developed adult brain, there is growing consensus that com- prehensive models in neuroscience must account for the facts that neural phenotypes and 15–18 behavior vary widely across the population and change over time within individuals . The road to prediction is less traveled. Recently, however, the use of machine-learning methods to predict behavior from brain measures has become increasingly common, due in part to the emergence of large data sets and new analytic and computational tools . Representing a critical avenue to understanding, these approaches provide new ways to account for 1 2 Department of Psychology, Yale University, New Haven, CT 06520, USA. Department of Psychiatry, Yale University, New Haven, CT 06511, USA. Correspondence and requests for materials should be addressed to M.D.R. (email: monica.rosenberg@yale.edu) NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications 1 | | | 1234567890():,; REVIEW ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 22,23 developmental changes in behavior, dynamic brain systems, and experiment with drugs than children or adults . These 8,20 associated individual differences while offering statistical rigor behaviors are thought to stem from adolescents’ increased 25,26 27,28 and clinical and translational beneﬁts for personalized medicine sensation-seeking and reward-sensitivity , as well as 5,6 29 and education . decreased self-control and emotional regulation abilities, 30–34 Although much predictive modeling research has focused on especially in social contexts . The prevalence of anxiety adults, forecasting outcomes in childhood and adolescence pre- disorders also peaks in adolescence, underscoring this develop- sents unique opportunities for scientiﬁc discovery and clinical mental period as a time of both vulnerability and opportunity for application. First and foremost, biological and statistical models intervention . that account for developmental change are necessary for truly Given that risky behavior during development has potentially understanding how neural circuits emerge to give rise to cogni- dire consequences, why does it persist across generations and tion and behavior. Predicting behavior from brain features during species ? Fear learning provides a useful example of potential development represents initial progress towards this goal, and evolutionary beneﬁts of seemingly costly behavior during 36,37 predicting future outcomes from past developmental changes adolescence . Across altricial species, whose young rely on represents an important next step. Predictive models of current parents for survival, fear learning is suppressed in early infancy, and future behavior may be especially beneﬁcial in adolescence, a presumably to ensure caregiver attachment even in cases of 38,39 developmental period of rapid social, emotional, psychological, neglect or abuse . In adolescence, fear of previously aversive and physical change characterized by mental and physical health environmental contexts is diminished whereas fear of previously vulnerability, but also opportunities for growth and aversive cues (i.e., conditioned stimuli) is ampliﬁed, a pattern that intervention . may facilitate exploration and independence but also safety from 40,41 In this forward-looking review, we highlight how predictive immediate threat . Importantly, these survival-relevant beha- modeling in developmental neuroscience can account for devel- viors do not develop in a vacuum. Rather, common genetic 42 43–45 opmental trajectories in behavior, dynamic brain systems, and variations and early life stressors affect how fear learning individual differences in both. After discussing these concepts in changes over time, inﬂuencing risk for negative outcomes such as the context of adolescence, we introduce predictive modeling and anxiety disorders . Just as developmental changes in fear its applications in developmental populations. Using adolescence learning confer both costs and beneﬁts, changes in risk taking as a case study, we address two complementary questions. First, during adolescence are advantageous at the group level but in how can prediction inform models of risk taking, a phenotype some contexts may be detrimental for the individual . The same that is strikingly elevated in adolescence? Second, how can con- is likely true for other processes following their own nonlinear sidering adolescence inform predictive modeling techniques and trajectories across development, including decision making , 47 48 motivate future research? We conclude by emphasizing the reward learning , and sensitivity to motivational , appetitive, 49,50 importance of approaches that predict current and future beha- and aversive cues . vior from developmental trajectories of brain structure and Although stereotypes can paint teenagers in an unﬂattering function. In doing so, we discuss how these methods complement light, recognizing that adolescent behaviors are single points and extend ongoing research on the neurodevelopmental pro- along broader, evolutionarily advantageous developmental tra- cesses that underlie the emergence and disruption of cognition jectories provides a more accurate, nuanced (and perhaps and behavior. sympathetic) picture. Analytic perspectives that consider beha- vioral shifts during the transitions into and out of adolescence, as well as their differential expression across environmental and Changes in behavior and brain systems across adolescence social contexts, are necessary for understanding how the brain Human abilities and behavior change dramatically across the gives rise to behavior over time. lifespan, emerging over development from the dynamic interplay between genes and experience. Developmental changes reﬂect Dynamic brain systems. The nonlinear behavioral trajectories neurobiological constraints shaped by evolution to meet the observed across adolescence emerge from a cascade of hier- unique challenges of each stage of life, including adolescence . archical changes in brain circuitry that were themselves shaped That evolutionary pressures have presumably tailored adolescent over the course of our evolutionary lineage . First to mature are behavior to facilitate the transition to independence, however, is connections within subcortical-limbic circuits, followed by con- frequently overlooked. Instead, adolescents, whose behavior is nections between cerebral cortex and subcortical-limbic circuits, sometimes judged as immature relative to their physical devel- 51,52 and, ﬁnally, connections across cortex . 22–24 opment, are often considered impaired mini-adults . In the Evidence for this developmental cascade comes from observa- following section, we emphasize the importance of considering tions of earlier changes in synaptic morphology and neurotrans- developmental changes, dynamic brain systems that unfold over mitter systems in subcortical relative to cortical regions and an time, and interindividual variability when seeking to establish earlier plateau in synaptic formation and subsequent pruning in descriptive and predictive models of behavior. unimodal sensory, motor, and subcortical regions relative to 53,54 multimodal association areas . These processes likely con- Developmental trajectories in behavior. When we think of the tribute to gray matter volume and cortical thickness changes 55–58 prototypical adolescent (or recall our own teenage years), a observed during adolescence and early adulthood that end in 59–61 number of quintessential traits may come to mind. We might the association cortices . Selective degradation of excitatory consider (or remember) risky behaviors like dangerous driving, synapses also affects the excitatory-inhibitory balance across illegal substance use, irresponsible sexual behavior; preoccupa- cortex, an equilibrium related to shifts in cognitive abilities and 51,62 tions with peer groups and social hierarchies; an uptick in feelings behavior . The relative decrease in prefrontal behavioral of anxiety; and heated conﬂicts with parents, teachers, or other regulation is reﬂected in changes in dopamine receptor density, well-meaning ﬁgures of authority. related to learning and reward prediction, that peak in the Epidemiological studies conﬁrm that our stereotypes largely striatum during adolescence but not until early adulthood in the 63–65 reﬂect typical adolescent behavior. Adolescents are more likely to prefrontal cortex . be injured or killed in motor vehicle accidents, contract sexually Structural and functional brain connections follow similar transmitted infections, engage in criminal activity, and patterns of development, providing additional evidence for a 2 NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 REVIEW ARTICLE YOUth / CID NeuroIMAGE Gen R IMAGEN SYS PING HBN ABCD NCANDA ENIGMA PNC HCP-D ABIDE CoRR ADHD-200 c-VEDA HRC Fig. 1 Existing, ongoing, or planned data sets including structural and/or functional neuroimaging data from ~500 or more children or adolescents. These data sets, which represent both prospective and retrospective samples, include the Adolescent Brain Cognitive Development study (ABCD; USA), 82 80 Healthy Brain Network (HBN; USA), Lifespan Human Connectome Project Development (HCP-D; USA), National Consortium on Alcohol and 149 150 NeuroDevelopment in Adolescence (NCANDA; USA), Pediatric Imaging, Neurocognition, and Genetics study (PING; USA), Philadelphia 151 152 Neurodevelopmental Cohort (PNC; USA), Saguenay Youth Study (SYS; Canada), High Risk Cohort Study for the Development of Childhood 153 81 Psychiatric Disorders (HRC; Brazil), Autism Brain Imaging Data Exchange (ABIDE; USA, Germany, Ireland, Belgium, Netherlands), Enhancing 154 79 NeuroImaging Genetics through Meta-Analysis (ENIGMA; worldwide), IMAGEN (England, Ireland, France, Germany), Dutch YOUth cohort (part of 155 156 the Consortium on Individual Development, or CID; Netherlands), Generation R Study (Gen R; Netherlands), NeuroIMAGE (follow-up of the Dutch arm of the International Multicenter ADHD Genetics, or IMAGE, project; Netherlands), Consortium on Vulnerability to Externalizing Disorders and 157 108 Addictions (c-VEDA; UK, India), Consortium for Reliability and Reproducibility (CoRR; China, USA, Canada, Germany), and ADHD-200 (USA, China). Although samples are distributed across the globe, African, Middle Eastern, South Asian, Oceanian, and Central and South American populations are underrepresented. Data collection efforts in these regions and others will be important for ensuring diverse, representative samples that will allow researchers to uncover general principles of the developing brain. (Map outline courtesy of Wikimedia user ‘Loadﬁle’ and is licensed under a CC BY SA 3.0 license) hierarchically emerging system ﬁrst dominated by mature Individual differences. Although neurobiology and behavior subcortical circuits and then balanced through interactions with tend to unfold in predictable ways across development, signiﬁcant 51,66 late-maturing prefrontal systems . As early as 1920, Flechsig’s individual differences lie atop this scaffolding. This variability histological studies revealed protracted myelin development in applies not only to an adolescent’s current behavioral and neural 67,68 association cortex .Reﬂecting this property of brain matura- characteristics, but also to their past and future phenotypes. That tion, diffusion tensor imaging studies, which measure water is, while one stereotype of adolescents is that they engage in risky diffusion modulated in part by axon myelination, suggest that the behaviors such as binge drinking, there are plenty of young development of posterior cortical-subcortical tracts precedes that people who do not ﬁt this mold. Even among adolescents who of fronto-subcortical tracts supporting top-down control of drink excessively, some may go on to develop substance use 69–72 behavior . Functional brain connectivity studies support disorders, while others may never progress to disordered these results, observing a general pattern of weakening short- drinking. range functional connections followed by strengthening long- Despite recognizing these individual differences, in research, 73–76 range cortical connections across adolescence . clinical, legal, and educational practice, we often treat variance Altogether, this work provides evidence for the progressive around average behavioral and neural phenotypes and trajectories development of connectivity within and between subcortical and as noise, or collapse it into discrete categories (e.g., patients vs. cortical brain regions, and offers a plausible neurobiological controls, adults vs. minors, etc.). Although these groups can be account of nonlinear trajectories in risk-related processes such as useful in practice, they do not necessarily represent biologically self-control, reward sensitivity, and emotion regulation. Emo- plausible or informative qualitative distinctions. Instead, tional reactivity, for example, may arise from the early dominance approaches that characterize the normative trajectories of of subcortical over cortical circuitry, later waning as cortical- dimensional behavioral and neural phenotypes, and investigate subcortical circuits related to top-down control, and then cortical how genetics and experience affect the timing and shape of these circuits involved in processes such as cognitive reappraisal, curves, are necessary for understanding how these processes 52 18,66,77 mature during adolescence and adulthood . More broadly, these unfold in development . In addition to informing models in ﬁndings highlight how approaching the study of adolescence basic science, individual differences approaches can provide from a dynamic, multimodal, circuit-based perspective (rather clinically applicable insight into the factors that confer risk for than a view that focuses on snapshots of individual brain regions and resilience to psychopathology and guide personalized in isolation) can inform our understanding of self-regulation and treatments . 51,52 risk-taking behavior during development . NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications 3 | | | REVIEW ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 Box 1 | Predictive modeling Whereas descriptive modeling is the process of learning associations between features and outcomes, predictive modeling leverages these relationships to make predictions from previously unseen data. Here, we reserve the term “prediction” for the output of models applied to novel individuals rather than to describe brain–behavior correlations . Although prediction pipelines are diverse, they typically involve four primary steps: feature selection, model building, model testing, and prediction evaluation (Fig. 2). Importantly, both feature selection and model building are performed using only training data. The resulting model is then applied unaltered to data from previously unseen individuals. Feature selection: Methods for feature selection, the process of identifying model predictors, fall into two broad categories: hypothesis-driven and data- driven approaches. Hypothesis-driven methods, which leverage existing knowledge to select features, are useful for testing predictions of existing scientiﬁc models. Data-driven methods rely on statistical techniques to identify the features most relevant to individual differences in behavior. These include ﬁlter methods (selecting features based relationships with behavior), wrapper methods (considering the predictive power of different feature combinations, e.g., by systematically eliminating the least predictive features from a model), and embedded methods (incorporating feature selection into model building, such as in lasso, elastic net, and ridge regression) . Both hypothesis- and data-driven approaches can incorporate predictors from multiple domains, including genetics, brain structure and function, and behavior. The developmental trajectories of these measures, such as slope, intercept, or inﬂection point, may also be included. Systematically removing a predictor or predictor class from a model can identify its unique contribution to predicting outcomes or behavior . Although there is no theoretical limit to the number of model features, it is best practice that they not exceed the number of observations to avoid modeling noise (overﬁtting) . Furthermore, it is important to consider the inherent tensions between interpretability, generalizability, and variance explained. While models with fewer features may be easier to interpret, models with more features may capture additional variance in behavior and better characterize complex multimodal neural phenotypes. Model building: Following feature selection, the relationship between predictors and behavior is formalized with a classiﬁer or regression model. The goal of a classiﬁer, such as a support vector machine or logistic regression, is to make discrete predictions. In neuroimaging research, classiﬁers represent the vast majority of predictive models: Of all multivariate models in translational neuroimaging, 75% were built to distinguish patients from control participants, whereas <3% were used to predict continuous symptom scores . Regression models, including linear and support vector regression algorithms, make continuous rather than categorical predictions, and can facilitate the development of transdiagnostic proﬁles of risk or resilience for psychopathology . Both classiﬁers and regression models can be applied to cross-sectional or longitudinal data, and the latter may incorporate techniques such as growth–curve modeling to predict past or future change . Model testing: Model testing, or applying a predictive algorithm to test data to evaluate its generalizability, distinguishes predictive from descriptive 5,8 models. The utility of out-of-sample validation for protecting against overﬁtting and false positives has been discussed in detail elsewhere . Here we highlight one dimension along which potential predictive models vary: how far out of sample they generalize. Internal validation (i.e., k-fold or leave-one-subject-out cross-validation) tests whether a model generalizes to novel individuals from a single data set. Although internal validation is useful for optimizing models and conferring statistical rigor when multiple data sets are not available, it may generate biased estimates of predictive power even when evaluated with permutation testing. Despite this limitation, the vast majority of predictive models in neuroimaging have been tested with internal validation alone . External validation tests whether a model generalizes beyond an initial training data set to individuals from completely independent samples. Curated data sets and platforms such as OpenfMRI that encourage data and model sharing can facilitate external validation and model reﬁnement. Prediction evaluation: Methods of model evaluation depend on whether predictions are discrete or continuous. Classiﬁer output can be evaluated with percent accuracy; sensitivity (the true positive rate, or percent of correctly identiﬁed patients) and speciﬁcity (the true negative rate, or percent of correctly identiﬁed controls); and/or the positive predictive value (percent of individuals called patients who are true patients) and negative predictive value (percent of individuals called controls who are true controls), which depend on disease prevalence. Regression model predictions can be assessed with measures such as correlation or mean-squared error . In all cases it may be useful to visualize all data points to fully evaluate relationships between behavior and predicted scores or category labels. Predictive modeling and its importance in developmental hundreds or thousands, of participants. These small samples, with neuroscience tightly controlled demographics and circumscribed behavioral Studying developmental trajectories, dynamic brain systems, and phenotypes, are not always conducive to studying population individual differences is becoming increasingly feasible with the variability. Larger samples that capture a broad range of pheno- rise of high-throughput data collection efforts . Longitudinal and types provide opportunities not only to describe brain–behavior cross-sectional samples of neuroimaging data from children and relationships, but to predict behavior from brain features at the 79 90,91 adolescents, such as the IMAGEN study , Lifespan Human level of single individuals . In this vein, researchers are Connectome Project Development , Brain Imaging Data searching for neuromarkers, or brain features that predict beha- 81 82 Exchange , Healthy Brain Network Biobank , and Adolescence vior, clinical symptoms, risk for or resilience against illness, or 83 5,6,92 Brain Cognitive Development Study , have accelerated advances treatment response . The pursuit of generalizable neuro- in basic and applied neuroscience (Fig. 1). Collaborative initia- markers goes hand-in-hand with predictive modeling, a techni- tives have also helped democratize data access, improve statistical que that leverages brain–behavior relationships to predict power, and facilitate transparent, reproducible research. The outcomes in novel individuals (Box 1 and Fig. 2). unique challenges posed by large-scale imaging samples, such as The statistician George Box famously claimed that “all models 84 93 how to perform adequate quality control , account for scanner are wrong but some are useful” . Models that predict outcomes 85,86 and site effects , and disentangle meaningful explanatory from previously unseen observations can be especially useful for 87 5,6 power from statistical signiﬁcance , are also motivating the both scientiﬁc discovery and clinical decision-making . From a 88 84 development of new data collection , preprocessing , and basic science perspective, predicting brain–behavior relationships analytic approaches. at the level of single subjects represents progress towards under- Large neuroimaging data sets are not only advancing under- standing how individual differences in brain features relate to standing of how brain features relate to behavior at the group individual differences in cognition and behavior . In addition, level, but are also renewing focus on the individual. Although because predictive models are by deﬁnition validated on inde- cognitive and developmental neuroscientists have long been pendent data, they can help foster robust, reproducible discoveries. interested in interindividual differences in abilities and behavior, The beneﬁts of individualized predictions of current and future traditional experiments have focused on tens, rather than behavior are especially pronounced in developmental populations 4 NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications | | | Predicted label NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 REVIEW ARTICLE ab c d Feature selection Model building Model validation Prediction evaluation Hypothesis-driven Classifier Internal validation Evaluate classification Select based on prior knowledge Categorial outcome Test on left-out individuals Category label Predicted label Category label 12 12 12 Brain feature(s) Brain feature(s) Behavior Data-driven Regression model External validation Evaluate dimensional prediction Select most predictive features Continuous outcome Test on new sites/populations Fig. 2 Schema of key concepts in predicting individual differences in behavior from brain features. a Feature selection. Feature selection techniques fall into two broad categories: hypothesis-driven (top-down) and data-driven (bottom up) approaches. b Model building. Machine-learning algorithms can be used to predict categorical measures, such as clinical diagnoses, or dimensional measures, such as task performance or symptom severity. Here, the dark blue line shows the relationship between a single hypothetical brain feature and a behavioral score. The light blue line illustrates a classiﬁer that divides individuals into categories based on this brain feature. (Note that, unlike in this condensed visualization, behavioral scores are typically related to category labels.) c Model validation. Predictive models are evaluated on previously unseen data—either left-out individuals from the initial data set (internal validation) or individuals from a completely new sample (external validation). d Prediction evaluation. Continuous predictions (bottom and left axes) are evaluated by comparing observed and predicted behavioral measures, e.g., with correlation or mean-squared error. Categorical predictions (top and right axes) are evaluated with percent correct; binary predictions can be assessed with sensitivity and speciﬁcity and/or positive and negative predictive value including adolescents. Because behavior and psychopathology are unemotional situations relative to social or emotionally charged 23,96 best viewed as the result of developmental processes that unfold contexts . 21,66 97 across the lifespan , characterizing individual arcs in Recent work from Rudolph and colleagues used predictive brain–behavior relationships over time can move us even closer modeling to identify the neural basis of this phenomenon, asking to understanding targets for change. Addressing the unique whether functional brain organization looks less mature in challenges presented by prediction in adolescence, including the emotional contexts, and whether this effect relates to individual complex dynamics linking neurobiological, behavioral, and differences in risky behavior. To this end, the authors calculated environmental change, can also help us better model periods such functional connectivity patterns from fMRI data collected while as prenatal development, infancy, aging, and illness course. 212 individuals aged 10–25 performed a go/no-go task in neutral Predictive models may not only contribute to progress in basic and emotional contexts. During emotional contexts, participants developmental neuroscience, but may also have implications for anticipated an aversive noise or a reward; during neutral contexts education, mental health, and legal policy. For example, early there was no anticipation of noise or reward. Using partial least predictions of behavioral impairments could facilitate earlier squares regression and 10-fold cross-validation, the authors ﬁrst treatments and improved health or educational outcomes . Pre- built a model to predict chronological age from functional con- diction can also inform pressing policy questions, such as char- nectivity patterns observed in the neutral context, and then acterizing the maturity of a particular individual in speciﬁc applied the same model to connectivity observed during the contexts to inform whether they should be treated as an adult in emotion manipulation. They found that a prediction made from 95,96 the justice system . Thus, although machine-learning models an individual’s neutral context pattern (their “neutral brain age”) of behavior in development may be “wrong” in the sense that they was closer to their chronological age than a prediction made from (necessarily) simplify complex neurobiological systems, they are their emotional context pattern (their “emotional brain age”). useful in that they can inform theories of how cognition and Further, both predictions tended to be younger than chron- behavior emerge from dynamic brain systems and speak to ological age in teens. Interestingly, there was a trend such that general educational, medical, and social policies. adolescents were more likely to look younger in emotional rela- tive to neutral contexts, but young adults who showed this pattern had greater risk preference and lower risk perception (Fig. 3). Predictive modeling and risk preference These ﬁndings illustrate the power of predictive modeling in One of the most common reprimands of wayward youths is: “Act delineating dynamic developmental changes and individual dif- your age!” The phrase — immortalized in the English-language ferences in risk taking behaviors. idiom “act your age, not your shoe size”— is so ubiquitous that it In addition to helping explain why adolescents may not “act even makes an appearance in song lyrics from the musical artist their age” under emotional arousal, the Rudolph et al. ﬁndings Prince. Its sentiment, however, is not straightforward. What does raise two notable points about predictive brain-based models in it mean for an adolescent or young adult to act his or her age? general. What counts as typical adolescent behavior? One possibility is First, when the model of chronological age was wrong, it was that “act your age” means, “make the most responsible decision wrong in interesting ways: A young adult incorrectly predicted you have the capacity to make”. What this entreaty fails to younger in an emotional context was more likely to show a “risky recognize, however, is that there is a discrepancy between how phenotype” than an individual incorrectly predicted older. Thus, responsibly adolescents and young adults can act in nonsocial, in some cases, model errors may be as informative as model NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications 5 | | | Training sample Behavior (current, future, or change over time) Predicted behavior Predicted behavior REVIEW ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 ab c Positive emotional context Negative emotional context Age Age Fig. 3 Adolescents’ functional connectivity patterns look younger in emotional contexts. Adapted with permission from Rudolph et al. . a Chronological age is plotted against age predicted from functional connectivity patterns observed in positive and negative emotional contexts. Individual points (participants) are ﬁt with polynomial curves. On average, adolescents are predicted younger in emotional contexts. b Adolescents (age 12–18; numerical difference) and young adults (age 18–21; p < 0.1 in negative emotional contexts and p < 0.05 in positive emotional contexts) who are predicted younger in emotional contexts tend to show greater risk preference. This trend is most pronounced in young adulthood. Open bars represent individuals predicted younger in emotional contexts, and ﬁlled bars represent individuals predicted older. Red bars show participants grouped by age predictions in positive emotional contexts; blue bars show participants grouped by age predictions in negative emotional contexts. c Functional network nodes, scaled by their importance in the age-prediction model, are grouped into the following functional networks deﬁned previously : default mode (red), dorsal attention (green), frontoparietal (yellow), salience (black), cingulo-opercular (purple), visual (blue), subcortical (orange), and ventral attention (teal). successes in unraveling the brain bases of individual differences in understanding the neurobiological basis of adolescent behavior. behavior. In particular, we encourage researchers to bridge data sets and Second, this model—along with many in cognitive and levels of analyses to develop generalizable, trajectory-based developmental neuroscience—predicts outcomes from functional models that predict current and future outcomes. connectivity data. Given that functional connectivity patterns can be affected by cognitive state , such models may not generalize across contexts as well as models based on state-independent Leverage multiple data sets to build and validate predictive features such as structural connectivity. (There is evidence, models. Predictive models will be most theoretically and practi- however, that functional connectome-based models generalize cally useful when they generalize beyond a single data set. across task-engaged and resting states to predict abilities such as Although historically replication and external validation samples attention .) Thus, researchers hoping to build an age-prediction were rare in fMRI due to cost and time constraints, open-access model with optimal predictive power and generalizability may data sets and a growing culture of data sharing are removing consider including structural features that may capture more barriers to access. Consider, for example, a group of investigators “trait”-related than state-related variance as predictors (see the interested in predicting impulsivity from resting-state functional section entitled “Include multimodal predictors”). connectivity data. These researchers could download data from Finally, it is important to note that although here maturity was the Human Connectome Project , model the relationship assessed with a single number—akin to the difference between an between impulsivity and functional connectivity, and then apply individual’s functional connectivity pattern and the age-typical their model to completely independent data from the Brain pattern—maturity does not lie on one continuum from “less” Genomics Superstruct Project to evaluate its generalizability. (in emotional states) to “more” (in unemotional states). Training and testing predictive models with open data sets has Rather, temporal differences in the ﬁne-tuning of interacting obvious beneﬁts. For our hypothetical investigators, downloading neural systems with age and experience impact behavioral data may cost a fraction as much as running their own, smaller phenotypes differently across development and vary across fMRI study. Open data sets also tend to offer relatively large individuals and contexts . For example, Rudolph and colleagues sample sizes, capturing a wide range of behavior and allowing show that, on average, adolescents’ functional connectivity researchers to ﬁt complex models and reﬁne model parameters proﬁles look younger in emotional contexts, and that young with nested cross-validation techniques. In addition, open adults who maintain this proﬁle show riskier choices. This samples can provide opportunities to validate models across work suggests that future studies can characterize each indivi- unique behavioral measures. Although this approach can be dual’s unique multivariate maturational proﬁle, that is, the age- challenging given that different-but-related measures may index typicality of both their trait- and state-dependent neural similar-but-not-identical mental processes, it is a useful way to phenotypes. investigate whether a model is capturing individual differences in a speciﬁc performance metric or a general cognitive function. For example, imagine that researchers build a model to predict The road ahead impulsivity questionnaire scores. If they apply this model to a Just as building predictive models can inform how we understand new sample in which impulsivity is measured with task risk taking in adolescence, studying adolescence can inform how performance, predictive power will be limited by the ground- we approach behavioral prediction. Here, motivated by predictive truth relationship between questionnaire scores and task and descriptive models of development, we suggest eight direc- performance. Successful generalization would provide additional tions for future research and highlight their importance for evidence that the model is related to individual differences in 6 NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications | | | Predicted younger Predicted older Risk preference NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 REVIEW ARTICLE impulsivity per se rather than individual differences in ques- Cerebellar volume Group average tionnaire scores alone. Thus, validating models in open data sets (Castellanos et al. ) can help establish their speciﬁcity and generalizability. Healthy This should not be taken imply that targeted studies are ADHD obsolete. Instead, experiments designed to probe speciﬁc Cortical thickness (Shaw et al. ) behavioral phenotypes with carefully designed psychological Healthy tasks are crucial complements to open data analyses. Because ADHD targeted studies have greater ﬂexibility in the participants they Functional connectome distinctiveness recruit, the behavioral measures they collect, and the tasks they (Kaufmann et al. ) administer, they can help elucidate brain–behavior relationships Healthy across populations and cognitive states. The impulsivity research Psychopathology including ADHD group, for example, could use data from a targeted study to ask Childhood Adolescence Adulthood whether the same functional network that predicts impulsivity in adults emerges in development to support children’s impulse Hypothetical individual control. (In fact, they may not even need to collect their own data with attention deficit to do so: Relevant targeted samples may be available on data- Developmental trajectories sharing platforms such as OpenfMRI .) Recent work examining Remission the heritability of the functional connectome used a similar approach, building a model of siblingship in a locally acquired Persistence data set, and validating it in the Human Connectome Project sample . The sustained attention connectome-based predictive model is Current multivariate another recent example of a model validated across multiple 98,103–107 neural phenotype imaging data sets . This model was deﬁned to predict Childhood Adolescence Adulthood individual differences in the ability to maintain focus from patterns of task-evoked and resting-state functional connectiv- Fig. 4 a Developmental changes in cerebellar volume, cortical thickness, ity . During fMRI, adult participants performed a challenging and functional connectome distinctiveness in healthy individuals and sustained attention task, which presumably perturbed attention- individuals with attention deﬁcits. Curves are based on data from relevant neural circuitry and ampliﬁed behaviorally relevant 109,113,119 refs. . b Developmental changes in a hypothetical adolescent with individual differences in functional connectivity. Models deﬁned attention deﬁcits. A model trained to use the developmental trajectories of on task-based data generalized to predict left-out participants’ multiple brain measures to predict future outcomes may best characterize task performance not only from data acquired as they were whether this individual’sdeﬁcits will improve, persist, or worsen. These engaged in the task, but also from data collected as they simply predictions may have implications for future treatment or cessation of rested. External validation with data from the ADHD-200 treatment Consortium revealed that the same functional networks that index attention task performance in adulthood predict ADHD symptoms in childhood. Together these results suggest that a Develop trajectory-based models with longitudinal data. Neu- common functional architecture supports sustained attention robiology is inherently dynamic, and understanding any dynamic across developmental stage (adults vs. children and adolescents), process in terms of both description and prediction requires clinical population (ADHD vs. control), and cognitive state (task appreciating changes over time. Atmospheric models, for exam- 103 111 vs. rest) . ple, rely on dynamical equations to predict the weather , and Another targeted study provided insights into potential stock forecasting models use measures of how a stock’s perfor- mechanisms of the model’s predictive networks. That is, the mance has changed in the past to predict how it will perform in same sustained attention connectome-based predictive model the future. We often use longitudinal data to make folk psycho- distinguished individuals who had taken a single dose of logical predictions, such as when we consider how quickly a methylphenidate (Ritalin) from controls, raising the possibility young tennis player climbed the rankings to estimate her shot at that networks reﬂect the expression of neurotransmitters whose winning Wimbledon, or use what we know about a friend’s recent extracellular concentration is modulated by methylphenidate . stress levels to predict how he will react in an emotional situation. Although the anatomy of the sustained attention model is Models that predict behavior from brain features can also complex, broad trends align with previous ﬁndings and suggest beneﬁt from longitudinal measures. Consider again the case of new targets for intervention . Functional connections between attention deﬁcits. Pioneering work applied growth–curve models sensorimotor and cerebellar regions predict more successful to cross-sectional and longitudinal data to establish delays in 112,113 sustained attention, whereas intra-cerebellar, intra-temporal, and cortical thickness and brain surface area maturation , as well temporal-parietal connections predict less successful attention. as a down-shifted trajectory of cerebellar growth in children 109,110 114–117 The participation of the cerebellum, implicated in ADHD , and adolescents with ADHD (Fig. 4; but see refs. for provides convergent evidence of its importance for attention. methodological considerations related to effects of head motion). Frontal and parietal regions traditionally related to attention and Recent work also suggests that the age-typicality of a child’sor attention impairments do appear in the predictive networks, but adolescent’s functional connectivity patterns is related to their 118,119 they represent >35% of all connections in the model, accentuat- psychiatric symptoms, including attention deﬁcits . In other ing the importance of data-driven approaches to feature selection. words, children and adolescents with attention deﬁcits show In light of the sustained attention model’s out-of-sample delayed maturational patterns of cortical thickness and functional generalizability—a recent proof-of-principle example—we are connectivity on average, and single snapshots of functional optimistic that, moving forward, a combination of high- connectivity predict single snapshots of attentional abilities in throughput data sets, targeted experiments, and “green science” novel individuals. It follows that a teenager’s unique trajectory of data sharing initiatives will facilitate robust, generalizable models functional connectivity and cortical thickness development may of cognitive abilities and behavior across development. provide more nuanced information about his or her attentional NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications 7 | | | Maturity Maturity REVIEW ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 abilities, predicting not only deﬁcit severity, but also perhaps could, for example, be titrated off of medication sooner than symptom persistence or abatement. Developmental neuroscien- otherwise possible (Fig. 4). tists pursuing trajectory-based predictive models can take advantage of large longitudinal samples such as IMAGEN or Include multimodal predictors. Models in human neuroscience the open-access ABCD collection effort, and of biostatistical often focus on a single type of brain feature, such as functional techniques developed to predict clinical outcomes from long- connectivity, to predict behavior. Although this approach is useful 120–123 itudinal biomarkers . for targeting speciﬁc neural mechanisms, constraining a model’s In addition to potentially increasing predictive power, feature space to a single modality may limit predictive power. individualized trajectory-based models can inform theories of Open-access data sets including a variety of scan types (e.g., T1- how neural phenotypes give rise to typical and atypical weighted, T2-weighted, proton density, T2-FLAIR, DTI, BOLD) development. For example, although we know that delayed facilitate the construction of models incorporating a range of cortical maturation trajectories characterize ADHD patients at features (e.g., myelination patterns, structural connectivity, task- the group level , it is not yet known whether a delayed based and resting-state functional connectivity) to maximize trajectory confers risk for attention deﬁcits at the level of predictive power and uncover the unique contributions of dif- individual subjects. Extending similar group-level ﬁndings to the ferent neural systems to current and future behavior. level of individual subjects can enhance the clinical utility of Researchers have demonstrated that including multiple feature research ﬁndings and inform novel interventions . classes improves individualized predictions in development not just in theory, but also in practice. For example, in the ﬁrst-ever Predict future outcomes. Models that predict current behavioral example of predictive modeling in developmental neuroscience, tendencies are technically postdictive in that they make retro- Dosenbach and colleagues asked whether resting-state func- spective, rather than prospective, predictions. Although these tional connectivity patterns can predict an individual subject’s models can inform relationships between neural and behavioral chronological age. Using data from 238 individuals aged 7–30, phenotypes, models that predict future outcomes may be most they trained models to predict categorical (child vs. adult) and useful in clinical and translational contexts, allowing for earlier dimensional (chronological age) measures of maturity. A support intervention, treatment, or cessation of treatment. vector machine classiﬁer correctly predicted whether an indivi- Recent work demonstrates that models that make future dual was a child or an adult 91% of the time, and a support vector forecasts are possible in the context of development. For example, regression algorithm accounted for 55% of the variance in Whelan and colleagues modeled neural and psychological chronological age. Motivated to ﬁnd the unique contribution of proﬁles of alcohol misuse before its onset in adolescence. Using multiple neuroanatomical features to age predictions, Brown and measures of brain structure and function, personality, cognitive colleagues built a model based on structural features that abilities, environmental factors, life experiences, and genetic accounted for 92% of the variance in age in a sample of 885 variants, the authors built a model that distinguished adolescents individuals aged 3–20. Interestingly, different features contributed who go on to binge drink from those who do not. New ﬁndings to model performance at different ages: Whereas T2 signal suggest that brain features can predict clinically relevant intensity in subcortical ROIs was most diagnostic of age in outcomes even earlier in development: cortical surface area and childhood, ﬁber tract diffusivity and subcortical structure volume functional connectivity observed at 6–12 months, for example, were most informative in adolescence, and ROI diffusivity was 125,126 130 131 predict autism diagnosis at age two . most informative in adulthood . Franke et al. additionally In addition to models that predict the onset of clinical found that predictions of a “brain age” model based on multiple symptoms or risky behavior, models that predict improvements neuroanatomical features were signiﬁcantly younger for adoles- in clinical outcomes can help identify resilience factors for cents born preterm than full term. Although the variance psychopathology. Recently, Plitt and colleagues used func- explained by functional connectivity and neuroanatomy cannot tional connectivity patterns to predict improvements in adoles- be directly compared given differences in methodology and cents’ and young adults’ autism symptoms. They found that participant samples across studies, multimodal approaches may functional connectivity in the salience, default mode, and capture more variance in individual differences than do unimodal frontoparietal networks, implicated in attention and goal- ones. 11,12 directed cognition , predicted symptom changes over time, Black box models that use an assortment of brain features to even when accounting for age, IQ, baseline symptoms, and predict outcomes may not necessarily provide interpretable links follow-up latency. Hoeft and colleagues also showed that between neurobiology and behavior. How can researchers unravel prefrontal activity and right superior longitudinal fasciculus the unique contributions of different feature classes to individual fractional anisotropy, but not reading or language test scores, differences? In modeling alcohol misuse in adolescence, Whelan predicted which children with dyslexia would show reading skill and colleagues provide one example of how this may be achieved. improvement over the course of 2.5 years. Functional and To start, their model of future binge drinking included neural, structural measures, therefore, can predict not only the onset of behavioral, lifestyle, and genetic factors. They then systematically clinical symptoms, but also the abatement. In addition, brain removed each feature class from the model to isolate its features can predict future outcomes over and above behavioral contribution to predictive power. The approach revealed that life measures alone—an important check when evaluating the utility history, personality, and brain variables were most uniquely of predictive models. predictive . Other approaches, such as dimensionality reduction In the future, trajectory-based approaches may better char- strategies and penalized regression methods, can also help acterize not just where a child or adolescent has been or is eliminate redundant predictor variables and identify those most currently, but where he or she is going. For example, models that tightly coupled with individual differences in behavior. use a child’s developmental growth curve (e.g., precocious, Although hypothesis-driven modeling approaches that use a 18,77, delayed, deviant, regressive, or resilient ) to predict the single feature or feature type to predict outcomes can advance persistence or worsening of clinical symptoms could have knowledge about the neurobiological bases of behavior, multi- implications for treatment. Such models may also have implica- variate models that incorporate both structural and functional tions for the cessation of treatment. A child with attentional features may improve prediction accuracy and offer converging impairments but a resilient developmental trajectory for ADHD insights (e.g., Fig. 4). 8 NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 REVIEW ARTICLE Continue to focus on dimensional outcomes. Centuries of that the functional architecture of attention in adolescence may observation and research tell us that cognitive abilities and differ from that in childhood and adulthood. These ﬁndings also behavior vary along a continuum at every stage of life. Consider underscore the importance of understanding the nonlinear the example of impulsivity: Although variability in impulsivity expression of dynamic and hierarchical changes in brain features can be captured, to some degree, with a categorical measure like and behavior with development. an ADHD diagnosis, it may be better quantiﬁed with dimensional Moving forward, it will be important to tailor predictive measures such as symptom severity or impulse-control task models to particular scientiﬁc questions and/or practical goals. performance. For example, models trained on one developmental period and When characterizing brain–behavior relationships in develop- tested on another can inform questions about common functional ment, the costs and beneﬁts of using categorical vs. dimensional mechanisms, whereas models trained on a range of age groups measures should be carefully considered. Although categorical may better characterize trajectories in brain–behavior relation- labels align with the current diagnostic system in clinical ships and offer greater predictive power across the lifespan. medicine and can help make results easier to interpret and Further, it is important to keep in mind that because children and present, dichotomizing continuous variables can reduce statistical adolescents are not simply “little adults” in terms of either power, obscure nonlinear relationships between variables and neurocircuitry or behavior, predictions in these populations will 132,133 outcomes, and increase the risk of false positive results .In likely often rely on development-speciﬁc models rather than addition, categorical models are often built on balanced samples models deﬁned in adults and applied to developmental data. of patients and control participants to avoid biased predictions, Future work testing whether models are valid across develop- but this ratio rarely reﬂects real-world illness prevalence. Thus, mental stages, clinical populations, and cognitive or affective reported measures of a model’s sensitivity and speciﬁcity may states can provide additional insight into the scope of their exaggerate its translational utility, and positive and negative generalizability. predictive values may be more useful measures of performance (see Box 1). Dimensional measures may better characterize the Bridge statistical predictability and biological plausibility. full range of behavior and clinically relevant outcomes, especially Predictive modeling in developmental neuroscience has two in development, when small differences in behavioral or neural parallel goals: to discover how the brain gives rise to behavior phenotypes can have important implications for treatment. across development, and to identify practically useful neuro- Although dimensional measures are frequently used to markers of behavior and clinically relevant outcomes. It is not characterize individual differences in cognitive and developmen- always obvious, however, how models that achieve the second 18,135 tal neuroscience and psychology , such approaches are goal can help make progress toward the ﬁrst. Instead, predictive infrequent in predictive modeling . Models that predict chron- models are sometimes considered opaque “black boxes” far 97,129–131 136–138 103,118,139 ological age , ﬂuid intelligence , attention , removed from biology and uninformative about the neural cir- 140 127 and improvements in math skills and autism symptoms , cuits supporting behavior. Some models are more susceptible to however, demonstrate that such approaches are powerful ways to this concern than others. Supekar and colleagues , for example, identify robust transdiagnostic biomarkers of abilities and used hippocampal volume and functional connectivity to predict behavior. children’s response to math tutoring. In doing so, they provide Looking ahead, modeling approaches that consider multiple clear evidence of the role of learning- and memory-relevant brain dimensional approaches at once, or those that identify latent regions in math skill improvements. On the other hand, models distributions from which behavior emerges, may help delineate that use deep neural networks to generate predictions may subtypes of clinical disorders and improve outcome predictions sometimes preclude easy (linguistic) interpretations of relation- and treatment . Regression models that predict dimensional ships between predictors and outcomes. As bigger data sets and outcomes and consider subgroups that make up heterogeneous more sophisticated algorithms result in greater and greater pre- patient populations will also continue to be valuable complements dictive power, it will be important for researchers to keep the ﬁrst to classiﬁers that predict group membership. goal of modeling—advances in basic science—in their sights. Large-scale data sets that include behavioral, neuroimaging, Establish boundary conditions. Given that brain structure and and genetic data provide exciting opportunities for researchers to function change dramatically across development, models trained explore the biological plausibility of predictive models. To in one developmental period (e.g., adulthood) should not always illustrate the promise of approaches that link levels of analysis, generalize to others (e.g., adolescence). Rather, upper bounds on let’s return to the hypothetical research group interested in models’ predictive power will be inﬂuenced by the reliability of impulsivity. Imagine that the research team identiﬁes a pattern of the brain and behavioral measures, their stability across devel- functional connectivity that predicts impulsivity across indivi- opment, and the developmental trajectories of the underlying duals. A subsequent issue of clear importance is the extent to neurobiology. which this network reﬂects the underlying function of molecular- Testing models across different developmental periods can help genetic mechanisms. To get initial traction on this question, the identify critical change points in the relationship between researchers could ask whether this network is heritable, or neurobiological processes and behavior. As a concrete example, whether structural genetic variants predict its function, which in the sustained attention connectome-based predictive model turn predicts impulsive behavior. They could also pursue cross- introduced earlier generalized from an adult to a developmental species work, asking whether important model features map on to 98,103 sample . The model, however, did not perform equally well known anatomical circuits, or whether hypothetical genes that in all age groups. Although predictions were signiﬁcantly related impact network function in humans affect behavior in rodents. to ADHD symptoms in children 8–9(n = 30), 10–11 (n = 28), and Thus, approaches that combine data sets to bridge levels of 12–13 (n = 41), they did not reach signiﬁcance in adolescents analysis and link genotype, neural phenotype, and behavior 14–16 (n = 14; unpublished results). Although certainly not across development may suggest new etiological hypotheses and conclusive given the exploratory nature of this analysis and the possible treatment targets of cognitive function and dysfunction. fact that predictive power is inﬂuenced by factors including sample size, data quality, and group variance in ADHD scores, Acknowledge limitations to advance understanding. Enthu- this outcome motivates future research by tentatively suggesting siasm for large neuroimaging data sets and individualized NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications 9 | | | REVIEW ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 predictions should be coupled with realistic assessments of Conclusions potential pitfalls related to methodology, interpretation, and A rich tradition of research in human neuroimaging has made implementation. Carefully considering these limitations, progress in explaining the neurobiology of cognition and beha- researchers are already beginning to develop new analytic vior. Less attention, however, has been devoted to predicting approaches and ﬁeld-wide standards to address cognitive abilities and behavior from brain features. Here we 84,87,89,141,142 them . argue that predictive modeling approaches that forecast outcomes Methodological pitfalls can erode the impact of predictive at the level of individuals are important complements to work models. For example, because head motion introduces signiﬁcant describing brain–behavior relationships at the group level, espe- confounds in both structural and functional imaging data, cially in the context of adolescence. 115,143 especially in developmental populations , up-to-date data- Not only can predictive models enhance the clinical and collection and preprocessing techniques are necessary for translational utility of neuroimaging research, they can also ensuring that predictions do not rely on motion-induced artifacts. account for critical features of behavior often overlooked in In addition, just as descriptive models may reﬂect sample noise, cross-sectional studies of the developed adult brain: develop- predictive models may be overﬁt to training data. Although mental trajectories, hierarchically emerging brain systems, and nested cross-validation techniques can help protect against individual differences in both. So far, investigating when models overﬁtting, external validation is critical for testing model accurately predict behavior and when they fail to do so has generalizability (Box 1). Finally, it is important that methodolo- illuminated potentially adolescent-speciﬁc changes in beha- gical choices be tailored to research goals. For example, is the goal vioral and neural phenotypes related to risk-taking and attention. to predict current phenotypes or future change? To make Looking ahead, models that offer probabilistic insights into absolute predictions (e.g., that a child will grow to be six feet individuals’ current and future behavior from their past devel- tall) or relative ones (that he or she will be in the 95th percentile opmental brain trajectories have the potential to provide deep for height)? To predict behavior from functional brain features insights into human brain development and function in both observed during task engagement, or to test whether a cognitive health and disease. process can be measured in the absence of an explicit task ?To maximize subgroup-level accuracy, or population-level general- Received: 6 October 2017 Accepted: 5 January 2018 izability? To prioritize statistical predictability, biological plausi- bility, or feature weight interpretability? Mismatches between study methods and goals can undermine the usefulness of predictive models. References Working with large data sets also poses several challenges to 1. Hofstadter, A. Explanation and necessity. Philos. Phenomenol. Res. 11, interpretation. For example, when samples are large enough, 339–347 (1951). brain–behavior relationships with even tiny effects sizes may 2. Shmueli, G. To explain or to predict? Stat. Sci. 25, 289–310 (2010). reach statistical signiﬁcance. Although such effects may be 3. Craver, C. F. When mechanistic models explain. Synthese 153, 355–376 interesting from a basic science perspective if robust to noise (2006). 4. Salmon, W. C. Why ask, ‘Why?’? An inquiry concerning scientiﬁc explanation. and replicable, they are unlikely to offer much clinical or practical 87 Proc. Address. Am. Philos. Assoc. 51, 683–705 (1978). beneﬁt in the near term . In addition, although large data sets 5. Woo, C.-W., Chang, L. J., Lindquist, M. A. & Wager, T. D. Building better may be more representative of the general population than biomarkers: brain models in translational neuroimaging. Nat. Neurosci. 20, smaller samples, they may offer a “false sense of security” since 365–377 (2017). they can still suffer from selection bias and skewed demo- Reviews advances in biomarker discovery in human neuroscience. 6. Gabrieli, J. D. E., Ghosh, S. S. & Whitﬁeld-Gabrieli, S. Prediction as a graphics and are not equally distributed across the globe (See humanitarian and pragmatic contribution from human cognitive Fig. 1). The careful evaluation of both statistical and clinical neuroscience. Neuron 85,11–26 (2015). signiﬁcance will be important for establishing scientiﬁcally valid, Highlights beneﬁts of predictive brain-based models for personalized practically useful models. medicine and education. Although neuroimaging-based predictive models have the 7. Douglas, H. E. Reintroducing prediction to explanation. Philos. Sci. 76, 444–463 (2009). potential to offer signiﬁcant beneﬁts, challenges can arise when 8. Yarkoni, T. & Westfall, J. Choosing prediction over explanation in psychology: applying them in clinical contexts. Recent work has highlighted lessons from machine learning. Perspect. Psychol. Sci. 12, 1100–1122 (2017). the “perilous path from publication to practice”, outlining a Argues for an increased focus on prediction over explanation for variety of scientiﬁc, implementation, and business-related obsta- understanding human behavior. cles . As scientists pursue individualized predictions, close 9. Goldman-Rakic, P. S. Topography of cognition: parallel distributed networks in primate association cortex. Annu. Rev. Neurosci. 11, 137–156 (1988). collaborations with clinicians, bioethicists, industry professionals, 10. Price, J. L. & Drevets, W. C. Neurocircuitry of mood disorders. and regulatory bodies will be necessary for effectively translating Neuropsychopharmacology 35, 192–216 (2009). 141,146 models to real-life patient-care settings . 11. Buckner, R. L., Andrews-Hanna, J. R. & Schacter, D. L. The brain’s default Finally, as models begin to make their way from bench to network: anatomy, function, and relevance to disease. Ann. N. Y. Acad. Sci. bedside, it is important to consider their ethical implications, 1124,1–38 (2008). 12. Corbetta, M. & Shulman, G. L. Control of goal-directed and stimulus-driven especially in the context of development. First, researchers need attention in the brain. Nat. Rev. Neurosci. 3, 201–215 (2002). to keep in mind—and clearly communicate to participants, 13. Power, J. D. et al. Functional network organization of the human brain. patients, and the public—that predictive models are probabilistic Neuron 72, 665–678 (2011). rather than deterministic in nature, and, just like traditional pen- 14. Yeo, B. T. T. et al. The organization of the human cerebral cortex estimated by and-paper tests, will never perfectly predict abilities or behavior. intrinsic functional connectivity. J. Neurophysiol. 106, 1125–1165 (2011). 15. Insel, T. et al. Research domain criteria (RDoC): toward a new classiﬁcation Predictions are thus best considered tools for scientiﬁc discovery framework for research on mental disorders. Am. J. Psychiatry 167, 748–751 and opportunities for informing clinical decision-making rather (2010). than portents of the future. Moving forward, the potential 16. Hyman, S. E. The diagnosis of mental disorders: the problem of reiﬁcation. beneﬁts of predictive modeling in development must be Annu. Rev. Clin. Psychol. 6, 155–179 (2010). continuously evaluated in light any of potential risks to privacy 17. Fair, D. A., Bathula, D., Nikolas, M. A. & Nigg, J. T. Distinct neuropsychological subgroups in typically developing youth inform or hyperbolic claims about our ability to predict the future. 10 NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 REVIEW ARTICLE heterogeneity in children with ADHD. Proc. Natl Acad. Sci. USA 109, development: relevance for dopamine’s role in adolescent decision making. 6769–6774 (2012). Dev. Cogn. Neurosci. 1, 578–590 (2011). 18. Casey, B. J., Oliveri, M. E. & Insel, T. A neurodevelopmental perspective on 48. Teslovich, T. et al. Adolescents let sufﬁcient evidence accumulate before the research domain criteria (RDoC) framework. Biol. Psychiatry 76, 350–353 making a decision when large incentives are at stake. Dev. Sci. 17,59–70 (2014). (2014). 19. Cohen, J. D. et al. Computational approaches to fMRI analysis. Nat. Neurosci. 49. Somerville, L. H., Jones, R. M. & Casey, B. J. A time of change: behavioral and 20, 304–313 (2017). neural correlates of adolescent sensitivity to appetitive and aversive 20. Shen, X. et al. Using connectome-based predictive modeling to predict environmental cues. Brain. Cogn. 72, 124–133 (2010). individual behavior from brain connectivity. Nat. Protoc. 12, 506–518 (2017). 50. Dreyfuss, M. et al. Teens impulsively react rather than retreat from threat. 21. Lee, F. S. et al. Adolescent mental health—Opportunity and obligation: Dev. Neurosci. 36, 220–227 (2014). emerging neuroscience offers hope for treatments. Science 346, 547–549 51. Casey, B., Galván, A. & Somerville, L. H. Beyond simple models of adolescence (2014). to an integrated circuit-based account: a commentary. Dev. Cogn. Neurosci. 22. Casey, B. J. Beyond simple models of self-control to circuit-based accounts of 17, 128–130 (2016). adolescent behavior. Annu. Rev. Psychol. 66, 295–319 (2015). 52. Casey, B., Heller, A. S., Gee, D. G. & Cohen, A. O. Development of the Presents evidence that adolescent-speciﬁc changes in self-control emerge emotional brain. Neurosci. Lett. https://doi.org/10.1016/j.neulet.2017.11.055 from the ﬁne-tuning of connections within and between subcortical, cortical (2017) prefrontal, and limbic circuits. 53. Bourgeois, J.-P., Goldman-Rakic, P. S. & Rakic, P. Synaptogenesis in the 23. Steinberg, L. Risk taking in adolescence. Curr. Dir. Psychol. Sci. 16,55–59 prefrontal cortex of rhesus monkeys. Cereb. Cortex. 4,78–96 (1994). (2007). 54. Huttenlocher, P. R. & Dabholkar, A. S. Regional differences in synaptogenesis 24. Ellis, B. J. et al. The evolutionary basis of risky adolescent behavior: in human cerebral cortex. J. Comp. Neurol. 387, 167–178 (1997). implications for science, policy, and practice. Dev. Psychol. 48, 598–623 55. Østby, Y. et al. Heterogeneity in subcortical brain development: a structural (2012). magnetic resonance imaging study of brain maturation from 8 to 30 years. J. 25. Arnett, J. J. Sensation seeking, aggressiveness, and adolescent reckless Neurosci. 29, 11772–11782 (2009). behavior. Pers. Individ. Dif. 20, 693–702 (1996). 56. Giedd, J. N. et al. Brain development during childhood and adolescence: a 26. Greene, K. et al. Targeting adolescent risk-taking behaviors: the contributions longitudinal MRI study. Nat. Neurosci. 2, 861–863 (1999). of egocentrism and sensation-seeking. J. Adolesc. 23, 439–461 (2000). 57. Tamnes, C. K. et al. Brain maturation in adolescence and young adulthood: 27. Van Leijenhorst, L. et al. What motivates the adolescent? brain regions regional age-related changes in cortical thickness and white matter volume mediating reward sensitivity across adolescence. Cereb. Cortex. 20,61–69 and microstructure. Cereb. Cortex. 20, 534–548 (2010). (2010). 58. Shaw, P. et al. Neurodevelopmental trajectories of the human cerebral cortex. 28. Galvan, A. et al. Earlier development of the accumbens relative to J. Neurosci. 28, 3586 LP–3583594 (2008). orbitofrontal cortex might underlie risk-taking behavior in adolescents. J. 59. Gogtay, N. et al. Dynamic mapping of human cortical development during Neurosci. 26, 6885–6892 (2006). childhood through early adulthood. Proc. Natl Acad. Sci. USA 101, 8174–8179 29. Somerville, L. H., Hare, T. & Casey, B. J. Frontostriatal maturation predicts (2004). cognitive control failure to appetitive cues in adolescents. J. Cogn. Neurosci. 60. Sowell, E. R., Thompson, P. M., Holmes, C. J., Jernigan, T. L. & Toga, A. W. In 23, 2123–2134 (2010). vivo evidence for post-adolescent brain maturation in frontal and striatal 30. Silvers, J. A. et al. Age-related differences in emotional reactivity, regulation, regions. Nat. Neurosci. 2, 859–861 (1999). and rejection sensitivity in adolescence. Emotion 12, 1235–1247 (2012). 61. Gogtay, N. & Thompson, P. M. Mapping gray matter development: 31. Chein, J., Albert, D., O’Brien, L., Uckert, K. & Steinberg, L. Peers increase Implications for typical development and vulnerability to psychopathology. adolescent risk taking by enhancing activity in the brain’s reward circuitry. Brain. Cogn. 72,6–15 (2010). Dev. Sci. 14,F1–F10 (2011). 62. Selemon, L. D. A role for synaptic plasticity in the adolescent development of 32. Steinberg, L. A social neuroscience perspective on adolescent risk-taking. Dev. executive function. Transl. Psychiatry 3, e238 (2013). Rev. 28,78–106 (2008). 63. Benes, F. M., Taylor, J. B. & Cunningham, M. C. Convergence and plasticity of 33. Blakemore, S.-J. & Robbins, T. W. Decision-making in the adolescent brain. monoaminergic systems in the medial prefrontal cortex during the postnatal Nat. Neurosci. 15, 1184–1191 (2012). period: implications for the development of psychopathology. Cereb. Cortex. 34. Somerville, L. H. The teenage brain: sensitivity to social evaluation. Curr. Dir. 10, 1014–1027 (2000). Psychol. Sci. 22, 121–127 (2013). 64. Brenhouse, H. C., Sonntag, K. C. & Andersen, S. L. Transient D1 dopamine 35. Spear, L. P. The adolescent brain and age-related behavioral manifestations. receptor expression on prefrontal cortex projection neurons: relationship to Neurosci. Biobehav. Rev. 24, 417–463 (2000). enhanced motivational salience of drug cues in adolescence. J. Neurosci. 28, 36. Hartley, C. A. & Lee, F. S. Sensitive periods in affective development: 2375–2382 (2008). nonlinear maturation of fear learning. Neuropsychopharmacology 40,50–60 65. Tseng, K. Y. & O’Donnell, P. Dopamine modulation of prefrontal cortical (2015). interneurons changes during adolescence. Cereb. Cortex. 17, 1235–1240 37. King, E. C., Pattwell, S. S., Sun, A., Glatt, C. E. & Lee, F. S. Nonlinear (2007). developmental trajectory of fear learning and memory. Ann. N. Y. Acad. Sci. 66. Casey, B. J., Getz, S. & Galvan, A. The adolescent brain. Dev. Rev. 28,62–77 1304,62–69 (2013). (2008). 38. Insel, T. R. & Young, L. J. The neurobiology of attachment. Nat. Rev. Neurosci. 67. Flechsig, P. Anatomie des menschlichen Gehirns und Rückenmarks auf 2, 129–136 (2001). myelogenetischer Grundlage. 121 (1920). 39. Landers, M. S. & Sullivan, R. M. The development and neurobiology of infant 68. Buckner, R. L. & Krienen, F. M. The evolution of distributed association attachment and fear. Dev. Neurosci. 34, 101–114 (2012). networks in the human brain. Trends Cogn. Sci. 17, 648–665 (2017). 40. Pattwell, S. S. et al. Altered fear learning across development in both mouse 69. Asato, M. R., Terwilliger, R., Woo, J. & Luna, B. White matter development in and human. Proc. Natl Acad. Sci. USA 109, 16318–16323 (2012). adolescence: a DTI study. Cereb. Cortex. 20, 2122–2131 (2010). 41. McCallum, J., Kim, J. H. & Richardson, R. Impaired extinction retention in 70. Simmonds, D. J., Hallquist, M. N., Asato, M. & Luna, B. Developmental stages adolescent rats: effects of D-cycloserine. Neuropsychopharmacology 35, and sex differences of white matter and behavioral development through 2134–2142 (2010). adolescence: a longitudinal diffusion tensor imaging (DTI) study. Neuroimage 42. Gee, D. G. et al. Individual differences in frontolimbic circuitry and anxiety 92, 356–368 (2014). emerge with adolescent changes in endocannabinoid signaling across species. 71. Barnea-Goraly, N. et al. White matter development during childhood and Proc. Natl Acad. Sci. USA 113, 4500–4505 (2016). adolescence: a cross-sectional diffusion tensor imaging study. Cereb. Cortex. 43. Malter Cohen, M. et al. Early-life stress has persistent effects on amygdala 15, 1848–1854 (2005). function and development in mice and humans. Proc. Natl Acad. Sci. USA 72. Liston, C. et al. Frontostriatal microstructure modulates efﬁcient recruitment 110, 18274–18278 (2013). of cognitive control. Cereb. Cortex. 16, 553–560 (2006). 44. Gee, D. G. et al. Early developmental emergence of human 73. Power, J. D., Fair, D. A., Schlaggar, B. L. & Petersen, S. E. The development of amygdala–prefrontal connectivity after maternal deprivation. Proc. Natl Acad. human functional brain networks. Neuron 67, 735–748 (2010). Sci. USA 110, 15638–15643 (2013). 74. Fair, D. A. et al. The maturing architecture of the brain’s default network. 45. Tottenham, N. et al. Elevated amygdala response to faces following early Proc. Natl Acad. Sci. USA 105, 4028–4032 (2008). deprivation. Dev. Sci. 14, 190–204 (2011). 75. Supekar, K., Musen, M. & Menon, V. Development of large-scale functional 46. Hartley, C. A. & Somerville, L. H. The neuroscience of adolescent decision- brain networks in children. PLOS Biol. 7, e1000157 (2009). making. Curr. Opin. Behav. Sci. 5, 108–115 (2015). 76. Kelly, A. M. C. et al. Development of anterior cingulate functional 47. van der Schaaf, M. E., Warmerdam, E., Crone, E. A. & Cools, R. Distinct linear connectivity from late childhood to early adulthood. Cereb. Cortex. 19, and non-linear trajectories of reward and punishment reversal learning during 640–657 (2009). NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications 11 | | | REVIEW ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 77. Di Martino, A. et al. Unraveling the miswired connectome: a developmental 105. Yoo, K. et al. Connectome-based predictive modeling of attention: comparing perspective. Neuron 83, 1335–1353 (2014). different functional connectivity features and prediction methods across 78. Sejnowski, T. J., Churchland, P. S. & Movshon, J. A. Putting big data to good datasets. Neuroimage 167,11–22 (2018). use in neuroscience. Nat. Neurosci. 17, 1440–1441 (2014). 106. Jangraw, D. C. et al. A functional connectivity-based neuromarker of sustained 79. Schumann, G. et al. The IMAGEN study: reinforcement-related behaviour in attention generalizes to predict recall in a reading task. Neuroimage 166, normal brain function and psychopathology. Mol. Psychiatry 15, 1128–1139 99–109 (2018). (2010). 107. Rosenberg, M. D. et al. Connectome-based fMRI models predict separable 80. Van Essen, D. C. et al. The WU-Minn Human Connectome Project: an components of attention in novel individuals. J. Cogn. Neurosci. 30, 160–173 overview. Neuroimage 80,62–79 (2013). (2018). 81. Di Martino, A. et al. The autism brain imaging data exchange: towards a large- 108. The ADHD-200 Consortium. The ADHD-200 consortium: a model to scale evaluation of the intrinsic brain architecture in autism. Mol. Psychiatry advance the translational potential of neuroimaging in clinical neuroscience. 19, 659–667 (2014). Front. Syst. Neurosci. 6, 62 (2012). 82. Alexander, L. M. et al. An open resource for transdiagnostic research in 109. Castellanos, F. X. et al. Developmental trajectories of brain volume pediatric mental health and learning disorders. Sci. Data 4, 170181 (2017). abnormalities in children and adolescents with attention-deﬁcit/hyperactivity 83. Casey, B. et al. The ABCD Study: functional imaging acquisition across disorder. J. Am. Med. Assoc. 288, 1740–1748 (2002). 21 sites. Dev. Cogn. Neurosci. (in press) (2018). 110. Casey, B. J., Nigg, J. T., Durston, S. New potential leads in the biology and 84. Alfaro-Almagro, F. et al. Image processing and quality control for the ﬁrst treatment of attention deﬁcit-hyperactivity disorder. Curr. Opin. Neurol. 20, 10,000 brain imaging datasets from UK Biobank. Neuroimage 166, 400–424 119–124 (2007). (2018). 111. Lorenz, E. N. Empirical orthogonal functions and statistical weather 85. Noble, S. et al. Inﬂuences on the test–retest reliability of functional prediction. Scientiﬁc Report No. 1. Statistical Forecasting Project, Dept. of connectivity MRI and its relationship with behavioral utility. Cereb. Cortex. Meteorology, MIT (1956). 27,1–15 (2017). 112. Shaw, P. et al. Development of cortical surface area and gyriﬁcation in 86. Noble, S. et al. Multisite reliability of MR-based functional connectivity. attention-deﬁcit/hyperactivity disorder. Biol. Psychiatry 72, 191–197 (2012). Neuroimage 146, 959–970 (2017). 113. Shaw, P. et al Attention-deﬁcit/hyperactivity disorder is characterized by a 87. Miller, K. L. et al. Multimodal population brain imaging in the UK Biobank delay in cortical maturation. Proc. Natl Acad. Sci. USA 104, 19649–19654 prospective epidemiological study. Nat. Neurosci. 19, 1523–1536 (2016). (2007). 88. Dosenbach, N. U. F. et al. Real-time motion analytics during brain MRI Provides evidence for delayed cortical thickness maturation in ADHD. improve data quality and reduce costs. Neuroimage 161,80–93 (2017). 114. Fair, D. A. et al. Distinct neural signatures detected for ADHD subtypes after 89. Bzdok, D. & Yeo, B. T. T. Inference in the age of big data: Future perspectives controlling for micro-movements in resting state functional connectivity MRI on neuroscience. Neuroimage 155, 549–564 (2017). data. Front. Syst. Neurosci. 6, 80 (2012). Suggests statistical tools useful for analyzing big data in human 115. Pardoe, H. R., Kucharsky Hiess, R. & Kuzniecky, R. Motion and neuroscience. morphometry in clinical and nonclinical populations. Neuroimage 135, 90. Dubois, J. & Adolphs, R. Building a science of individual differences from 177–185 (2016). fMRI. Trends Cogn. Sci. 20, 425–443 (2016). 116. Couvy-Duchesne, B. et al. Head motion and inattention/hyperactivity share 91. Finn, E. S. & Constable, R. T. Individual variation in functional brain common genetic inﬂuences: implications for fMRI studies of ADHD. PLoS connectivity: implications for personalized approaches to psychiatric disease. ONE 11, e0146271 (2016). Dialogues Clin. Neurosci. 18, 277–287 (2016). 117. Engelhardt, L. E. et al. Children’s head motion during fMRI tasks is heritable 92. Jollans, L. & Whelan, R. The clinical added value of imaging: a perspective and stable over time. Dev. Cogn. Neurosci. 25,58–68 (2017). from outcome prediction. Biol. Psychiatry Cogn. Neurosci. Neuroimaging 1, 118. Kessler, T., Angstadt, M. & Sripada, C. Growth charting of brain connectivity 423–432 (2016). networks and the identiﬁcation of attention impairment in youth. JAMA 93. Box, G. E. P. in Robustness in Statistics (eds Launer, R. & Wilderson, G.) Psychiatry 73, 481–489 (2016). 201–236 (Academic Press, New York, 1979). 119. Kaufmann, T. et al Delayed stabilization and individualization in connectome 94. Finn, E. S. et al. Can brain state be manipulated to emphasize individual development are related to psychiatric disorders. Nat. Neurosci. 20, 513–515 differences in functional connectivity? Neuroimage 160, 140–151 (2017). (2017). 95. Cohen, A. O. & Casey, B. J. Rewiring juvenile justice: the intersection of Demonstrates that functional connectivity patterns become more stable and developmental neuroscience and legal policy. Trends Cogn. Sci. 18,63–65 unique during adolescence, and that delays in this pattern of development (2014). are related to psychiatric symptoms. 96. Cohen, A. O. et al When is an adolescent an adult? assessing cognitive control 120. Albert, P. S. A linear mixed model for predicting a binary event from in emotional and nonemotional contexts. Psychol. Sci. 27, 549–562 (2016). longitudinal data under random effects misspeciﬁcation. Stat. Med. 31, Provides evidence for a shift in cognitive capacity in emotional contexts that 143–154 (2012). accompanies dynamic changes in prefrontal circuitry during adolescence. 121. Liu, D. & Albert, P. S. Combination of longitudinal biomarkers in predicting 97. Rudolph, M. D. et al At risk of being risky: the relationship between ‘brain age’ binary events. Biostatistics 15, 706–718 (2014). under emotional states and risk preference. Dev. Cogn. Neurosci. 24,93–106 122. Sweeting, M. J. & Thompson, S. G. Making predictions from complex (2017). longitudinal data, with application to planning monitoring intervals in a Shows that, on average, adolescents’ functional connectivity proﬁles look national screening programme. J. R. Stat. Soc. Ser. A. Stat. Soc. 175, 569–586 younger in emotional contexts, and that young adults who maintain this (2012). proﬁle show riskier choices. 123. Li, L., Luo, S., Hu, B. & Greene, T. Dynamic prediction of renal failure using 98. Rosenberg, M. D., Finn, E. S., Scheinost, D., Constable, R. T. & Chun, M. M. longitudinal biomarkers in a cohort study of chronic kidney disease. Stat. Characterizing attention with predictive network models. Trends Cogn. Sci. 21, Biosci. 9, 357–378 (2016). 290–302 (2017). 124. Whelan, R. et al Neuropsychosocial proﬁles of current and future adolescent 99. Van Essen, D. C. & Glasser, M. F. The Human Connectome Project: progress alcohol misusers. Nature 512, 185–189 (2014). and prospects. Cerebrum 2016,10–16 (2016). Predicts future alcohol misuse from neural, behavioral, environmental, and 100. Holmes, A. J. et al Brain Genomics Superstruct Project initial data release with genetic data in a large sample of adolescents. structural, functional, and behavioral measures. Sci. Data 2, 150031 (2015). 125. Emerson, R. W. et al. Functional neuroimaging of high-risk 6-month-old Details the Brain Genomics Superstruct Project repository, which includes infants predicts a diagnosis of autism at 24 months of age. Sci. Transl. Med. 9 neuroimaging and self-report behavioral and cognitive data from 1,570 eaag2882 (2017). adult participants. 126. Hazlett, H. C. et al. Early brain development in infants at high risk for autism 101. Poldrack, R. A. et al. Toward open sharing of task-based fMRI data: the spectrum disorder. Nature 542, 348–351 (2017). OpenfMRI project. Front. Neuroinf. 7, 12 (2013). 127. Plitt, M., Barnes, K. A., Wallace, G. L., Kenworthy, L. & Martin, A. Resting- 102. Miranda-Domínguez, Ó. et al. Heritability of the human connectome: a state functional connectivity predicts longitudinal change in autistic traits and connectotyping study. Netw. Neurosci.1–48. https://doi.org/10.1162/ adaptive functioning in autism. Proc. Natl Acad. Sci. USA 112, E6699–E6706 NETN_a_00029 (2017). (2015). 103. Rosenberg, M. D. et al. A neuromarker of sustained attention from whole- 128. Hoeft, F. et al. Neural systems predicting long-term outcome in dyslexia. Proc. brain functional connectivity. Nat. Neurosci. 19, 165–171 (2016). Natl Acad. Sci. USA 108, 361–366 (2011). Introduces a model that generalizes across datasets to predict sustained 129. Dosenbach, N. U. F. et al. Prediction of individual brain maturity using fMRI. attention from task-based and resting-state functional connectivity. Science 329, 1358–1361 (2010). 104. Rosenberg, M. D. et al. Methylphenidate modulates functional network Predicted chronological age from functional connectivity patterns in the ﬁrst connectivity to enhance attention. J. Neurosci. 36, 9547–9557 (2016). example of predictive modeling in developmental cognitive neuroscience. 12 NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 REVIEW ARTICLE 130. Brown, T. T. et al. Neuroanatomical assessment of biological maturity. Curr. 152. Pausova, Z. et al. Genes, maternal smoking, and the offspring brain and body Biol. 22, 1693–1698 (2012). during adolescence: design of the Saguenay Youth Study. Hum. Brain. Mapp. 131. Franke, K., Luders, E., May, A., Wilke, M. & Gaser, C. Brain maturation: 28, 502–518 (2007). predicting individual BrainAGE in children and adolescents using structural 153. Salum, G. A. et al. High risk cohort study for psychiatric disorders in MRI. Neuroimage 63, 1305–1312 (2012). childhood: rationale, design, methods and preliminary results. Int. J. Methods 132. Altman, D. G. & Royston, P. The cost of dichotomising continuous variables. Psychiatr. Res. 24,58–73 (2015). BMJ 332, 1080 (2006). 154. Thompson, P. M. et al. The ENIGMA Consortium: large-scale collaborative 133. Austin, P. C. & Brunner, L. J. Inﬂation of the type I error rate when a analyses of neuroimaging and genetic data. Brain. Imaging Behav. 8, 153–182 continuous confounding variable is categorized in logistics regression analyses. (2014). Statist. Med. 23, 1159–1178 (2004). 155. Kooijman, M. N. et al. The Generation R Study: design and cohort update 134. Castellanos, F. X., Di Martino, A., Craddock, R. C., Mehta, A. D. & Milham, 2017. Eur. J. Epidemiol. 31, 1243–1264 (2016). M. P. Clinical applications of the functional connectome. Neuroimage 80, 156. von Rhein, D. et al. The NeuroIMAGE study: a prospective phenotypic, 527–540 (2013). cognitive, genetic and MRI study in children with attention-deﬁcit/ 135. Hudziak, J. J., Achenbach, T. M., Althoff, R. R. & Pine, D. S. A dimensional hyperactivity disorder. Design and descriptives. Eur. Child. Adolesc. Psychiatry approach to developmental psychopathology. Int. J. Methods Psychiatr. Res. 24, 265–281 (2015). 16, S16–S23 (2007). 157. Zuo, X. N. et al. An open science resource for establishing reliability and 136. Finn, E. S. et al. Functional connectome ﬁngerprinting: identifying individuals reproducibility in functional connectomics. Sci. Data 1, 140049 (2014). using patterns of brain connectivity. Nat. Neurosci. 18, 1664–1671 (2015). 137. Smith, S. M. et al. Functional connectomics from resting-state fMRI. Trends Acknowledgements Cogn. Sci. 17, 666–682 (2013). This work was supported by the National Institute of Mental Health (K01 MH099232 to 138. Ferguson, M. A., Anderson, J. S. & Spreng, R. N. Fluid and ﬂexible minds: A.J.H.), National Institute on Drug Abuse (U01 DA041174 to B.J.C.), and MacArthur intelligence reﬂects synchrony in the brain’s intrinsic network architecture. Foundation Research Network on Law and Neuroscience (B.J.C.). M.D.R. is supported by Netw. Neurosci. 1, 192–207 (2017). a Theresa Seessel Postdoctoral Fellowship at Yale University. We thank Kevin Anderson, 139. Poole, V. N. et al. Intrinsic functional connectivity predicts individual Adam Chekroud, Emily Finn, and David Gruskin for helpful discussion and/or com- differences in distractibility. Neuropsychologia 86, 176–182 (2016). ments on earlier drafts. 140. Supekar, K. et al. Neural predictors of individual differences in response to math tutoring in primary-grade school children. Proc. Natl Acad. Sci. USA 110, 8230–8235 (2013). Author contributions 141. AM, C. Bigger data, harder questions—opportunities throughout mental M.D.R. drafted the text and graphics with input and edits from B.J.C. and A.J.H. All health care. JAMA Psychiatry 74, 1183–1184 (2017). authors revised and ﬁnalized the manuscript for publication. 142. Smith, S. et al A positive-negative mode of population covariation links brain connectivity, demographics and behavior. Nat. Neurosci. 18, 1565–1567 (2015). Additional information Identiﬁes a pattern of functional brain connectivity related to a 'positive- Competing interests: The authors declare no competing ﬁnancial interests. negative' axis of lifestyle, demographic and psychometric factors. 143. Ciric, R. et al Benchmarking of participant-level confound regression Reprints and permission information is available online at http://npg.nature.com/ strategies for the control of motion artifact in studies of functional reprintsandpermissions/ connectivity. Neuroimage 154, 174–187 (2017). Compares current methods for controlling for motion artifacts in functional Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in connectivity data. published maps and institutional afﬁliations. 144. Swanson, J. M. The UK Biobank and selection bias. Lancet 380, 110 (2017). 145. Chekroud, A. M. & Koutsouleris, N. The perilous path from publication to practice. Mol. Psychiatry 23,24–25 (2017). 146. MP, P. Evidence-based pragmatic psychiatry—a call to action. JAMA Open Access This article is licensed under a Creative Commons Psychiatry 74, 1185–1186 (2017). Attribution 4.0 International License, which permits use, sharing, 147. Guyon, I. & Elisseeff, A. An introduction to variable and feature selection. J. adaptation, distribution and reproduction in any medium or format, as long as you give Mach. Learn. Res. 3, 1157–1182 (2003). appropriate credit to the original author(s) and the source, provide a link to the Creative 148. Kievit, R. et al. Developmental cognitive neuroscience using latent change Commons license, and indicate if changes were made. The images or other third party score models: a tutorial and applications. Dev. Cogn. Neurosci. (2017). material in this article are included in the article’s Creative Commons license, unless 149. Brown, S. A. et al. The National Consortium on Alcohol and indicated otherwise in a credit line to the material. If material is not included in the NeuroDevelopment in Adolescence (NCANDA): a multisite study of adolescent development and substance use. J. Stud. Alcohol. Drugs 76, 895–908 article’s Creative Commons license and your intended use is not permitted by statutory (2015). regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/ 150. Jernigan, T. L. et al. The pediatric imaging, neurocognition, and genetics (PING) data repository. Neuroimage 124, 1149–1154 (2016). licenses/by/4.0/. 151. Satterthwaite, T. D. et al. The Philadelphia Neurodevelopmental Cohort: a publicly available resource for the study of normal and abnormal brain © The Author(s) 2018 development in youth. Neuroimage 124, 1115–1119 (2016). NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications 13 | | | http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Nature Communications Springer Journals http://www.deepdyve.com/lp/springer-journals/prediction-complements-explanation-in-understanding-the-developing-00J6PES4eo

Loading next page...

References (185)

J. Silvers, Kateri McRae, J. Gabrieli, J. Gross, Katherine Remy, K. Ochsner (2012)
Age-related differences in emotional reactivity, regulation, and rejection sensitivity in adolescence.
Emotion, 12 6
Adam Chekroud, N. Koutsouleris (2018)
The perilous path from publication to practice
Molecular Psychiatry, 23
P. Thompson, J. Stein, S. Medland, D. Hibar, A. Vasquez, M. Rentería, R. Toro, N. Jahanshad, G. Schumann, B. Franke, M. Wright, N. Martin, I. Agartz, M. Alda, S. Alhusaini, L. Almasy, J. Almeida, K. Alpert, N. Andreasen, O. Andreassen, L. Apostolova, Katja Appel, N. Armstrong, Benjamin Aribisala, M. Bastin, M. Bauer, C. Bearden, Ørjan Bergmann, E. Binder, J. Blangero, H. Bockholt, E. Bøen, C. Bois, D. Boomsma, T. Booth, Ian Bowman, J. Bralten, R. Brouwer, H. Brunner, David Brohawn, R. Buckner, J. Buitelaar, K. Bulayeva, J. Bustillo, V. Calhoun, D. Cannon, R. Cantor, M. Carless, X. Caseras, G. Cavalleri, M. Chakravarty, K. Chang, C. Ching, A. Christoforou, S. Cichon, V. Clark, P. Conrod, G. Coppola, B. Crespo-Facorro, J. Curran, M. Czisch, I. Deary, E. Geus, A. Braber, G. Delvecchio, C. Depondt, L. Haan, G. Zubicaray, D. Dima, R. Dimitrova, S. Djurovic, Hong-wei Dong, G. Donohoe, R. Duggirala, T. Dyer, S. Ehrlich, C. Ekman, T. Elvsåshagen, L. Emsell, S. Erk, T. Espeseth, J. Fagerness, S. Fears, I. Fedko, G. Fernández, S. Fisher, T. Foroud, P. Fox, C. Francks, S. Frangou, Eva-Maria Frey, T. Frodl, V. Frouin, H. Garavan, S. Giddaluru, D. Glahn, B. Godlewska, R. Goldstein, R. Gollub, H. Grabe, O. Grimm, O. Gruber, Tulio Guadalupe, R. Gur, R. Gur, H. Göring, S. Hagenaars, T. Hajek, G. Hall, J. Hall, J. Hardy, C. Hartman, J. Hass, S. Hatton, U. Haukvik, K. Hegenscheid, A. Heinz, I. Hickie, B. Ho, D. Hoehn, P. Hoekstra, Marisa Hollinshead, A. Holmes, G. Homuth, M. Hoogman, L. Hong, N. Hosten, J. Hottenga, H. Pol, K. Hwang, C. Jack, M. Jenkinson, Caroline Johnston, E. Jönsson, R. Kahn, D. Kasperavičiūtė, S. Kelly, Sungeun Kim, P. Kochunov, L. Koenders, B. Krämer, J. Kwok, J. Lagopoulos, G. Laje, M. Landén, B. Landman, J. Lauriello, S. Lawrie, Phil Lee, S. Hellard, H. Lemaître, Cassandra Leonardo, C. Li, B. Liberg, D. Liewald, Xinmin Liu, Lorna Lopez, E. Loth, A. Lourdusamy, M. Luciano, F. Macciardi, M. Machielsen, G. MacQueen, U. Malt, R. Mandl, D. Manoach, J. Martinot, M. Matarin, K. Mather, M. Mattheisen, M. Mattingsdal, A. Meyer-Lindenberg, C. Mcdonald, A. McIntosh, F. McMahon, K. Mcmahon, E. Meisenzahl, I. Melle, Y. Milaneschi, S. Mohnke, G. Montgomery, D. Morris, E. Moses, B. Mueller, S. Maniega, Thomas Mühleisen, B. Müller-Myhsok, B. Mwangi, M. Nauck, K. Nho, Thomas Nichols, L. Nilsson, A. Nugent, L. Nyberg, R. Olvera, J. Oosterlaan, R. Ophoff, M. Pandolfo, Melina Papalampropoulou-Tsiridou, M. Papmeyer, T. Paus, Z. Pausova, G. Pearlson, B. Penninx, Charles Peterson, A. Pfennig, M. Phillips, G. Pike, J. Poline, S. Potkin, B. Pütz, A. Ramasamy, J. Rasmussen, M. Rietschel, M. Rijpkema, S. Risacher, J. Roffman, R. Roiz-Santiañez, N. Romanczuk-Seiferth, E. Rose, N. Royle, D. Rujescu, M. Ryten, P. Sachdev, Alireza Salami, T. Satterthwaite, J. Savitz, A. Saykin, C. Scanlon, L. Schmaal, H. Schnack, A. Schork, S. Schulz, Remmelt Schür, L. Seidman, Li Shen, J. Shoemaker, A. Simmons, S. Sisodiya, Colin Smith, J. Smoller, J. Soares, S. Sponheim, E. Sprooten, J. Starr, V. Steen, S. Strakowski, Lachlan Strike, J. Sussmann, P. Sämann, A. Teumer, A. Toga, D. Tordesillas-Gutiérrez, D. Trabzuni, S. Trost, J. Turner, M. Heuvel, N. Wee, K. Eijk, T. Erp, N. Haren, D. Ent, M. Tol, M. Hernández, D. Veltman, A. Versace, H. Völzke, R. Walker, H. Walter, Lei Wang, J. Wardlaw, M. Weale, M. Weiner, W. Wen, L. Westlye, H. Whalley, C. Whelan, T. White, A. Winkler, K. Wittfeld, G. Woldehawariat, C. Wolf, D. Zilles, M. Zwiers, A. Thalamuthu, P. Schofield, N. Freimer, N. Lawrence, W. Drevets (2014)
The ENIGMA Consortium: large-scale collaborative analyses of neuroimaging and genetic data
Brain Imaging and Behavior, 8
R. Ciric, D. Wolf, Jonathan Power, D. Roalf, Graham Baum, K. Ruparel, R. Shinohara, M. Elliott, S. Eickhoff, C. Davatzikos, R. Gur, R. Gur, D. Bassett, T. Satterthwaite (2017)
Benchmarking of participant-level confound regression strategies for the control of motion artifact in studies of functional connectivity
NeuroImage, 154
D. Simmonds, M. Hallquist, Miya Asato, B. Luna (2014)
Developmental stages and sex differences of white matter and behavioral development through adolescence: A longitudinal diffusion tensor imaging (DTI) study
NeuroImage, 92
Bruce Ellis, M. Giudice, T. Dishion, A. Figueredo, P. Gray, Vladas Griskevicius, Patricia Hawley, W. Jacobs, J. James, Anthony Volk, David Wilson (2012)
The evolutionary basis of risky adolescent behavior: implications for science, policy, and practice.
Developmental psychology, 48 3
CURRENT DIRECTIONS IN PSYCHOLOGICAL SCIENCE Risk Taking in Adolescence New Perspectives From Brain and Behavioral Science
A. Cohen, B. Casey (2014)
Rewiring juvenile justice: the intersection of developmental neuroscience and legal policy
Trends in Cognitive Sciences, 18
Mark Plitt, K. Barnes, G. Wallace, L. Kenworthy, Alex Martin (2015)
Resting-state functional connectivity predicts longitudinal change in autistic traits and adaptive functioning in autism
Proceedings of the National Academy of Sciences, 112
S. Noble, Marisa Spann, F. Tokoglu, X. Shen, T. Constable, D. Scheinost (2017)
Influences on the Test–Retest Reliability of Functional Connectivity MRI and its Relationship with Behavioral Utility
Cerebral Cortex (New York, NY), 27
M. Schaaf, E. Warmerdam, E. Crone, R. Cools (2011)
Distinct linear and non-linear trajectories of reward and punishment reversal learning during development: Relevance for dopamine's role in adolescent decision making
Developmental Cognitive Neuroscience, 1
T Insel (2010)
10.1176/appi.ajp.2010.09091379
Am. J. Psychiatry, 167
Heather Douglas (2009)
Reintroducing Prediction to Explanation
Philosophy of Science, 76
A. Galván, T. Hare, Cindy Parra, Jackie Penn, H. Voss, G. Glover, B. Casey (2006)
Earlier Development of the Accumbens Relative to Orbitofrontal Cortex Might Underlie Risk-Taking Behavior in Adolescents
The Journal of Neuroscience, 26
Kaustubh Supekar, A. Swigart, Caitlin Tenison, D. Jolles, M. Rosenberg-Lee, L. Fuchs, V. Menon (2013)
Neural predictors of individual differences in response to math tutoring in primary-grade school children
Proceedings of the National Academy of Sciences, 110
P. Goldman-Rakic (1988)
Topography of cognition: parallel distributed networks in primate association cortex.
Annual review of neuroscience, 11
L. Somerville, T. Hare, B. Casey (2011)
Frontostriatal Maturation Predicts Cognitive Control Failure to Appetitive Cues in Adolescents
Journal of Cognitive Neuroscience, 23
D. Fair, Deepti Bathula, M. Nikolas, J. Nigg (2012)
Distinct neuropsychological subgroups in typically developing youth inform heterogeneity in children with ADHD
Proceedings of the National Academy of Sciences, 109
Michael Dreyfuss, Kristina Caudle, Andrew Drysdale, Natalie Johnston, A. Cohen, L. Somerville, A. Galván, N. Tottenham, T. Hare, B. Casey (2014)
Teens Impulsively React rather than Retreat from Threat
Developmental Neuroscience, 36
B. Casey, M. Oliveri, T. Insel (2014)
A Neurodevelopmental Perspective on the Research Domain Criteria (RDoC) Framework
Biological Psychiatry, 76
Predicted chronological age from functional connectivity patterns in the first example of predictive modeling in developmental cognitive neuroscience
Choong-Wan Woo, Luke Chang, M. Lindquist, T. Wager (2017)
Building better biomarkers: brain models in translational neuroimaging
Nature Neuroscience, 20
N. Dosenbach, Binyam Nardos, A. Cohen, D. Fair, Jonathan Power, J. Church, S. Nelson, G. Wig, Alecia Vogel, C. Lessov-Schlaggar, K. Barnes, Joseph Dubis, E. Feczko, R. Coalson, J. Pruett, D. Barch, S. Petersen, B. Schlaggar (2010)
Prediction of Individual Brain Maturity Using fMRI
Science, 329
L Steinberg (2007)
10.1111/j.1467-8721.2007.00475.x
Curr. Dir. Psychol. Sci., 16
Matthew Cohen, D. Jing, Ruirong Yang, N. Tottenham, F. Lee, B. Casey (2013)
Early-life stress has persistent effects on amygdala function and development in mice and humans
Proceedings of the National Academy of Sciences, 110
C. Liston, R. Watts, N. Tottenham, M. Davidson, S. Niogi, A. Uluğ, B. Casey (2006)
Frontostriatal microstructure modulates efficient recruitment of cognitive control.
Cerebral cortex, 16 4
T. Insel, L. Young (2001)
The neurobiology of attachment
Nature Reviews Neuroscience, 2
J. Chein, Dustin Albert, Lia O'brien, Kaitlyn Uckert, L. Steinberg (2011)
Peers increase adolescent risk taking by enhancing activity in the brain's reward circuitry.
Developmental science, 14 2
H. Pardoe, R. Hiess, R. Kuzniecky (2016)
Motion and morphometry in clinical and nonclinical populations
NeuroImage, 135
Marjolein Kooijman, Claudia Kruithof, C. Duijn, L. Duijts, O. Franco, Marinus IJzendoorn, J. Jongste, C. Klaver, A. Lugt, J. Mackenbach, H. Moll, R. Peeters, H. Raat, E. Rings, F. Rivadeneira, M. Schroeff, E. Steegers, H. Tiemeier, A. Uitterlinden, F. Verhulst, E. Wolvius, J. Felix, V. Jaddoe (2016)
The Generation R Study: design and cohort update 2017
European Journal of Epidemiology, 31
D. Essen, Stephen Smith, D. Barch, Timothy Behrens, E. Yacoub, K. Uğurbil (2013)
The WU-Minn Human Connectome Project: An overview
NeuroImage, 80
Óscar Miranda-Domínguez, E. Feczko, D. Grayson, H. Walum, J. Nigg, D. Fair (2017)
Heritability of the human connectome: A connectotyping study
Network Neuroscience, 2
D. Bzdok, B. Yeo (2017)
Inference in the age of big data: Future perspectives on neuroscience
NeuroImage, 155
J. McCallum, J. Kim, R. Richardson (2010)
Impaired Extinction Retention in Adolescent Rats: Effects of D-Cycloserine
Neuropsychopharmacology, 35
G. Shmueli (2010)
To Explain or To Predict?
Indian School of Business Research Paper Series
Miya Asato, R. Terwilliger, J. Woo, B. Luna (2010)
White matter development in adolescence: a DTI study.
Cerebral cortex, 20 9
R. Poldrack, D. Barch, Jason Mitchell, T. Wager, A. Wagner, J. Devlin, Chad Cumba, Oluwasanmi Koyejo, M. Milham (2013)
Toward open sharing of task-based fMRI data: the OpenfMRI project
Frontiers in Neuroinformatics, 7
T Insel (2010)
Research domain criteria (RDoC): toward a new classification framework for research on mental disorders
Am. J. Psychiatry, 167
C. Tamnes, Ylva Ostby, A. Fjell, L. Westlye, P. Due-Tønnessen, K. Walhovd (2010)
Brain maturation in adolescence and young adulthood: regional age-related changes in cortical thickness and white matter volume and microstructure.
Cerebral cortex, 20 3
Marc Rudolph, Oscar Miranda-Dominguez, A. Cohen, Kaitlyn Breiner, L. Steinberg, R. Bonnie, Elizabeth Scott, Kim Taylor-Thompson, J. Chein, Karla Fettich, J. Richeson, Danielle Dellarco, A. Galván, B. Casey, D. Fair (2017)
At risk of being risky: The relationship between “brain age” under emotional states and risk preference
Developmental Cognitive Neuroscience, 24
FS Lee (2014)
Adolescent mental health—Opportunity and obligation: emerging neuroscience offers hope for treatments
Science, 346
H. Hazlett, H. Gu, B. Munsell, S. Kim, M. Styner, J. Wolff, J. Elison, M. Swanson, Hongtu Zhu, K. Botteron, D. Collins, J. Constantino, S. Dager, A. Estes, Alan Evans, V. Fonov, G. Gerig, P. Kostopoulos, R. McKinstry, J. Pandey, S. Paterson, J. Pruett, R. Schultz, D. Shaw, L. Zwaigenbaum, J. Piven (2017)
Early brain development in infants at high risk for autism spectrum disorder
Nature, 542
C. Craver (2006)
When mechanistic models explain
Synthese, 153
R. Emerson, C. Adams, T. Nishino, H. Hazlett, J. Wolff, L. Zwaigenbaum, J. Constantino, M. Shen, M. Swanson, J. Elison, S. Kandala, A. Estes, K. Botteron, D. Collins, S. Dager, Alan Evans, G. Gerig, H. Gu, R. McKinstry, S. Paterson, R. Schultz, M. Styner, B. Schlaggar, J. Pruett, J. Piven (2017)
Functional neuroimaging of high-risk 6-month-old infants predicts a diagnosis of autism at 24 months of age
Science Translational Medicine, 9
L. Spear (2000)
The adolescent brain and age-related behavioral manifestations
Neuroscience & Biobehavioral Reviews, 24
S. Morgenthaler (2001)
Robustness in Statistics
G Shmueli (2010)
To explain or to predict?
Stat. Sci., 25
X. Zuo, Jeffrey Anderson, P. Bellec, R. Birn, B. Biswal, J. Blautzik, J. Breitner, R. Buckner, V. Calhoun, F. Castellanos, Antao Chen, Bing Chen, Jiangtao Chen, Xu Chen, S. Colcombe, W. Courtney, R. Craddock, A. Martino, Hao-Ming Dong, Xiaolan Fu, Q. Gong, Krzysztof Gorgolewski, Ying Han, Ye He, Yong He, Erica Ho, A. Holmes, Xiaohui Hou, Jeremy Huckins, T. Jiang, Yi Jiang, William Kelley, C. Kelly, M. King, S. LaConte, J. Lainhart, X. Lei, Huijie Li, Kaiming Li, Kuncheng Li, Qixiang Lin, Dongqiang Liu, Jia Liu, Xun Liu, Yijun Liu, G. Lu, Jie Lu, B. Luna, Jing Luo, Daniel Lurie, Y. Mao, D. Margulies, A. Mayer, T. Meindl, M. Meyerand, Weizhi Nan, Jared Nielsen, David O’Connor, David Paulsen, V. Prabhakaran, Zhigang Qi, J. Qiu, C. Shao, Z. Shehzad, Weijun Tang, A. Villringer, Hui-ling Wang, K. Wang, D. Wei, Gao-Xia Wei, X. Weng, Xuehai Wu, Ting-Fa Xu, N. Yang, Zhi Yang, Y. Zang, Lei Zhang, Qinglin Zhang, Zhe Zhang, Zhiqiang Zhang, Ke Zhao, Zonglei Zhen, Yuan Zhou, Xing-Ting Zhu, M. Milham (2014)
An open science resource for establishing reliability and reproducibility in functional connectomics
Scientific Data, 1
S. Blakemore, T. Robbins (2012)
Decision-making in the adolescent brain
Nature Neuroscience, 15
Kaustubh Supekar, M. Musen, V. Menon (2009)
Development of Large-Scale Functional Brain Networks in Children
PLoS Biology, 7
N. Barnea-Goraly, V. Menon, M. Eckert, L. Tamm, R. Bammer, Asya Karchemskiy, C. Dant, A. Reiss (2005)
White matter development during childhood and adolescence: a cross-sectional diffusion tensor imaging study.
Cerebral cortex, 15 12
M. Rosenberg, E. Finn, D. Scheinost, R. Constable, M. Chun (2017)
Characterizing Attention with Predictive Network Models
Trends in Cognitive Sciences, 21
Isabelle Guyon, A. Elisseeff (2003)
An Introduction to Variable and Feature Selection
J. Mach. Learn. Res., 3
L. Somerville, R. Jones, B. Casey (2010)
A time of change: Behavioral and neural correlates of adolescent sensitivity to appetitive and aversive environmental cues
Brain and Cognition, 72
K. Miller, F. Alfaro-Almagro, N. Bangerter, D. Thomas, E. Yacoub, Junqian Xu, A. Bartsch, S. Jbabdi, S. Sotiropoulos, J. Andersson, L. Griffanti, G. Douaud, T. Okell, P. Weale, I. Dragonu, Steve Garratt, Sarah Hudson, R. Collins, M. Jenkinson, P. Matthews, Stephen Smith (2016)
Multimodal population brain imaging in the UK Biobank prospective epidemiological study
Nature neuroscience, 19
J. Price, W. Drevets (2010)
Neurocircuitry of Mood Disorders
Neuropsychopharmacology, 35
Elizabeth King, Siobhan Pattwell, Alice Sun, C. Glatt, F. Lee (2013)
Nonlinear developmental trajectory of fear learning and memory
Annals of the New York Academy of Sciences, 1304
A. Moran, Edward Stein, Hadass Tischler, K. Belelovsky, I. Bar-Gad (2011)
Dynamic Stereotypic Responses of Basal Ganglia Neurons to Subthalamic Nucleus High-Frequency Stimulation in the Parkinsonian Primate
Frontiers in Systems Neuroscience, 5
V. Jaddoe, C. Duijn, A. Heijden, J. Mackenbach, H. Moll, E. Steegers, H. Tiemeier, A. Uitterlinden, F. Verhulst, A. Hofman (2010)
The Generation R Study: design and cohort update 2010
European Journal of Epidemiology, 25
D. Altman, P. Royston (2006)
The cost of dichotomising continuous variables
BMJ : British Medical Journal, 332
D. Gee, L. Gabard-Durnam, Jessica Flannery, Bonnie Goff, K. Humphreys, Eva Telzer, T. Hare, S. Bookheimer, N. Tottenham (2013)
Early developmental emergence of human amygdala–prefrontal connectivity after maternal deprivation
Proceedings of the National Academy of Sciences, 110
N. Gogtay, P. Thompson
Mapping Gray Matter Development: Implications for Typical Development and Vulnerability to Psychopathology
F. Castellanos, Patti Lee, W. Sharp, N. Jeffries, D. Greenstein, L. Clasen, J. Blumenthal, R. James, C. Ebens, J. Walter, A. Zijdenbos, Alan Evans, J. Giedd, J. Rapoport (2002)
Developmental trajectories of brain volume abnormalities in children and adolescents with attention-deficit/hyperactivity disorder.
JAMA, 288 14
TheADHD-200 Consortium (2012)
The ADHD-200 consortium: a model to advance the translational potential of neuroimaging in clinical neuroscience. Frontiers in Systems
Neuroscience, 6
M. Paulus (2017)
Evidence-Based Pragmatic Psychiatry-A Call to Action.
JAMA psychiatry, 74 12
T. Satterthwaite, J. Connolly, K. Ruparel, M. Calkins, Chad Jackson, M. Elliott, D. Roalf, K. Prabhakaran, Ryan Hopson, Meckenzie Behr, H. Qiu, F. Mentch, R. Chiavacci, P. Sleiman, R. Gur, H. Hakonarson, R. Gur (2016)
The Philadelphia Neurodevelopmental Cohort: A publicly available resource for the study of normal and abnormal brain development in youth
NeuroImage, 124
W. Salmon (1978)
Why Ask, ‘Why?’? An Inquiry Concerning Scientific Explanation
, 51
L. Steinberg (2008)
A Social Neuroscience Perspective on Adolescent Risk-Taking.
Developmental review : DR, 28 1
K. Greene, M. Krcmar, L. Walters, D. Rubin, Jerold, Lauren Hale (2000)
Targeting adolescent risk-taking behaviors: the contributions of egocentrism and sensation-seeking.
Journal of adolescence, 23 4
J. Bourgeois, P. Goldman-Rakic, P. Rakić (1994)
Synaptogenesis in the prefrontal cortex of rhesus monkeys.
Cerebral cortex, 4 1
Laura Engelhardt, Mary Roe, J. Juranek, Dana DeMaster, K. Harden, E. Tucker-Drob, J. Church (2017)
Children’s head motion during fMRI tasks is heritable and stable over time
Developmental Cognitive Neuroscience, 25
G. Box (1979)
Robustness in the Strategy of Scientific Model Building.
V. Poole, M. Robinson, Omar Singleton, J. DeGutis, W. Milberg, R. McGlinchey, D. Salat, M. Esterman (2016)
Intrinsic functional connectivity predicts individual differences in distractibility
Neuropsychologia, 86
Shows that, on average, adolescents' functional connectivity profiles look younger in emotional contexts, and that young adults who maintain this profile show riskier choices
Liang Li, S. Luo, Bo Hu, T. Greene (2016)
Dynamic Prediction of Renal Failure Using Longitudinal Biomarkers in a Cohort Study of Chronic Kidney Disease
Statistics in Biosciences, 9
Kuei Tseng, P. O’Donnell (2006)
Dopamine modulation of prefrontal cortical interneurons changes during adolescence.
Cerebral cortex, 17 5
G. Schumann, E. Loth, T. Banaschewski, A. Barbot, G. Barker, C. Büchel, P. Conrod, J. Dalley, Herta Flor, Juergen Gallinat, H. Garavan, A. Heinz, Bernd Itterman, M. Lathrop, C. Mallik, Karl Mann, J. Martinot, T. Paus, T. Paus, T. Paus, J. Poline, T. Robbins, M. Rietschel, Laurence Reed, M. Smolka, R. Spanagel, C. Speiser, D. Stephens, A. Ströhle, M. Struve (2010)
The IMAGEN study: reinforcement-related behaviour in normal brain function and psychopathology
Molecular Psychiatry, 15
(2015)
Details the Brain Genomics Superstruct Project repository, which includes neuroimaging and self-report behavioral and cognitive data from 1
N. Tottenham, T. Hare, A. Millner, T. Gilhooly, J. Zevin, B. Casey (2011)
Elevated amygdala response to faces following early deprivation.
Developmental science, 14 2
D. Fair, J. Nigg, Swathi Iyer, Deepti Bathula, K. Mills, N. Dosenbach, B. Schlaggar, M. Mennes, David Gutman, S. Bangaru, J. Buitelaar, D. Dickstein, A. Martino, D. Kennedy, C. Kelly, B. Luna, J. Schweitzer, K. Velanova, Yu-Feng Wang, S. Mostofsky, F. Castellanos, M. Milham (2012)
Distinct neural signatures detected for ADHD subtypes after controlling for micro-movements in resting state functional connectivity MRI data
Frontiers in Systems Neuroscience, 6
P. Shaw, K. Eckstrand, W. Sharp, J. Blumenthal, J. Lerch, Dede Greenstein, L. Clasen, Alan Evans, J. Giedd, J. Rapoport (2007)
Attention-deficit/hyperactivity disorder is characterized by a delay in cortical maturation
Proceedings of the National Academy of Sciences, 104
Predicts future alcohol misuse from neural, behavioral, environmental, and genetic data in a large sample of adolescents
H. Brenhouse, K. Sonntag, S. Andersen (2008)
Transient D1 Dopamine Receptor Expression on Prefrontal Cortex Projection Neurons: Relationship to Enhanced Motivational Salience of Drug Cues in Adolescence
The Journal of Neuroscience, 28
P. Huttenlocher, A. Dabholkar (1997)
Regional differences in synaptogenesis in human cerebral cortex
Journal of Comparative Neurology, 387
Adriana Di Martino, D. Fair, C. Kelly, T. Satterthwaite, F. Castellanos, M. Thomason, R. Craddock, B. Luna, Bennett L. Leventhal, X. Zuo, M. Milham (2014)
Unraveling the Miswired Connectome: A Developmental Perspective
Neuron, 83
Catherine Hartley, F. Lee (2015)
Sensitive Periods in Affective Development: Nonlinear Maturation of Fear Learning
Neuropsychopharmacology, 40
D. Gee, R. Fetcho, D. Jing, Anfei Li, C. Glatt, Andrew Drysdale, A. Cohen, Danielle Dellarco, Ruirong Yang, A. Dale, T. Jernigan, F. Lee, B. Casey (2016)
Individual differences in frontolimbic circuitry and anxiety emerge with adolescent changes in endocannabinoid signaling across species
Proceedings of the National Academy of Sciences, 113
L. Selemon (2013)
A role for synaptic plasticity in the adolescent development of executive function
Translational Psychiatry, 3
BJ Casey, J. Nigg, S. Durston (2007)
New potential leads in the biology and treatment of attention deficit-hyperactivity disorder
Current Opinion in Neurology, 20
A. Kelly, A. Martino, L. Uddin, Z. Shehzad, D. Gee, P. Reiss, D. Margulies, F. Castellanos, M. Milham (2009)
Development of anterior cingulate functional connectivity from late childhood to early adulthood.
Cerebral cortex, 19 3
Stephen Smith, Thomas Nichols, D. Vidaurre, A. Winkler, T. Behrens, M. Glasser, K. Uğurbil, D. Barch, D. Essen, K. Miller (2015)
A positive-negative mode of population covariation links brain connectivity, demographics and behavior
Nature neuroscience, 18
P. Shaw, M. Malek, B. Watson, W. Sharp, A. Evans, D. Greenstein (2012)
Development of Cortical Surface Area and Gyrification in Attention-Deficit/Hyperactivity Disorder
Biological Psychiatry, 72
B. Casey, T. Cannonier, M. Conley, A. Cohen, D. Barch, M. Heitzeg, M. Soules, Theresa Teslovich, Danielle Dellarco, H. Garavan, C. Orr, T. Wager, M. Banich, N. Speer, M. Sutherland, M. Riedel, A. Dick, J. Bjork, K. Thomas, B. Chaarani, M. Mejia, D. Hagler, M. Cornejo, Chelsea Sicat, M. Harms, N. Dosenbach, M. Rosenberg, E. Earl, H. Bartsch, R. Watts, J. Polimeni, J. Kuperman, D. Fair, A. Dale (2018)
The Adolescent Brain Cognitive Development (ABCD) study: Imaging acquisition across 21 sites
Developmental Cognitive Neuroscience, 32
M. Corbetta, G. Shulman (2002)
Control of goal-directed and stimulus-driven attention in the brain
Nature Reviews Neuroscience, 3
Stephen Smith, D. Vidaurre, C. Beckmann, M. Glasser, M. Jenkinson, K. Miller, Thomas Nichols, E. Robinson, G. Salimi-Khorshidi, M. Woolrich, D. Barch, K. Uğurbil, D. Essen (2013)
Functional connectomics from resting-state fMRI
Trends in Cognitive Sciences, 17
Adam Chekroud (2017)
Bigger Data, Harder Questions-Opportunities Throughout Mental Health Care.
JAMA psychiatry, 74 12
B. Casey, S. Getz, A. Galván (2008)
The adolescent brain.
Developmental review : DR, 28 1
D. Jangraw, J. Gonzalez-Castillo, D. Handwerker, Merage Ghane, M. Rosenberg, Puja Panwar, P. Bandettini (2018)
A functional connectivity-based neuromarker of sustained attention generalizes to predict recall in a reading task
NeuroImage, 166
M. Rosenberg, Shenmin Zhang, W. Hsu, D. Scheinost, E. Finn, X. Shen, R. Constable, C. Li, M. Chun (2016)
Methylphenidate Modulates Functional Network Connectivity to Enhance Attention
The Journal of Neuroscience, 36
K. Franke, E. Luders, A. May, M. Wilke, Christian Gaser (2012)
Brain maturation: Predicting individual BrainAGE in children and adolescents using structural MRI
NeuroImage, 63
Thomas Insel, Bruce Cuthbert, Marjorie Garvey, R. Heinssen, Daniel Pine (2018)
Wesleyan University From the SelectedWorks of Charles A . Sanislow , Ph . D . 2010 Research Domain Criteria ( RDoC ) : Toward a New Classification Framework for Research on Mental Disorders
B. Yeo, Fenna Krienen, J. Sepulcre, M. Sabuncu, D. Lashkari, Marisa Hollinshead, J. Roffman, J. Smoller, Lilla Zöllei, J. Polimeni, B. Fischl, Hesheng Liu, R. Buckner (2011)
The organization of the human cerebral cortex estimated by intrinsic functional connectivity
D. Rhein, M. Mennes, H. Ewijk, A. Groenman, M. Zwiers, J. Oosterlaan, D. Heslenfeld, B. Franke, P. Hoekstra, S. Faraone, C. Hartman, J. Buitelaar (2015)
The NeuroIMAGE study: a prospective phenotypic, cognitive, genetic and MRI study in children with attention-deficit/hyperactivity disorder. Design and descriptives
European Child & Adolescent Psychiatry, 24
Daniel Kessler, Mike Angstadt, C. Sripada (2016)
Growth Charting of Brain Connectivity Networks and the Identification of Attention Impairment in Youth.
JAMA psychiatry, 73 5
E. Finn, D. Scheinost, Daniel Finn, X. Shen, X. Papademetris, T. Constable (2017)
Can brain state be manipulated to emphasize individual differences in functional connectivity?
NeuroImage, 160
BTT Yeo (2011)
10.1152/jn.00338.2011
J. Neurophysiol., 106
M. Landers, R. Sullivan (2012)
The Development and Neurobiology of Infant Attachment and Fear
Developmental Neuroscience, 34
T. Yarkoni, Jacob Westfall (2017)
Choosing Prediction Over Explanation in Psychology: Lessons From Machine Learning
Perspectives on Psychological Science, 12
(1956)
Empirical orthogonal functions and statistical weather prediction
Demonstrates that functional connectivity patterns become more stable and unique during adolescence, and that delays in this pattern of development are related to psychiatric symptoms
F. Castellanos, A. Martino, R. Craddock, A. Mehta, M. Milham (2013)
Clinical applications of the functional connectome
NeuroImage, 80
J. Giedd, J. Blumenthal, N. Jeffries, F. Castellanos, H. Liu, A. Zijdenbos, T. Paus, Alan Evans, J. Rapoport (1999)
Brain development during childhood and adolescence: a longitudinal MRI study
Nature Neuroscience, 2
N. Dosenbach, J. Koller, E. Earl, Oscar Miranda-Dominguez, Rachel Klein, A. Van, A. Snyder, B. Nagel, J. Nigg, Annie Nguyen, Victoria Wesevich, D. Greene, D. Fair (2017)
Real-time motion analytics during brain MRI improve data quality and reduce costs
NeuroImage, 161
F. Hoeft, Bruce McCandliss, J. Black, Alexander Gantman, N. Zakerani, C. Hulme, H. Lyytinen, S. Whitfield-Gabrieli, G. Glover, A. Reiss, J. Gabrieli (2010)
Neural systems predicting long-term outcome in dyslexia
Proceedings of the National Academy of Sciences, 108
Ylva Østby, C. Tamnes, A. Fjell, L. Westlye, P. Due-Tønnessen, K. Walhovd (2009)
Heterogeneity in Subcortical Brain Development: A Structural Magnetic Resonance Imaging Study of Brain Maturation from 8 to 30 Years
The Journal of Neuroscience, 29
K. Yoo, M. Rosenberg, W. Hsu, S. Zhang, C. Li, D. Scheinost, T. Constable, M. Chun (2018)
Connectome-based predictive modeling of attention: Comparing different functional connectivity features and prediction methods across datasets
NeuroImage, 167
G. Salum, A. Gadelha, P. Pan, Tais Moriyama, A. Graeff-Martins, Ana Tamanaha, P. Alvarenga, F. Krieger, B. Fleitlich-Bilyk, A. Jackowski, J. Sato, E. Brietzke, G. Polanczyk, H. Brentani, J. Mari, M. Rosário, G. Manfro, R. Bressan, M. Mercadante, E. Miguel, L. Rohde (2015)
High risk cohort study for psychiatric disorders in childhood: rationale, design, methods and preliminary results
International Journal of Methods in Psychiatric Research, 24
J. Swanson (2012)
The UK Biobank and selection bias
The Lancet, 380
E. Finn, R. Constable (2016)
Individual variation in functional brain connectivity: implications for personalized approaches to psychiatric disease
Dialogues in Clinical Neuroscience, 18
Jonathan Power, D. Fair, B. Schlaggar, S. Petersen (2010)
The Development of Human Functional Brain Networks
Neuron, 67
Jonathan Power, A. Cohen, S. Nelson, G. Wig, K. Barnes, J. Church, Alecia Vogel, Timothy Laumann, F. Miezin, B. Schlaggar, S. Petersen (2011)
Functional Network Organization of the Human Brain
Neuron, 72
L. Jollans, R. Whelan (2016)
The Clinical Added Value of Imaging: A Perspective From Outcome Prediction.
Biological psychiatry. Cognitive neuroscience and neuroimaging, 1 5
M. Sweeting, S. Thompson (2012)
Making predictions from complex longitudinal data, with application to planning monitoring intervals in a national screening programme
Journal of the Royal Statistical Society. Series A, (Statistics in Society), 175
(2017)
Suggests statistical tools useful for analyzing big data in human neuroscience
D. Fair, A. Cohen, N. Dosenbach, J. Church, F. Miezin, D. Barch, M. Raichle, S. Petersen, B. Schlaggar (2008)
The maturing architecture of the brain's default network
Proceedings of the National Academy of Sciences, 105
B. Couvy-Duchesne, J. Ebejer, N. Gillespie, D. Duffy, I. Hickie, P. Thompson, N. Martin, G. Zubicaray, K. Mcmahon, S. Medland, M. Wright (2016)
Head Motion and Inattention/Hyperactivity Share Common Genetic Influences: Implications for fMRI Studies of ADHD
PLoS ONE, 11
E. Finn, X. Shen, D. Scheinost, M. Rosenberg, Jessica Huang, M. Chun, X. Papademetris, R. Constable (2015)
Functional connectome fingerprinting: Identifying individuals based on patterns of brain connectivity
Nature neuroscience, 18
J. Cohen, N. Daw, Barbara Engelhardt, U. Hasson, Kai Li, Y. Niv, K. Norman, Jonathan Pillow, P. Ramadge, N. Turk-Browne, Theodore Willke (2017)
Computational approaches to fMRI analysis
Nature Neuroscience, 20
B. Casey, A. Heller, D. Gee, A. Cohen (2017)
Development of the emotional brain
Neuroscience Letters, 693
M. Rosenberg, W. Hsu, D. Scheinost, R. Constable, M. Chun (2018)
Connectome-based Models Predict Separable Components of Attention in Novel Individuals
Journal of Cognitive Neuroscience, 30
10.1101/110429
X. Shen, E. Finn, D. Scheinost, M. Rosenberg, M. Chun, X. Papademetris, R. Constable (2017)
Using connectome-based predictive modeling to predict individual behavior from brain connectivity
Nature Protocols, 12
S. Brown, T. Brumback, K. Tomlinson, Kevin Cummins, W. Thompson, B. Nagel, M. Bellis, S. Hooper, D. Clark, T. Chung, Brant Hasler, I. Colrain, F. Baker, D. Prouty, A. Pfefferbaum, E. Sullivan, K. Pohl, T. Rohlfing, B. Nichols, Weiwei Chu, S. Tapert (2015)
The National Consortium on Alcohol and NeuroDevelopment in Adolescence (NCANDA): A Multisite Study of Adolescent Development and Substance Use.
Journal of studies on alcohol and drugs, 76 6
J. Gabrieli, Satrajit Ghosh, S. Whitfield-Gabrieli (2015)
Prediction as a Humanitarian and Pragmatic Contribution from Human Cognitive Neuroscience
Neuron, 85
S. Noble, D. Scheinost, E. Finn, X. Shen, X. Papademetris, S. McEwen, C. Bearden, J. Addington, B. Goodyear, K. Cadenhead, Heline Mirzakhanian, B. Cornblatt, D. Olvet, D. Mathalon, T. McGlashan, D. Perkins, A. Belger, L. Seidman, H. Thermenos, M. Tsuang, T. Erp (2017)
Multisite reliability of MR-based functional connectivity
NeuroImage, 146
F. Lee, H. Heimer, J. Giedd, E. Lein, N. Šestan, D. Weinberger, B. Casey (2014)
Adolescent mental health—Opportunity and obligation
Science, 346
Samantha Strong, E. Silson, A. Gouws, A. Morland, D. McKeefry (2016)
A Direct Demonstration of Functional Differences between Subdivisions of Human V5/MT+
Cerebral Cortex (New York, NY), 27
A. Martino, Chaogan Yan, Qingyang Li, Erin Denio, F. Castellanos, K. Alaerts, Jeffrey Anderson, M. Assaf, S. Bookheimer, M. Dapretto, Ben Deen, S. Delmonte, I. Dinstein, B. Ertl-Wagner, D. Fair, L. Gallagher, Daniel Kennedy, C. Keown, C. Keysers, J. Lainhart, C. Lord, B. Luna, V. Menon, N. Minshew, Christopher Monk, S. Mueller, R. Müller, M. Nebel, J. Nigg, K. O'Hearn, K. Pelphrey, S. Peltier, J. Rudie, S. Sunaert, M. Thioux, J. Tyszka, L. Uddin, J. Verhoeven, N. Wenderoth, J. Wiggins, S. Mostofsky, M. Milham (2013)
The Autism Brain Imaging Data Exchange: Towards Large-Scale Evaluation of the Intrinsic Brain Architecture in Autism
Molecular psychiatry, 19
(2017)
Argues for an increased focus on prediction over explanation for understanding human behavior
F. Alfaro-Almagro, M. Jenkinson, N. Bangerter, J. Andersson, L. Griffanti, G. Douaud, S. Sotiropoulos, S. Jbabdi, Moisés Hernández-Fernández, Emmanuel Vallée, D. Vidaurre, Matthew Webster, Paul McCarthy, C. Rorden, Alessandro Daducci, D. Alexander, Hui Zhang, I. Dragonu, P. Matthews, K. Miller, Stephen Smith (2017)
Image processing and Quality Control for the first 10,000 brain imaging datasets from UK Biobank
Neuroimage, 166
P. Flechsig
Anatomie des menschlichen Gehirns und Rückenmarks : auf myelogenetischer Grundlage
T. Kaufmann, D. Alnæs, N. Doan, C. Brandt, O. Andreassen, L. Westlye (2017)
Delayed stabilization and individualization in connectome development are related to psychiatric disorders
Nature Neuroscience, 20
Siobhan Pattwell, S. Duhoux, Catherine Hartley, David Johnson, D. Jing, Mark Elliott, Erika Ruberry, A. Powers, Natasha Mehta, Ruirong Yang, Fatima Soliman, C. Glatt, B. Casey, I. Ninan, F. Lee (2012)
Altered fear learning across development in both mouse and human
Proceedings of the National Academy of Sciences, 109
L. Leijenhorst, Kiki Zanolie, Kiki Zanolie, C. Meel, C. Meel, P. Westenberg, S. Rombouts, S. Rombouts, E. Crone (2010)
What motivates the adolescent? Brain regions mediating reward sensitivity across adolescence.
Cerebral cortex, 20 1
T. Sejnowski, T. Sejnowski, P. Churchland, P. Churchland, J. Movshon (2014)
Putting big data to good use in neuroscience
Nature Neuroscience, 17
D. Essen, M. Glasser (2016)
The Human Connectome Project: Progress and Prospects
Cerebrum: the Dana Forum on Brain Science, 2016
R. Whelan, R. Watts, C. Orr, R. Althoff, E. Artiges, T. Banaschewski, G. Barker, A. Bokde, C. Büchel, F. Carvalho, P. Conrod, H. Flor, M. Fauth-Bühler, V. Frouin, J. Gallinat, G. Gan, P. Gowland, A. Heinz, B. Ittermann, C. Lawrence, K. Mann, J. Martinot, F. Nees, Nick Ortiz, M. Paillère-Martinot, T. Paus, Z. Pausova, M. Rietschel, T. Robbins, M. Smolka, A. Ströhle, G. Schumann, H. Garavan (2014)
Neuropsychosocial profiles of current and future adolescent alcohol misusers
Nature, 512
N Gogtay (2010)
10.1016/j.bandc.2009.08.009
Brain. Cogn., 72
(2016)
Introduces a model that generalizes across datasets to predict sustained attention from task-based and resting-state functional connectivity
(2015)
Identifies a pattern of functional brain connectivity related to a 'positivenegative' axis of lifestyle, demographic and psychometric factors
T. Brown, J. Kuperman, Yoonho Chung, M. Erhart, Connor McCabe, D. Hagler, Vijay Venkatraman, N. Akshoomoff, D. Amaral, C. Bloss, B. Casey, Linda Chang, T. Ernst, J. Frazier, J. Gruen, W. Kaufmann, T. Kenet, D. Kennedy, S. Murray, E. Sowell, E. Sowell, T. Jernigan, A. Dale (2012)
Neuroanatomical Assessment of Biological Maturity
Current Biology, 22
Catherine Hartley, L. Somerville (2015)
The neuroscience of adolescent decision-making
Current Opinion in Behavioral Sciences, 5
Theresa Teslovich, M. Mulder, Nicholas Franklin, Erika Ruberry, A. Millner, L. Somerville, P. Simen, S. Durston, S. Durston, B. Casey (2014)
Adolescents let sufficient evidence accumulate before making a decision when large incentives are at stake.
Developmental science, 17 1
P. Austin, L. Brunner (2004)
Inflation of the type I error rate when a continuous confounding variable is categorized in logistic regression analyses
Statistics in Medicine, 23
P. Shaw, N. Kabani, J. Lerch, K. Eckstrand, R. Lenroot, N. Gogtay, D. Greenstein, L. Clasen, Alan Evans, J. Rapoport, J. Giedd, S. Wise (2008)
Neurodevelopmental Trajectories of the Human Cerebral Cortex
The Journal of Neuroscience, 28
J. Arnett (1996)
Sensation seeking, aggressiveness, and adolescent reckless behavior.
Personality and Individual Differences, 20
M. Rosenberg, E. Finn, D. Scheinost, X. Papademetris, X. Shen, R. Constable, M. Chun (2015)
A neuromarker of sustained attention from whole-brain functional connectivity
Nature neuroscience, 19
J. Hudziak, T. Achenbach, R. Althoff, D. Pine (2007)
A dimensional approach to developmental psychopathology
International Journal of Methods in Psychiatric Research, 16
T. Jernigan, T. Brown, D. Hagler, N. Akshoomoff, H. Bartsch, Erik Newman, W. Thompson, C. Bloss, S. Murray, N. Schork, David Kennedy, J. Kuperman, Connor McCabe, Yoonho Chung, Ondrej Libiger, M. Maddox, B. Casey, Linda Chang, T. Ernst, J. Frazier, J. Gruen, E. Sowell, T. Kenet, W. Kaufmann, S. Mostofsky, D. Amaral, A. Dale (2016)
The Pediatric Imaging, Neurocognition, and Genetics (PING) Data Repository
NeuroImage, 124
Compares current methods for controlling for motion artifacts in functional connectivity data
Z. Pausova, T. Paus, M. Abrahamowicz, J. Almerigi, Nadine Arbour, M. Bernard, D. Gaudet, P. Hanzálek, P. Hamet, Alan Evans, Michael Kramer, L. Laberge, S. Leal, G. Leonard, J. Lerner, R. Lerner, Jean Mathieu, M. Perron, B. Pike, A. Pitiot, L. Richer, J. Séguin, Catriona Syme, R. Toro, R. Tremblay, S. Veillette, K. Watkins (2007)
Genes, maternal smoking, and the offspring brain and body during adolescence: Design of the Saguenay Youth Study
Human Brain Mapping, 28
B. Casey, A. Galván, L. Somerville (2015)
Beyond simple models of adolescence to an integrated circuit-based account: A commentary
Developmental Cognitive Neuroscience, 17
A. Meyer (2003)
Risk-Taking, Adolescence
R. Buckner, Fenna Krienen (2013)
The evolution of distributed association networks in the human brain
Trends in Cognitive Sciences, 17
A. Cohen, Kaitlyn Breiner, L. Steinberg, R. Bonnie, Elizabeth Scott, Kim Taylor-Thompson, Marc Rudolph, J. Chein, J. Richeson, A. Heller, Melanie Silverman, Danielle Dellarco, D. Fair, A. Galván, B. Casey (2016)
When Is an Adolescent an Adult? Assessing Cognitive Control in Emotional and Nonemotional Contexts
Psychological Science, 27
RL Buckner, JR Andrews-Hanna, DL Schacter (2008)
The brain’s default network: anatomy, function, and relevance to disease
Ann. N. Y. Acad. Sci., 1124
Michael Ferguson, Michael Ferguson, Jeffrey Anderson, R. Spreng (2017)
Fluid and flexible minds: Intelligence reflects synchrony in the brain’s intrinsic network architecture
Network Neuroscience, 1
Bruce Perry, Marsha White, Steve Dorman, Gogtay Md, J. Giedd, Leslie Ba, Kiralee Bs, D. Greenstein, A. Vaituzis, David Bs, Thomas Nugent, Iii Ab, L. Clasen, A. Toga, J. Rapoport, P. Thompson, Johnson Sb, Blum Rw, Tiemeier H, Lenroot Rk, Greenstein Dk, Tran L, Pierson Giedd, Jn Cerebellum, Giedd Jn, Lalonde Fm, Celano Mj, White Sl, Wallace Gl, Lee Nr, Lenroot Magnetic, Shaw P, Addington A, Evans A, Rapoport J (2004)
Dynamic mapping of human cortical development during childhood through early adulthood.
Proceedings of the National Academy of Sciences of the United States of America, 101 21
(2016)
Provides evidence for a shift in cognitive capacity in emotional contexts that accompanies dynamic changes
(2015)
Presents evidence that adolescent-specific changes in self-control emerge from the fine-tuning of connections within and between subcortical
F. Benes, J. Taylor, Miles Cunningham (2000)
Convergence and plasticity of monoaminergic systems in the medial prefrontal cortex during the postnatal period: implications for the development of psychopathology.
Cerebral cortex, 10 10
Highlights benefits of predictive brain-based models for personalized medicine and education
Lindsay Alexander, J. Escalera, Lei Ai, Charissa Andreotti, Karina Febre, Alexander Mangone, Natan, Vega-Potler, N. Langer, A. Alexander, Meagan Kovacs, S. Litke, B. O’Hagan, Jennifer, Andersen, Batya Bronstein, Anastasia Bui, Marijayne Bushey, H. Butler, V. Castagna, Nicolas, Camacho, E. Chan, Danielle Citera, J. Clucas, Samantha Cohen, Sarah Dufek, Megan Eaves, Brian, Fradera, J. Gardner, Natalie Grant-Villegas, Gabrielle Green, Camille Gregory, Emily Hart, Shana Harris, M. Horton, D. Kahn, Katherine Kabotyanski, B. Karmel, S. Kelly, Kayla Kleinman, B. Koo, Eliza Kramer, Elizabeth Lennon, C. Lord, G. Mantello, A. Margolis, R. Kathleen, Merikangas, Judith Milham, G. Minniti, Rebecca Neuhaus, Alexandra Levine, Yael Osman, C. Lucas, Parra, K. Pugh, Amy Racanello, Anita Restrepo, Tian Saltzman, Batya Septimus, Rachel, Waltz, Anna Yeo, F. Castellanos, A. Klein, T. Paus, B. Leventhal, R., Cameron Craddock, H. Koplewicz, M. Milham (2019)
Descriptor : An open resource for transdiagnostic research in pediatric mental health and learning disorders
J. Dubois, R. Adolphs (2016)
Building a Science of Individual Differences from fMRI
Trends in Cognitive Sciences, 20
R. Buckner, J. Andrews-Hanna, D. Schacter (2008)
The Brain's Default Network
Annals of the New York Academy of Sciences, 1124
R. Kievit, A. Brandmaier, G. Ziegler, A. Harmelen, S. Mooij, M. Moutoussis, I. Goodyer, E. Bullmore, Peter Jones, P. Fonagy, U. Lindenberger, R. Dolan (2017)
Developmental cognitive neuroscience using latent change score models: A tutorial and applications
Developmental Cognitive Neuroscience, 33
S. Hyman (2010)
The diagnosis of mental disorders: the problem of reification.
Annual review of clinical psychology, 6
B. Casey (2015)
Beyond simple models of self-control to circuit-based accounts of adolescent behavior.
Annual review of psychology, 66
A. Holmes, Marisa Hollinshead, Timothy O’Keefe, Victor Petrov, Gabriele Fariello, L. Wald, B. Fischl, B. Rosen, R. Mair, J. Roffman, J. Smoller, R. Buckner (2015)
Brain Genomics Superstruct Project initial data release with structural, functional, and behavioral measures
Scientific Data, 2
E. Sowell, P. Thompson, C. Holmes, T. Jernigan, A. Toga (1999)
In vivo evidence for post-adolescent brain maturation in frontal and striatal regions
Nature Neuroscience, 2
A. Hofstadter (1951)
Explanation and Necessity
Philosophy and Phenomenological Research, 11
LH Somerville (2013)
The teenage brain: sensitivity to social evaluation
Curr. Dir. Psychol. Sci., 22
Danping Liu, P. Albert (2014)
Combination of longitudinal biomarkers in predicting binary events.
Biostatistics, 15 4
P. Albert (2012)
A linear mixed model for predicting a binary event from longitudinal data under random effects misspecification
Statistics in Medicine, 31
L. Somerville (2013)
The Teenage Brain
Current Directions in Psychological Science, 22

Publisher: Springer Journals
Copyright: Copyright © 2018 by The Author(s)
Subject: Science, Humanities and Social Sciences, multidisciplinary; Science, Humanities and Social Sciences, multidisciplinary; Science, multidisciplinary
eISSN: 2041-1723
DOI: 10.1038/s41467-018-02887-9
pmid: 29467408
Publisher site: See Article on Publisher Site

Abstract

REVIEW ARTICLE DOI: 10.1038/s41467-018-02887-9 OPEN Prediction complements explanation in understanding the developing brain 1 1 1,2 Monica D. Rosenberg , B.J. Casey & Avram J. Holmes A central aim of human neuroscience is understanding the neurobiology of cognition and behavior. Although we have made signiﬁcant progress towards this goal, reliance on group- level studies of the developed adult brain has limited our ability to explain population variability and developmental changes in neural circuitry and behavior. In this review, we suggest that predictive modeling, a method for predicting individual differences in behavior from brain features, can complement descriptive approaches and provide new ways to account for this variability. Highlighting the outsized scientiﬁc and clinical beneﬁts of pre- diction in developmental populations including adolescence, we show that predictive brain- based models are already providing new insights on adolescent-speciﬁc risk-related beha- viors. Together with large-scale developmental neuroimaging datasets and complementary analytic approaches, predictive modeling affords us the opportunity and obligation to identify novel treatment targets and individually tailor the course of interventions for developmental psychopathologies that impact so many young people today. nderstanding how the brain gives rise to cognition and behavior is a fundamental goal of human neuroscience. Scientists, philosophers, and statisticians have long debated the Unature of understanding, but tend to agree that there are two routes to achieving it: 1–4 explanation and prediction . Despite the historical dominance of explanation as a route to understanding, scientists and philosophers of science have emphasized the importance of both 5–8 these approaches . As noted by the philosopher Heather Douglas, “explanation and prediction are best understood in light of each other and thus … should not be viewed as competing goals but rather as two goals wherein the achievement of one should facilitate the achievement of the other” . Foundational cross-species research has made signiﬁcant progress on the path towards neu- roscientiﬁc explanation. Researchers have described neural bases of cognition, characterizing how patterns of brain organization from neural circuits to functional networks relate to behavior 9–14 and psychopathology . Although this work has traditionally taken a cross-sectional, group- level approach to studying the developed adult brain, there is growing consensus that com- prehensive models in neuroscience must account for the facts that neural phenotypes and 15–18 behavior vary widely across the population and change over time within individuals . The road to prediction is less traveled. Recently, however, the use of machine-learning methods to predict behavior from brain measures has become increasingly common, due in part to the emergence of large data sets and new analytic and computational tools . Representing a critical avenue to understanding, these approaches provide new ways to account for 1 2 Department of Psychology, Yale University, New Haven, CT 06520, USA. Department of Psychiatry, Yale University, New Haven, CT 06511, USA. Correspondence and requests for materials should be addressed to M.D.R. (email: monica.rosenberg@yale.edu) NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications 1 | | | 1234567890():,; REVIEW ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 22,23 developmental changes in behavior, dynamic brain systems, and experiment with drugs than children or adults . These 8,20 associated individual differences while offering statistical rigor behaviors are thought to stem from adolescents’ increased 25,26 27,28 and clinical and translational beneﬁts for personalized medicine sensation-seeking and reward-sensitivity , as well as 5,6 29 and education . decreased self-control and emotional regulation abilities, 30–34 Although much predictive modeling research has focused on especially in social contexts . The prevalence of anxiety adults, forecasting outcomes in childhood and adolescence pre- disorders also peaks in adolescence, underscoring this develop- sents unique opportunities for scientiﬁc discovery and clinical mental period as a time of both vulnerability and opportunity for application. First and foremost, biological and statistical models intervention . that account for developmental change are necessary for truly Given that risky behavior during development has potentially understanding how neural circuits emerge to give rise to cogni- dire consequences, why does it persist across generations and tion and behavior. Predicting behavior from brain features during species ? Fear learning provides a useful example of potential development represents initial progress towards this goal, and evolutionary beneﬁts of seemingly costly behavior during 36,37 predicting future outcomes from past developmental changes adolescence . Across altricial species, whose young rely on represents an important next step. Predictive models of current parents for survival, fear learning is suppressed in early infancy, and future behavior may be especially beneﬁcial in adolescence, a presumably to ensure caregiver attachment even in cases of 38,39 developmental period of rapid social, emotional, psychological, neglect or abuse . In adolescence, fear of previously aversive and physical change characterized by mental and physical health environmental contexts is diminished whereas fear of previously vulnerability, but also opportunities for growth and aversive cues (i.e., conditioned stimuli) is ampliﬁed, a pattern that intervention . may facilitate exploration and independence but also safety from 40,41 In this forward-looking review, we highlight how predictive immediate threat . Importantly, these survival-relevant beha- modeling in developmental neuroscience can account for devel- viors do not develop in a vacuum. Rather, common genetic 42 43–45 opmental trajectories in behavior, dynamic brain systems, and variations and early life stressors affect how fear learning individual differences in both. After discussing these concepts in changes over time, inﬂuencing risk for negative outcomes such as the context of adolescence, we introduce predictive modeling and anxiety disorders . Just as developmental changes in fear its applications in developmental populations. Using adolescence learning confer both costs and beneﬁts, changes in risk taking as a case study, we address two complementary questions. First, during adolescence are advantageous at the group level but in how can prediction inform models of risk taking, a phenotype some contexts may be detrimental for the individual . The same that is strikingly elevated in adolescence? Second, how can con- is likely true for other processes following their own nonlinear sidering adolescence inform predictive modeling techniques and trajectories across development, including decision making , 47 48 motivate future research? We conclude by emphasizing the reward learning , and sensitivity to motivational , appetitive, 49,50 importance of approaches that predict current and future beha- and aversive cues . vior from developmental trajectories of brain structure and Although stereotypes can paint teenagers in an unﬂattering function. In doing so, we discuss how these methods complement light, recognizing that adolescent behaviors are single points and extend ongoing research on the neurodevelopmental pro- along broader, evolutionarily advantageous developmental tra- cesses that underlie the emergence and disruption of cognition jectories provides a more accurate, nuanced (and perhaps and behavior. sympathetic) picture. Analytic perspectives that consider beha- vioral shifts during the transitions into and out of adolescence, as well as their differential expression across environmental and Changes in behavior and brain systems across adolescence social contexts, are necessary for understanding how the brain Human abilities and behavior change dramatically across the gives rise to behavior over time. lifespan, emerging over development from the dynamic interplay between genes and experience. Developmental changes reﬂect Dynamic brain systems. The nonlinear behavioral trajectories neurobiological constraints shaped by evolution to meet the observed across adolescence emerge from a cascade of hier- unique challenges of each stage of life, including adolescence . archical changes in brain circuitry that were themselves shaped That evolutionary pressures have presumably tailored adolescent over the course of our evolutionary lineage . First to mature are behavior to facilitate the transition to independence, however, is connections within subcortical-limbic circuits, followed by con- frequently overlooked. Instead, adolescents, whose behavior is nections between cerebral cortex and subcortical-limbic circuits, sometimes judged as immature relative to their physical devel- 51,52 and, ﬁnally, connections across cortex . 22–24 opment, are often considered impaired mini-adults . In the Evidence for this developmental cascade comes from observa- following section, we emphasize the importance of considering tions of earlier changes in synaptic morphology and neurotrans- developmental changes, dynamic brain systems that unfold over mitter systems in subcortical relative to cortical regions and an time, and interindividual variability when seeking to establish earlier plateau in synaptic formation and subsequent pruning in descriptive and predictive models of behavior. unimodal sensory, motor, and subcortical regions relative to 53,54 multimodal association areas . These processes likely con- Developmental trajectories in behavior. When we think of the tribute to gray matter volume and cortical thickness changes 55–58 prototypical adolescent (or recall our own teenage years), a observed during adolescence and early adulthood that end in 59–61 number of quintessential traits may come to mind. We might the association cortices . Selective degradation of excitatory consider (or remember) risky behaviors like dangerous driving, synapses also affects the excitatory-inhibitory balance across illegal substance use, irresponsible sexual behavior; preoccupa- cortex, an equilibrium related to shifts in cognitive abilities and 51,62 tions with peer groups and social hierarchies; an uptick in feelings behavior . The relative decrease in prefrontal behavioral of anxiety; and heated conﬂicts with parents, teachers, or other regulation is reﬂected in changes in dopamine receptor density, well-meaning ﬁgures of authority. related to learning and reward prediction, that peak in the Epidemiological studies conﬁrm that our stereotypes largely striatum during adolescence but not until early adulthood in the 63–65 reﬂect typical adolescent behavior. Adolescents are more likely to prefrontal cortex . be injured or killed in motor vehicle accidents, contract sexually Structural and functional brain connections follow similar transmitted infections, engage in criminal activity, and patterns of development, providing additional evidence for a 2 NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 REVIEW ARTICLE YOUth / CID NeuroIMAGE Gen R IMAGEN SYS PING HBN ABCD NCANDA ENIGMA PNC HCP-D ABIDE CoRR ADHD-200 c-VEDA HRC Fig. 1 Existing, ongoing, or planned data sets including structural and/or functional neuroimaging data from ~500 or more children or adolescents. These data sets, which represent both prospective and retrospective samples, include the Adolescent Brain Cognitive Development study (ABCD; USA), 82 80 Healthy Brain Network (HBN; USA), Lifespan Human Connectome Project Development (HCP-D; USA), National Consortium on Alcohol and 149 150 NeuroDevelopment in Adolescence (NCANDA; USA), Pediatric Imaging, Neurocognition, and Genetics study (PING; USA), Philadelphia 151 152 Neurodevelopmental Cohort (PNC; USA), Saguenay Youth Study (SYS; Canada), High Risk Cohort Study for the Development of Childhood 153 81 Psychiatric Disorders (HRC; Brazil), Autism Brain Imaging Data Exchange (ABIDE; USA, Germany, Ireland, Belgium, Netherlands), Enhancing 154 79 NeuroImaging Genetics through Meta-Analysis (ENIGMA; worldwide), IMAGEN (England, Ireland, France, Germany), Dutch YOUth cohort (part of 155 156 the Consortium on Individual Development, or CID; Netherlands), Generation R Study (Gen R; Netherlands), NeuroIMAGE (follow-up of the Dutch arm of the International Multicenter ADHD Genetics, or IMAGE, project; Netherlands), Consortium on Vulnerability to Externalizing Disorders and 157 108 Addictions (c-VEDA; UK, India), Consortium for Reliability and Reproducibility (CoRR; China, USA, Canada, Germany), and ADHD-200 (USA, China). Although samples are distributed across the globe, African, Middle Eastern, South Asian, Oceanian, and Central and South American populations are underrepresented. Data collection efforts in these regions and others will be important for ensuring diverse, representative samples that will allow researchers to uncover general principles of the developing brain. (Map outline courtesy of Wikimedia user ‘Loadﬁle’ and is licensed under a CC BY SA 3.0 license) hierarchically emerging system ﬁrst dominated by mature Individual differences. Although neurobiology and behavior subcortical circuits and then balanced through interactions with tend to unfold in predictable ways across development, signiﬁcant 51,66 late-maturing prefrontal systems . As early as 1920, Flechsig’s individual differences lie atop this scaffolding. This variability histological studies revealed protracted myelin development in applies not only to an adolescent’s current behavioral and neural 67,68 association cortex .Reﬂecting this property of brain matura- characteristics, but also to their past and future phenotypes. That tion, diffusion tensor imaging studies, which measure water is, while one stereotype of adolescents is that they engage in risky diffusion modulated in part by axon myelination, suggest that the behaviors such as binge drinking, there are plenty of young development of posterior cortical-subcortical tracts precedes that people who do not ﬁt this mold. Even among adolescents who of fronto-subcortical tracts supporting top-down control of drink excessively, some may go on to develop substance use 69–72 behavior . Functional brain connectivity studies support disorders, while others may never progress to disordered these results, observing a general pattern of weakening short- drinking. range functional connections followed by strengthening long- Despite recognizing these individual differences, in research, 73–76 range cortical connections across adolescence . clinical, legal, and educational practice, we often treat variance Altogether, this work provides evidence for the progressive around average behavioral and neural phenotypes and trajectories development of connectivity within and between subcortical and as noise, or collapse it into discrete categories (e.g., patients vs. cortical brain regions, and offers a plausible neurobiological controls, adults vs. minors, etc.). Although these groups can be account of nonlinear trajectories in risk-related processes such as useful in practice, they do not necessarily represent biologically self-control, reward sensitivity, and emotion regulation. Emo- plausible or informative qualitative distinctions. Instead, tional reactivity, for example, may arise from the early dominance approaches that characterize the normative trajectories of of subcortical over cortical circuitry, later waning as cortical- dimensional behavioral and neural phenotypes, and investigate subcortical circuits related to top-down control, and then cortical how genetics and experience affect the timing and shape of these circuits involved in processes such as cognitive reappraisal, curves, are necessary for understanding how these processes 52 18,66,77 mature during adolescence and adulthood . More broadly, these unfold in development . In addition to informing models in ﬁndings highlight how approaching the study of adolescence basic science, individual differences approaches can provide from a dynamic, multimodal, circuit-based perspective (rather clinically applicable insight into the factors that confer risk for than a view that focuses on snapshots of individual brain regions and resilience to psychopathology and guide personalized in isolation) can inform our understanding of self-regulation and treatments . 51,52 risk-taking behavior during development . NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications 3 | | | REVIEW ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 Box 1 | Predictive modeling Whereas descriptive modeling is the process of learning associations between features and outcomes, predictive modeling leverages these relationships to make predictions from previously unseen data. Here, we reserve the term “prediction” for the output of models applied to novel individuals rather than to describe brain–behavior correlations . Although prediction pipelines are diverse, they typically involve four primary steps: feature selection, model building, model testing, and prediction evaluation (Fig. 2). Importantly, both feature selection and model building are performed using only training data. The resulting model is then applied unaltered to data from previously unseen individuals. Feature selection: Methods for feature selection, the process of identifying model predictors, fall into two broad categories: hypothesis-driven and data- driven approaches. Hypothesis-driven methods, which leverage existing knowledge to select features, are useful for testing predictions of existing scientiﬁc models. Data-driven methods rely on statistical techniques to identify the features most relevant to individual differences in behavior. These include ﬁlter methods (selecting features based relationships with behavior), wrapper methods (considering the predictive power of different feature combinations, e.g., by systematically eliminating the least predictive features from a model), and embedded methods (incorporating feature selection into model building, such as in lasso, elastic net, and ridge regression) . Both hypothesis- and data-driven approaches can incorporate predictors from multiple domains, including genetics, brain structure and function, and behavior. The developmental trajectories of these measures, such as slope, intercept, or inﬂection point, may also be included. Systematically removing a predictor or predictor class from a model can identify its unique contribution to predicting outcomes or behavior . Although there is no theoretical limit to the number of model features, it is best practice that they not exceed the number of observations to avoid modeling noise (overﬁtting) . Furthermore, it is important to consider the inherent tensions between interpretability, generalizability, and variance explained. While models with fewer features may be easier to interpret, models with more features may capture additional variance in behavior and better characterize complex multimodal neural phenotypes. Model building: Following feature selection, the relationship between predictors and behavior is formalized with a classiﬁer or regression model. The goal of a classiﬁer, such as a support vector machine or logistic regression, is to make discrete predictions. In neuroimaging research, classiﬁers represent the vast majority of predictive models: Of all multivariate models in translational neuroimaging, 75% were built to distinguish patients from control participants, whereas <3% were used to predict continuous symptom scores . Regression models, including linear and support vector regression algorithms, make continuous rather than categorical predictions, and can facilitate the development of transdiagnostic proﬁles of risk or resilience for psychopathology . Both classiﬁers and regression models can be applied to cross-sectional or longitudinal data, and the latter may incorporate techniques such as growth–curve modeling to predict past or future change . Model testing: Model testing, or applying a predictive algorithm to test data to evaluate its generalizability, distinguishes predictive from descriptive 5,8 models. The utility of out-of-sample validation for protecting against overﬁtting and false positives has been discussed in detail elsewhere . Here we highlight one dimension along which potential predictive models vary: how far out of sample they generalize. Internal validation (i.e., k-fold or leave-one-subject-out cross-validation) tests whether a model generalizes to novel individuals from a single data set. Although internal validation is useful for optimizing models and conferring statistical rigor when multiple data sets are not available, it may generate biased estimates of predictive power even when evaluated with permutation testing. Despite this limitation, the vast majority of predictive models in neuroimaging have been tested with internal validation alone . External validation tests whether a model generalizes beyond an initial training data set to individuals from completely independent samples. Curated data sets and platforms such as OpenfMRI that encourage data and model sharing can facilitate external validation and model reﬁnement. Prediction evaluation: Methods of model evaluation depend on whether predictions are discrete or continuous. Classiﬁer output can be evaluated with percent accuracy; sensitivity (the true positive rate, or percent of correctly identiﬁed patients) and speciﬁcity (the true negative rate, or percent of correctly identiﬁed controls); and/or the positive predictive value (percent of individuals called patients who are true patients) and negative predictive value (percent of individuals called controls who are true controls), which depend on disease prevalence. Regression model predictions can be assessed with measures such as correlation or mean-squared error . In all cases it may be useful to visualize all data points to fully evaluate relationships between behavior and predicted scores or category labels. Predictive modeling and its importance in developmental hundreds or thousands, of participants. These small samples, with neuroscience tightly controlled demographics and circumscribed behavioral Studying developmental trajectories, dynamic brain systems, and phenotypes, are not always conducive to studying population individual differences is becoming increasingly feasible with the variability. Larger samples that capture a broad range of pheno- rise of high-throughput data collection efforts . Longitudinal and types provide opportunities not only to describe brain–behavior cross-sectional samples of neuroimaging data from children and relationships, but to predict behavior from brain features at the 79 90,91 adolescents, such as the IMAGEN study , Lifespan Human level of single individuals . In this vein, researchers are Connectome Project Development , Brain Imaging Data searching for neuromarkers, or brain features that predict beha- 81 82 Exchange , Healthy Brain Network Biobank , and Adolescence vior, clinical symptoms, risk for or resilience against illness, or 83 5,6,92 Brain Cognitive Development Study , have accelerated advances treatment response . The pursuit of generalizable neuro- in basic and applied neuroscience (Fig. 1). Collaborative initia- markers goes hand-in-hand with predictive modeling, a techni- tives have also helped democratize data access, improve statistical que that leverages brain–behavior relationships to predict power, and facilitate transparent, reproducible research. The outcomes in novel individuals (Box 1 and Fig. 2). unique challenges posed by large-scale imaging samples, such as The statistician George Box famously claimed that “all models 84 93 how to perform adequate quality control , account for scanner are wrong but some are useful” . Models that predict outcomes 85,86 and site effects , and disentangle meaningful explanatory from previously unseen observations can be especially useful for 87 5,6 power from statistical signiﬁcance , are also motivating the both scientiﬁc discovery and clinical decision-making . From a 88 84 development of new data collection , preprocessing , and basic science perspective, predicting brain–behavior relationships analytic approaches. at the level of single subjects represents progress towards under- Large neuroimaging data sets are not only advancing under- standing how individual differences in brain features relate to standing of how brain features relate to behavior at the group individual differences in cognition and behavior . In addition, level, but are also renewing focus on the individual. Although because predictive models are by deﬁnition validated on inde- cognitive and developmental neuroscientists have long been pendent data, they can help foster robust, reproducible discoveries. interested in interindividual differences in abilities and behavior, The beneﬁts of individualized predictions of current and future traditional experiments have focused on tens, rather than behavior are especially pronounced in developmental populations 4 NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications | | | Predicted label NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 REVIEW ARTICLE ab c d Feature selection Model building Model validation Prediction evaluation Hypothesis-driven Classifier Internal validation Evaluate classification Select based on prior knowledge Categorial outcome Test on left-out individuals Category label Predicted label Category label 12 12 12 Brain feature(s) Brain feature(s) Behavior Data-driven Regression model External validation Evaluate dimensional prediction Select most predictive features Continuous outcome Test on new sites/populations Fig. 2 Schema of key concepts in predicting individual differences in behavior from brain features. a Feature selection. Feature selection techniques fall into two broad categories: hypothesis-driven (top-down) and data-driven (bottom up) approaches. b Model building. Machine-learning algorithms can be used to predict categorical measures, such as clinical diagnoses, or dimensional measures, such as task performance or symptom severity. Here, the dark blue line shows the relationship between a single hypothetical brain feature and a behavioral score. The light blue line illustrates a classiﬁer that divides individuals into categories based on this brain feature. (Note that, unlike in this condensed visualization, behavioral scores are typically related to category labels.) c Model validation. Predictive models are evaluated on previously unseen data—either left-out individuals from the initial data set (internal validation) or individuals from a completely new sample (external validation). d Prediction evaluation. Continuous predictions (bottom and left axes) are evaluated by comparing observed and predicted behavioral measures, e.g., with correlation or mean-squared error. Categorical predictions (top and right axes) are evaluated with percent correct; binary predictions can be assessed with sensitivity and speciﬁcity and/or positive and negative predictive value including adolescents. Because behavior and psychopathology are unemotional situations relative to social or emotionally charged 23,96 best viewed as the result of developmental processes that unfold contexts . 21,66 97 across the lifespan , characterizing individual arcs in Recent work from Rudolph and colleagues used predictive brain–behavior relationships over time can move us even closer modeling to identify the neural basis of this phenomenon, asking to understanding targets for change. Addressing the unique whether functional brain organization looks less mature in challenges presented by prediction in adolescence, including the emotional contexts, and whether this effect relates to individual complex dynamics linking neurobiological, behavioral, and differences in risky behavior. To this end, the authors calculated environmental change, can also help us better model periods such functional connectivity patterns from fMRI data collected while as prenatal development, infancy, aging, and illness course. 212 individuals aged 10–25 performed a go/no-go task in neutral Predictive models may not only contribute to progress in basic and emotional contexts. During emotional contexts, participants developmental neuroscience, but may also have implications for anticipated an aversive noise or a reward; during neutral contexts education, mental health, and legal policy. For example, early there was no anticipation of noise or reward. Using partial least predictions of behavioral impairments could facilitate earlier squares regression and 10-fold cross-validation, the authors ﬁrst treatments and improved health or educational outcomes . Pre- built a model to predict chronological age from functional con- diction can also inform pressing policy questions, such as char- nectivity patterns observed in the neutral context, and then acterizing the maturity of a particular individual in speciﬁc applied the same model to connectivity observed during the contexts to inform whether they should be treated as an adult in emotion manipulation. They found that a prediction made from 95,96 the justice system . Thus, although machine-learning models an individual’s neutral context pattern (their “neutral brain age”) of behavior in development may be “wrong” in the sense that they was closer to their chronological age than a prediction made from (necessarily) simplify complex neurobiological systems, they are their emotional context pattern (their “emotional brain age”). useful in that they can inform theories of how cognition and Further, both predictions tended to be younger than chron- behavior emerge from dynamic brain systems and speak to ological age in teens. Interestingly, there was a trend such that general educational, medical, and social policies. adolescents were more likely to look younger in emotional rela- tive to neutral contexts, but young adults who showed this pattern had greater risk preference and lower risk perception (Fig. 3). Predictive modeling and risk preference These ﬁndings illustrate the power of predictive modeling in One of the most common reprimands of wayward youths is: “Act delineating dynamic developmental changes and individual dif- your age!” The phrase — immortalized in the English-language ferences in risk taking behaviors. idiom “act your age, not your shoe size”— is so ubiquitous that it In addition to helping explain why adolescents may not “act even makes an appearance in song lyrics from the musical artist their age” under emotional arousal, the Rudolph et al. ﬁndings Prince. Its sentiment, however, is not straightforward. What does raise two notable points about predictive brain-based models in it mean for an adolescent or young adult to act his or her age? general. What counts as typical adolescent behavior? One possibility is First, when the model of chronological age was wrong, it was that “act your age” means, “make the most responsible decision wrong in interesting ways: A young adult incorrectly predicted you have the capacity to make”. What this entreaty fails to younger in an emotional context was more likely to show a “risky recognize, however, is that there is a discrepancy between how phenotype” than an individual incorrectly predicted older. Thus, responsibly adolescents and young adults can act in nonsocial, in some cases, model errors may be as informative as model NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications 5 | | | Training sample Behavior (current, future, or change over time) Predicted behavior Predicted behavior REVIEW ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 ab c Positive emotional context Negative emotional context Age Age Fig. 3 Adolescents’ functional connectivity patterns look younger in emotional contexts. Adapted with permission from Rudolph et al. . a Chronological age is plotted against age predicted from functional connectivity patterns observed in positive and negative emotional contexts. Individual points (participants) are ﬁt with polynomial curves. On average, adolescents are predicted younger in emotional contexts. b Adolescents (age 12–18; numerical difference) and young adults (age 18–21; p < 0.1 in negative emotional contexts and p < 0.05 in positive emotional contexts) who are predicted younger in emotional contexts tend to show greater risk preference. This trend is most pronounced in young adulthood. Open bars represent individuals predicted younger in emotional contexts, and ﬁlled bars represent individuals predicted older. Red bars show participants grouped by age predictions in positive emotional contexts; blue bars show participants grouped by age predictions in negative emotional contexts. c Functional network nodes, scaled by their importance in the age-prediction model, are grouped into the following functional networks deﬁned previously : default mode (red), dorsal attention (green), frontoparietal (yellow), salience (black), cingulo-opercular (purple), visual (blue), subcortical (orange), and ventral attention (teal). successes in unraveling the brain bases of individual differences in understanding the neurobiological basis of adolescent behavior. behavior. In particular, we encourage researchers to bridge data sets and Second, this model—along with many in cognitive and levels of analyses to develop generalizable, trajectory-based developmental neuroscience—predicts outcomes from functional models that predict current and future outcomes. connectivity data. Given that functional connectivity patterns can be affected by cognitive state , such models may not generalize across contexts as well as models based on state-independent Leverage multiple data sets to build and validate predictive features such as structural connectivity. (There is evidence, models. Predictive models will be most theoretically and practi- however, that functional connectome-based models generalize cally useful when they generalize beyond a single data set. across task-engaged and resting states to predict abilities such as Although historically replication and external validation samples attention .) Thus, researchers hoping to build an age-prediction were rare in fMRI due to cost and time constraints, open-access model with optimal predictive power and generalizability may data sets and a growing culture of data sharing are removing consider including structural features that may capture more barriers to access. Consider, for example, a group of investigators “trait”-related than state-related variance as predictors (see the interested in predicting impulsivity from resting-state functional section entitled “Include multimodal predictors”). connectivity data. These researchers could download data from Finally, it is important to note that although here maturity was the Human Connectome Project , model the relationship assessed with a single number—akin to the difference between an between impulsivity and functional connectivity, and then apply individual’s functional connectivity pattern and the age-typical their model to completely independent data from the Brain pattern—maturity does not lie on one continuum from “less” Genomics Superstruct Project to evaluate its generalizability. (in emotional states) to “more” (in unemotional states). Training and testing predictive models with open data sets has Rather, temporal differences in the ﬁne-tuning of interacting obvious beneﬁts. For our hypothetical investigators, downloading neural systems with age and experience impact behavioral data may cost a fraction as much as running their own, smaller phenotypes differently across development and vary across fMRI study. Open data sets also tend to offer relatively large individuals and contexts . For example, Rudolph and colleagues sample sizes, capturing a wide range of behavior and allowing show that, on average, adolescents’ functional connectivity researchers to ﬁt complex models and reﬁne model parameters proﬁles look younger in emotional contexts, and that young with nested cross-validation techniques. In addition, open adults who maintain this proﬁle show riskier choices. This samples can provide opportunities to validate models across work suggests that future studies can characterize each indivi- unique behavioral measures. Although this approach can be dual’s unique multivariate maturational proﬁle, that is, the age- challenging given that different-but-related measures may index typicality of both their trait- and state-dependent neural similar-but-not-identical mental processes, it is a useful way to phenotypes. investigate whether a model is capturing individual differences in a speciﬁc performance metric or a general cognitive function. For example, imagine that researchers build a model to predict The road ahead impulsivity questionnaire scores. If they apply this model to a Just as building predictive models can inform how we understand new sample in which impulsivity is measured with task risk taking in adolescence, studying adolescence can inform how performance, predictive power will be limited by the ground- we approach behavioral prediction. Here, motivated by predictive truth relationship between questionnaire scores and task and descriptive models of development, we suggest eight direc- performance. Successful generalization would provide additional tions for future research and highlight their importance for evidence that the model is related to individual differences in 6 NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications | | | Predicted younger Predicted older Risk preference NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 REVIEW ARTICLE impulsivity per se rather than individual differences in ques- Cerebellar volume Group average tionnaire scores alone. Thus, validating models in open data sets (Castellanos et al. ) can help establish their speciﬁcity and generalizability. Healthy This should not be taken imply that targeted studies are ADHD obsolete. Instead, experiments designed to probe speciﬁc Cortical thickness (Shaw et al. ) behavioral phenotypes with carefully designed psychological Healthy tasks are crucial complements to open data analyses. Because ADHD targeted studies have greater ﬂexibility in the participants they Functional connectome distinctiveness recruit, the behavioral measures they collect, and the tasks they (Kaufmann et al. ) administer, they can help elucidate brain–behavior relationships Healthy across populations and cognitive states. The impulsivity research Psychopathology including ADHD group, for example, could use data from a targeted study to ask Childhood Adolescence Adulthood whether the same functional network that predicts impulsivity in adults emerges in development to support children’s impulse Hypothetical individual control. (In fact, they may not even need to collect their own data with attention deficit to do so: Relevant targeted samples may be available on data- Developmental trajectories sharing platforms such as OpenfMRI .) Recent work examining Remission the heritability of the functional connectome used a similar approach, building a model of siblingship in a locally acquired Persistence data set, and validating it in the Human Connectome Project sample . The sustained attention connectome-based predictive model is Current multivariate another recent example of a model validated across multiple 98,103–107 neural phenotype imaging data sets . This model was deﬁned to predict Childhood Adolescence Adulthood individual differences in the ability to maintain focus from patterns of task-evoked and resting-state functional connectiv- Fig. 4 a Developmental changes in cerebellar volume, cortical thickness, ity . During fMRI, adult participants performed a challenging and functional connectome distinctiveness in healthy individuals and sustained attention task, which presumably perturbed attention- individuals with attention deﬁcits. Curves are based on data from relevant neural circuitry and ampliﬁed behaviorally relevant 109,113,119 refs. . b Developmental changes in a hypothetical adolescent with individual differences in functional connectivity. Models deﬁned attention deﬁcits. A model trained to use the developmental trajectories of on task-based data generalized to predict left-out participants’ multiple brain measures to predict future outcomes may best characterize task performance not only from data acquired as they were whether this individual’sdeﬁcits will improve, persist, or worsen. These engaged in the task, but also from data collected as they simply predictions may have implications for future treatment or cessation of rested. External validation with data from the ADHD-200 treatment Consortium revealed that the same functional networks that index attention task performance in adulthood predict ADHD symptoms in childhood. Together these results suggest that a Develop trajectory-based models with longitudinal data. Neu- common functional architecture supports sustained attention robiology is inherently dynamic, and understanding any dynamic across developmental stage (adults vs. children and adolescents), process in terms of both description and prediction requires clinical population (ADHD vs. control), and cognitive state (task appreciating changes over time. Atmospheric models, for exam- 103 111 vs. rest) . ple, rely on dynamical equations to predict the weather , and Another targeted study provided insights into potential stock forecasting models use measures of how a stock’s perfor- mechanisms of the model’s predictive networks. That is, the mance has changed in the past to predict how it will perform in same sustained attention connectome-based predictive model the future. We often use longitudinal data to make folk psycho- distinguished individuals who had taken a single dose of logical predictions, such as when we consider how quickly a methylphenidate (Ritalin) from controls, raising the possibility young tennis player climbed the rankings to estimate her shot at that networks reﬂect the expression of neurotransmitters whose winning Wimbledon, or use what we know about a friend’s recent extracellular concentration is modulated by methylphenidate . stress levels to predict how he will react in an emotional situation. Although the anatomy of the sustained attention model is Models that predict behavior from brain features can also complex, broad trends align with previous ﬁndings and suggest beneﬁt from longitudinal measures. Consider again the case of new targets for intervention . Functional connections between attention deﬁcits. Pioneering work applied growth–curve models sensorimotor and cerebellar regions predict more successful to cross-sectional and longitudinal data to establish delays in 112,113 sustained attention, whereas intra-cerebellar, intra-temporal, and cortical thickness and brain surface area maturation , as well temporal-parietal connections predict less successful attention. as a down-shifted trajectory of cerebellar growth in children 109,110 114–117 The participation of the cerebellum, implicated in ADHD , and adolescents with ADHD (Fig. 4; but see refs. for provides convergent evidence of its importance for attention. methodological considerations related to effects of head motion). Frontal and parietal regions traditionally related to attention and Recent work also suggests that the age-typicality of a child’sor attention impairments do appear in the predictive networks, but adolescent’s functional connectivity patterns is related to their 118,119 they represent >35% of all connections in the model, accentuat- psychiatric symptoms, including attention deﬁcits . In other ing the importance of data-driven approaches to feature selection. words, children and adolescents with attention deﬁcits show In light of the sustained attention model’s out-of-sample delayed maturational patterns of cortical thickness and functional generalizability—a recent proof-of-principle example—we are connectivity on average, and single snapshots of functional optimistic that, moving forward, a combination of high- connectivity predict single snapshots of attentional abilities in throughput data sets, targeted experiments, and “green science” novel individuals. It follows that a teenager’s unique trajectory of data sharing initiatives will facilitate robust, generalizable models functional connectivity and cortical thickness development may of cognitive abilities and behavior across development. provide more nuanced information about his or her attentional NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications 7 | | | Maturity Maturity REVIEW ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 abilities, predicting not only deﬁcit severity, but also perhaps could, for example, be titrated off of medication sooner than symptom persistence or abatement. Developmental neuroscien- otherwise possible (Fig. 4). tists pursuing trajectory-based predictive models can take advantage of large longitudinal samples such as IMAGEN or Include multimodal predictors. Models in human neuroscience the open-access ABCD collection effort, and of biostatistical often focus on a single type of brain feature, such as functional techniques developed to predict clinical outcomes from long- connectivity, to predict behavior. Although this approach is useful 120–123 itudinal biomarkers . for targeting speciﬁc neural mechanisms, constraining a model’s In addition to potentially increasing predictive power, feature space to a single modality may limit predictive power. individualized trajectory-based models can inform theories of Open-access data sets including a variety of scan types (e.g., T1- how neural phenotypes give rise to typical and atypical weighted, T2-weighted, proton density, T2-FLAIR, DTI, BOLD) development. For example, although we know that delayed facilitate the construction of models incorporating a range of cortical maturation trajectories characterize ADHD patients at features (e.g., myelination patterns, structural connectivity, task- the group level , it is not yet known whether a delayed based and resting-state functional connectivity) to maximize trajectory confers risk for attention deﬁcits at the level of predictive power and uncover the unique contributions of dif- individual subjects. Extending similar group-level ﬁndings to the ferent neural systems to current and future behavior. level of individual subjects can enhance the clinical utility of Researchers have demonstrated that including multiple feature research ﬁndings and inform novel interventions . classes improves individualized predictions in development not just in theory, but also in practice. For example, in the ﬁrst-ever Predict future outcomes. Models that predict current behavioral example of predictive modeling in developmental neuroscience, tendencies are technically postdictive in that they make retro- Dosenbach and colleagues asked whether resting-state func- spective, rather than prospective, predictions. Although these tional connectivity patterns can predict an individual subject’s models can inform relationships between neural and behavioral chronological age. Using data from 238 individuals aged 7–30, phenotypes, models that predict future outcomes may be most they trained models to predict categorical (child vs. adult) and useful in clinical and translational contexts, allowing for earlier dimensional (chronological age) measures of maturity. A support intervention, treatment, or cessation of treatment. vector machine classiﬁer correctly predicted whether an indivi- Recent work demonstrates that models that make future dual was a child or an adult 91% of the time, and a support vector forecasts are possible in the context of development. For example, regression algorithm accounted for 55% of the variance in Whelan and colleagues modeled neural and psychological chronological age. Motivated to ﬁnd the unique contribution of proﬁles of alcohol misuse before its onset in adolescence. Using multiple neuroanatomical features to age predictions, Brown and measures of brain structure and function, personality, cognitive colleagues built a model based on structural features that abilities, environmental factors, life experiences, and genetic accounted for 92% of the variance in age in a sample of 885 variants, the authors built a model that distinguished adolescents individuals aged 3–20. Interestingly, different features contributed who go on to binge drink from those who do not. New ﬁndings to model performance at different ages: Whereas T2 signal suggest that brain features can predict clinically relevant intensity in subcortical ROIs was most diagnostic of age in outcomes even earlier in development: cortical surface area and childhood, ﬁber tract diffusivity and subcortical structure volume functional connectivity observed at 6–12 months, for example, were most informative in adolescence, and ROI diffusivity was 125,126 130 131 predict autism diagnosis at age two . most informative in adulthood . Franke et al. additionally In addition to models that predict the onset of clinical found that predictions of a “brain age” model based on multiple symptoms or risky behavior, models that predict improvements neuroanatomical features were signiﬁcantly younger for adoles- in clinical outcomes can help identify resilience factors for cents born preterm than full term. Although the variance psychopathology. Recently, Plitt and colleagues used func- explained by functional connectivity and neuroanatomy cannot tional connectivity patterns to predict improvements in adoles- be directly compared given differences in methodology and cents’ and young adults’ autism symptoms. They found that participant samples across studies, multimodal approaches may functional connectivity in the salience, default mode, and capture more variance in individual differences than do unimodal frontoparietal networks, implicated in attention and goal- ones. 11,12 directed cognition , predicted symptom changes over time, Black box models that use an assortment of brain features to even when accounting for age, IQ, baseline symptoms, and predict outcomes may not necessarily provide interpretable links follow-up latency. Hoeft and colleagues also showed that between neurobiology and behavior. How can researchers unravel prefrontal activity and right superior longitudinal fasciculus the unique contributions of different feature classes to individual fractional anisotropy, but not reading or language test scores, differences? In modeling alcohol misuse in adolescence, Whelan predicted which children with dyslexia would show reading skill and colleagues provide one example of how this may be achieved. improvement over the course of 2.5 years. Functional and To start, their model of future binge drinking included neural, structural measures, therefore, can predict not only the onset of behavioral, lifestyle, and genetic factors. They then systematically clinical symptoms, but also the abatement. In addition, brain removed each feature class from the model to isolate its features can predict future outcomes over and above behavioral contribution to predictive power. The approach revealed that life measures alone—an important check when evaluating the utility history, personality, and brain variables were most uniquely of predictive models. predictive . Other approaches, such as dimensionality reduction In the future, trajectory-based approaches may better char- strategies and penalized regression methods, can also help acterize not just where a child or adolescent has been or is eliminate redundant predictor variables and identify those most currently, but where he or she is going. For example, models that tightly coupled with individual differences in behavior. use a child’s developmental growth curve (e.g., precocious, Although hypothesis-driven modeling approaches that use a 18,77, delayed, deviant, regressive, or resilient ) to predict the single feature or feature type to predict outcomes can advance persistence or worsening of clinical symptoms could have knowledge about the neurobiological bases of behavior, multi- implications for treatment. Such models may also have implica- variate models that incorporate both structural and functional tions for the cessation of treatment. A child with attentional features may improve prediction accuracy and offer converging impairments but a resilient developmental trajectory for ADHD insights (e.g., Fig. 4). 8 NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 REVIEW ARTICLE Continue to focus on dimensional outcomes. Centuries of that the functional architecture of attention in adolescence may observation and research tell us that cognitive abilities and differ from that in childhood and adulthood. These ﬁndings also behavior vary along a continuum at every stage of life. Consider underscore the importance of understanding the nonlinear the example of impulsivity: Although variability in impulsivity expression of dynamic and hierarchical changes in brain features can be captured, to some degree, with a categorical measure like and behavior with development. an ADHD diagnosis, it may be better quantiﬁed with dimensional Moving forward, it will be important to tailor predictive measures such as symptom severity or impulse-control task models to particular scientiﬁc questions and/or practical goals. performance. For example, models trained on one developmental period and When characterizing brain–behavior relationships in develop- tested on another can inform questions about common functional ment, the costs and beneﬁts of using categorical vs. dimensional mechanisms, whereas models trained on a range of age groups measures should be carefully considered. Although categorical may better characterize trajectories in brain–behavior relation- labels align with the current diagnostic system in clinical ships and offer greater predictive power across the lifespan. medicine and can help make results easier to interpret and Further, it is important to keep in mind that because children and present, dichotomizing continuous variables can reduce statistical adolescents are not simply “little adults” in terms of either power, obscure nonlinear relationships between variables and neurocircuitry or behavior, predictions in these populations will 132,133 outcomes, and increase the risk of false positive results .In likely often rely on development-speciﬁc models rather than addition, categorical models are often built on balanced samples models deﬁned in adults and applied to developmental data. of patients and control participants to avoid biased predictions, Future work testing whether models are valid across develop- but this ratio rarely reﬂects real-world illness prevalence. Thus, mental stages, clinical populations, and cognitive or affective reported measures of a model’s sensitivity and speciﬁcity may states can provide additional insight into the scope of their exaggerate its translational utility, and positive and negative generalizability. predictive values may be more useful measures of performance (see Box 1). Dimensional measures may better characterize the Bridge statistical predictability and biological plausibility. full range of behavior and clinically relevant outcomes, especially Predictive modeling in developmental neuroscience has two in development, when small differences in behavioral or neural parallel goals: to discover how the brain gives rise to behavior phenotypes can have important implications for treatment. across development, and to identify practically useful neuro- Although dimensional measures are frequently used to markers of behavior and clinically relevant outcomes. It is not characterize individual differences in cognitive and developmen- always obvious, however, how models that achieve the second 18,135 tal neuroscience and psychology , such approaches are goal can help make progress toward the ﬁrst. Instead, predictive infrequent in predictive modeling . Models that predict chron- models are sometimes considered opaque “black boxes” far 97,129–131 136–138 103,118,139 ological age , ﬂuid intelligence , attention , removed from biology and uninformative about the neural cir- 140 127 and improvements in math skills and autism symptoms , cuits supporting behavior. Some models are more susceptible to however, demonstrate that such approaches are powerful ways to this concern than others. Supekar and colleagues , for example, identify robust transdiagnostic biomarkers of abilities and used hippocampal volume and functional connectivity to predict behavior. children’s response to math tutoring. In doing so, they provide Looking ahead, modeling approaches that consider multiple clear evidence of the role of learning- and memory-relevant brain dimensional approaches at once, or those that identify latent regions in math skill improvements. On the other hand, models distributions from which behavior emerges, may help delineate that use deep neural networks to generate predictions may subtypes of clinical disorders and improve outcome predictions sometimes preclude easy (linguistic) interpretations of relation- and treatment . Regression models that predict dimensional ships between predictors and outcomes. As bigger data sets and outcomes and consider subgroups that make up heterogeneous more sophisticated algorithms result in greater and greater pre- patient populations will also continue to be valuable complements dictive power, it will be important for researchers to keep the ﬁrst to classiﬁers that predict group membership. goal of modeling—advances in basic science—in their sights. Large-scale data sets that include behavioral, neuroimaging, Establish boundary conditions. Given that brain structure and and genetic data provide exciting opportunities for researchers to function change dramatically across development, models trained explore the biological plausibility of predictive models. To in one developmental period (e.g., adulthood) should not always illustrate the promise of approaches that link levels of analysis, generalize to others (e.g., adolescence). Rather, upper bounds on let’s return to the hypothetical research group interested in models’ predictive power will be inﬂuenced by the reliability of impulsivity. Imagine that the research team identiﬁes a pattern of the brain and behavioral measures, their stability across devel- functional connectivity that predicts impulsivity across indivi- opment, and the developmental trajectories of the underlying duals. A subsequent issue of clear importance is the extent to neurobiology. which this network reﬂects the underlying function of molecular- Testing models across different developmental periods can help genetic mechanisms. To get initial traction on this question, the identify critical change points in the relationship between researchers could ask whether this network is heritable, or neurobiological processes and behavior. As a concrete example, whether structural genetic variants predict its function, which in the sustained attention connectome-based predictive model turn predicts impulsive behavior. They could also pursue cross- introduced earlier generalized from an adult to a developmental species work, asking whether important model features map on to 98,103 sample . The model, however, did not perform equally well known anatomical circuits, or whether hypothetical genes that in all age groups. Although predictions were signiﬁcantly related impact network function in humans affect behavior in rodents. to ADHD symptoms in children 8–9(n = 30), 10–11 (n = 28), and Thus, approaches that combine data sets to bridge levels of 12–13 (n = 41), they did not reach signiﬁcance in adolescents analysis and link genotype, neural phenotype, and behavior 14–16 (n = 14; unpublished results). Although certainly not across development may suggest new etiological hypotheses and conclusive given the exploratory nature of this analysis and the possible treatment targets of cognitive function and dysfunction. fact that predictive power is inﬂuenced by factors including sample size, data quality, and group variance in ADHD scores, Acknowledge limitations to advance understanding. Enthu- this outcome motivates future research by tentatively suggesting siasm for large neuroimaging data sets and individualized NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications 9 | | | REVIEW ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 predictions should be coupled with realistic assessments of Conclusions potential pitfalls related to methodology, interpretation, and A rich tradition of research in human neuroimaging has made implementation. Carefully considering these limitations, progress in explaining the neurobiology of cognition and beha- researchers are already beginning to develop new analytic vior. Less attention, however, has been devoted to predicting approaches and ﬁeld-wide standards to address cognitive abilities and behavior from brain features. Here we 84,87,89,141,142 them . argue that predictive modeling approaches that forecast outcomes Methodological pitfalls can erode the impact of predictive at the level of individuals are important complements to work models. For example, because head motion introduces signiﬁcant describing brain–behavior relationships at the group level, espe- confounds in both structural and functional imaging data, cially in the context of adolescence. 115,143 especially in developmental populations , up-to-date data- Not only can predictive models enhance the clinical and collection and preprocessing techniques are necessary for translational utility of neuroimaging research, they can also ensuring that predictions do not rely on motion-induced artifacts. account for critical features of behavior often overlooked in In addition, just as descriptive models may reﬂect sample noise, cross-sectional studies of the developed adult brain: develop- predictive models may be overﬁt to training data. Although mental trajectories, hierarchically emerging brain systems, and nested cross-validation techniques can help protect against individual differences in both. So far, investigating when models overﬁtting, external validation is critical for testing model accurately predict behavior and when they fail to do so has generalizability (Box 1). Finally, it is important that methodolo- illuminated potentially adolescent-speciﬁc changes in beha- gical choices be tailored to research goals. For example, is the goal vioral and neural phenotypes related to risk-taking and attention. to predict current phenotypes or future change? To make Looking ahead, models that offer probabilistic insights into absolute predictions (e.g., that a child will grow to be six feet individuals’ current and future behavior from their past devel- tall) or relative ones (that he or she will be in the 95th percentile opmental brain trajectories have the potential to provide deep for height)? To predict behavior from functional brain features insights into human brain development and function in both observed during task engagement, or to test whether a cognitive health and disease. process can be measured in the absence of an explicit task ?To maximize subgroup-level accuracy, or population-level general- Received: 6 October 2017 Accepted: 5 January 2018 izability? To prioritize statistical predictability, biological plausi- bility, or feature weight interpretability? Mismatches between study methods and goals can undermine the usefulness of predictive models. References Working with large data sets also poses several challenges to 1. Hofstadter, A. Explanation and necessity. Philos. Phenomenol. Res. 11, interpretation. For example, when samples are large enough, 339–347 (1951). brain–behavior relationships with even tiny effects sizes may 2. Shmueli, G. To explain or to predict? Stat. Sci. 25, 289–310 (2010). reach statistical signiﬁcance. Although such effects may be 3. Craver, C. F. When mechanistic models explain. Synthese 153, 355–376 interesting from a basic science perspective if robust to noise (2006). 4. Salmon, W. C. Why ask, ‘Why?’? An inquiry concerning scientiﬁc explanation. and replicable, they are unlikely to offer much clinical or practical 87 Proc. Address. Am. Philos. Assoc. 51, 683–705 (1978). beneﬁt in the near term . In addition, although large data sets 5. Woo, C.-W., Chang, L. J., Lindquist, M. A. & Wager, T. D. Building better may be more representative of the general population than biomarkers: brain models in translational neuroimaging. Nat. Neurosci. 20, smaller samples, they may offer a “false sense of security” since 365–377 (2017). they can still suffer from selection bias and skewed demo- Reviews advances in biomarker discovery in human neuroscience. 6. Gabrieli, J. D. E., Ghosh, S. S. & Whitﬁeld-Gabrieli, S. Prediction as a graphics and are not equally distributed across the globe (See humanitarian and pragmatic contribution from human cognitive Fig. 1). The careful evaluation of both statistical and clinical neuroscience. Neuron 85,11–26 (2015). signiﬁcance will be important for establishing scientiﬁcally valid, Highlights beneﬁts of predictive brain-based models for personalized practically useful models. medicine and education. Although neuroimaging-based predictive models have the 7. Douglas, H. E. Reintroducing prediction to explanation. Philos. Sci. 76, 444–463 (2009). potential to offer signiﬁcant beneﬁts, challenges can arise when 8. Yarkoni, T. & Westfall, J. Choosing prediction over explanation in psychology: applying them in clinical contexts. Recent work has highlighted lessons from machine learning. Perspect. Psychol. Sci. 12, 1100–1122 (2017). the “perilous path from publication to practice”, outlining a Argues for an increased focus on prediction over explanation for variety of scientiﬁc, implementation, and business-related obsta- understanding human behavior. cles . As scientists pursue individualized predictions, close 9. Goldman-Rakic, P. S. Topography of cognition: parallel distributed networks in primate association cortex. Annu. Rev. Neurosci. 11, 137–156 (1988). collaborations with clinicians, bioethicists, industry professionals, 10. Price, J. L. & Drevets, W. C. Neurocircuitry of mood disorders. and regulatory bodies will be necessary for effectively translating Neuropsychopharmacology 35, 192–216 (2009). 141,146 models to real-life patient-care settings . 11. Buckner, R. L., Andrews-Hanna, J. R. & Schacter, D. L. The brain’s default Finally, as models begin to make their way from bench to network: anatomy, function, and relevance to disease. Ann. N. Y. Acad. Sci. bedside, it is important to consider their ethical implications, 1124,1–38 (2008). 12. Corbetta, M. & Shulman, G. L. Control of goal-directed and stimulus-driven especially in the context of development. First, researchers need attention in the brain. Nat. Rev. Neurosci. 3, 201–215 (2002). to keep in mind—and clearly communicate to participants, 13. Power, J. D. et al. Functional network organization of the human brain. patients, and the public—that predictive models are probabilistic Neuron 72, 665–678 (2011). rather than deterministic in nature, and, just like traditional pen- 14. Yeo, B. T. T. et al. The organization of the human cerebral cortex estimated by and-paper tests, will never perfectly predict abilities or behavior. intrinsic functional connectivity. J. Neurophysiol. 106, 1125–1165 (2011). 15. Insel, T. et al. Research domain criteria (RDoC): toward a new classiﬁcation Predictions are thus best considered tools for scientiﬁc discovery framework for research on mental disorders. Am. J. Psychiatry 167, 748–751 and opportunities for informing clinical decision-making rather (2010). than portents of the future. Moving forward, the potential 16. Hyman, S. E. The diagnosis of mental disorders: the problem of reiﬁcation. beneﬁts of predictive modeling in development must be Annu. Rev. Clin. Psychol. 6, 155–179 (2010). continuously evaluated in light any of potential risks to privacy 17. Fair, D. A., Bathula, D., Nikolas, M. A. & Nigg, J. T. Distinct neuropsychological subgroups in typically developing youth inform or hyperbolic claims about our ability to predict the future. 10 NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 REVIEW ARTICLE heterogeneity in children with ADHD. Proc. Natl Acad. Sci. USA 109, development: relevance for dopamine’s role in adolescent decision making. 6769–6774 (2012). Dev. Cogn. Neurosci. 1, 578–590 (2011). 18. Casey, B. J., Oliveri, M. E. & Insel, T. A neurodevelopmental perspective on 48. Teslovich, T. et al. Adolescents let sufﬁcient evidence accumulate before the research domain criteria (RDoC) framework. Biol. Psychiatry 76, 350–353 making a decision when large incentives are at stake. Dev. Sci. 17,59–70 (2014). (2014). 19. Cohen, J. D. et al. Computational approaches to fMRI analysis. Nat. Neurosci. 49. Somerville, L. H., Jones, R. M. & Casey, B. J. A time of change: behavioral and 20, 304–313 (2017). neural correlates of adolescent sensitivity to appetitive and aversive 20. Shen, X. et al. Using connectome-based predictive modeling to predict environmental cues. Brain. Cogn. 72, 124–133 (2010). individual behavior from brain connectivity. Nat. Protoc. 12, 506–518 (2017). 50. Dreyfuss, M. et al. Teens impulsively react rather than retreat from threat. 21. Lee, F. S. et al. Adolescent mental health—Opportunity and obligation: Dev. Neurosci. 36, 220–227 (2014). emerging neuroscience offers hope for treatments. Science 346, 547–549 51. Casey, B., Galván, A. & Somerville, L. H. Beyond simple models of adolescence (2014). to an integrated circuit-based account: a commentary. Dev. Cogn. Neurosci. 22. Casey, B. J. Beyond simple models of self-control to circuit-based accounts of 17, 128–130 (2016). adolescent behavior. Annu. Rev. Psychol. 66, 295–319 (2015). 52. Casey, B., Heller, A. S., Gee, D. G. & Cohen, A. O. Development of the Presents evidence that adolescent-speciﬁc changes in self-control emerge emotional brain. Neurosci. Lett. https://doi.org/10.1016/j.neulet.2017.11.055 from the ﬁne-tuning of connections within and between subcortical, cortical (2017) prefrontal, and limbic circuits. 53. Bourgeois, J.-P., Goldman-Rakic, P. S. & Rakic, P. Synaptogenesis in the 23. Steinberg, L. Risk taking in adolescence. Curr. Dir. Psychol. Sci. 16,55–59 prefrontal cortex of rhesus monkeys. Cereb. Cortex. 4,78–96 (1994). (2007). 54. Huttenlocher, P. R. & Dabholkar, A. S. Regional differences in synaptogenesis 24. Ellis, B. J. et al. The evolutionary basis of risky adolescent behavior: in human cerebral cortex. J. Comp. Neurol. 387, 167–178 (1997). implications for science, policy, and practice. Dev. Psychol. 48, 598–623 55. Østby, Y. et al. Heterogeneity in subcortical brain development: a structural (2012). magnetic resonance imaging study of brain maturation from 8 to 30 years. J. 25. Arnett, J. J. Sensation seeking, aggressiveness, and adolescent reckless Neurosci. 29, 11772–11782 (2009). behavior. Pers. Individ. Dif. 20, 693–702 (1996). 56. Giedd, J. N. et al. Brain development during childhood and adolescence: a 26. Greene, K. et al. Targeting adolescent risk-taking behaviors: the contributions longitudinal MRI study. Nat. Neurosci. 2, 861–863 (1999). of egocentrism and sensation-seeking. J. Adolesc. 23, 439–461 (2000). 57. Tamnes, C. K. et al. Brain maturation in adolescence and young adulthood: 27. Van Leijenhorst, L. et al. What motivates the adolescent? brain regions regional age-related changes in cortical thickness and white matter volume mediating reward sensitivity across adolescence. Cereb. Cortex. 20,61–69 and microstructure. Cereb. Cortex. 20, 534–548 (2010). (2010). 58. Shaw, P. et al. Neurodevelopmental trajectories of the human cerebral cortex. 28. Galvan, A. et al. Earlier development of the accumbens relative to J. Neurosci. 28, 3586 LP–3583594 (2008). orbitofrontal cortex might underlie risk-taking behavior in adolescents. J. 59. Gogtay, N. et al. Dynamic mapping of human cortical development during Neurosci. 26, 6885–6892 (2006). childhood through early adulthood. Proc. Natl Acad. Sci. USA 101, 8174–8179 29. Somerville, L. H., Hare, T. & Casey, B. J. Frontostriatal maturation predicts (2004). cognitive control failure to appetitive cues in adolescents. J. Cogn. Neurosci. 60. Sowell, E. R., Thompson, P. M., Holmes, C. J., Jernigan, T. L. & Toga, A. W. In 23, 2123–2134 (2010). vivo evidence for post-adolescent brain maturation in frontal and striatal 30. Silvers, J. A. et al. Age-related differences in emotional reactivity, regulation, regions. Nat. Neurosci. 2, 859–861 (1999). and rejection sensitivity in adolescence. Emotion 12, 1235–1247 (2012). 61. Gogtay, N. & Thompson, P. M. Mapping gray matter development: 31. Chein, J., Albert, D., O’Brien, L., Uckert, K. & Steinberg, L. Peers increase Implications for typical development and vulnerability to psychopathology. adolescent risk taking by enhancing activity in the brain’s reward circuitry. Brain. Cogn. 72,6–15 (2010). Dev. Sci. 14,F1–F10 (2011). 62. Selemon, L. D. A role for synaptic plasticity in the adolescent development of 32. Steinberg, L. A social neuroscience perspective on adolescent risk-taking. Dev. executive function. Transl. Psychiatry 3, e238 (2013). Rev. 28,78–106 (2008). 63. Benes, F. M., Taylor, J. B. & Cunningham, M. C. Convergence and plasticity of 33. Blakemore, S.-J. & Robbins, T. W. Decision-making in the adolescent brain. monoaminergic systems in the medial prefrontal cortex during the postnatal Nat. Neurosci. 15, 1184–1191 (2012). period: implications for the development of psychopathology. Cereb. Cortex. 34. Somerville, L. H. The teenage brain: sensitivity to social evaluation. Curr. Dir. 10, 1014–1027 (2000). Psychol. Sci. 22, 121–127 (2013). 64. Brenhouse, H. C., Sonntag, K. C. & Andersen, S. L. Transient D1 dopamine 35. Spear, L. P. The adolescent brain and age-related behavioral manifestations. receptor expression on prefrontal cortex projection neurons: relationship to Neurosci. Biobehav. Rev. 24, 417–463 (2000). enhanced motivational salience of drug cues in adolescence. J. Neurosci. 28, 36. Hartley, C. A. & Lee, F. S. Sensitive periods in affective development: 2375–2382 (2008). nonlinear maturation of fear learning. Neuropsychopharmacology 40,50–60 65. Tseng, K. Y. & O’Donnell, P. Dopamine modulation of prefrontal cortical (2015). interneurons changes during adolescence. Cereb. Cortex. 17, 1235–1240 37. King, E. C., Pattwell, S. S., Sun, A., Glatt, C. E. & Lee, F. S. Nonlinear (2007). developmental trajectory of fear learning and memory. Ann. N. Y. Acad. Sci. 66. Casey, B. J., Getz, S. & Galvan, A. The adolescent brain. Dev. Rev. 28,62–77 1304,62–69 (2013). (2008). 38. Insel, T. R. & Young, L. J. The neurobiology of attachment. Nat. Rev. Neurosci. 67. Flechsig, P. Anatomie des menschlichen Gehirns und Rückenmarks auf 2, 129–136 (2001). myelogenetischer Grundlage. 121 (1920). 39. Landers, M. S. & Sullivan, R. M. The development and neurobiology of infant 68. Buckner, R. L. & Krienen, F. M. The evolution of distributed association attachment and fear. Dev. Neurosci. 34, 101–114 (2012). networks in the human brain. Trends Cogn. Sci. 17, 648–665 (2017). 40. Pattwell, S. S. et al. Altered fear learning across development in both mouse 69. Asato, M. R., Terwilliger, R., Woo, J. & Luna, B. White matter development in and human. Proc. Natl Acad. Sci. USA 109, 16318–16323 (2012). adolescence: a DTI study. Cereb. Cortex. 20, 2122–2131 (2010). 41. McCallum, J., Kim, J. H. & Richardson, R. Impaired extinction retention in 70. Simmonds, D. J., Hallquist, M. N., Asato, M. & Luna, B. Developmental stages adolescent rats: effects of D-cycloserine. Neuropsychopharmacology 35, and sex differences of white matter and behavioral development through 2134–2142 (2010). adolescence: a longitudinal diffusion tensor imaging (DTI) study. Neuroimage 42. Gee, D. G. et al. Individual differences in frontolimbic circuitry and anxiety 92, 356–368 (2014). emerge with adolescent changes in endocannabinoid signaling across species. 71. Barnea-Goraly, N. et al. White matter development during childhood and Proc. Natl Acad. Sci. USA 113, 4500–4505 (2016). adolescence: a cross-sectional diffusion tensor imaging study. Cereb. Cortex. 43. Malter Cohen, M. et al. Early-life stress has persistent effects on amygdala 15, 1848–1854 (2005). function and development in mice and humans. Proc. Natl Acad. Sci. USA 72. Liston, C. et al. Frontostriatal microstructure modulates efﬁcient recruitment 110, 18274–18278 (2013). of cognitive control. Cereb. Cortex. 16, 553–560 (2006). 44. Gee, D. G. et al. Early developmental emergence of human 73. Power, J. D., Fair, D. A., Schlaggar, B. L. & Petersen, S. E. The development of amygdala–prefrontal connectivity after maternal deprivation. Proc. Natl Acad. human functional brain networks. Neuron 67, 735–748 (2010). Sci. USA 110, 15638–15643 (2013). 74. Fair, D. A. et al. The maturing architecture of the brain’s default network. 45. Tottenham, N. et al. Elevated amygdala response to faces following early Proc. Natl Acad. Sci. USA 105, 4028–4032 (2008). deprivation. Dev. Sci. 14, 190–204 (2011). 75. Supekar, K., Musen, M. & Menon, V. Development of large-scale functional 46. Hartley, C. A. & Somerville, L. H. The neuroscience of adolescent decision- brain networks in children. PLOS Biol. 7, e1000157 (2009). making. Curr. Opin. Behav. Sci. 5, 108–115 (2015). 76. Kelly, A. M. C. et al. Development of anterior cingulate functional 47. van der Schaaf, M. E., Warmerdam, E., Crone, E. A. & Cools, R. Distinct linear connectivity from late childhood to early adulthood. Cereb. Cortex. 19, and non-linear trajectories of reward and punishment reversal learning during 640–657 (2009). NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications 11 | | | REVIEW ARTICLE NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 77. Di Martino, A. et al. Unraveling the miswired connectome: a developmental 105. Yoo, K. et al. Connectome-based predictive modeling of attention: comparing perspective. Neuron 83, 1335–1353 (2014). different functional connectivity features and prediction methods across 78. Sejnowski, T. J., Churchland, P. S. & Movshon, J. A. Putting big data to good datasets. Neuroimage 167,11–22 (2018). use in neuroscience. Nat. Neurosci. 17, 1440–1441 (2014). 106. Jangraw, D. C. et al. A functional connectivity-based neuromarker of sustained 79. Schumann, G. et al. The IMAGEN study: reinforcement-related behaviour in attention generalizes to predict recall in a reading task. Neuroimage 166, normal brain function and psychopathology. Mol. Psychiatry 15, 1128–1139 99–109 (2018). (2010). 107. Rosenberg, M. D. et al. Connectome-based fMRI models predict separable 80. Van Essen, D. C. et al. The WU-Minn Human Connectome Project: an components of attention in novel individuals. J. Cogn. Neurosci. 30, 160–173 overview. Neuroimage 80,62–79 (2013). (2018). 81. Di Martino, A. et al. The autism brain imaging data exchange: towards a large- 108. The ADHD-200 Consortium. The ADHD-200 consortium: a model to scale evaluation of the intrinsic brain architecture in autism. Mol. Psychiatry advance the translational potential of neuroimaging in clinical neuroscience. 19, 659–667 (2014). Front. Syst. Neurosci. 6, 62 (2012). 82. Alexander, L. M. et al. An open resource for transdiagnostic research in 109. Castellanos, F. X. et al. Developmental trajectories of brain volume pediatric mental health and learning disorders. Sci. Data 4, 170181 (2017). abnormalities in children and adolescents with attention-deﬁcit/hyperactivity 83. Casey, B. et al. The ABCD Study: functional imaging acquisition across disorder. J. Am. Med. Assoc. 288, 1740–1748 (2002). 21 sites. Dev. Cogn. Neurosci. (in press) (2018). 110. Casey, B. J., Nigg, J. T., Durston, S. New potential leads in the biology and 84. Alfaro-Almagro, F. et al. Image processing and quality control for the ﬁrst treatment of attention deﬁcit-hyperactivity disorder. Curr. Opin. Neurol. 20, 10,000 brain imaging datasets from UK Biobank. Neuroimage 166, 400–424 119–124 (2007). (2018). 111. Lorenz, E. N. Empirical orthogonal functions and statistical weather 85. Noble, S. et al. Inﬂuences on the test–retest reliability of functional prediction. Scientiﬁc Report No. 1. Statistical Forecasting Project, Dept. of connectivity MRI and its relationship with behavioral utility. Cereb. Cortex. Meteorology, MIT (1956). 27,1–15 (2017). 112. Shaw, P. et al. Development of cortical surface area and gyriﬁcation in 86. Noble, S. et al. Multisite reliability of MR-based functional connectivity. attention-deﬁcit/hyperactivity disorder. Biol. Psychiatry 72, 191–197 (2012). Neuroimage 146, 959–970 (2017). 113. Shaw, P. et al Attention-deﬁcit/hyperactivity disorder is characterized by a 87. Miller, K. L. et al. Multimodal population brain imaging in the UK Biobank delay in cortical maturation. Proc. Natl Acad. Sci. USA 104, 19649–19654 prospective epidemiological study. Nat. Neurosci. 19, 1523–1536 (2016). (2007). 88. Dosenbach, N. U. F. et al. Real-time motion analytics during brain MRI Provides evidence for delayed cortical thickness maturation in ADHD. improve data quality and reduce costs. Neuroimage 161,80–93 (2017). 114. Fair, D. A. et al. Distinct neural signatures detected for ADHD subtypes after 89. Bzdok, D. & Yeo, B. T. T. Inference in the age of big data: Future perspectives controlling for micro-movements in resting state functional connectivity MRI on neuroscience. Neuroimage 155, 549–564 (2017). data. Front. Syst. Neurosci. 6, 80 (2012). Suggests statistical tools useful for analyzing big data in human 115. Pardoe, H. R., Kucharsky Hiess, R. & Kuzniecky, R. Motion and neuroscience. morphometry in clinical and nonclinical populations. Neuroimage 135, 90. Dubois, J. & Adolphs, R. Building a science of individual differences from 177–185 (2016). fMRI. Trends Cogn. Sci. 20, 425–443 (2016). 116. Couvy-Duchesne, B. et al. Head motion and inattention/hyperactivity share 91. Finn, E. S. & Constable, R. T. Individual variation in functional brain common genetic inﬂuences: implications for fMRI studies of ADHD. PLoS connectivity: implications for personalized approaches to psychiatric disease. ONE 11, e0146271 (2016). Dialogues Clin. Neurosci. 18, 277–287 (2016). 117. Engelhardt, L. E. et al. Children’s head motion during fMRI tasks is heritable 92. Jollans, L. & Whelan, R. The clinical added value of imaging: a perspective and stable over time. Dev. Cogn. Neurosci. 25,58–68 (2017). from outcome prediction. Biol. Psychiatry Cogn. Neurosci. Neuroimaging 1, 118. Kessler, T., Angstadt, M. & Sripada, C. Growth charting of brain connectivity 423–432 (2016). networks and the identiﬁcation of attention impairment in youth. JAMA 93. Box, G. E. P. in Robustness in Statistics (eds Launer, R. & Wilderson, G.) Psychiatry 73, 481–489 (2016). 201–236 (Academic Press, New York, 1979). 119. Kaufmann, T. et al Delayed stabilization and individualization in connectome 94. Finn, E. S. et al. Can brain state be manipulated to emphasize individual development are related to psychiatric disorders. Nat. Neurosci. 20, 513–515 differences in functional connectivity? Neuroimage 160, 140–151 (2017). (2017). 95. Cohen, A. O. & Casey, B. J. Rewiring juvenile justice: the intersection of Demonstrates that functional connectivity patterns become more stable and developmental neuroscience and legal policy. Trends Cogn. Sci. 18,63–65 unique during adolescence, and that delays in this pattern of development (2014). are related to psychiatric symptoms. 96. Cohen, A. O. et al When is an adolescent an adult? assessing cognitive control 120. Albert, P. S. A linear mixed model for predicting a binary event from in emotional and nonemotional contexts. Psychol. Sci. 27, 549–562 (2016). longitudinal data under random effects misspeciﬁcation. Stat. Med. 31, Provides evidence for a shift in cognitive capacity in emotional contexts that 143–154 (2012). accompanies dynamic changes in prefrontal circuitry during adolescence. 121. Liu, D. & Albert, P. S. Combination of longitudinal biomarkers in predicting 97. Rudolph, M. D. et al At risk of being risky: the relationship between ‘brain age’ binary events. Biostatistics 15, 706–718 (2014). under emotional states and risk preference. Dev. Cogn. Neurosci. 24,93–106 122. Sweeting, M. J. & Thompson, S. G. Making predictions from complex (2017). longitudinal data, with application to planning monitoring intervals in a Shows that, on average, adolescents’ functional connectivity proﬁles look national screening programme. J. R. Stat. Soc. Ser. A. Stat. Soc. 175, 569–586 younger in emotional contexts, and that young adults who maintain this (2012). proﬁle show riskier choices. 123. Li, L., Luo, S., Hu, B. & Greene, T. Dynamic prediction of renal failure using 98. Rosenberg, M. D., Finn, E. S., Scheinost, D., Constable, R. T. & Chun, M. M. longitudinal biomarkers in a cohort study of chronic kidney disease. Stat. Characterizing attention with predictive network models. Trends Cogn. Sci. 21, Biosci. 9, 357–378 (2016). 290–302 (2017). 124. Whelan, R. et al Neuropsychosocial proﬁles of current and future adolescent 99. Van Essen, D. C. & Glasser, M. F. The Human Connectome Project: progress alcohol misusers. Nature 512, 185–189 (2014). and prospects. Cerebrum 2016,10–16 (2016). Predicts future alcohol misuse from neural, behavioral, environmental, and 100. Holmes, A. J. et al Brain Genomics Superstruct Project initial data release with genetic data in a large sample of adolescents. structural, functional, and behavioral measures. Sci. Data 2, 150031 (2015). 125. Emerson, R. W. et al. Functional neuroimaging of high-risk 6-month-old Details the Brain Genomics Superstruct Project repository, which includes infants predicts a diagnosis of autism at 24 months of age. Sci. Transl. Med. 9 neuroimaging and self-report behavioral and cognitive data from 1,570 eaag2882 (2017). adult participants. 126. Hazlett, H. C. et al. Early brain development in infants at high risk for autism 101. Poldrack, R. A. et al. Toward open sharing of task-based fMRI data: the spectrum disorder. Nature 542, 348–351 (2017). OpenfMRI project. Front. Neuroinf. 7, 12 (2013). 127. Plitt, M., Barnes, K. A., Wallace, G. L., Kenworthy, L. & Martin, A. Resting- 102. Miranda-Domínguez, Ó. et al. Heritability of the human connectome: a state functional connectivity predicts longitudinal change in autistic traits and connectotyping study. Netw. Neurosci.1–48. https://doi.org/10.1162/ adaptive functioning in autism. Proc. Natl Acad. Sci. USA 112, E6699–E6706 NETN_a_00029 (2017). (2015). 103. Rosenberg, M. D. et al. A neuromarker of sustained attention from whole- 128. Hoeft, F. et al. Neural systems predicting long-term outcome in dyslexia. Proc. brain functional connectivity. Nat. Neurosci. 19, 165–171 (2016). Natl Acad. Sci. USA 108, 361–366 (2011). Introduces a model that generalizes across datasets to predict sustained 129. Dosenbach, N. U. F. et al. Prediction of individual brain maturity using fMRI. attention from task-based and resting-state functional connectivity. Science 329, 1358–1361 (2010). 104. Rosenberg, M. D. et al. Methylphenidate modulates functional network Predicted chronological age from functional connectivity patterns in the ﬁrst connectivity to enhance attention. J. Neurosci. 36, 9547–9557 (2016). example of predictive modeling in developmental cognitive neuroscience. 12 NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications | | | NATURE COMMUNICATIONS | DOI: 10.1038/s41467-018-02887-9 REVIEW ARTICLE 130. Brown, T. T. et al. Neuroanatomical assessment of biological maturity. Curr. 152. Pausova, Z. et al. Genes, maternal smoking, and the offspring brain and body Biol. 22, 1693–1698 (2012). during adolescence: design of the Saguenay Youth Study. Hum. Brain. Mapp. 131. Franke, K., Luders, E., May, A., Wilke, M. & Gaser, C. Brain maturation: 28, 502–518 (2007). predicting individual BrainAGE in children and adolescents using structural 153. Salum, G. A. et al. High risk cohort study for psychiatric disorders in MRI. Neuroimage 63, 1305–1312 (2012). childhood: rationale, design, methods and preliminary results. Int. J. Methods 132. Altman, D. G. & Royston, P. The cost of dichotomising continuous variables. Psychiatr. Res. 24,58–73 (2015). BMJ 332, 1080 (2006). 154. Thompson, P. M. et al. The ENIGMA Consortium: large-scale collaborative 133. Austin, P. C. & Brunner, L. J. Inﬂation of the type I error rate when a analyses of neuroimaging and genetic data. Brain. Imaging Behav. 8, 153–182 continuous confounding variable is categorized in logistics regression analyses. (2014). Statist. Med. 23, 1159–1178 (2004). 155. Kooijman, M. N. et al. The Generation R Study: design and cohort update 134. Castellanos, F. X., Di Martino, A., Craddock, R. C., Mehta, A. D. & Milham, 2017. Eur. J. Epidemiol. 31, 1243–1264 (2016). M. P. Clinical applications of the functional connectome. Neuroimage 80, 156. von Rhein, D. et al. The NeuroIMAGE study: a prospective phenotypic, 527–540 (2013). cognitive, genetic and MRI study in children with attention-deﬁcit/ 135. Hudziak, J. J., Achenbach, T. M., Althoff, R. R. & Pine, D. S. A dimensional hyperactivity disorder. Design and descriptives. Eur. Child. Adolesc. Psychiatry approach to developmental psychopathology. Int. J. Methods Psychiatr. Res. 24, 265–281 (2015). 16, S16–S23 (2007). 157. Zuo, X. N. et al. An open science resource for establishing reliability and 136. Finn, E. S. et al. Functional connectome ﬁngerprinting: identifying individuals reproducibility in functional connectomics. Sci. Data 1, 140049 (2014). using patterns of brain connectivity. Nat. Neurosci. 18, 1664–1671 (2015). 137. Smith, S. M. et al. Functional connectomics from resting-state fMRI. Trends Acknowledgements Cogn. Sci. 17, 666–682 (2013). This work was supported by the National Institute of Mental Health (K01 MH099232 to 138. Ferguson, M. A., Anderson, J. S. & Spreng, R. N. Fluid and ﬂexible minds: A.J.H.), National Institute on Drug Abuse (U01 DA041174 to B.J.C.), and MacArthur intelligence reﬂects synchrony in the brain’s intrinsic network architecture. Foundation Research Network on Law and Neuroscience (B.J.C.). M.D.R. is supported by Netw. Neurosci. 1, 192–207 (2017). a Theresa Seessel Postdoctoral Fellowship at Yale University. We thank Kevin Anderson, 139. Poole, V. N. et al. Intrinsic functional connectivity predicts individual Adam Chekroud, Emily Finn, and David Gruskin for helpful discussion and/or com- differences in distractibility. Neuropsychologia 86, 176–182 (2016). ments on earlier drafts. 140. Supekar, K. et al. Neural predictors of individual differences in response to math tutoring in primary-grade school children. Proc. Natl Acad. Sci. USA 110, 8230–8235 (2013). Author contributions 141. AM, C. Bigger data, harder questions—opportunities throughout mental M.D.R. drafted the text and graphics with input and edits from B.J.C. and A.J.H. All health care. JAMA Psychiatry 74, 1183–1184 (2017). authors revised and ﬁnalized the manuscript for publication. 142. Smith, S. et al A positive-negative mode of population covariation links brain connectivity, demographics and behavior. Nat. Neurosci. 18, 1565–1567 (2015). Additional information Identiﬁes a pattern of functional brain connectivity related to a 'positive- Competing interests: The authors declare no competing ﬁnancial interests. negative' axis of lifestyle, demographic and psychometric factors. 143. Ciric, R. et al Benchmarking of participant-level confound regression Reprints and permission information is available online at http://npg.nature.com/ strategies for the control of motion artifact in studies of functional reprintsandpermissions/ connectivity. Neuroimage 154, 174–187 (2017). Compares current methods for controlling for motion artifacts in functional Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in connectivity data. published maps and institutional afﬁliations. 144. Swanson, J. M. The UK Biobank and selection bias. Lancet 380, 110 (2017). 145. Chekroud, A. M. & Koutsouleris, N. The perilous path from publication to practice. Mol. Psychiatry 23,24–25 (2017). 146. MP, P. Evidence-based pragmatic psychiatry—a call to action. JAMA Open Access This article is licensed under a Creative Commons Psychiatry 74, 1185–1186 (2017). Attribution 4.0 International License, which permits use, sharing, 147. Guyon, I. & Elisseeff, A. An introduction to variable and feature selection. J. adaptation, distribution and reproduction in any medium or format, as long as you give Mach. Learn. Res. 3, 1157–1182 (2003). appropriate credit to the original author(s) and the source, provide a link to the Creative 148. Kievit, R. et al. Developmental cognitive neuroscience using latent change Commons license, and indicate if changes were made. The images or other third party score models: a tutorial and applications. Dev. Cogn. Neurosci. (2017). material in this article are included in the article’s Creative Commons license, unless 149. Brown, S. A. et al. The National Consortium on Alcohol and indicated otherwise in a credit line to the material. If material is not included in the NeuroDevelopment in Adolescence (NCANDA): a multisite study of adolescent development and substance use. J. Stud. Alcohol. Drugs 76, 895–908 article’s Creative Commons license and your intended use is not permitted by statutory (2015). regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/ 150. Jernigan, T. L. et al. The pediatric imaging, neurocognition, and genetics (PING) data repository. Neuroimage 124, 1149–1154 (2016). licenses/by/4.0/. 151. Satterthwaite, T. D. et al. The Philadelphia Neurodevelopmental Cohort: a publicly available resource for the study of normal and abnormal brain © The Author(s) 2018 development in youth. Neuroimage 124, 1115–1119 (2016). NATURE COMMUNICATIONS (2018) 9:589 DOI: 10.1038/s41467-018-02887-9 www.nature.com/naturecommunications 13 | | |

Journal

Nature Communications – Springer Journals

Published: Feb 21, 2018

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Prediction complements explanation in understanding the developing brain

Prediction complements explanation in understanding the developing brain

Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Prediction complements explanation in understanding the developing brain

Prediction complements explanation in understanding the developing brain

References (185)

Abstract

Journal

Recommended Articles

There are no references for this article.

Our policy towards the use of cookies