elifeeLifeeLife2050-084XeLife Sciences Publications, Ltd9715010.7554/eLife.9715010.7554/eLife.97150.3version of recordResearch ArticleNeuroscienceA robust brain network for sustained attention from adolescence to adulthood that predicts later substance useWengYihehttps://orcid.org/0009-0005-2628-94921KruschwitzJohann23Rueda-DelgadoLaura Mhttps://orcid.org/0000-0002-8730-89771RuddyKathy Lhttps://orcid.org/0000-0001-5501-042314BoyleRory1FranzenLuisa15SerinEminhttps://orcid.org/0000-0002-3570-3027267NwezeTochukwu8HansonJamie9SmythAlannah1FarnanTomhttps://orcid.org/0000-0002-7214-40411BanaschewskiTobias10BokdeArun LW11DesrivièresSylvane12FlorHerta1314GrigisAntoine15GaravanHugh16GowlandPenny A17HeinzAndreas2BrühlRüdigerhttps://orcid.org/0000-0003-0111-599618MartinotJean-Luchttps://orcid.org/0000-0002-0136-038819MartinotMarie-Laure Paillère1920ArtigesEric1921McGrathJane11NeesFrauke101322Papadopoulos OrfanosDimitrihttps://orcid.org/0000-0002-1242-899015PausTomas2324PoustkaLuise25HolzNathalie10FröhnerJuliane26SmolkaMichael Nhttps://orcid.org/0000-0001-5398-556927VaidyaNilakshi28SchumannGunter2728WalterHenrik2WhelanRoberthttps://orcid.org/0000-0002-2790-7281robert.whelan@tcd.ie1IMAGEN Consortiumhttps://ror.org/02tyrky19School of Psychology and Global Brain Health Institute, Trinity College DublinDublinIrelandhttps://ror.org/001w7jn25Department of Psychiatry and Psychotherapy CCM, Charité – Universitätsmedizin Berlin, corporate member of Freie Universität Berlin, Humboldt-Universität zu Berlin, and Berlin Institute of HealthBerlinGermanyhttps://ror.org/042aqky30Collaborative Research Centre (SFB 940) 'Volition and Cognitive Control', Technische Universität DresdenDresdenGermanyhttps://ror.org/00hswnk62School of Psychology, Queens University BelfastBelfastUnited Kingdomhttps://ror.org/02jz4aj89Faculty of Psychology and Neuroscience, Maastricht UniversityMaastrichtNetherlandshttps://ror.org/001w7jn25Charité – Universitätsmedizin Berlin, Einstein Center for Neurosciences BerlinBerlinGermanyhttps://ror.org/05ewdps05Bernstein Center for Computational NeuroscienceBerlinGermanyhttps://ror.org/03r0ha626Department of Psychology, University of UtahSalt Lake CityUnited Stateshttps://ror.org/01an3r305Department of Psychology, Learning Research & Development Center, University of PittsburghPittsburghUnited Stateshttps://ror.org/01hynnt93Department of Child and Adolescent Psychiatry and Psychotherapy, Central Institute of Mental Health, Medical Faculty Mannheim, Heidelberg UniversityMannheimGermanyhttps://ror.org/02tyrky19Discipline of Psychiatry, School of Medicine and Trinity College Institute of Neuroscience, Trinity College DublinDublinIrelandhttps://ror.org/0220mzb33Centre for Population Neuroscience and Precision Medicine (PONS), Institute of Psychiatry, Psychology, & Neuroscience, SGDP Centre, King’s College LondonLondonUnited Kingdomhttps://ror.org/01hynnt93Institute of Cognitive and Clinical Neuroscience, Central Institute of Mental Health, Mannheim, Heidelberg UniversityMannheimGermanyhttps://ror.org/031bsb921Department of Psychology, School of Social Sciences, University of MannheimMannheimGermanyhttps://ror.org/03xjwb503NeuroSpin, CEA, Université Paris-SaclayGif-sur-YvetteFrancehttps://ror.org/0155zta11Departments of Psychiatry and Psychology, University of VermontBurlingtonUnited Stateshttps://ror.org/01ee9ar58Sir Peter Mansfield Imaging Centre School of Physics and Astronomy, University of Nottingham, University ParkNottinghamUnited Kingdomhttps://ror.org/05r3f7h03Physikalisch-Technische Bundesanstalt (PTB)Braunschweig and BerlinGermanyhttps://ror.org/00hx6zz33Institut National de la Santé et de la Recherche Médicale, INSERM U 1299 'Trajectoires développementales & psychiatrie', University Paris-Saclay, CNRS; Ecole Normale Supérieure Paris-Saclay, Centre BorelliGif-sur-YvetteFrancehttps://ror.org/02mh9a093AP-HP Sorbonne University, Department of Child and Adolescent Psychiatry, Pitié-Salpêtrière HospitalParisFrancePsychiatry Department, EPS Barthélémy DurandEtampesFrancehttps://ror.org/04v76ef78Institute of Medical Psychology and Medical Sociology, University Medical Center Schleswig Holstein, Kiel UniversityKielGermanyhttps://ror.org/0161xgx34Departments of Psychiatry and Neuroscience, Faculty of Medicine and Centre Hosptalier Universitaire Sainte-Justine, University of MontrealMontrealCanadahttps://ror.org/03dbr7087Departments of Psychiatry and Psychology, University of TorontoTorontoCanadahttps://ror.org/021ft0n22Department of Child and Adolescent Psychiatry and Psychotherapy, University Medical Centre GöttingenGöttingenGermanyhttps://ror.org/042aqky30Department of Psychiatry and Neuroimaging Center, Technische Universität DresdenDresdenGermanyhttps://ror.org/001w7jn25Centre for Population Neuroscience and Stratified Medicine (PONS), Department of Psychiatry and Neuroscience, Charité Universitätsmedizin BerlinBerlinGermanyhttps://ror.org/013q1eq08Centre for Population Neuroscience and Precision Medicine (PONS), Institute for Science and Technology of Brain-inspired Intelligence (ISTBI), Fudan UniversityShanghaiChinaZhangXilinReviewing Editorhttps://ror.org/01kq0pv72South China Normal UniversityChinade LangeFloris PSenior EditorDonders Institute for Brain, Cognition and BehaviourNetherlands0509202413RP9715003042024This manuscript was published as a preprint.07042024This manuscript was published as a reviewed preprint.05062024The reviewed preprint was revised.18072024© 2024, Weng et al2024Weng et alhttp://creativecommons.org/licenses/by/4.0/This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Substance use, including cigarettes and cannabis, is associated with poorer sustained attention in late adolescence and early adulthood. Previous studies were predominantly cross-sectional or under-powered and could not indicate if impairment in sustained attention was a predictor of substance use or a marker of the inclination to engage in such behavior. This study explored the relationship between sustained attention and substance use across a longitudinal span from ages 14 to 23 in over 1000 participants. Behaviors and brain connectivity associated with diminished sustained attention at age 14 predicted subsequent increases in cannabis and cigarette smoking, establishing sustained attention as a robust biomarker for vulnerability to substance use. Individual differences in network strength relevant to sustained attention were preserved across developmental stages and sustained attention networks generalized to participants in an external dataset. In summary, brain networks of sustained attention are robust, consistent, and able to predict aspects of later substance use.

neuroimagingadolescencesubstance uselongitudinalcannabisconnectivityResearch organismHumanhttp://dx.doi.org/10.13039/501100004543China Scholarship Council - Trinity College Dublin Joint Scholarship Programme202006750028WengYihehttp://dx.doi.org/10.13039/501100000780European Union-funded FP6 Integrated Project IMAGENLSHM-CT- 2007-037286SchumannGunterhttp://dx.doi.org/10.13039/501100007601Horizon 2020ERC Advanced Grant 'STRATIFY' 695313SchumannGunterhttp://dx.doi.org/10.13039/501100007155Medical Research Council'c-VEDA' MR/N000390/1SchumannGunterhttp://dx.doi.org/10.13039/100000002National Institutes of HealthR01DA049238SchumannGunterhttp://dx.doi.org/10.13039/501100007155Medical Research Foundation and Medical Research CouncilMR/R00465X/1SchumannGunterhttp://dx.doi.org/10.13039/501100007155Medical Research Foundation and Medical Research CouncilMR/S020306/1SchumannGunterhttp://dx.doi.org/10.13039/501100000780European Union funded project 'environMENTAL'101057429SchumannGunterhttp://dx.doi.org/10.13039/501100001665Agence Nationale de la RechercheANR-12-SAMA-0004MartinotMarie-Laure PaillèreSchumannGunterhttp://dx.doi.org/10.13039/501100001602Science Foundation Ireland16/ERCD/3797MartinotMarie-Laure PaillèreSchumannGunterWhelanRoberthttp://dx.doi.org/10.13039/501100001665Agence Nationale de la RechercheAAPG2019 - GeBraMartinotMarie-Laure PaillèreSchumannGunterThe funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.Author impact statementDifferences in behavior and brain connectivity in young adolescents predicted future substance use.publishing-routeprcIntroduction

Sustained attention is a critical cognitive process in daily life, playing a significant role in academic achievement, social communication, and mental health (Esterman and Rothlein, 2019) and can be defined as “the focus on performance on a single task over time, with the goal of explaining both the fluctuations within an individual as well as the individual differences in overall ability to maintain stable task performance” (p. 174) (Esterman and Rothlein, 2019). Sustained attention notably improves between the ages of 9 and 16 (Thomson et al., 2022), concomitant with cognitive maturation and brain development during adolescence (Paus, 2005). The functional neuroanatomy of sustained attention involves cingulate, prefrontal, and parietal cortices; supplementary motor area; frontal eye field; and cerebellum (Bauer et al., 2020; Pinar et al., 2018).

Cross-sectional studies suggest that substance use during adolescence, including cigarette smoking (Treur et al., 2015), alcohol consumption (Ueno et al., 2022), and cannabis use (Wallace et al., 2019), is associated with poorer sustained attention. For instance, adolescents (14–17 years of age) using cannabis a minimum of 4 days per week for at least the last 6 months showed impaired sustained attention in the rapid visual information processing (RVP) task, and in the immediate memory task versus non-users (Dougherty et al., 2013). Adolescents (12–17 years of age) in a high tetrahydrocannabinol (THC, the primary psychoactive component in cannabis) group exhibited lower accuracy on the RVP task than a low THC group (Shannon et al., 2010). Cigarette users aged 18–29 showed significant cognitive impairments in sustained attention than non-smokers in the RVP task (Chamberlain et al., 2012). A systematic review of the next-day cognitive effects of heavy alcohol consumption demonstrated impairments in sustained attention during alcohol hangovers using meta-analysis (Yakir et al., 2007). These findings highlight the negative associations between substance use and sustained attention.

Given the cross-sectional nature of the behavioral and neuroimaging studies above, it remains unclear if impaired sustained attention predates the initiation of substance use and/or if it is a consequence of substance use. Only one longitudinal study (Harakeh et al., 2012) has examined the association between sustained attention and cigarette smoking, employing measurements across three waves and involving a large sample of 1797 adolescents. Poor sustained attention, unlike other neurocognitive functions such as working memory, attention flexibility, or perceptual sensitivity, was associated with the increased probability of adolescents subsequently initiating cigarette smoking between ages 11 and 13 and with a higher chance of being a daily smoker by age 16. Harakeh and colleagues’ findings suggest that poor sustained attention may precede the onset of cigarette smoking. However, as their study was based on a behavioral level, the neural correlates underlying these associations remain untested.

Although lower sustained attention has been associated with subsequent cigarette smoking, individuals commonly engage in the concurrent use of multiple substances (Crummy et al., 2020), perhaps due to shared pathological substrates for substance use. A meta-analysis identified common neural alterations in primary dorsal striatal, and frontal circuits, engaged in reward/salience processing, habit formation, and executive control across various substances (nicotine, cannabis, alcohol, and cocaine) (Thiele and Bellgrove, 2018). Those involved in substance use often co-use both cannabis and cigarettes (Agrawal et al., 2012; Hindocha et al., 2016; Weinberger et al., 2018). Agrawal et al., 2012, reported that 90% of cannabis users smoke cigarettes during their lifetime, and the widespread co-use of the two may be attributed to genetic sharing (Agrawal et al., 2010; Yadav et al., 2016) and similar neural mechanisms (Klugah-Brown et al., 2020).

Functional brain networks can predict various behavioral traits, such as substance use (Yip et al., 2019) and sustained attention (Rosenberg et al., 2016). Previous studies (e.g. Rosenberg et al., 2018) have used brain connectivity to develop predictive models of sustained attention that can be generalized to healthy and clinical populations. However, while behavioral changes in sustained attention have been documented and functional brain networks that predict substance use have been identified (Yip et al., 2019), the underlying change in sustained attention brain networks from adolescence to adulthood and their relation to substance use are relatively less well described. Lower sustained attention has been accompanied by both stronger reductions in neural activity in the visual cortex and stronger recruitment of the right supramarginal gyrus with increasing time on a sustained attention task with central cues in cigarette smokers as opposed to non-smokers (Vossel et al., 2011). In a resting-state functional magnetic resonance imaging (fMRI) paradigm, cannabis users aged 16–26 had stronger connectivity between the left posterior cingulate cortex and the cerebellum, correlated with poorer performance on sustained attention/working memory and verbal learning measures (Ritchay et al., 2021). Although most brain connectomic research has utilized resting-state fMRI data, functional connectivity (FC) during task performance has demonstrated superiority in predicting individual behaviors and traits, due to its potential to capture more behaviorally relevant information (Dhamala et al., 2023; Greene et al., 2018; Yoo et al., 2018). Specifically, Zhao et al., 2023, suggested that task-related FC outperforms both typical task-based and resting-state FC in predicting individual differences. Hence, we applied task-related FC to predict sustained attention over time.

Previous studies found that FC patterns predicted individual differences in sustained attention (Chen et al., 2022; O’Halloran et al., 2018; Sripada et al., 2020), yet relatively little is known about the relationship between brain activity related to sustained attention and substance use over time. A latent change score model can quantify bidirectional longitudinal relations between substance use and both behaviors and brain activity associated with sustained attention, shedding light on how substance use impacts sustained attention and its associated brain activity, and vice versa. In this study, we used task-fMRI from the IMAGEN dataset, a longitudinal study with >1000 participants at each timepoint (ages 14, 19, and 23 years). We first obtained task-related whole-brain connectivity and then used connectome-based predictive modeling (CPM) to predict sustained attention from ages 14 to 23. Additionally, previous cross-sectional and longitudinal studies (Broyd et al., 2016; Harakeh et al., 2012; Lisdahl and Price, 2012) have shown that there are linear relationships between substance use and sustained attention over time. We therefore employed correlation analyses and a latent change score model to estimate the relationship between substance use and both behaviors and brain activity associated with sustained attention. Given the substantial sample size and longitudinal design of Harakeh et al.’s study, we hypothesized that behavioral and predictive networks associated with lower sustained attention would predict increased substance use (particularly cigarette smoking) over time.

ResultsBehavioral changes over time

Reaction time (RT) variability is a straightforward measure of sustained attention, with increasing variability thought to reflect poor sustained attention. RT variability can be defined as the intra-individual coefficient of variation (ICV), calculated as the standard deviation of Go RT divided by the mean Go RT from Go trials in the stop signal task. Lower ICV indicates better sustained attention. Participants’ demographic information for all analyses is shown in Table 1 (see also Supplementary file 1a and b). A linear mixed model analysis showed significant fixed effects of age (i.e. timepoint) on ICV (F1895.3 = 51.14, p<0.001) (Figure 1A). Post hoc analysis showed that ICV decreased with age: ICV at age 14 was significantly higher than ICV at ages 19 (t=6.535, p<0.001) and 23 (t=10.109, p<0.001). ICV at age 19 was also significantly higher than that at age 23 (t=4.768, p<0.001). The full results of the linear mixed model analysis are shown in Supplementary file 1c and d. In addition, we found that individual differences in ICV were significantly correlated between the three timepoints (Figure 1B and Supplementary file 1e, all p<2.8e–7).

Demographic information of adolescents in the linear mixed model across three timepoints.
Age 14Age 19Age 23
N (three timepoints)2148
Sex (M/F)1055/1093
Age (years)14.4±0.419±0.722.6±0.7
Mean FD (mm)0.28±0.320.18±0.170.18±0.12
GO RT (ms)466.6±80400.7±71.8403.9±73.8
ICV0.234±0.0380.224±0.0510.217±0.052
Stop RT (ms)461.5±114.8360±82.4363.6±78.2
SSD (ms)319.3±148.1188.1±132.4190±158.4
SSRT (ms)217.8±37.2213.3±43.3216.2±42.6
pOmission (%)4.4±10.52.6±8.63.7±11.1
pChoiceError (%)4.7±6.64.8±4.75.2±7.6
pCommission (%)47.9±6.347.5±647.2±6.9

Note: These data pertain to the participants included in the behavioural analyses. N, number of subjects; FD, framewise displacement of MR images; ICV, intra-individual coefficient of variation (assay for sustained attention); SSRT, stop signal reaction time; GO RT, reaction time in Go trials; Stop RT, reaction time in stop fail trials; SSD, stop signal delay; pOmisssion, probability of go omissions (no response); pChoiceError, probability of choice errors on Go trials; pCommission, probability of commission on Stop trials.

Intra-individual coefficient of variation (ICV) changes over time.

(A) ICV changes over time. (B) Correlation of ICV between timepoints within participants. †, p<0.001.

Cross-sectional brain connectivity

This study employed CPM, a data-driven neuroscience approach, to identify three predictive networks – positive, negative, and combined – to predict ICV from brain connectivity. CPM typically uses the strength of the predictive networks to predict individual differences in traits and behaviors. The predictive networks were obtained based on connectivity analyses of the whole brain. Specifically, we assessed whether connections between brain areas (i.e. edges) in a task-related FC matrix derived from generalized psychophysiological interaction (gPPI) analysis were positively or negatively correlated with ICV using a significance threshold of p<0.01. These positively or negatively correlated connections were regarded as positive or negative networks, respectively. The network strength of positive networks (or negative networks) was determined for each individual by summing the connection strength of each positively (or negatively) correlated edge. The combined network was determined by subtracting the strength of the negative network from the positive network. We then built a linear model between network strength and ICV in the training set and applied these predictive networks to yield network strength and a linear model in the test set to calculate predicted ICV using k-fold cross-validation (CV).

Positive, negative, and combined networks derived from Go trials significantly predicted ICV: at age 14 (r=0.25, r=0.25, and r=0.28, respectively, all p<0.001) (Figure 2A), at age 19 (r=0.27, r=0.25, r=0.28, respectively, all p<0.001) (Figure 2B), and at age 23 (r=0.38, r=0.33, and r=0.37, respectively, all p<0.001) (Figure 2C). The connectome patterns of predictive networks are shown in Figure 2D–I. Figure 2—figure supplement 1 summarizes the connectivity within and between functional networks and depicts their respective contribution to the predictive network. The above results were validated using 10-fold CV; similar results were obtained when using 5-fold CV and leave-site-out CV (Supplementary file 1f). The predictive networks had similar connectome patterns when different exclusion criteria for head motion were used (mean framewise displacement, mean FD <0.2–0.4 mm) (Figure 3—figure supplements 2–4A). In addition, we found that network strength of positive, negative, and combined networks derived from Go trials was significantly correlated between the three timepoints (Supplementary file 1g , all p<0.003).

The predictive performances and networks of intra-individual coefficient of variation (ICV) per timepoint derived from Go trials.

Correlation between observed and predicted ICV in positive, negative, and combined networks at (A) age 14, (B) age 19, and (C) age 23. Predictive networks for ICV are at (D) age 14, (E) age 19, and (F) age 23. Connectome of positive and negative networks of ICV at (G) age 14, (H) age 19, and (I) age 23. The edges depicted above are those selected in at least 95% of cross-validation folds. Red, blue, and green spheres/lines/scatters represent positive, negative, and combined networks respectively. MF, medial frontal; FP, frontoparietal; DMN, default mode; MOT, motor; VI, visual I; VII, visual II; VAs, visual association; SAL, salience; SC, subcortical; CBL, cerebellar. R/L, right/left hemisphere. ***, p<0.001.

The predictive networks predicting intra-individual coefficient of variation (ICV) per timepoint derived from Go trials.

The heatmaps show predictive networks with non-zero values at (A) age 14, (B) age 19, (C) age 23. Each heatmap cell shows the number of edges between or within functional networks. Radar plots show the proportion of the functional networks involved in predictive networks. The edges depicted above are those selected in at least 95% of cross-validation folds. Red, blue, and green represent positive, negative, and combined networks respectively. MF, medial frontal; FP, frontoparietal; DMN, default mode; MOT, motor; VI, visual I network; VII, visual II network; VAs, visual association network; SAL, salience; SC, subcortical; CBL, cerebellar. R/L, right/left hemisphere.

Positive, negative, and combined networks derived from Successful stop trials significantly predicted ICV: at age 14 (r=0.22, p<0.001; r=0.12, p=0.017; and r=0.20, p<0.001, respectively) (Figure 3A), at age 19 (r=0.19, p<0.001; r=0.15, p=0.001; and r=0.18, p<0.001, respectively) (Figure 3B), and at age 23 (r=0.24, r=0.21, and r=0.23, respectively, all p<0.001) (Figure 3C). The connectome patterns of predictive networks are shown in Figure 3D–I. Figure 3—figure supplement 1 summarizes the connectivity within and between functional networks and the proportion of brain networks involved in the predictive network. We obtained similar results using a 5-fold CV and leave-site-out CV (Supplementary file 1e). The predictive networks had similar connectome patterns when different exclusion criteria for head motion were used (mean FD <0.2–0.4 mm) (Figure 3—figure supplements 2–4B). In addition, we found that network strength of positive, negative, and combined networks derived from Successful stop trials was significantly correlated between the three timepoints (Supplementary file 1f, all p<0.001).

The predictive performances and networks of intra-individual coefficient of variation (ICV) per timepoint derived from Successful stop trials.

Correlation between observed and predicted ICV in positive, negative, and combined networks at (A) age 14, (B) age 19, and (C) age 23. Predictive networks for ICV are at (D) age 14, (E) age 19, and (F) age 23. Connectome of positive and negative networks of ICV at (G) age 14, (H) age 19, and (I) age 23. The edges depicted above are those selected in at least 95% of cross-validation folds. Red, blue, and green spheres/lines/scatters represent positive, negative, and combined networks respectively. MF, medial frontal; FP, frontoparietal; DMN, default mode; MOT, motor; VI, visual I; VII, visual II; VAs, visual association; SAL, salience; SC, subcortical; CBL, cerebellar. R/L, right/left hemisphere. *, p<0.05; **, p<0.01; ***, p<0.001.

The predictive networks predicting intra-individual coefficient of variation (ICV) per timepoint derived from Successful stop trials.

The heatmaps show predictive networks with non-zero values at (A) age 14, (B) age 19, and (C) age 23. Each heatmap cell shows the number of edges between or within functional networks. Radar plots show the proportion of the functional networks involved in predictive networks. The edges depicted above are those selected in at least 95% of cross-validation folds. Red, blue, and green represent positive, negative, and combined networks respectively. MF, medial frontal; FP, frontoparietal; DMN, default mode; MOT, motor; VI, visual I network; VII, visual II network; VAs, visual association network; SAL, salience; SC, subcortical; CBL, cerebellar. R/L, right/left hemisphere.

Connectome of positive and negative networks predicting intra-individual coefficient of variation (ICV) at age 14 with mean framewise displacement (meanFD) from 0.2 mm to 0.5 mm.

Red and blue lines represent positive and negative networks respectively. MF, medial frontal; FP, frontoparietal; DMN, default mode; MOT, motor; VI, visual I; VII, visual II; VAs, visual association; SAL, salience; SC, subcortical; CBL, cerebellar. R/L, right/left hemisphere.

Connectome of positive and negative networks predicting intra-individual coefficient of variation (ICV) at age 19 with mean framewise displacement (meanFD) from 0.2 mm to 0.5 mm.

Red and blue lines represent positive and negative networks respectively. MF, medial frontal; FP, frontoparietal; DMN, default mode; MOT, motor; VI, visual I; VII, visual II; VAs, visual association; SAL, salience; SC, subcortical; CBL, cerebellar. R/L, right/left hemisphere.

Connectome of positive and negative networks predicting intra-individual coefficient of variation (ICV) at age 23 with mean framewise displacement (meanFD) from 0.2 mm to 0.5 mm.

Red and blue lines represent positive and negative networks respectively. MF, medial frontal; FP, frontoparietal; DMN, default mode; MOT, motor; VI, visual I; VII, visual II; VAs, visual association; SAL, salience; SC, subcortical; CBL, cerebellar. R/L, right/left hemisphere.

To examine the specificity of sustained attention networks identified from CPM analysis, the correlations between the network strength of positive and negative networks and performances from a neuropsychology battery (Cambridge Neuropsychological Test Automated Battery [CANTAB]) (Fray et al., 1996) were calculated at each timepoint separately. All positive and negative networks derived from Go and Successful stop trials were significantly correlated with a behavioral assay of sustained attention – the RVP task – at ages 14 and 19 (all p<0.028). Age 23 had no RVP task data in the IMAGEN study. There were sporadic significant correlations between constructs such as delay aversion/impulsivity and negative network strength, for example, but the most robust correlations were with the RVP. Detailed information is shown in Appendix 1 and Supplementary file 1l.

ICV prediction across time

Positive, negative, and combined networks derived from Go trials defined at age 14 predicted ICV at ages 19 (r=0.16, r=0.14, and r=0.16, all p<0.001) (Figure 4A, top row) and 23 (r=0.20, r=0.12, and r=0.17, all p<0.001) (Figure 4A, middle row). Likewise, positive, negative, and combined networks derived from Go trials defined at age 19 predicted ICV at age 23 (r=0.30, r=0.26, and r=0.31, respectively, all p<0.001) (Figure 4A, bottom row).

The predictive performances of intra-individual coefficient of variation (ICV) across timepoints and generalization in STRATIFY.

Predictive performances of ICV (A) derived from Go trials and (B) derived from Successful stop trials. The top, middle, and bottom rows of (A) and (B) panels show the predictive performance: using models defined at age 14 to predict age 19 (i.e. 14 years → 19 years), using models defined at age 14 to predict age 23 (i.e. 14 years → 23 years), and using models defined at age 19 to predict age 23 (i.e. 19 years → 23 years) respectively. Generalization of predictive networks predicting ICV defined at age 23 in STRATIFY (i.e. 23 years → STRATIFY) derived from (C) Go trials and (D) Successful stop trials. The red, blue, and green scatter represent positive, negative, and combined networks. †, p<0.001.

Generalization in subgroups in STRATIFY.

(A) Predictive performance in distinct patient groups in STRATIFY derived from Go and Successful stop trials. (B) The correlation between network strength and intra-individual coefficient of variation (ICV) across patient cohorts in STRATIFY derived from Go and Successful stop trials. AUD, alcohol use disorder; MDD, major depression disorder; BN, bulimia nervosa; AN, anorexia nervosa; HC, healthy controls. *, p<0.05; **, p<0.01; ***, p<0.001.

Positive, negative, and combined networks derived from Successful stop trials defined at age 14 predicted ICV at age 19 (r=0.11, r=0.12, and r=0.13, all p<0.001) (Figure 4B, top row) and 23 (r=0.14, r=0.15, and r=0.15, all p<0.001) (Figure 4B, middle row). Positive, negative, and combined networks derived from Successful stop trials defined at age 19 predicted ICV at age 23 (r=0.17, r=0.16, and r=0.17, respectively, all p<0.001) (Figure 4B, bottom row).

Generalization of ICV brain networks

We tested if the predictive networks defined at age 23 in IMAGEN would generalize to an external dataset, namely STRATIFY (N = ~300), comprising individuals also aged 23. When applied to the whole STRATIFY sample, positive, negative, and combined networks derived from Go trials at age 23 in IMAGEN predicted ICV in STRATIFY (r=0.34, r=0.34, and r=0.35, respectively, all p<0.001) (Figure 4C), as did networks derived from Successful stop trials (r=0.26, r=0.22, and r=0.26, respectively, all p<0.001) (Figure 4D).

Factor analysis of substance use

Exploratory factor analysis on data from the Timeline Followback (TLFB) (Sobell et al., 1996), an instrument for measuring the consumption of alcohol, drugs, and smoking for participants, yielded two common factors at age 14 and three common factors at ages 19 and 23. According to the rotated factor loading analysis, at age 14, two common factors were identified, which we labeled as (i) alcohol and (ii) cigarette and cannabis use (Cig+CB). At ages 19 and 23, three common factors were identified, which we labeled as (i) alcohol, (ii) Cig+CB, and (iii) drug (including cocaine, ecstasy, and ketamine) use. Additional details about this data reduction step are shown in Figure 5—figure supplement 1 and Supplementary file 1k.

Correlation between behavior and brain to cannabis and cigarette use

We calculated the Spearman correlation between ICV/sustained brain activity and TLFB factor score per timepoint and across timepoints. Brain activity was measured by the strength of positive and negative networks predicting sustained attention. The p values were corrected by false discovery rate (FDR) correction (q<0.05). Figure 5A–C summarizes the results showing the correlation between ICV/brain activity and Cig+CB per timepoint and across timepoints. Figure 5A shows correlations between ICV and Cig+CB (Supplementary file 1n-o). ICV was correlated with Cig+CB at ages 19 (Rho = 0.13, p<0.001) and 23 (Rho = 0.17, p<0.001). ICV at ages 14 (Rho = 0.13, p=0.007) and 19 (Rho = 0.13, p=0.0003) were correlated with Cig+CB at age 23. Cig+CB at age 19 was correlated with ICV at age 23 (Rho = 0.13, p=9.38E-05). Figure 5B shows correlations between brain activity derived from Go trials and Cig+CB (Supplementary file 1r-s). Brain activities of positive and negative networks derived from Go trials were correlated with Cig+CB at age 23 (positive network: Rhop = 0.12, p<0.001; negative network: Rhon = –0.11, p<0.001). Brain activity of the negative network derived from Go trials at age 14 was correlated with Cig+CB at age 23 (Rhon = –0.16, p=0.001). Cig+CB at age 19 was correlated with brain activity of the positive network derived from Go trials at age 23 (Rhop = 0.10, p=0.002). Figure 5C shows the correlations between brain activity derived from Successful stop and Cig+CB (Supplementary file 1r-s). Brain activities of positive and negative networks derived from Successful stop were correlated with Cig+CB at ages 19 (positive network: Rhop = 0.10, p=0.001; negative network: Rhon = –0.08, p=0.013) and 23 (positive network: Rhop = 0.13, p<0.001; negative network: Rhon = –0.11, p=0.001). No correlation between alcohol use and ICV/brain activity was found after FDR correction. Detailed results on the correlation between ICV/brain activity and substance use can be found in the Supplementary file 1n-u.

Significant correlations between sustained attention and substance use across timepoints (false discovery rate [FDR] correction, q<0.05).

(A) Correlations between the intra-individual coefficient of variation (ICV) and cigarette and cannabis use (Cig+CB) across timepoints. Correlations between sustained attention network strength and Cig+CB across timepoints (B) derived from Go trials and (C) derived from Successful stop trials. Rhop: r value between network strength of the positive network. Rhon: r value between network strength of the negative network.

Exploratory factor analysis of Timeline Followback (TLFB) at each timepoint.

(A) Total variance explained for exploratory factor analysis of TLFB items. (B) Rotated component matrix for exploratory factor analysis. Extraction method: Principal component analysis. Rotation method: Varimax with Kaiser normalization.

Significant correlations between sustained attention and substance use across timepoints (false discovery rate [FDR] correction, q<0.05).

(A) Correlations between the intra-individual coefficient of variation (ICV) and cigarette and cannabis use (Cig+CB) across timepoints. Correlations between sustained attention network strength and Cig+CB across timepoints (B) derived from Go trials and (C) derived from Successful stop trials. Rhop: r value between network strength of the positive network. Rhon: r value between network strength of the negative network.

Bivariate latent change score model

We used a bivariate latent change score model to explore the relationship between substance use (specifically Cig+CB and alcohol use) and ICV/brain activity. This approach tests for bidirectional associations, examining how substance use at age 14 predicts changes in ICV/brain activity from ages 14 to 23 and vice versa (Figure 6). Below, we present the findings regarding the lagged effects of substance use on ICV/brain activity and the lagged effects of ICV/brain activity on substance use (Table 2). The p values were corrected by FDR correction (q<0.05).

A simplified bivariate latent change score model for substance use and ICV/brain activity.

SUB, substance use (alcohol, cigarette, and cannabis use); Brain, brain network strength of positive/negative network of sustained attention derived from Go trials/Successful stop trials. ICV, intra-individual coefficient of variation. T1, timepoint 1 (age 14); T2, timepoint 2 (age 19); T3, timepoint 3 (age 23). γ1, lagged effects of substance use on ICV or brain activity. γ2, lagged effects of ICV or brain activity on substance use. The square/circle represents the observation/true score in the model.

Bivariate latent change score model showing the bidirectional association between substance use and ICV/brain networks (false discovery rate corrected).
Cig+CBAlcohol use
Lagged effects of Cig+CB (γ1)Lagged effects of ICV/brain networks (γ2)Lagged effects of alcohol use (γ1)Lagged effects of ICV/brain networks (γ2)
Std. β (SE)Std. β (SE)Std. β (SE)Std. β (SE)
ICV0.017 (0.039)0.117 (0.031)***0.005 (0.029)0.057 (0.030)
SA GT PosNet–0.026 (0.030)0.087 (0.032)**0.025 (0.030)0.022 (0.036)
SA GT NegNet0.012 (0.026)–0.094 (0.035)**–0.012 (0.030)–0.059 (0.034)
SA SS PosNet0.005 (0.025)0.070 (0.036)0.101 (0.040)0.046 (0.039)
SA SS NegNet0.038 (0.028)–0.061 (0.031)–0.003 (0.035)–0.069 (0.031)
Lagged effects of Cig+CB on changes in ICV and brain activity

We examined if Cig+CB use at age 14 predicted the changes in ICV or brain activity (i.e. predictive network strength) associated with sustained attention across ages 14–23. No significance was observed in the lagged effects of Cig+CB on changes in ICV and brain activity (all p>0.172).

Lagged effects of ICV and brain activity on changes in Cig+CB

We examined if ICV or brain activity associated with sustained attention at age 14 predicted changes in Cig+CB use across ages 14–23. Behaviors and brain activity associated with poor sustained attention predicted a greater increase in subsequent cigarette and cannabis use. Specifically, higher ICV at age 14 predicted a greater increase in Cig+CB from ages 14 to 23 (Std. β=0.12, p<0.001). Higher sustained attention network strength for positive network derived from Go trials at age 14 predicted a greater increase in Cig+CB from ages 14 to 23 (Std. β=0.09, p=0.006). Lower sustained attention network strength for the negative network, also derived from Go trials at age 14, predicted a greater increase in Cig+CB from ages 14 to 23 (Std. β=–0.09, p=0.006). No other lagged effects of brain activity on changes in Cig+CB remained significant after FDR correction (all p>0.047). Figure 7 illustrates the changes in raw scores of cigarette and cannabis use from the TLFB for individuals at age 14 with higher sustained attention (i.e. lower ICV, lower strength of positive network, or higher strength of negative network) and lower sustained attention (i.e. higher ICV, higher strength of positive network, or lower strength of negative network).

Cigarette and cannabis score in Timeline Followback changes in individuals with high sustained attention (High SA) and low sustained attention (Low SA) from ages 14 to 23.

Participants were categorized into five equal groups based on the intra-individual coefficient of variation (ICV), strength of positive network, and strength of negative network at age 14. (A) Top ICV (Low SA) and bottom ICV (High SA) groups. (B) The top strength of the positive network (Low SA) and bottom strength of the positive network (High SA) groups derived from Go trials. (C) The top strength of the negative network (High SA) and bottom strength of the negative network (Low SA) groups derived from Go trials. Note that the higher strength of the negative network reflects lower ICV and higher sustained attention.

Cigarette and cannabis score in Timeline Followback change in individuals with good working memory (Good WM) and poor working memory (Poor WM) from ages 14 to 23.

Participants were categorized into five equal groups based on the performance of strategy working memory task at age 14.

Association between alcohol use and ICV/brain activity

We examined if alcohol use at age 14 predicted changes in ICV or brain activity associated with sustained attention across ages 14–23, or vice versa. No significant results were found for the lagged effects of alcohol use on changes in ICV and brain activity, nor the lagged effects of ICV and brain activity on changes in alcohol use. The p values were insignificant after FDR correction (all p>0.011).

Discussion

It is well known that increased substance use, including cigarettes and cannabis, is associated with poorer sustained attention in late adolescence and early adulthood (Chamberlain et al., 2012; Dougherty et al., 2013). However, previous studies, which were predominantly cross-sectional or under-powered, left a critical question unanswered. That is, was the impairment in sustained attention a predictor of substance use or a marker of the inclination to engage in such behavior? Using a substantial sample size, our results indicate that behavior and brain connectivity associated with poorer sustained attention at age 14 predicted a larger increase in cannabis and cigarette smoking from ages 14 to 23. Furthermore, our findings highlight the robustness of the brain network associated with sustained attention over time, making the latter a potentially useful biomarker for vulnerability to substance use.

Substance use and the sustained attention network

Our study applied a latent change score model on a large longitudinal dataset, testing the precedence between substance use and sustained attention. In contrast to prior research suggesting that substance use impaired sustained attention (Broyd et al., 2016; Figueiredo et al., 2020), our results indicate that lower sustained attention also predates substance use. A link between substance use and sustained attention is plausible, given the underlying neurobiology of this sustained attention. Substantial evidence from neuropharmacological studies in rats and humans has shown the modulatory role of neurotransmitters in sustained attention (Bloomfield et al., 2016; Granon et al., 2000; Marshall et al., 2019). Elevated dopamine and noradrenaline levels in the prefrontal cortex lead to improved sustained attention in a dose-dependent manner (Marshall et al., 2019). In humans, methylphenidate, a psychostimulant commonly used to treat ADHD, increases both noradrenaline and dopamine signaling and improves sustained attention (Dockree et al., 2017). Thus, poorer sustained attention may reflect a lower basal level of dopamine and noradrenaline. More importantly, studies in primates (Morgan et al., 2002; Nader et al., 2006), rodents (Dalley et al., 2007; Trifilieff et al., 2017), and humans (Casey et al., 2014; Trifilieff and Martinez, 2014; Volkow et al., 2006) have indicated that low basal dopamine levels are markers of vulnerability for increased drug administration. For example, Casey et al., 2014, demonstrated that blunted dopamine release may precede the development of addiction in humans. Nader et al., 2006, found a negative correlation between baseline D2 receptor availability and rates of cocaine self-administration in monkeys. Thus, these findings collectively suggest that sustained attention and its brain network could serve as a biomarker of vulnerability to substance use.

These results emphasize the specificity of sustained attention and its associated brain networks, rather than other cognitive abilities, for predicting substance use over time. Unlike sustained attention, no significant differences in cigarette and cannabis use were observed between individuals with lower and higher working memory at baseline during the strategy working memory (SWM) task (Supplementary file 1w and Figure 7—figure supplement 1). Our results support the behavioral-only findings of a previous study (Harakeh et al., 2012), which found that individuals with poorer sustained attention, rather than other cognitive functions, were more likely to initiate smoking cigarettes. Our study goes further by showing that sustained attention brain networks can predict substance use in the future.

Neural associations between cigarette and cannabis use

We constructed composite scores of substance use. An exploratory factor analysis identified cigarettes and cannabis items as a common factor, aligning with previous studies (Ferland and Hurd, 2020; Hindocha et al., 2016; Weinberger et al., 2018) that indicate concurrent cannabis and cigarette use among users. A national survey in America indicated that 18–23% of cigarette smokers aged 12–17 met the criteria for cannabis use disorder, in contrast to only 2% of non-smoking youth (Weinberger et al., 2018). Another national online survey in the UK reported that 80.8% of cigarette smokers engage in cannabis consumption, indicating a prevalent practice of co-administering cannabis and tobacco through smoking (Hindocha et al., 2021). Shared genetic factors (Agrawal et al., 2010; Yadav et al., 2016) and similar neural associations (Wetherill et al., 2015) contribute to the co-use of cannabis and cigarettes. Yadav et al., 2016, demonstrated a strong and significant genetic correlation between lifetime cannabis use and lifetime cigarette smoking within a large cohort of 32,330 participants, suggesting a high degree of genetic sharing between the two. Using neuroimaging techniques, Wetherill et al., 2015, indicated that individuals who used cannabis, smoked tobacco, or engaged in co-use exhibited larger gray matter volumes in the left putamen compared to healthy controls. Both nicotine and cannabis have similar effects on mesolimbic dopaminergic pathways engaged, modulating dopamine release in the striatum (Bossong et al., 2009; Dongelmans et al., 2021). Collectively, these findings suggest a similar neural association between cigarette and cannabis use.

Specificity and robustness of sustained attention networks

The brain networks we describe were specific to sustained attention. The strength of the sustained attention brain network was robustly correlated with RVP task performance, a typical sustained attention task, rather than other cognitive measures (Supplementary file 1l). Importantly, as highlighted in a previous study (Cwiek et al., 2022), emphasizing the importance of generalization in an external dataset, our study found that the sustained attention network derived from Go trials and Successful stop trials generalized to an external dataset (see further discussion on the generalization in subgroups in STRATIFY in Appendix 1).

We also replicated and extended the developmental pattern of sustained attention and its networks from mid-adolescence to young adulthood. A notable enhancement in sustained attention (i.e. decreased ICV) was observed from ages 14 to 23, as expected (Fortenbaugh et al., 2015; Williams et al., 2005). Sustained attention networks derived from Go and Successful stop trials predicted behavior at different timepoints, implying that individual differences in sustained attention and associated networks were preserved throughout development. Previously, in neurodiverse youth, attention networks in individuals remained stable across months to years (Sisk et al., 2022). Rosenberg et al., 2020, also illustrated that the same functional connections predicting overall sustained attention ability also forecasted attentional changes observed over minutes, days, weeks, and months. Here, we contribute to these insights by extending the understanding that attention network stability is not only applicable to neurodiverse populations but also holds in a sizeable cohort of healthy participants. Furthermore, our findings indicate that sustained attention networks remain stable over several years, providing valuable insights into the potential for sustained attention to function as a robust and efficient biomarker for substance use. However, there are still some individual variabilities not captured in this study, which could be attributed to the diversity in genetic, environmental, and developmental factors influencing sustained attention and substance use. Future research should aim to explore these variabilities in greater depth to gain better understanding of the relationship between sustained attention and substance use.

In conclusion, robust sustained attention networks were identifiable from ages 14 to 23. Individual differences in sustained attention network strength were predictable across time. Poorer sustained attention and strength of the associated brain networks at age 14 predicted greater increases in cannabis and cigarette smoking from ages 14 to 23.

Materials and methodsParticipants

All neuroimaging data and behavioral data were obtained from the IMAGEN study. IMAGEN is a large longitudinal study that recruited over 2000 participants aged 14–23 in Europe (Kaiser et al., 2022). This study used the stop signal task fMRI data at ages 14, 19, and 23. In addition, we used an independent dataset STRATIFY as external validation for age 23. STRATIFY (N = ~300) is a sub-dataset within IMAGEN that recruits fMRI data from patients aged 23. Written and informed consent was obtained from all participants by the IMAGEN consortium and the study was approved by the institutional ethics committee of King’s College London (PNM/10/11-126), University of Nottingham (D/11/2007), Trinity College Dublin (SPREC092007-01), Technische Universitat Dresden (EK 235092007), Commissariat a l'Energie Atomique et aux Energies Alternatives, INSERM (2007-A00778-45), University Medical Center at the University of Hamburg (M-191/07) and in Germany at medical ethics committee of the University of Heidelberg (2007-024N-MA) in accordance with the Declaration of Helsinki. We followed the exclusion criteria outlined in previous studies (O’Halloran et al., 2018; Whelan et al., 2014). Participants were excluded from the CPM analysis if they had more than 20% errors on the Go trials (incorrect responses or responses that were too late) or if they had a mean framewise displacement (mean FD)>0.5 mm. Finally, 717 participants at age 14, 1081 participants at age 19, and 1120 participants at age 23 were used to predict ICV. In STRATIFY, 304 participants were used to predict ICV.

Stop signal task

The stop signal task required participants to respond to a Go signal (arrows pointing left/right) by pressing the left/right button while withholding their response if the Go signal was unpredictably followed by a Stop signal (arrows pointing upward). The Go signal was displayed on the screen for 1000 ms in the Go trials, while the Stop signal appeared for 100–300 ms following the Go signal on average 300 ms later in unpredictable Stop trials. To adjust task difficulty dynamically, we used a tracking algorithm on the delay between the Go signal and Stop signal (stop signal delay, 250–900 ms in 50 ms increments) (Verbruggen et al., 2019), to produce 50% successful and 50% unsuccessful inhibition trials. The task at age 14 included 400 Go trials and 80 variable delay Stop trials, with 3 and 7 Go trials between successive Stop trials. The task at ages 19 and 23 consisted of 300 Go trials and 60 variable delay Stop trials. Before the MRI scan, participants also performed a practice session with a block of 60 trials to become familiar with the task. ICV is used to assess sustained attention in this task for each participant. ICV reflects short-term within-person variations in task performance (O’Halloran et al., 2018). Specifically, ICV is computed by dividing the standard deviation of Go RT by the mean Go RT. Lower ICV indicates better sustained attention.

Self-report questionnairesPuberty development scale

The puberty development scale (PDS), an 8-item self-report assessment, measures the pubertal development of adolescents (Petersen et al., 1988). The PDS evaluates physical development using a 5-point scale where 1 corresponds to prepubertal, 2 to beginning pubertal, 3 to mid-pubertal, 4 to advanced pubertal, and 5 to postpubertal. In addition, the items are adapted for sex, such as voice changes for males or menarche for females.

Timeline Followback

We used the TLFB, a retrospective self-report instrument that uses a calendar method to evaluate prior substance use consumption over the past 30 days (Sobell et al., 1996). The TLFB has strong reliability and validity for assessing alcohol consumption, and we used it to measure the use of alcohol, drugs, and smoking for participants.

MRI acquisition and pre-processing

Functional MRI data of the stop signal task in the IMAGEN study were collected at eight scan sites (London, Nottingham, Dublin, Mannheim, Dresden, Berlin, Hamburg, and Paris), and data in STRATIFY were collected at three scan sites (Berlin, two scanners in London) with 3T MRI scanners. The MR scanning protocols, cross-site standardization, and quality checks are further described in Whelan et al., 2012. All images were obtained using echo-planar imaging (EPI) sequence with the following parameters: repetition time=2.2 s, echo time=30 ms, flip angle = 75°, field of view=224 mm × 224 mm, data matrix = 64 × 64, slice thickness = 2.4 mm with 1 mm slice gap, voxel size = 3.5 mm × 3.5 mm × 4.38 mm, 40 transversal interleaved slices. The MRI data has 444 volumes at age 14 and 320–350 volumes at ages 19 and 23. Standardized hardware was used for visual stimulus presentation (Nordic Neurolab, Bergen, Norway) at all scan sites.

All fMRI data from the IMAGEN study were pre-processed centrally using SPM12 (Statistical Parametric Mapping, http://www.fil.ion.ucl.ac.uk/spm/) with an automated pipeline. The images were corrected for slice timing and then realigned to the first volumes to correct head motions. Participants were excluded from the study if they had a mean FD >0.5 mm. Subsequently, the data were non-linearly transformed to the Montreal Neurological Institute Coordinate System space using a custom EPI template with the voxels resampled at 3  mm× 3 mm ×3 mm resolution. Finally, the images were smoothed with a Gaussian kernel at a full-width-at-half-maximum of 5 mm.

Generalized psychophysiological interaction analysis

In this study, we adopted gPPI analysis to generate task-related FC matrices and applied CPM analysis to investigate predictive brain networks from adolescents to young adults. PPI analysis describes task-dependent FC between brain regions, traditionally examining connectivity between a seed region of interest (ROI) and the voxels of the whole rest brain. However, this study conducted a gPPI analysis, which is on ROI-to-ROI basis (Di et al., 2021), to yield a gPPI matrix across the whole brain instead of just a single seed region. First, we conducted a general linear model (GLM) analysis on the pre-processed fMRI data to examine brain activity during the stop signal task. Two separate GLMs were created for Go trials and Successful stop trials. The Go trials model included three task regressors (Go trials, Failed stop trials, and Successful stop trials) and 36 nuisance regressors, which accounted for factors such as head motion and the signal from white matter and cerebrospinal fluid. The 36 nuisance regressors are 3 translations, 3 rotations, mean white matter signal, mean cerebrospinal fluid signal, mean gray matter signal, their derivatives, and the squares of all these variables. Given the high frequency of Go trials in SST, it is common to treat Go trials as an implicit baseline, as in previous IMAGEN studies (D’Alberto et al., 2018; Whelan et al., 2012). Hence, we built a separate GLM for Successful stop trials, which included two task regressors (Failed and Successful stop trials) and 36 nuisance regressors. All task regressors were modeled by convolving with the canonical hemodynamic response function (HRF) and high pass filtered (128 s). We then conducted a gPPI analysis across the entire brain using the Shen atlas with 268 regions (Shen et al., 2013) for both Go and Successful stop trials. The gPPI analysis involved deconvolving the time series of each ROI with the HRF, multiplying it by the psychological variables of interest to yield a neural level PPI term, and convolving the resulting PPI term with the HRF to obtain the BOLD level PPI effects (Di and Biswal, 2019). Separate GLM models were used to estimate the PPI effect of each ROI for Go trials and Successful stop trials, regressing the eigenvariate of the seed ROI. The GLM of the Go trials included one regressor of another ROI eigenvariate, three regressors of task condition, three regressors of the PPI effects, and one contrast term (Equation 1). The GLM of Successful stop trials included one regressor of another ROI eigenvariate, two regressors of task condition, two regressors of the PPI effects, and one contrast term (Equation 2), shown as follows:Y=β0+β1∗Xphysio+β2∗Xpsycho(SS)+β3∗Xpsycho(FS)+β4∗Xpsycho(GO)+β5∗Xphysio∗Xpsycho(SS)+β6∗Xphysio∗Xpsycho(FS)+β7∗Xphysio∗Xpsycho(GO)+εY=β0+β1∗Xphysio+β2∗Xpsycho(SS)+β3∗Xpsycho(FS)+β4∗Xphysio∗Xpsycho(SS)+β5∗Xphysio∗Xpsycho(FS)+ε

Note: SS, Successful stop trials; FS, Failed stop trials; GO, Go trials.

where Y is the time series of seed ROI, Xphysio is the time series of another ROI, Xpsycho is the task design term, and ε is the residual term. The gPPI analysis was performed across each ROI from the Shen atlas, resulting in a 268*268 gPPI matrix for each participant derived from Go trials and Successful stop trials separately. The matrices were transposed and averaged with the original matrices to yield symmetrical matrices (Di et al., 2021), and prepared for further analysis.

Connectome-based predictive modelingICV prediction

CPM is a data-driven method that can examine individual differences in brain connectivity (Shen et al., 2017). CPM identifies pairwise connections between brain regions most highly correlated with a given phenotype. Using the PPI matrix, we employed CPM to predict ICV, for ages 14, 19, and 23. The CPM analysis process includes feature selection, model building, and validation (Figure 8). We applied CV to divide all participants into training and testing sets. (i) First, we used partial correlation to calculate the relationship between each edge in the gPPI matrix and behavioral phenotype while controlling several covariates in the training set. These covariates included ages, genders, mode-centered PDS (at age 14 only), mean FD, and scan sites, regarded as a dummy variable. The r value with an associated p value for each edge was obtained, and a threshold p=0.01 (Feng et al., 2024; Ren et al., 2021; Yoo et al., 2018) was set to select edges. The positive or negative correlated edges in feature selection were regarded as positive or negative networks. (ii) Second, we calculated network strength for each participant in the training set by summing the selected edges in the gPPI matrix for both positive and negative networks. We also estimated the network strength of a combined network by subtracting the strength of the negative from the strength of the positive network. (iii) Finally, we constructed predictive models based on the assumption of a linear relationship between network strength of the positive, negative, and combined networks, and behavioral phenotype in the training set. The covariates were also adjusted in this linear model. The network strengths for each participant in the testing set were calculated and input into the predictive model along with the covariates to predict each network’s behavioral phenotypes.

Schematic of connectome-based predictive modeling.

(i) Feature selection. The correlation between each edge in the generalized psychophysiological interaction (gPPI) matrix and the behavioral phenotype is calculated while controlling for several covariates in the training set. These covariates include age, gender, mean framewise displacement (mean FD), scan sites, and mode-centered PDS (only for age 14). The r value with the associated p value for each edge is obtained using partial correlation, and a threshold of p=0.01 is used to select the edges. Positively or negatively correlated edges are regarded as positive or negative networks. Network strength is then calculated by summing the selected edges in the gPPI matrix for both positive and negative networks, as well as by subtracting the strength of the negative network from the strength of the positive network to obtain the combined network strength. (ii) Model building. Linear models are constructed between the network strength of the positive, negative, combined network, and behavioral phenotype in the training set. The network strength is then calculated for each participants in the testing set and input into the predictive model along with covariates to yield a predicted behavioral phenotype (e.g. predicted intra-individual coefficient of variation [ICV]) for each network. (iii) Model validation. The predictive performance is evaluated by calculating the correlation between predicted and observed values.

Three CV schemes

We used three CV schemes to test the robustness of predictive performance: k-fold (10-fold and 5-fold) and leave-site-out CV. For the k-fold CV, we randomly divided participants into 10 or 5 approximately equal-sized groups. For each fold, we trained the model on nine or four groups, respectively, and used it to predict the behavioral phenotype of the remaining group. We then assessed the predictive performance by comparing the predicted and observed values. For the leave-site-out CV, we divided participants into eight groups based on their scan site. To account for the random splits of the k-fold CV, we repeated the process 50 times and calculated the average predictive performance for both the 10-fold and 5-fold CV (Lichenstein et al., 2021). In addition, we set a 95% threshold for selecting edges present in at least 48 out of 50 iterations to visualize the results. We also ran the CPM analysis with mean FD thresholds of 0.2, 0.3, and 0.4 mm to account for the influence of head motion on the predictive performance. Furthermore, we conducted the CPM analysis using a range of thresholds for feature selection and observed similar results across different thresholds (see Appendix 1, Supplementary file 1h). The main text shows the results of the 10-fold CPM. The 5-fold CPM and leave-site-out CV results are shown in Appendix 1.

Prediction across timepoints and STRATIFY

To assess the ability of models developed at one timepoint to predict ICV at different timepoints, we applied predictive models developed at ages 14 and 19 to predict ICV at subsequent timepoints. Specifically, we used predictive models (including the parameters and selected edges) developed at age 14 to predict ICV at ages 19 and 23. We first calculated the network strength using the gPPI matrix at age 19 or 23 based on the selected edges identified from CPM analysis at age 14. We then used the linear model parameters (slope and intercept) from CPM analysis at age 14 to fit the network strength and predict ICV at age 19 or 23. Finally, we evaluated the predictive performance by calculating the correlation between the predicted and observed values at age 19 or 23. Similarly, we applied models developed at age 19 to predict ICV at age 23. In addition, we examined the generalizability of predictive models at age 23 by applying them to the STRATIFY dataset, which also includes participants who were 23 years of age. Furthermore, we estimated the predictive performances of ICV across patient groups in the STRATIFY. The correlation between the residual network strength of predictive networks and ICV was calculated across groups in the STRATIFY. The covariates, including age, sex, and mean FD, were regressed for network strength before the correlation analysis. It is worth noting that when applying models developed at one timepoint to predict at another timepoint or to generalize to a different dataset, the model was built using all participants from the timepoint at which the model was developed.

Statistical analysisExploratory factor analysis

To explore the underlying structure of adolescent substance use, we performed an exploratory factor analysis using principal component extraction (Gaskin and Happell, 2014) on TLFB using Predictive Analytics Software (SPSS) version 20. Factor analysis explores the underlying structure of a set of observed variables without imposing a preconceived structure on the outcome. We used six items at age 14 and nine items at ages 19 and 23 of TLFB, including alcohol, tobacco, cannabis, cocaine, ecstasy, and ketamine (as shown in Supplementary file 1k). We excluded items assessing the use of other drugs due to high proportions of missing data, standard deviations close to 0, or a Kaiser-Meyer-Olkin (KMO) statistic for individual variables below 0.5, considered the minimum value for a sample to be adequate. The KMO measure of sampling adequacy was 0.66 at age 14, 0.81 at age 19, and 0.77 at age 23. In addition, all Bartlett’s tests of sphericity were significant (age 14: χ2(15)=5137.067, p<0.001; age 19: χ2(36)=5031.641, p<0.001; age 23: χ2(36)=5106.265, p<0.001), indicating that there was an underlying correlation structure, and that factor analysis was appropriate. We rotated the factors using the varimax method with kaiser normalization to make it easier to discern the underlying measured constructs.

Linear mixed model

We constructed a linear mixed model to examine the change in ICV over time using the lme4 and lmerTest packages in RStudio (version: 1.4; http://www.rstudio.com/) and R (version 4.1.1; https://www.r-project.org/). The timepoint was the fixed effect of interest in the model, while the participants was a random effect. Several covariates, including sex, scan sites, mode-center PDS, and age at 14, were also included as fixed effects in the models. The linear mixed model is shown as follows:ICV∼Timepoint+Sex+Scan site+Mode_center PDS+Age at 14+(1∣Participant)

Correlation between network strength and substance use

To examine the relationship between ICV/brain activity and substance use, we correlated the network strength of predictive networks with the factor scores of substance use at each timepoint and across all three timepoints separately. To control for potential confounders, we calculated residual network strength and residual factor scores by regressing the effects of age, sex, scan sites, mean FD (for network strength), and mode-centered PDS (for age 14). We used Spearman correlation to assess the association between residual network strength and residual TLFB, as their distributions did not follow a normal distribution. We used an FDR correction (q<0.05) for the multiple correlations.

Furthermore, we employed a three-wave bivariate latent change score model using the lavvan package in R and RStudio to detect the linear change over time. This model allows us to quantify the longitudinal bidirectional influence between substance use and ICV over time (Nweze et al., 2023). Specifically, it facilitated an understanding of whether substance use predicted ICV and its brain activity, and vice versa. The key feature of this model is its ability to assess linear increases or decreases within the same construct across two adjacent waves. Change scores were calculated by regressing the observable score at a given timepoint from the previous timepoint (e.g. ΔCig+CB in T1–T2 or ΔCig+CB in T2–T3, where T1=timepoint 1, T2=timepoint 2, and T3=timepoint 3). Additionally, cross-lagged dynamic coupling (i.e. bidirectionality) was employed to explore individual differences in the relationships between substance use and linear changes in ICV/brain activity, as well as the relationship between ICV/brain activity and linear change in substance use. The model accounted for covariates such as age, sex, and scan sites. For more details about the latent change score model, refer to the reference Nweze et al., 2023.

As Figure 6 shows, the latent change score model was specifically applied to examine the association between substance use and behaviors and brain activity associated with sustained attention. We focused on the relationship between the network strength of positive and negative networks, derived from Go and Successful stop trials, and two types of substance use (Cig+CB and alcohol use). Notably, drug use data were excluded as adolescents at age 14 have no drug score. A total of 10 models were performed, and all model fit indices met the predefined criteria: CFI>0.92, RMSEA<0.05, and SRMR<0.03. An FDR correction (q<0.05) was applied for multiple correlations. It is worth noting that all the correlations between substance use and sustained attention were conducted using the same sample across three timepoints.

Permutation test

For the CPM analysis, we used a permutation test to assess the significance of the predictive performance, which is the correlation between the observed and predicted values. To generate a null distribution of these correlation values, we randomly shuffled the correspondence of the behavioral data and the PPI matrix of all participants and reran the CPM pipeline with the shuffled data 1000 times. Based on this distribution, we set a threshold of p<0.05 to determine the significance level at 95% for the predictive performance using 10-fold, 5-fold, and leave-site-out CV.

To estimate the significance of the predictive performance across timepoints and the external validation in the STRATIFY dataset, we shuffled the predictive values 1000 times. Then, we correlated the shuffled values with observed values to yield a null distribution of predictive correlation values. We also set a threshold of p<0.05 to determine the significance level at 95% for the predictive performance across timepoints and generalization in STRATIFY.

Additional informationCompeting interests

No competing interests declared

Served in an advisory or consultancy role for Actelion, Hexal Pharma, Lilly, Lundbeck, Medice, Novartis, Shire. He received conference support or speaker's fee by Lilly, Medice Novartis and Shire. Has been involved in clinical trials conducted by Shire & Viforpharma. Received royalities from Hogrefe, Kohlhammer, CIP Medien, Oxford University Press. The present work is unrelated to the above grants and relationships

Received a speaker honorarium from Servier (2014)

Author contributions

Conceptualization, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing – original draft, Writing – review and editing

Conceptualization, Methodology, Writing – review and editing

Conceptualization

Conceptualization

Methodology

Methodology

Methodology

Methodology, Writing – review and editing

Methodology, Writing – review and editing

Writing – review and editing

Writing – review and editing

Investigation

Investigation

Investigation

Investigation

Investigation

Investigation

Investigation

Investigation

Investigation

Investigation

Investigation

Investigation

Writing – review and editing

Investigation

Investigation

Investigation

Investigation

Investigation

Investigation

Investigation

Investigation

Investigation

Investigation

Conceptualization, Resources, Data curation, Supervision, Funding acquisition, Validation, Investigation, Methodology, Writing – original draft, Project administration, Writing – review and editing

Data curation, Project administration, Resources, Supervision

Ethics

Human subjects: Written and informed consent was obtained from all participants by the IMAGEN consortium and the study was approved by the institutional ethics committee of King's College London (PNM/10/11-126), University of Nottingham (D/11/2007), Trinity College Dublin (SPREC092007-01), Technische Universitat Dresden (EK 235092007), Commissariat a l'Energie Atomique et aux Energies Alternatives, INSERM (2007-A00778-45), University Medical Center at the University of Hamburg (M-191/07) and in Germany at medical ethics committee of the University of Heidelberg (2007-024N-MA) in accordance with the Declaration of Helsinki.

Additional filesParticipants' demographic information.Data availability

IMAGEN data are available from a dedicated database: https://imagen2.cea.fr. Due to participant consent restrictions, IMAGEN data cannot be made fully open access. Code for CPM analysis is available at https://osf.io/6ejpd/. Custom code that supports the findings of this study is available at https://github.com/YiheWeng/Weng_eLife_2024_scripts (copy archived at Weng, 2024). All data needed to evaluate the conclusions in the paper are present in the paper and/or Appendix 1.

The following previously published dataset was used:

BoyleRWengYWhelanR20234.4 Studying the connectome at a large scaleOpen Science Framework6ejpd

ReferencesAgrawalASilbergJLLynskeyMTMaesHHEavesLJ2010Mechanisms underlying the lifetime co-occurrence of tobacco and cannabis use in adolescent and young adult twinsDrug and Alcohol Dependence108495510.1016/j.drugalcdep.2009.11.016AgrawalABudneyAJLynskeyMT2012The co‐occurring use and misuse of cannabis and tobacco: A reviewAddiction1071221123310.1111/j.1360-0443.2012.03837.xBauerCCCRozenkrantzLCaballeroCNieto-CastanonASchererEWestMRMrazekMPhillipsDTGabrieliJDEWhitfield-GabrieliS2020Mindfulness training preserves sustained attention and resting state anticorrelation between default-mode network and dorsolateral prefrontal cortex: A randomized controlled trialHuman Brain Mapping415356536910.1002/hbm.2519732969562BloomfieldMAPAshokAHVolkowNDHowesOD2016The effects of Δ9-tetrahydrocannabinol on the dopamine systemNature53936937710.1038/nature20153BossongMGvan BerckelBNMBoellaardRZuurmanLSchuitRCWindhorstADvan GervenJMARamseyNFLammertsmaAAKahnRS2009Delta 9-tetrahydrocannabinol induces dopamine release in the human striatumNeuropsychopharmacology3475976610.1038/npp.2008.13818754005BroydSJvan HellHHBealeCYücelMSolowijN2016Acute and chronic effects of cannabinoids on human cognition-a systematic reviewBiological Psychiatry7955756710.1016/j.biopsych.2015.12.00226858214CaiWDubergKPadmanabhanARehertRBradleyTCarrionVMenonV2019Hyperdirect insula-basal-ganglia pathway and adult-like maturity of global brain responses predict inhibitory control in childrenNature Communications10479810.1038/s41467-019-12756-8CaseyKFBenkelfatCCherkasovaMVBakerGBDagherALeytonM2014Reduced dopamine response to amphetamine in subjects at ultra-high risk for addictionBiological Psychiatry76233010.1016/j.biopsych.2013.08.03324138922ChamberlainSROdlaugBLSchreiberLRNGrantJE2012Association between tobacco smoking and cognitive functioning in young adultsThe American Journal on Addictions21 Suppl 1S14S1910.1111/j.1521-0391.2012.00290.x23786505ChanraudSPitelA-LMüller-OehringEMPfefferbaumASullivanEV2013Remapping the brain to compensate for impairment in recovering alcoholicsCerebral Cortex239710410.1093/cercor/bhr38122275479ChenJTamAKebetsVOrbanCOoiLQRAsplundCLMarekSDosenbachNUFEickhoffSBBzdokDHolmesAJYeoBTT2022Shared and unique brain network features predict cognitive, personality, and mental health scores in the ABCD studyNature Communications13221710.1038/s41467-022-29766-8CrummyEAO’NealTJBaskinBMFergusonSM2020One is not enough: understanding and modeling polysubstance useFrontiers in Neuroscience1456910.3389/fnins.2020.0056932612502CwiekARajtmajerSMWybleBHonavarVGrossnerEHillaryFG2022Feeding the machine: Challenges to reproducible predictive modeling in resting-state connectomicsNetwork Neuroscience6294810.1162/netn_a_0021235350584D’AlbertoNChaaraniBOrrCASpechlerPAAlbaughMDAllgaierNWonnellABanaschewskiTBokdeALWBrombergUBüchelCQuinlanEBConrodPJDesrivièresSFlorHFröhnerJHFrouinVGowlandPHeinzAIttermanBMartinotJ-LPaillère MartinotM-LArtigesENeesFPapadopoulos OrfanosDPoustkaLRobbinsTWSmolkaMNWalterHWhelanRSchumannGPotterASGaravanH2018Individual differences in stop-related activity are inflated by the adaptive algorithm in the stop signal taskHuman Brain Mapping393263327610.1002/hbm.2407529656430DalleyJWFryerTDBrichardLRobinsonESJTheobaldDEHLääneKPeñaYMurphyERShahYProbstKAbakumovaIAigbirhioFIRichardsHKHongYBaronJ-CEverittBJRobbinsTW2007Nucleus accumbens D2/3 receptors predict trait impulsivity and cocaine reinforcementScience3151267127010.1126/science.113707317332411DhamalaEYeoBTTHolmesAJ2023One size does not fit all: Methodological considerations for brain-based predictive modeling in psychiatryBiological Psychiatry9371772810.1016/j.biopsych.2022.09.02436577634DiXBiswalBB2019Toward task connectomics: Examining whole-brain task modulated connectivity in different task domainsCerebral Cortex291572158310.1093/cercor/bhy05529931116DiXZhangZGBiswalBB2021Understanding psychophysiological interaction and its relations to beta series correlationBrain Imaging and Behavior1595897310.1007/s11682-020-00304-832710336DockreePMBarnesJJMatthewsNDeanAJAbeRNandamLSKellySPBellgroveMAO’ConnellRG2017The effects of methylphenidate on the neural signatures of sustained attentionBiological Psychiatry8268769410.1016/j.biopsych.2017.04.01628599833DongelmansMDurand-de CuttoliRNguyenCComeMDurantéEKLemoineDBritoRAhmed YahiaTMondoloniSDidienneSBousseyrolEHannesseBReynoldsLMTorquetNDalkaraDMartiFMourotANaudéJFaureP2021Chronic nicotine increases midbrain dopamine neuron activity and biases individual strategies towards reduced exploration in miceNature Communications12694510.1038/s41467-021-27268-734836948DoughertyDMMathiasCWDawesMAFurrRMCharlesNELiguoriAShannonEEAchesonA2013Impulsivity, attention, memory, and decision-making among adolescent marijuana usersPsychopharmacology22630731910.1007/s00213-012-2908-523138434EstermanMRothleinD2019Models of sustained attentionCurrent Opinion in Psychology2917418010.1016/j.copsyc.2019.03.00530986621FengQRenZWeiDLiuCWangXLiXTieBTangSQiuJ2024Connectome-based predictive modeling of Internet addiction symptomatologySocial Cognitive and Affective Neuroscience19nsae00710.1093/scan/nsae007FerlandJ-MNHurdYL2020Deconstructing the neurobiology of cannabis use disorderNature Neuroscience2360061010.1038/s41593-020-0611-032251385FigueiredoPRTolomeoSSteeleJDBaldacchinoA2020Neurocognitive consequences of chronic cannabis use: A systematic review and meta-analysisNeuroscience and Biobehavioral Reviews10835836910.1016/j.neubiorev.2019.10.01431715191FortenbaughFCDeGutisJGermineLWilmerJBGrossoMRussoKEstermanM2015Sustained Attention across the life span in a sample of 10,000: Dissociating ability and strategyPsychological Science261497151010.1177/095679761559489626253551FrayPJRobbinsTWSahakianBJ1996Neuorpsychiatyric applications of CANTABInternational Journal of Geriatric Psychiatry1132933610.1002/(SICI)1099-1166(199604)11:4<329::AID-GPS453>3.0.CO;2-6FröhnerJHTeckentrupVSmolkaMNKroemerNB2019Addressing the reliability fallacy in fMRI: Similar group effects may arise from unreliable individual effectsNeuroImage19517418910.1016/j.neuroimage.2019.03.05330930312GaskinCJHappellB2014On exploratory factor analysis: A review of recent evidence, an assessment of current practice, and recommendations for future useInternational Journal of Nursing Studies5151152110.1016/j.ijnurstu.2013.10.005GranonSPassettiFThomasKLDalleyJWEverittBJRobbinsTW2000Enhanced and impaired attentional performance after infusion of D1 dopaminergic receptor agents into rat prefrontal cortexThe Journal of Neuroscience201208121510.1523/JNEUROSCI.20-03-01208.200010648725GreeneASGaoSScheinostDConstableRT2018Task-induced brain state manipulation improves prediction of individual traitsNature Communications9280710.1038/s41467-018-04920-330022026GunnCMackusMGriffinCMunafòMRAdamsS2018A systematic review of the next-day effects of heavy alcohol consumption on cognitive performanceAddiction1132182219310.1111/add.1440430144191HarakehZde SonnevilleLvan den EijndenRJJMHuizinkACReijneveldSAOrmelJVerhulstFCMonshouwerKVolleberghWAM2012The association between neurocognitive functioning and smoking in adolescence: The TRAILS studyNeuropsychology2654155010.1037/a002921722775342HindochaCFreemanTPFerrisJALynskeyMTWinstockAR2016No smoke without tobacco: A global overview of cannabis and tobacco routes of administration and their association with intention to quitFrontiers in Psychiatry710410.3389/fpsyt.2016.0010427458388HindochaCBroseLSWalshHCheesemanH2021Cannabis use and co‐use in tobacco smokers and non‐smokers: Prevalence and associations with mental health in a cross‐sectional, nationally representative sample of adults in Great Britain, 2020Addiction1162209221910.1111/add.15381KaiserAHolzNEBanaschewskiTBaumeisterSBokdeALWDesrivièresSFlorHFröhnerJHGrigisAGaravanHGowlandPHeinzAIttermannBMartinotJLPaillère MartinotMLArtigesEMillenetSOrfanosDPPoustkaLSchwarzESmolkaMNWalterHWhelanRSchumannGBrandeisDNeesFBanaschewskiTBarkerGJBokdeALWDesrivièresSFlorHGrigisAGaravanHGowlandPHeinzABrühlRMartinotJLPaillère MartinotMLArtigesENeesFOrfanosDPLemaitreHPausTPoustkaLHohmannSMillenetSFröhnerJHRobinsonLSmolkaMNWalterHWintererJMWhelanRSchumannG2022A developmental perspective on facets of impulsivity and brain activity correlates from adolescence to adulthoodBiological Psychiatry71103111510.1016/j.bpsc.2022.02.003Klugah-BrownBDiXZweeringsJMathiakKBeckerBBiswalB2020Common and separable neural alterations in substance use disorders: A coordinate-based meta-analyses of functional neuroimaging studies in humansHuman Brain Mapping414459447710.1002/hbm.2508532964613KühnSLisofskyNBanaschewskiTBarkerGBokdeALWBrombergUBüchelCBrühlRQuinlanEBDesrivièresSFlorHGrigisAGaravanHGowlandPHeinzAIttermannBMartinotJ-LMartinotM-LPNeesFOrfanosDPPausTPoustkaLFröhnerJHSmolkaMNWalterHWhelanRSchumannGGallinatJIMAGEN Consortium2020Hierarchical associations of alcohol use disorder symptoms in late adolescence with markers during early adolescenceAddictive Behaviors10010613010.1016/j.addbeh.2019.10613031622946LiGChenYTangXLiC-SR2021Alcohol use severity and the neural correlates of the effects of sleep disturbance on sustained visual attentionJournal of Psychiatric Research14230231110.1016/j.jpsychires.2021.08.01834416549LichensteinSDScheinostDPotenzaMNCarrollKMYipSW2021Dissociable neural substrates of opioid and cocaine use identified via connectome-based modellingMolecular Psychiatry264383439310.1038/s41380-019-0586-y31719641LisdahlKMPriceJS2012Increased marijuana use and gender predict poorer cognitive functioning in adolescents and emerging adultsJournal of the International Neuropsychological Society1867868810.1017/S135561771200027622613255MarshallCABrodnikZDMortensenOVReithMEAShumskyJSWaterhouseBDEspañaRAKortagereS2019Selective activation of Dopamine D3 receptors and norepinephrine transporter blockade enhances sustained attentionNeuropharmacology14817818810.1016/j.neuropharm.2019.01.00330633928MorganDGrantKAGageHDMachRHKaplanJRPrioleauONaderSHBuchheimerNEhrenkauferRLNaderMA2002Social dominance in monkeys: dopamine D2 receptors and cocaine self-administrationNature Neuroscience516917410.1038/nn79811802171MyersNEStokesMGNobreAC2017Prioritizing information during working memory: Beyond sustained internal attentionTrends in Cognitive Sciences2144946110.1016/j.tics.2017.03.010NaderMAMorganDGageHDNaderSHCalhounTLBuchheimerNEhrenkauferRMachRH2006PET imaging of dopamine D2 receptors during chronic cocaine self-administration in monkeysNature Neuroscience91050105610.1038/nn173716829955NarayanAJAitkenBDowneyLAHayleyAC2021The effects of amphetamines alone and in combination with alcohol on functional neurocognition: A systematic reviewNeuroscience and Biobehavioral Reviews13186588110.1016/j.neubiorev.2021.10.00334626687NwezeTBanaschewskiTAjaeluCOkoyeCEzenwaMWhelanRPapadopoulos OrfanosDBokdeALWDesrivièresSGrigisAGaravanHGowlandPHeinzABrühlRMartinotJMartinotMPArtigesENeesFPausTPoustkaLHohmannSMillenetSFröhnerJHSmolkaMNWalterHSchumannGHansonJLIMAGEN Consortium2023Trajectories of cortical structures associated with stress across adolescence: a bivariate latent change score approachJournal of Child Psychology and Psychiatry641159117510.1111/jcpp.13793O’HalloranLCaoZRuddyKJollansLAlbaughMDAleniAPotterASVaheyNBanaschewskiTHohmannSBokdeALWBrombergUBüchelCQuinlanEBDesrivièresSFlorHFrouinVGowlandPHeinzAIttermannBNeesFOrfanosDPPausTSmolkaMNWalterHSchumannGGaravanHKellyCWhelanR2018Neural circuitry underlying sustained attention in healthy adolescents and in ADHD symptomatologyNeuroImage16939540610.1016/j.neuroimage.2017.12.03029274748PausT2005Mapping brain maturation and cognitive development during adolescenceTrends in Cognitive Sciences9606810.1016/j.tics.2004.12.008PetersenACCrockettLRichardsMBoxerA1988A self-report measure of pubertal status: Reliability, validity, and initial normsJournal of Youth and Adolescence1711713310.1007/BF0153796224277579PinarAHawiZCumminsTJohnsonBPauperMTongJTiegoJFinlayAKleinMFrankeBFornitoABellgroveMA2018Genome-wide association study reveals novel genetic locus associated with intra-individual variability in response timeTranslational Psychiatry820710.1038/s41398-018-0262-z30287865RenZDakerRJShiLSunJBeatyREWuXChenQYangWLyonsIMGreenAEQiuJ2021Connectome-based predictive modeling of creativity anxietyNeuroImage22511746910.1016/j.neuroimage.2020.11746933099006RitchayMMHugginsAAWallaceALLarsonCLLisdahlKM2021Resting state functional connectivity in the default mode network: Relationships between cannabis use, gender, and cognition in adolescents and young adultsNeuroImage. Clinical3010266410.1016/j.nicl.2021.10266433872994RosenbergMDFinnESScheinostDPapademetrisXShenXConstableRTChunMM2016A neuromarker of sustained attention from whole-brain functional connectivityNature Neuroscience1916517110.1038/nn.417926595653RosenbergMDHsuWTScheinostDTodd ConstableRChunMM2018Connectome-based models predict separable components of attention in novel individualsJournal of Cognitive Neuroscience3016017310.1162/jocn_a_0119729040013RosenbergMDScheinostDGreeneASAveryEWKwonYHFinnESRamaniRQiuMConstableRTChunMM2020Functional connectivity predicts changes in attention observed across minutes, days, and monthsPNAS1173797380710.1073/pnas.191222611732019892ShannonEEMathiasCWDoughertyDMLiguoriA2010Cognitive impairments in adolescent cannabis users are related to THC levelsAddictive Disorders & Their Treatment915816310.1097/ADT.0b013e3181c8c667ShenXTokogluFPapademetrisXConstableRT2013Groupwise whole-brain parcellation from resting-state fMRI data for network node identificationNeuroImage8240341510.1016/j.neuroimage.2013.05.08123747961ShenXFinnESScheinostDRosenbergMDChunMMPapademetrisXConstableRT2017Using connectome-based predictive modeling to predict individual behavior from brain connectivityNature Protocols1250651810.1038/nprot.2016.17828182017SiskLMRapuanoKMConleyMIGreeneASHorienCRosenbergMDScheinostDConstableRTGlattCECaseyBJGeeDG2022Genetic variation in endocannabinoid signaling is associated with differential network-level functional connectivity in youthJournal of Neuroscience Research10073174310.1002/jnr.2494634496065SobellLCBrownJLeoGISobellMB1996The reliability of the Alcohol Timeline Followback when administered by telephone and by computerDrug and Alcohol Dependence42495410.1016/0376-8716(96)01263-x8889403SpearLP2018Effects of adolescent alcohol consumption on the brain and behaviourNature Reviews. Neuroscience1919721410.1038/nrn.2018.1029467469SquegliaLMJacobusJTapertSF2009The influence of substance use on adolescent brain developmentClinical EEG and Neuroscience40313810.1177/15500594090400011019278130SripadaCRutherfordSAngstadtMThompsonWKLucianaMWeigardAHydeLHHeitzegM2020Prediction of neurocognition in youth from resting state fMRIMolecular Psychiatry253413342110.1038/s41380-019-0481-631427753TapertSFSchweinsburgADBarlettVCBrownSAFrankLRBrownGGMeloyMJ2004Blood oxygen level dependent response and spatial working memory in adolescents with alcohol use disordersAlcoholism, Clinical and Experimental Research281577158610.1097/01.alc.0000141812.81234.a615597092ThieleABellgroveMA2018Neuromodulation of attentionNeuron9776978510.1016/j.neuron.2018.01.00829470969ThomsonPMalpasCBVijayakumarNJohnsonKAAndersonVEfronDHazellPSilkTJ2022Longitudinal maturation of resting state networks: Relevance to sustained attention and attention deficit/hyperactivity disorderCognitive, Affective & Behavioral Neuroscience221432144610.3758/s13415-022-01017-935676491TreurJLWillemsenGBartelsMGeelsLMvan BeekJHDAHuppertzCvan BeijsterveldtCEMBoomsmaDIVinkJM2015Smoking during adolescence as a risk factor for attention problemsBiological Psychiatry7865666310.1016/j.biopsych.2014.06.01925092631TrifilieffPMartinezD2014Blunted dopamine release as a biomarker for vulnerability for substance use disordersBiological Psychiatry764510.1016/j.biopsych.2014.04.01724925889TrifilieffPDucrocqFvan der VeldtSMartinezD2017Blunted dopamine transmission in addiction: Potential mechanisms and implications for behaviorSeminars in Nuclear Medicine47647410.1053/j.semnuclmed.2016.09.00327987559UenoFMatsushitaSHaraSOshimaSRohSRamchandaniVAMimuraMUchidaH2022Influence of alcohol and acetaldehyde on cognitive function: Findings from an alcohol clamp study in healthy young adultsAddiction11793494510.1111/add.15733VerbruggenFAronARBandGPBesteCBissettPGBrockettATBrownJWChamberlainSRChambersCDColoniusHColzatoLSCorneilBDCoxonJPDupuisAEagleDMGaravanHGreenhouseIHeathcoteAHusterRJJahfariSKenemansJLLeunissenILiC-SRLoganGDMatzkeDMorein-ZamirSMurthyAParéMPoldrackRARidderinkhofKRRobbinsTWRoeschMRubiaKSchacharRJSchallJDStockA-KSwannNCThakkarKNvan der MolenMWVermeylenLVinkMWesselJRWhelanRZandbeltBBBoehlerCN2019A consensus guide to capturing the ability to inhibit actions and impulsive behaviors in the stop-signal taskeLife8e4632310.7554/eLife.4632331033438VolkowNDWangGJBegleiterHPorjeszBFowlerJSTelangFWongCMaYLoganJGoldsteinRAlexoffDThanosPK2006High Levels of Dopamine D2 receptors in unaffected members of alcoholic familiesArchives of General Psychiatry6399910.1001/archpsyc.63.9.999VosselSWarbrickTMobascherAWintererGFinkGR2011Spatial and sustained attention in relation to smoking status: behavioural performance and brain activation patternsJournal of Psychopharmacology251485149510.1177/026988111039183021300633WallaceALWadeNEHatcherKFLisdahlKM2019Effects of cannabis use and subclinical ADHD symptomology on attention based tasks in adolescents and young adultsArchives of Clinical Neuropsychology3470070510.1093/arclin/acy080WeinbergerAHPacekLRWallMMZvolenskyMJCopelandJGaleaSNahviSMoellerSJHasinDSGoodwinRD2018Trends in cannabis use disorder by cigarette smoking status in the United States, 2002-2016Drug and Alcohol Dependence191455110.1016/j.drugalcdep.2018.06.01630077055WengY2024Weng_eLife_2024_scriptsswh:1:rev:8af8f2e64e72df3e860b4aa23501d445d49f21b7Software Heritagehttps://archive.softwareheritage.org/swh:1:dir:50af38d7961d7951acf7f8080c48d054df968dc1;origin=https://github.com/YiheWeng/Weng_eLife_2024_scripts;visit=swh:1:snp:2392e9b416e787fbeea71512d12948d99ed66f3a;anchor=swh:1:rev:8af8f2e64e72df3e860b4aa23501d445d49f21b7WetherillRRJagannathanKHagerNChildressARRaoHFranklinTR2015Cannabis, cigarettes, and their co-occurring use: Disentangling differences in gray matter volumeThe International Journal of Neuropsychopharmacology18yv06110.1093/ijnp/pyv06126045474WhelanRConrodPJPolineJ-BLourdusamyABanaschewskiTBarkerGJBellgroveMABüchelCByrneMCumminsTDRFauth-BühlerMFlorHGallinatJHeinzAIttermannBMannKMartinotJ-LLalorECLathropMLothENeesFPausTRietschelMSmolkaMNSpanagelRStephensDNStruveMThyreauBVollstaedt-KleinSRobbinsTWSchumannGGaravanHthe IMAGEN Consortium2012Adolescent impulsivity phenotypes characterized by distinct brain networksNature Neuroscience1592092510.1038/nn.3092WhelanRWattsROrrCAAlthoffRRArtigesEBanaschewskiTBarkerGJBokdeALWBüchelCCarvalhoFMConrodPJFlorHFauth-BühlerMFrouinVGallinatJGanGGowlandPHeinzAIttermannBLawrenceCMannKMartinotJ-LNeesFOrtizNPaillère-MartinotM-LPausTPausovaZRietschelMRobbinsTWSmolkaMNStröhleASchumannGGaravanHIMAGEN Consortium2014Neuropsychosocial profiles of current and future adolescent alcohol misusersNature51218518910.1038/nature1340225043041WilliamsBRHultschDFStraussEHHunterMATannockR2005Inconsistency in reaction time across the life spanNeuropsychology19889610.1037/0894-4105.19.1.8815656766YadavPHutchisonCABasnayakeKStringerSJeskyMFiferLSnellKPinneyJDraysonMTCookMCockwellP2016Patients with multiple myeloma have excellent long-term outcomes after recovery from dialysis-dependent acute kidney injuryEuropean Journal of Haematology9661061710.1111/ejh.1264426248588YakirARigbiAKanyasKPollakYKahanaGKarniOEitanRKertzmanSLererB2007Why do young women smoke? III. Attention and impulsivity as neurocognitive predisposing factorsEuropean Neuropsychopharmacology1733935110.1016/j.euroneuro.2006.09.00417141485YipSWScheinostDPotenzaMNCarrollKM2019Connectome-based prediction of cocaine abstinenceThe American Journal of Psychiatry17615616410.1176/appi.ajp.2018.1710114730606049YooKRosenbergMDHsuW-TZhangSLiC-SRScheinostDConstableRTChunMM2018Connectome-based predictive modeling of attention: Comparing different functional connectivity features and prediction methods across datasetsNeuroImage167112210.1016/j.neuroimage.2017.11.01029122720ZehraALindgrenEWiersCEFreemanCMillerGRamirezVShokri-KojoriEWangG-JTalagalaLTomasiDVolkowND2019Neural correlates of visual attention in alcohol use disorderDrug and Alcohol Dependence19443043710.1016/j.drugalcdep.2018.10.03230502544ZhaoWPalmerCEThompsonWKChaaraniBGaravanHPCaseyBJJerniganTLDaleAMFanCC2021Individual differences in cognitive performance are better predicted by global rather than localized BOLD activity patterns across the cortexCerebral Cortex311478148810.1093/cercor/bhaa29033145600ZhaoWMakowskiCHaglerDJGaravanHPThompsonWKGreeneDJJerniganTLDaleAM2023Task fMRI paradigms may capture more behaviorally relevant information than resting-state functional connectivityNeuroImage27011994610.1016/j.neuroimage.2023.11994636801369Appendix 1MethodCambridge Neuropsychological Test Automated Battery

Several CANTAB tasks were used to examine cognitive abilities: the affective go/no-go task (AGN); the RVP task, the SWM task, and the Cambridge guessing task (CGT). Only the participants at age 23 completed the CGT task. The RVP measures sustained attention. The AGN measures inhibitory control in the context of emotionally salient information, the CGT assesses impulsivity, the SWM assesses working memory. Detailed information on CANTAB was described in Kühn et al., 2020.

To examine the sustained attention network specificity, we correlated the network strength of predictive networks predicting ICV with CANTAB task at each timepoint separately. To control for potential confounders, we calculated residual network strength and residual performances of three CANTAB tasks by regressing the effects of age, sex, scan sites, mean FD for network strength and mode-centered PDS for age 14. Finally, we used Spearman correlation to assess the association between residual sustained attention network strength and CANTAB performances.

To examine the specificity of sustained attention at baseline in influencing substance use, a two-sample t-test was performed to detect the significant difference in cigarette and cannabis use between high and lower cognition groups at baseline. Participants were categorized into five groups based on the ICV, network strength of positive and negative networks at age 14. The top of participants with the highest ICV/network strength of positive network, or lowest network strength of negative strength comprised the low sustained attention group, while the bottom of participants with lowest ICV/network strength of positive network, or highest network strength of negative strength constituted the high sustained attention group. Cig+CB were then compared between the higher and lower sustained attention groups at each timepoint. We found the significant coupling effect between Cig+CB and network strength derived from Go trials, instead of Successful stop trials. Here, we only tested the difference in Cig+CB using positive and negative network derived from Go trials. We performed the similar analysis by stratifying the participants into higher strategy working memory between error (SWM_BE) group and lower SWM_BE group according to the between error value from the SWM task.

Generalization in subgroups from STRATIFY

We tested if the predictive networks defined at age 23 in IMAGEN would generalize to distinct patient groups in STRATIFY. STRATIFY includes several subgroups of individuals aged 23 with alcohol use disorder (AUD), major depression disorder (MDD), bulimia nervosa (BN), anorexia nervosa (AN), and 19 healthy controls.

Dice coefficient

We calculated the Dice coefficient (DC) to quantify the similarity of predictive networks across the three timepoints. A permutation test was also performed to estimate the predictive network similarity’s significance. First, we shuffled the ICV at each timepoint and performed feature selection based on random behavioral phenotypes to yield random predictive networks, including positive and negative. Then we calculated DC from predictive networks between each pair of timepoint. These steps were iterated 1000 times to generate a null distribution of DC values. Finally, we set a threshold of p<0.05 to determine the significance level at 95% for the similarity of the predictive networks between each timepoint.

Comparison of predictive networks identified at one timepoint versus another

Steiger’s Z value was employed to compare predictive performances of networks identified at different timepoints. This analysis involved comparing the R values derived from networks defined at distinct ages to predict ICV at the same age. For example, we compared the r values of brain networks defined at age 14 when predicting ICV at 19 (i.e. positive network: r=0.25, negative network: r=0.25, combined network: r=0.28) with those R values of brain networks defined at age 19 itself (i.e. positive network: r=0.16, negative network: r=0.14, combined network: r=0.16) derived from Go trials using Steiger’s Z test (age 14 → age 19 vs. age 19 → 19). Similarly, comparisons were made between networks defined at age 14 predicting ICV at age 23 and those at age 23 predicting ICV at age 23 (age 14 → age 23 vs. age 23 → 23), as well as between networks defined at age 19 predicting ICV at age 23 and those at age 23 predicting ICV at age 23 (age 19 → age 23 vs. age 23 → age 23). These comparisons were performed separately for Go trials and Successful stop trials.

CPM analysis using Failed stop trials

We performed another CPM analysis using Failed stop trials using gPPI matrix obtained from the second GLM, described in the main text. The CPM analysis was conducted using 10-fold CV, 5-fold CV, and leave-site-out CV.

Prediction across timepoints controlling for ICV at age 14

To examine whether connectivity predictors shared variations of sustained attention across timepoints, we applied predictive models developed at ages 14 and 19 to predict ICV at subsequent timepoints controlling for ICV at age 14. Specifically, we used predictive models (including parameters and selected edges) developed at age 14 to predict ICV at ages 19 and 23 separately. First, we calculated the network strength using the gPPI matrix at ages 19 and 23 based on the selected edges identified from CPM analysis at age 14. We then estimated the predicted ICV at ages 19 and 23 by applying the linear model parameters (slope and intercept) obtained from CPM analysis at age 14 to the network strength. Finally, we evaluated the predictive performance by calculating the partial correlation between the predicted and observed values at ages 19 and 23, controlling for ICV at age 14. Similarly, we applied models developed at age 19 to predict ICV at age 23, also controlling for ICV at age 14. To assess the significance of the predictive performance, we used a permutation test, shuffling the predicted ICV values and calculating partial correlation to a general random distribution over 1000 iterations.

ResultsSpecificity of sustained attention network

Sustained attention network strength derived from Go trials and Successful stop trials was significantly correlated with the accuracy of the RVP task (all p<0.05) for both negative and positive networks at ages 14 and 19 but not with AGN task performance (all p>0.05) (Supplementary file 1l). These results suggest that the networks derived from Go trails and Successful stop trials are specific to sustained attention.

No significant difference in Cig+CB was found between high and low sustained attention groups (obtained from both behavior level and brain activity) at age 14 (all p>0.462). Higher Cig+CB use was found in low sustained attention group compared to high sustained attention group at age 19 (all p<0.021) and age 23 (all p<0.007) (Supplementary file 1v). In addition, no significant difference in Cig+CB was found at ages 14 (t=0.11, p=0.912), 19 (t=1.65, p=0.10), and 23 (t=1.43, p=0.154) between higher and lower SWM_BE groups (Supplementary file 1w).

CPM predictive performance derived from Failed stop trials

Positive, negative, and combined networks derived from Failed stop trials significantly predicted ICV: at age 14 (r=0.10, p=0.033; r=0.19, p<0.001; and r=0.17, p<0.001, respectively), at age 19 (r=0.21, r=0.18, and r=0.21, all p<0.001, respectively), and at age 23 (r=0.33, r=0.35, and r=0.36, respectively, all p<0.001). We obtained similar results using a 5-fold CV and leave-site-out CV (Supplementary file 1f).

Predictive network similarity

With respect to Go trials, the mean DC values for the positive and negative networks across all three timepoints were 0.06 and 0.03, respectively, for ICV. With respect to Successful stop trials, the mean DC values for positive and negative networks predicting ICV across all three timepoints were 0.01 and 0.01.

Positive and negative networks predicting ICV derived from Go trials were significantly similar between ages 14 and 19 (DC = 0.03, p=0.001 and DC = 0.03, p<0.001), and between ages 19 and 23 (DC = 0.06 and DC = 0.04, respectively, all p<0.001) (FDR correction, 0.05). The positive network predicting ICV derived from Go trials was significantly similar between ages 14 and 23 (DC = 0.07, p<0.001) (FDR correction, 0.05). The mean DC of the positive and negative networks predicting ICV across three timepoints derived from Go trials are 0.06 and 0.03, respectively.

The negative networks predicting ICV derived from Successful stop trials were significantly similar between ages 14 and 19 (DC = 0.03, p=0.001) (FDR correction, q<0.05). The mean DC of the positive and negative networks predicting ICV derived from Successful stop trials was 0.01 and 0.01. Detailed results about DC between each pair of timepoints can be found in Supplementary file 1m.

Generalization in subgroups in STRATIFY

We examined generalization to separate patient cohorts in STRATIFY. Brain networks predicting ICV derived from Go trials defined at age 23 generalized to almost all patient cohorts, including AUD, MDD, BN, and AN (all p<0.05). The prediction for the healthy controls was moderately accurate (r~0.4), although this was not statistically significant due to the small sample size (n=19) for Go trials (Figure 4—figure supplement 1A, left panel). However, brain networks predicting ICV derived from Successful stop trials failed to predict ICV in individuals with AUD (p>0.05), although they generalized to other patient groups (Figure 4—figure supplement 1A, right panel). Furthermore, the correlations between sustained attention network strength of positive, negative, and combined networks derived from Successful stop trials and ICV in the groups with AUD were in the opposite direction compared with all other groups (Figure 4—figure supplement 1B).

Comparison of predictive performance at different timepoints

Steiger’s Z value was used to test if the difference in R values obtained using predictive networks at one timepoint versus another. For positive, negative, and combined networks predicting ICV derived from Go trials at age 19, the R values were higher when using predictive networks defined at 19 than those defined at 14 (Z=3.79, Z=3.39, Z=3.99, all p<0.00071). Similarly, the R values for positive, negative, and combined networks predicting ICV derived from Go trials at age 23 were higher when using predictive networks defined at age 23 compared to those defined at ages 14 (Z=6.00, Z=5.96, Z=6.67, all p<3.47e–9) or 19 (Z=2.80, Z=2.36, Z=2.57, all p<0.005).

At age 19, the R value for the positive network predicting ICV derived from Successful stop trials was higher when using predictive networks defined at 19 compared to those defined at 14 (Z=1.54, p=0.022), while the negative and combined networks did not show a significant difference (Z=0.85, p=0.398; Z=2.29, p=0.123). At age 23, R values for the positive and combined networks predicting ICV derived from Successful stop trials were higher when using predictive networks defined at 23 compared to those defined at 14 (Z=3.00, Z=2.48, all p<3.47e–9) or 19 (Z=2.52, Z=1.99, all p<0.005). However, the R value for the negative network at age 23 did not significantly differ when using predictive networks defined at 14 (Z=1.80, p=0.072) or 19 (Z=1.48, p=0.138).

Correlation between drug use and behavior and brain activity

ICV negatively correlated with drug use at age 19 (Rho = –0.11, p=0.001) (Supplementary file 1n). ICV at age 23 negatively correlated with drug use at age 19 (Rho = –0.08, p=0.014) (Supplementary file 1q). Sustained attention network strength derived from Successful stop trials significantly correlated with drug use at age 19 for the positive network (Rho = –0.12, p<0.001; FDR correction, 0.05) (Supplementary file 1r). Sustained attention network strength derived from Successful stop trials at age 23 correlated with drug use at age 19 (positive network: Rho = –0.09, p=0.005; negative network: Rho = 0.09, p=0.007) (Supplementary file 1u).

Predictions across timepoints controlling for ICV at age 14

Positive and combined networks derived from Go trials defined at age 14 predicted ICV at ages 19 (r=0.10, p=0.028; r=0.08, p=0.047) but negative network did not (r=0.06, p=0.119). Positive network derived from Go trials defined at age 14 predicted ICV at age 23 (r=0.11, p=0.013) but negative and combined networks did not (r=0.04, p=0.187; r=0.08, p=0.056). Positive, negative, and combined networks derived from Go trials defined at age 19 predicted ICV at age 23 (r=0.22, r=0.19, and r=0.22, respectively, all p<0.001).

Positive, negative, and combined networks derived from Successful stop trials defined at age 14 predicted ICV at ages 19 (r=0.08, p=0.036; r=0.10, p=0.012; r=0.11, p=0.009) and 23 (r=0.11, p=0.005; r=0.13, p=0.005; r=0.13, p=0.017) respectively. Positive, negative, and combined networks derived from Successful stop trials defined at age 19 predicted ICV at age 23 (r=0.18, r=0.18, and r=0.17, respectively, all p<0.001).

DiscussionSustained attention networks functioning as global activation

Our findings strongly indicate that sustained attention relies on global brain activation (i.e. network strength) rather than specific regions or networks (see also Zhao et al., 2021). We observed brain networks associated with high or low sustained attention span in large-scale networks across the cortex, subcortex, and cerebellum across adolescence to adulthood (see Figure 2—figure supplement 1, Figure 3—figure supplement 1, consistent with Rosenberg et al., 2020), instead of being confined to a few key regions. In our study, however, although the edges in the sustained attention networks were significantly similar from ages 14 to 23 (Supplementary file 1m), there were relatively few overlapping edges in the predictive networks over time. It is worth noting that DC values depend heavily on the significance threshold applied to the data (Fröhner et al., 2019). However, sustained attention network patterns identified could efficiently predict sustained attention for the subsequent timepoints. A prior study (Cai et al., 2019) has shown that children aged 9–12 could recruit key nodes (e.g. rIFG and rMFG), eliciting an adult-like global activation pattern that predicted their inhibitory control abilities. Similarly, our findings suggest that adolescents likely exhibit adult-like global activation patterns predicting sustained attention.

Aberrant sustained attention network in AUD

A notable exception was the failure of the network derived from Successful stop trials to generalize to patients with AUD, requiring higher attention levels. We speculate that activity in the sustained attention network of individuals with AUD might be similar to healthy adults during low cognitive demands but abnormal compared to healthy adults when faced with higher cognitive demands. Evidence from past literature shows that alcohol misuse is associated with attention deficits and dysfunctional neural mechanisms (Gunn et al., 2018; Li et al., 2021; Narayan et al., 2021; Spear, 2018). Furthermore, lower activation of parietal and prefrontal cortices has been observed in abstinent patients with AUD compared with healthy controls during visual attention tasks (Zehra et al., 2019), suggesting that differences in the attention network of individuals with AUD might underlie attention deficits in AUD.

Although the sustained attention network derived from Successful stop trials seen in healthy controls failed to generalize to AUD, we observed no significant difference in behavioral measures – ICV – between healthy controls and those with AUD. Our results may reflect compensatory mechanisms in AUD that allow these individuals to complete sustained attention tasks, which is consistent with prior studies (Tapert et al., 2004; Zehra et al., 2019). Compensation manifests as abnormal brain activity while performing normally on the task (Chanraud et al., 2013). Zehra et al., 2019, found brain activation differences during attention tasks despite no differences in behavioral performance between the abstinent patients with AUD and healthy controls. Previous studies (Chanraud et al., 2013; Squeglia et al., 2009) pointed out that individuals with AUD might exhibit subtle neural reorganization or compensation to preserve normal cognitive abilities. Tapert et al., 2004, found that heavy and light drinkers had similar behavioral performance on a working memory task. Activation differences were found in the parietal and occipital lobes and the cerebellar, indicating subtle neuronal reorganization may occur in AUD. Similarly, a study found that individuals with AUD maintained standard working memory by recruiting other cerebellar-based functional networks to complete the task (Chanraud et al., 2013). Successful completion of working memory tasks requires sustained attention (Myers et al., 2017). These studies suggest that the failure of the sustained attention network derived from Successful stop trials to generalize AUD in the current study may be due to compensatory mechanisms employed by individuals with AUD while completing tasks requiring high-level sustained attention.

Specificity of the prediction of predictive networks

We found that task-related function connectivity derived from Go trials, Successful stop trials, and Failed stop trials successfully predicted sustained attention across three timepoints. However, predictive performances of predictive networks derived from Go trials were higher than those derived from Successful stop trials and Failed stop trials. These results suggest that sustained attention is particularly crucial during Go trials when participants need to respond to the Go signal. In contrast, although Successful Stop and Failed Stop trials also require sustained attention, these tasks primarily involve inhibitory control along with sustained attention.

10.7554/eLife.97150.3.sa0eLife assessmentZhangXilinReviewing EditorSouth China Normal UniversityChinaConvincingImportant

This study presents an important finding on the relationship between brain activity related to sustained attention and substance use in adolescence/early adulthood with a large longitudinal dataset. The evidence supporting the claims of the authors is convincing. The work will be of interest to cognitive neuroscientists, psychologists, and clinicians working on substance use or addiction.

10.7554/eLife.97150.3.sa1Reviewer #1 (Public review):Reviewer

This study explored the relationship between sustained attention and substance use from ages 14 to 23 in a large longitudinal dataset. They found behaviour and brain connectivity associated with poorer sustained attention at age 14 predicted subsequent increase in cannabis and cigarette smoking from ages 14-23. They concluded that the brain network of sustained attention is a robust biomarker for vulnerability to substance use. The big strength of the study is a substantial sample size and validation of the generalization to an external dataset. In addition, various methods/models were used to prove the relationship between sustained attention and substance use over time.

10.7554/eLife.97150.3.sa2Reviewer #2 (Public review):Reviewer

Weng and colleagues investigated the relationship between sustained attention and substance use in a large cohort across three longitudinal visits (ages 14, 19, and 23). They employed a stop signal task to assess sustained attention and utilized the Timeline Followback self-report questionnaire to measure substance use. They assessed the linear relationship between sustained attention-associated functional connections and substance use at an earlier visit (age 14 or 19). Subsequently, they utilized this relationship along with the functional connection profile at a later age (age 19 or 23) to predict substance use at those respective ages. The authors found that connections in association with reduced sustained attention predicted subsequent increases in substance use, a conclusion validated in an external dataset. Altogether, the authors suggest that sustained attention could serve as a robust biomarker for predicting future substance use.

This study by Weng and colleagues focused on an important topic of substance use prediction in adolescence/early adulthood.

10.7554/eLife.97150.3.sa3Reviewer #3 (Public review):Reviewer

Summary:

Weng and colleagues investigated the association between attention-related connectivity and substance use. They conducted a study with a sizable sample of over 1,000 participants, collecting longitudinal data at ages 14, 19, and 23. Their findings indicate that behaviors and brain connectivity linked to sustained attention at age 14 forecasted subsequent increases in cigarette and cannabis use from ages 14 to 23. However, early substance use did not predict future attention levels or attention-related connectivity strength.

Strengths:

The study's primary strength lies in its large sample size and longitudinal design spanning three time-points. A robust predictive analysis was employed, demonstrating that diminished sustained attention behavior and connectivity strength predict substance use, while early substance use does not forecast future attention-related behavior or connectivity strength.

10.7554/eLife.97150.3.sa4Author responseWengYiheAuthorTrinity College DublinDublinIrelandKruschwitzJohannAuthorCharite BerlinBerlinGermanyRueda-DelgadoLaura MAuthorBrainWaveBank, Ltd.DublinIrelandRuddyKathy LAuthorTrinity College DublinDublinIrelandBoyleRoryAuthorTrinity College DublinDublinIrelandFranzenLuisaAuthorMaastricht UniversityMaastrichtNetherlandsSerinEminAuthorBerlin Institute of Health at Charité - Universitätsmedizin BerlinBerlinGermanyNwezeTochukwuAuthorUniversity of UtahUtahUnited StatesHansonJamieAuthorUniversity of PittsburghPittsburghUnited StatesSmythAlannahAuthorTrinity College DublinDublinIrelandFarnanTomAuthorTrinity College DublinDublinIrelandBanaschewskiTobiasAuthorHeidelberg UniversityMannheimGermanyBokdeArun LWAuthorTrinity College DublinDublinIrelandDesrivièresSylvaneAuthorKing's College LondonLondonUnited KingdomFlorHertaAuthorHeidelberg UniversityMannheimGermanyGrigisAntoineAuthorUniversité Paris-SaclayParisFranceGaravanHughAuthorUniversity of VermontBurlingtonUnited StatesGowlandPenny AAuthorUniversity of NottinghamNottinghamUnited KingdomHeinzAndreasAuthorCharité - Universitätsmedizin BerlinBerlinGermanyBrühlRüdigerAuthorPhysikalisch-Technische BundesanstaltBerlinGermanyMartinotJean-LucAuthorUniversité Paris-Saclay, CNRS, INSERMParisFranceMartinotMarie-Laure PaillèreAuthorInstitut National de la Santé et de la Recherche MédicaleGif-sur-YvetteFranceArtigesEricAuthorUniversité Paris-Saclay, CNRS, INSERMParisFranceMcGrathJaneAuthorTrinity College DublinDublinIrelandNeesFraukeAuthorUniversity Medical Center Schleswig-Holstein, Kiel UniversityKielGermanyPapadopoulos OrfanosDimitriAuthorCEA SaclayGif-sur-YvetteFrancePausTomasAuthorUniversity of MontrealMontrealCanadaPoustkaLuiseAuthorUniversität GöttingenGöttingenGermanyNathalie HolzNathalieAuthorCentral Institute of Mental HealthMannheimGermanyFröhnerJulianeAuthorTU DresdenDresdenGermanySmolkaMichael NAuthorTU DresdenDresdenGermanyVaidyaNilakshiAuthorUniversity Medical Centre GöttingenGöttingenGermanySchumannGunterAuthorKing&apos;s College LondonLondonUnited KingdomWalterHenrikAuthorCharité - Universitätsmedizin BerlinBerlinGermanyWhelanRobertAuthorTrinity College DublinDublinIreland

The following is the authors’ response to the original reviews.

Reviewer #1 (Recommendations For The Authors):

Although the manuscript is well organized and written, it could be largely improved and therefore made more plausible and easier to read. See my point-by-point comments listed below:

(1) The introduction section is a bit overloaded with some unnecessary information. For example, the authors discussed the relationship between neurotransmitters in the prefrontal and striatum and substance use/sustained attention. However, the results are related to neither the neurotransmitters nor the striatum. In addition, there is a contradictory description about neurotransmitters there, Nicotine/THC leads to increased neurotransmitters, and decreased neurotransmitters is related to poor sustained attention. Does that mean that the use of Nicotine/THC could increase sustained attention?

Thanks for this insightful question. We understand your concern regarding the seemingly contradictory statements about neurotransmitters and sustained attention. Previous studies have shown that acute administration of nicotine can improve sustained attention (Lawrence et al., 2002; Potter and Newhouse, 2008; Valentine and Sofuoglu, 2018; Young et al., 2004). On the other hand, the acute effects of smoking cannabis on sustained attention are mixed and depend on factors such as dosage and individual differences (Crean et al., 2011). For instance, a previous study (Hart et al., 2001) found that performance on a tracking task, which requires sustained attention, was found to improve significantly after smoking cannabis with a high dose of THC, albeit in experienced cannabis users. However, chronic substance use, including nicotine and cannabis, has been associated with impaired sustained attention (Chamberlain et al., 2012; Dougherty et al., 2013).

To address your concerns and improve clarity and succinctness of the Introduction, we have removed the description of neurotransmitters from the Introduction. This revision should make the introduction more concise and focus on the direct relationships pertinent to our study.

(2) It is a bit hard to follow the story for the readers because the Results section went straight into detail. For example, the authors directly introduced that they used the ICV from the Go trials to index sustained attention without basic knowledge about the task. Why use the ICV of Go trials instead of other trials (i.e., successful stop trials) as an index of sustained attention? I suggest presenting the subjects and task details about the data before the detailed behavioral results. The results section should include enough information to understand the presenting results for the readers, rather than forcing the reader to find the answer in the later Methods section.

We appreciate your suggestion to provide more context about the task and ICV before diving into the detailed behavioural results.

We used the ICV derived from the Go trials instead of Success stop trials as an index of sustained attention, based on the nature of the stop-signal task and the specific data it generates. Previous studies have indicated that reaction time (RT) variability is a straightforward measure of sustained attention, with increasing variability thought to reflect poorer ability to sustain attention (Esterman and Rothlein, 2019). RT variability is defined as ICV, calculated as the standard deviation of mean Go RT divided by the mean Go RT from Go trials (O'Halloran et al., 2018). The stop signal task includes both Go trials and stop trials. During Go trials, participants are required to respond as quickly and accurately as possible to a Go signal, allowing for the recording of RT for calculating ICV. In contrast, stop trials are designed to measure inhibitory control, where successful response inhibition results in no RT or response recorded in the output. Therefore, Go trials are specifically used to assess sustained attention, while Stop trials primarily assess inhibitory control (Verbruggen et al., 2019).

We acknowledge the importance of providing this contextual information within the Results section to enhance reader understanding. We have added this information before presenting the behavioural results on Page 6.

Results

(1) Behavioural changes over time

Reaction time (RT) variability is a straightforward measure of sustained attention, with increasing variability thought to reflect poor sustained attention. RT variability is defined as intra-individual coefficient of variation (ICV), calculated as the standard deviation of mean Go RT divided by the mean Go RT from Go trials in the stop signal task. Lower ICV indicates better sustained attention.

(3) The same problem for section 2 in the Results. What are the predictive networks? Are the predictive networks the same as the networks constructed based on the correlation with ICV? My intuitive feeling is that they are the circular analyses here. The positive/negative/combined networks are calculated based on the correlation between the edges and ICV. Then the author used the network to predict the ICV again. The manipulation from the raw networks (I think they are based on PPI) to the predictive network, and the calculation of the predicted ICV are all missing. The direct exposure of the results to the readers without enough detailed knowledge made everything hard to digest.

We thank the Reviewer for the insightful comment. We agree with the need for more clarity regarding the predictive networks and the CPM analysis before presenting results. CPM, a data-driven neuroscience approach, is applied to predict individual behaviour from brain functional connectivity (Rosenberg et al., 2016; Shen et al., 2017). The CPM analysis used the strength of the predictive network to predict the individual difference in traits and behaviours. CPM includes several steps: feature selection, feature summarization, model building, and assessment of prediction significance (see Fig. S1).

During feature selection, we assessed whether connections between brain areas (i.e., edges) in a task-related functional connectivity matrix (derived from general psychophysiological interaction analysis) were positively or negatively correlated with ICV using a significance threshold of P < 0.01. These positively or negatively correlated connections are regarded as positive or negative network, respectively. The network strength of the positive network (or negative network) was determined in each individual by summing the connection strength of each positively (or negatively) correlated edge. The combined network was determined by subtracting the strength of the negative network from the positive network. Next, CPM built a linear model between the network strength of the predictive network and ICV. This model was initially developed using the training set. The predictive networks were then applied to the test set, where network strength was calculated again, and the linear model was used to predict ICV using k-fold cross-validation. Following your advice, we have updated it in the Results section to include these details on Page 7.

Results

(2) Cross-sectional brain connectivity

This study employed CPM, a data-driven neuroscience approach, to identify three predictive networks— positive, negative, and combined— that predict ICV from brain functional connectivity. CPM typically uses the strength of the predictive networks to predict individual differences in traits and behaviors. The predictive networks were obtained based on connectivity analyses of the whole brain. Specifically, we assessed whether connections between brain areas (i.e., edges) in a task-related functional connectivity matrix derived from generalized psychophysiological interaction analysis were positively or negatively correlated with ICV using a significance threshold of P < 0.01. These positively or negatively correlated connections were regarded as positive or negative network, respectively. The network strength of positive networks (or negative networks) was determined for each individual by summing the connection strength of each positively (or negatively) correlated edge. The combined network was determined by subtracting the strength of the negative network from the positive network. We then built a linear model between network strength and ICV in the training set and applied these predictive networks to yield network strength and a linear model in the test set to calculate predicted ICV using k-fold cross validation.

(4) The authors showed the positive/negative/combined networks from both Go trials and successful stop trials can predict the ICV. I am wondering how the author could validate the specificity of the prediction of these positive/negative/combined networks. For example, how about the networks from the failed stop trials?

We appreciate the opportunity to clarify the specificity of the predictive networks identified in our study. Here is a more detailed explanation of our findings and their implications.

To validate the specificity of the sustained attention network identified from CPM analysis, we calculated correlations between the network strength of positive and negative networks and performances from a neuropsychology battery (CANTAB) at each timepoint separately. CANTAB includes several tasks that measure various cognitive functions, such as sustained attention, inhibitory control, impulsivity, and working memory. We found that all positive and negative networks derived from Go and Successful stop trials significantly correlated with a behavioural assay of sustained attention – the rapid visual information processing (RVP) task – at ages 14 and 19 (all P values < 0.028). Age 23 had no RVP task data in the IMAGEN study. There were sporadic significant correlations between constructs such as delay aversion/impulsivity and negative network strength, for example, but the correlations with the RVP were always significant. This demonstrates that the strength of the sustained attention brain network was specifically and robustly correlated with a typical sustained attention task, rather than other cognitive measures. The results are described in the main text on Page 8 and shown in Supplementary materials (Pages 1 and 3) and Table S12.

In addition, we conducted a CPM analysis to predict ICV using gPPI under Failed stop trials. Our findings showed that positive, negative, and combined networks derived from Failed stop trials significantly predicted ICV: at age 14 (r = 0.10, P = 0.033; r = 0.19, P < 0.001; and r = 0.17, P < 0.001, respectively), at age 19 (r = 0.21; r = 0.18; and r = 0.21, all P < 0.001, respectively), and at age 23 (r = 0.33, r = 0.35, and r = 0.36, respectively, all P < 0.001). Similar results were obtained using a 5-fold CV and leave-site-out CV.

Our analysis further showed that task-related functional connectivity derived from Go trials, Successful Stop trials, and Failed Stop trials could predict sustained attention across three timepoints. However, the predictive performances of networks derived from Go trials were higher than those from Successful Stop and Failed Stop trials. This suggests that sustained attention is particularly crucial during Go trials when participants need to respond to the Go signal. In contrast, although Successful Stop and Failed Stop trials also require sustained attention, these tasks primarily involve inhibitory control along with sustained attention.

Taken together, these findings underscore the specificity of the predictive networks of sustained attention. We have updated these results in the Supplementary Materials (Pages 3-5 and Page 7):

Method

CPM analysis using Failed stop trials

We performed another CPM analysis using Failed stop trials using gPPI matrix obtained from the second GLM, described in the main text. The CPM analysis was conducted using 10-fold CV, 5-fold CV and leave-site-out CV.

Results

CPM predictive performance under Failed stop trials

Positive, negative, and combined networks derived from Failed stop trials significantly predicted ICV: at age 14 (r = 0.10, P = 0.033; r = 0.19, P < 0.001; and r = 0.17, P < 0.001, respectively), at age 19 (r = 0.21; r = 0.18; and r = 0.21, all P < 0.001, respectively), and at age 23 (r = 0.33, r = 0.35, and r = 0.36, respectively, all P < 0.001). We obtained similar results using a 5-fold CV and leave-site-out CV (Table S6).

Discussion

Specificity of the prediction of predictive networks

We found that task-related function connectivity derived from Go trials, Successful stop trials, and Failed stop trials successfully predicted sustained attention across three timepoints. However, predictive performances of predictive networks derived from Go trials were higher than those derived from Successful stop trials and Failed stop trials. These results suggest that sustained attention is particularly crucial during Go trials when participants need to respond to the Go signal. In contrast, although Successful Stop and Failed Stop trials also require sustained attention, these tasks primarily involve inhibitory control along with sustained attention.

(5) The author used PPI to define the connectivity of the network. I am not sure why the author used two GLMs for the PPI analysis separately. In the second GLM, Go trials were treated as an implicit baseline. What does this exactly mean? And the gPPI analysis across the entire brain using the Shen atlas is not clear. Normally, as I understand, the PPI/gPPI is conducted to test the task-modulated connectivity between one seed region and the voxels of the whole rest brain. Did the author perform the PPI for each ROI from Shen atlas? More details about how to use PPI to construct the network are required.

Thank you for your insightful questions. Here, we’d like to clarify how we applied generalized PPI across the whole brain using the Shen atlas and why we used two separate GLMs for the gPPI analysis.

Yes, PPI is conducted to test the task-modulated connectivity between one seed region and other brain areas. This method can be both voxel-based and ROI-based. In our study, we performed ROI-based gPPI analysis using Shen atlas with 268 regions. Specifically, we performed the PPI on each seed region of interest (ROI) to estimate the task-related FC between this ROI and the remaining ROI (267 regions) under a specific task condition. By performing this analysis across each ROI in the Shen atlas, we generated a 268 × 268 gPPI matrix for each task condition. The matrices were then transposed and averaged with the original matrices, which yielded symmetrical matrices, which were subsequently used for CPM analysis.

Regarding the use of two separate GLMs for the gPPI analysis, our study aimed to define the task-related FC under two conditions: Go trials and Successful stop trials. The first GLM including Go trials was built to estimate the gPPI during Go trials. However, due to the high frequency of Go trials in the stop signal task, it is common to regard the Go trials as an implicit baseline, as in previous IMAGEN studies (D'Alberto et al., 2018; Whelan et al., 2012). Therefore, to achieve a more accurate estimation of FC during Successful stop trials, we built a second GLM specifically for these trials. Accordingly, we have updated it in the Method Section in the main text on Page 16.

Method

2.5 Generalized psychophysiological interaction (gPPI) analysis

In this study, we adopted gPPI analysis to generate task-related FC matrices and applied CPM analysis to investigate predictive brain networks from adolescents to young adults. PPI analysis describes task-dependent FC between brain regions, traditionally examining connectivity between a seed region of interest (ROI) and the voxels of the whole rest brain. However, this study conducted a generalized PPI analysis, which is on ROI-to-ROI basis (Di et al., 2021), to yield a gPPI matrix across the whole brain instead of just a single seed region.

Given the high frequency of Go trials in SST, it is common to treat Go trials as an implicit baseline in previous IMAGEN studies (D'Alberto et al., 2018; Whelan et al., 2012). Hence, we built a separate GLM for Successful stop trials, which included two task regressors (Failed and Successful stop trials) and 36 nuisance regressors.

(6) Why did the author use PPI to construct the network, rather than the other similar methods, for example, beta series correlation (BSC)?

Thanks for your question. PPI is an approach used to calculate the functional connectivity (FC) under a specific task (i.e., task-related FC). Although most brain connectomic research has utilized resting-state FC (e.g., beta series correlation), FC during task performance has demonstrated superiority in predicting individual behaviours and traits, due to its potential to capture more behaviourally relevant information (Dhamala et al., 2022; Greene et al., 2018; Yoo et al., 2018). Specifically, Zhao et al. (2023) suggested that task-related FC outperforms both typical task-based and resting-state FC in predicting individual differences. Therefore, we chose to use task-related FC to predict sustained attention over time. We have updated it in the Introduction on Page 5.

Introduction

Although most brain connectomic research has utilized resting-state fMRI data, functional connectivity (FC) during task performance has demonstrated superiority in predicting individual behaviours and traits, due to its potential to capture more behaviourally relevant information (Dhamala et al., 2022; Greene et al., 2018; Yoo et al., 2018). Specifically, Zhao et al. (2023) suggested that task-related FC outperforms both typical task-based and resting-state FC in predicting individual differences. Hence, we applied task-related FC to predict sustained attention over time.

(7) In the section of 'Correlation analysis between the network strength and substance use', the author just described that 'the correlations between xx and xx are shown in Fig5X', and repeated it three times for three correlation results. What exactly are the results? The author should describe the results in detail. And I am wondering whether there are scatter plots for these correlation analyses?

We’d like to clarify the results in Fig. 5. Fig. 5 illustrates the significant correlations between behaviour and brain activity associated with sustained attention and Cigarette and cannabis use (Cig+CB) after FDR correction. Panel A shows the significant correlation between behaviour level of sustained attention and Cig+CB. Panels B and C show the correlations between brain activity associated with sustained attention and Cig+CB. While Panel B presents the brain activity derived from Go trials, Panel C presents brain activity derived from Successful stop trials. In response to your suggestion, we have described these results in detail on Page 9. We also have included scatter plots for the significant correlations, which are shown in Fig. 5 in Supplementary materials (Fig. S10).

Results

(6) Correlation between behaviour and brain to cannabis and cigarette use

Figs. 5A-C summarizes the results showing the correlation between ICV/brain activity and Cig+CB per timepoint and across timepoints. Fig. 5A shows correlations between ICV and Cig+CB (Tables S14-15). ICV was correlated with Cig+CB at ages 19 (Rho = 0.13, P < 0.001) and 23 (Rho = 0.17, P < 0.001). ICV at ages 14 (Rho = 0.13, P = 0.007) and 19 (Rho = 0.13, P = 0.0003) were correlated with Cig+CB at age 23. Cig+CB at age 19 was correlated with ICV at age 23 (Rho = 0.13, P = 9.38E-05). Fig. 5B shows correlations between brain activity derived from Go trials and Cig+CB (Tables S18-19). Brain activities of positive and negative networks derived from Go trials were correlated with Cig+CB at age 23 (positive network: Rhop = 0.12, P < 0.001; negative network: Rhon = -0.11, P < 0.001). Brain activity of the negative network derived from Go trials at age 14 was correlated with Cig+CB at age 23 (Rhon = -0.16, P = 0.001). Cig+CB at age 19 was correlated with brain activity of the positive network derived from Go trials at age 23 (Rhop = 0.10, P = 0.002). Fig. 5C shows the correlations between brain activity derived from Successful stop and Cig+CB (Tables S18-19). Brain activities of positive and negative networks derived from Successful stop were correlated with Cig+CB at ages 19 (positive network: Rhop = 0.10, P = 0.001; negative network: Rhon = -0.08, P = 0.013) and 23 (positive network: Rhop = 0.13, P < 0.001; negative network: Rhon = -0.11, P = 0.001).

(8) Lastly, the labels of (A), (B) ... in the figure captions are unclear. The authors should find a better way to place the labels in the caption and keep them consistent throughout all figures.

Thank you for this valuable comment. We have revised the figure captions in the main text to ensure the labels (A), (B), etc., are placed more clearly and consistently across all figures.

Reviewer #2 (Public Review):

While the study largely achieves its aims, several points merit further clarification:

(1) Regarding connectome-based predictive modeling, an assumption is that connections associated with sustained attention remain consistent across age groups. However, this assumption might be challenged by observed differences in the sustained attention network profile (i.e., connections and related connection strength) across age groups (Figures 2 G-I, Fig. 3 G_I). It's unclear how such differences might impact the prediction results.

Thank you for your insightful comment. We’d like to clarify that we did not assume that connections associated with sustained attention remain completely consistent across age groups. Indeed, we expected that connections would change across age groups, due to the developmental changes in brain function and structure from adolescence to adulthood. Our focus was on the consistency of individual differences in sustained attention networks over time, recognising that the actual connections within those networks may change. However, we did show that there is some consistency in the specific connections associated with sustained attention over time. Notably, this consistency markedly increases when comparing ages 19 and 23, when developmental factors are less relevant. We support our reasoning above with the following analyses:

(1) Supplementary materials (Pages 2 and 5), relevant sections highlighted here for emphasis.

Method

Comparison of predictive networks identified at one timepoint versus another

Steiger’s Z value was employed to compare predictive performances of networks identified at different timepoints. This analysis involved comparing the R values derived from networks defined at distinct ages to predict ICV at the same age. For example, we compared the r values of brain networks defined at age 14 when predicting ICV at 19 (i.e., positive network: r = 0.25, negative network: r = 0.25, combined network: r = 0.28) with those R values of brain networks defined at age 19 itself (i.e., positive network: r = 0.16, negative network: r = 0.14, combined network: r = 0.16) derived from Go trials using Steiger's Z test (age 14 → age 19 vs. age 19 → 19). Similarly, comparisons were made between networks defined at age 14 predicting ICV at age 23 and those at age 23 predicting ICV at age 23 (age 14 → age 23 vs. age 23 → 23), as well as between networks defined at age 19 predicting ICV at age 23 and those at age 23 predicting ICV at age 23 (age 19 -> age 23 vs. age 23 -> age 23). These comparisons were performed separately for Go trials and Successful Stop trials.

Results

Comparison of predictive performance at different timepoints

For positive, negative, and combined networks predicting ICV derived from Go trials at age 19, the R values were higher when using predictive networks defined at 19 than those defined at 14 (Z = 3.79, Z = 3.39, Z = 3.99, all P < 0.00071). Similarly, the R values for positive, negative, and combined networks predicting ICV derived from Go trials at age 23 were higher when using predictive networks defined at age 23 compared to those defined at ages 14 (Z = 6.00, Z = 5.96, Z = 6.67, all P < 3.47e-9) or 19 (Z = 2.80, Z = 2.36, Z = 2.57, all P < 0.005).

At age 19, the R value for the positive network predicting ICV derived from Successful stop trials was higher when using predictive networks defined at 19 compared to those defined at 14 (Z = 1.54, P = 0.022), while the negative and combined networks did not show a significant difference (Z = 0.85, P = 0.398; Z = 2.29, P = 0.123). At age 23, R values for the positive and combined networks predicting ICV derived from Successful stop trials were higher when using predictive networks defined at 23 compared to those defined at 14 (Z = 3.00, Z = 2.48, all P < 3.47e-9) or 19 (Z = 2.52, Z = 1.99, all P < 0.005). However, the R value for the negative network at age 23 did not significantly differ when using predictive networks defined at 14 (Z = 1.80, P = 0.072) or 19 (Z = 1.48, P = 0.138).

These results indicate that some specific pairwise connections associated with sustained attention at earlier ages, such as 14 and 19, are still relevant as individuals grow older. However, some connections are not optimal for good sustained attention at older ages. That is, the brain reorganizes its connection patterns to maintain optimal functionality for sustained attention as it matures.

(2) Consistency of Individual Differences:

We found individual differences in ICV were significantly correlated between the three timepoints (Fig. 1B). In addition, we calculated the correlations of network strength of predictive networks predicting sustained attention derived from Go trials and Successful trials between each timepoints. We found that the correlations of network strength for predictive networks (derived from Go trials and Successful trials) were also significant (all P < 0.003). We have updated these results in the main text (Pages 7-8) and Supplementary Materials (Table S7).

(2) Cross-sectional brain connectivity

In addition, we found that network strength of positive, negative, and combined networks derived from Go trials was significantly correlated between the three timepoints (Table S7, all P < 0.003).

In addition, we found that network strength of positive, negative, and combined networks derived from Successful stop trials was significantly correlated between the three timepoints (Table S7, all P < 0.001).

(3) Predictive networks across timepoints: Predictive networks defined at age 14 were successfully applied to predict ICV at ages 19 and 23. Similarly, predictive networks defined at age 19 were successfully applied to predict ICV at age 23 (Fig. 4). These results reflect the robustness of the brain network associated with sustained attention over time.

(4) Dice coefficient analysis: We calculated the Dice coefficient to quantify the similarity of predictive networks across the three timepoints. Connections in the sustained attention networks were significantly similar from ages 14 to 23 (Table S13), despite relatively few overlapping edges over time (as discussed in Supplementary Materials on Page 6).

(5) Global brain activation: Based on these findings, we indicate that sustained attention relies on global brain activation (i.e., network strength) rather than specific regions or networks (see also (Zhao et al., 2021)).

In summary, brain network connections undergo change and are not completely consistent across time. However, individual differences in sustained attention and its network are consistent across time, as we found that (1) the brain reorganizes its connection patterns to maintain optimal functionality for sustained attention as it matures. (2) ICV and network strength of sustained attention network were significantly correlated between each timepoint. (3) Sustained attention networks identified from previous timepoints could predict ICV in the subsequent timepoint. (4) Dice coefficient analysis indicated that the edges in the sustained attention networks were significantly similar from ages 14 to 23. (5) Sustained attention networks function as a global activation, rather than specific regions or networks.

(2) Another assumption of the connectome-based predictive modeling is that the relationship between sustained attention network and substance use is linear and remains linear over development. Such linear evidence from either the literature or their data would be of help.

Thanks for your valuable suggestion. We'd like to clarify that while CPM assumes a linear relationship between brain and behaviour (Shen et al., 2017), it does not assume that the relationship between the sustained attention network and substance use remains linear over development.

Our approach in applying CPM to predict sustained attention across different timepoints was based on previous neuroimaging studies (Rosenberg et al., 2016; Rosenberg et al., 2020), which indicated linear associations between brain connectivity patterns and sustained attention using CPM analysis. These findings support the notion of a linear relationship between brain connectivity and sustained attention. In this study, we performed CPM analysis to identify predictive networks predicting sustained attention, not substance use and used the network strength of these predictive networks to represent sustained attention activity.

To examine the relationship between substance use and sustained attention, as well as its associated brain activity, we conducted correlation analyses and utilized a latent change score model instead of CPM analysis. This decision was informed by cross-sectional studies (Broyd et al., 2016; Lisdahl and Price, 2012) that consistently reported linear associations between substance use and impairments in sustained attention. Additionally, longitudinal research by (Harakeh et al., 2012) indicated a linear relationship between poorer sustained attention and the initiation and escalation of substance use over time.

Given these previous findings, we assumed a linear relationship between sustained attention and substance use. Our analyses included calculating correlations between substance use and sustained attention, as well as its associated brain activity at each timepoint and across timepoints (Fig. 5). Furthermore, we employed a three-wave bivariable latent change score model, a longitudinal approach, to assess the relationship between substance use and behavirour and brain activity associated with sustained attention (Figs. 6-7). We have added more information in the Introduction to make it more clear on Page 6.

Introduction

Additionally, previous cross-sectional and longitudinal studies (Broyd et al., 2016; Harakeh et al., 2012; Lisdahl and Price, 2012) have shown that there are linear relationships between substance use and sustained attention over time. We therefore employed correlation analyses and a latent change score model to estimate the relationship between substance use and both behaviours and brain activity associated with sustained attention.

(3) Heterogeneity in results suggests individual variability that is not fully captured by group-level analyses. For instance, Figure 1A shows decreasing ICV (better-sustained attention) with age on the group level, while there are both increasing and decreasing patterns on the individual level via visual inspection. Figure 7 demonstrates another example in which the group with a high level of sustained attention has a lower risk of substance use at a later age compared to that in the group with a low level of sustained attention. However, there are individuals in the high sustained attention group who have substance use scores as high as those in the low sustained attention group. This is important to take into consideration and could be a potential future direction for research.

Thanks for this valuable comment. We appreciate your observation regarding the individual variability that is not fully captured by group-level analyses to some degree. Fig. 1A shows the results from a linear mixed model, which explains group-level changes over time while accounting for the random effect within subjects. Similarly, Fig. 7 shows the group-level association between substance use and sustained attention. We agree that future research could indeed consider individual variability. For example, participants could be categorized based on their consistent trajectories of ICV or substance use (i.e., keep decreasing/increasing) over multiple timepoints. We agree that incorporating individual-level analyses in the future could provide valuable insights and are grateful for your suggestion, which will inform our future research directions.

The above-mentioned points might partly explain the significant but low correlations between the observed and predicted ICV as shown in Figure 4. Addressing these limitations would help enhance the study's conclusions and guide future research efforts.

We have updated the text in the Discussion on Page 13:

Discussion

However, there are still some individual variabilities not captured in this study, which could be attributed to the diversity in genetic, environmental, and developmental factors influencing sustained attention and substance use. Future research should aim to explore these variabilities in greater depth to gain better understanding of the relationship between sustained attention and substance use.

Reviewer #3 (Public Review):

Weaknesses: It's questionable whether the prediction approach (i.e., CPM), even when combined with longitudinal data, can establish causality. I recommend removing the term 'consequence' in the abstract and replacing it with 'predict'. Additionally, the paper could benefit from enhanced rigor through additional analyses, such as testing various thresholds and conducting lagged effect analyses with covariate regression.

Thank you for your comment. We have replaced “consequence” by “predict” in the abstract.

Abstract

Previous studies were predominantly cross-sectional or under-powered and could not indicate if impairment in sustained attention was a predictor of substance-use or a marker of the inclination to engage in such behaviour.

Reviewer #3 (Recommendations For The Authors):

(1) The connectivity analysis predicts both baseline and longitudinal attention measures. However, given the high correlation in attention abilities across the three time-points, it's unclear whether the connectivity predicts shared variations of attention across three time points. It would be insightful to assess if predictions at the 2nd and 3rd-time points remained significant after controlling for attention abilities at the initial time point.

Thanks for your comments. We performed the CPM analysis to predict ICV at the 2nd and 3rd timepoint, controlling for ICV at age 14 as a covariate. We found that controlling for ICV at age 14, positive, negative, and combined networks derived from Successful stop trials defined at age 14 still predicted ICV at ages 19 and 23. In addition, positive, negative, and combined networks derived from Successful stop trials defined at age 19 predicted ICV at age 23. In addition, positive, negative, and combined networks derived from Go trials defined at age 19 still predicted ICV at age 23, after controlling for ICV at age 14. However, positive, negative, and combined networks derived from Go trials defined at age 14 had lower predictive performances in predicting ICV at ages 19 and 23, after controlling for ICV at age 14. Notably, controlling for ICV at the initial timepoint did not significantly impact the performances of predictive networks derived from Successful stop trials. Accordingly, we have added this analysis and the results in the Supplementary Materials (Pages 3 and 5).

Method

Prediction across timepoints controlling for ICV at age 14

To examine whether connectivity predictors shared variations of sustained attention across timepoints, we applied predictive models developed at ages 14 and 19 to predict ICV at subsequent timepoints controlling for ICV at age 14. Specifically, we used predictive models (including parameters and selected edges) developed at age 14 to predict ICV at ages 19 and 23 separately. First, we calculated the network strength using the gPPI matrix at ages 19 and 23 based on the selected edges identified from CPM analysis at age 14. We then estimated the predicted ICV at ages 19 and 23 by applying the linear model parameters (slope and intercept) obtained from CPM analysis at age 14 to the network strength. Finally, we evaluated the predictive performance by calculating the partial correlation between the predicted and observed values at ages 19 and 23, controlling for ICV at age 14. Similarly, we applied models developed at age 19 to predict ICV at age 23, also controlling for ICV at age 14. To assess the significance of the predictive performance, we used a permutation test, shuffling the predicted ICV values and calculating partial correlation to general a random distribution over 1,000 iterations.

Results

Predictions across timepoints controlling for ICV at age 14

Positive and combined networks derived from Go trials defined at age 14 predicted ICV at ages 19 (r = 0.10, P = 0.028; r = 0.08, P = 0.047) but negative network did not (r = 0.06, P = 0.119). Positive network derived from Go trials defined at age 14 predicted ICV at age 23 (r = 0.11, P = 0.013) but negative and combined networks did not (r = 0.04, P = 0.187; r = 0.08, P = 0.056). Positive, negative, and combined networks derived from Go trials defined at age 19 predicted ICV at age 23 (r = 0.22, r = 0.19, and r = 0.22, respectively, all P < 0.001).

Positive, negative, and combined networks derived from Successful stop trials defined at age 14 predicted ICV at age 19 (r = 0.08, P = 0.036; r = 0.10, P = 0.012; r = 0.11, P = 0.009) and 23 (r = 0.11, P = 0.005; r = 0.13, P = 0.005; r = 0.13, P = 0.017) respectively. Positive, negative, and combined networks derived from Successful stop trials defined at age 19 predicted ICV at age 23 (r = 0.18, r = 0.18, and r = 0.17, respectively, all P < 0.001).

(2) In the Results section, a significance threshold of p = 0.01 was used for the CPM analysis. It would be beneficial to test the stability of these findings using alternative thresholds such as p = 0.05 or p = 0.005.

We appreciate this insightful comment. We appreciate the suggestion to test the stability of our findings using alternative significance thresholds. Indeed, we have already conducted CPM analyses using a range of thresholds, including 0.1, 0.05, 0.01, 0.005, 0.001, 0.0005, and 0.0001 (see Table S8 in supplementary Materials). The results were similar across different thresholds. Following prior studies (Feng et al., 2024; Ren et al., 2021; Yoo et al., 2018) which used P < 0.01 for feature selection, we chose to focus on the threshold of P < 0.01 for our main analysis. Following your suggestion, we have highlighted this in the Method section on Pages 17-18.

Method

2.6.1 ICV prediction

The r value with an associated P value for each edge was obtained, and a threshold P = 0.01 (Feng et al., 2024; Ren et al., 2021; Yoo et al., 2018) was set to select edges.

2.6.2 Three cross-validation schemes

In addition, we conducted the CPM analysis using a range of thresholds for feature selection and observed similar results across different thresholds (See Supplementary Materials Table S8).

(3) Could you clarify if you used one sub-sample to extract connectivity related to sustained attention and then used another sub-sample to predict substance use with attention-related connectivity?

Thank you very much for the question. We used the same sample to extract the brain network strength and estimated the correlation with substance use using both the Spearman correlation and latent change score model across three timepoints. We controlled for covariates including sex, age, and scan site at the same time. Accordingly, we have clarified this in the Method section on Page 20. We note that the CPM analyses were conducted using cross-validation, plus a leave-site-out analysis.

Method

2.7.3 Correlation between network strength and substance use

It is worth noting that all the correlations between substance use and sustained attention were conducted using the same sample across three timepoints.

(4) Could you clarify whether you have regressed covariates in the lagged effects analysis of part 7?

Thanks for this question. Yes, we confirmed that we controlled the covariates including age, sex and scan sites in the latent change score model. We have described them more clearly now in the Method section (Page 18).

Method

2.7.3 Correlation between network strength and substance use

Additionally, cross-lagged dynamic coupling (i.e., bidirectionality) was employed to explore individual differences in the relationships between substance use and linear changes in ICV/brain activity, as well as the relationship between ICV/brain activity and linear change in substance use. The model accounted for covariates such as age, sex and scan sites.

References:

Broyd, S.J., van Hell, H.H., Beale, C., Yucel, M., Solowij, N., 2016. Acute and Chronic Effects of Cannabinoids on Human Cognition-A Systematic Review. Biol Psychiatry 79, 557-567.

Chamberlain, S.R., Odlaug, B.L., Schreiber, L.R.N., Grant, J.E., 2012. Association between Tobacco Smoking and Cognitive Functioning in Young Adults. The American Journal on Addictions 21, S14-S19.

Crean, R.D., Crane, N.A., Mason, B.J., 2011. An evidence based review of acute and long-term effects of cannabis use on executive cognitive functions. J Addict Med 5, 1-8.

D'Alberto, N., Chaarani, B., Orr, C.A., Spechler, P.A., Albaugh, M.D., Allgaier, N., Wonnell, A., Banaschewski, T., Bokde, A.L.W., Bromberg, U., Buchel, C., Quinlan, E.B., Conrod, P.J., Desrivieres, S., Flor, H., Frohner, J.H., Frouin, V., Gowland, P., Heinz, A., Itterman, B., Martinot, J.L., Paillere Martinot, M.L., Artiges, E., Nees, F., Papadopoulos Orfanos, D., Poustka, L., Robbins, T.W., Smolka, M.N., Walter, H., Whelan, R., Schumann, G., Potter, A.S., Garavan, H., 2018. Individual differences in stop-related activity are inflated by the adaptive algorithm in the stop signal task. Hum Brain Mapp 39, 3263-3276.

Dhamala, E., Yeo, B.T.T., Holmes, A.J., 2022. Methodological Considerations for Brain-Based Predictive Modelling in Psychiatry. Biological Psychiatry.

Di, X., Zhang, Z.G., Biswal, B.B., 2021. Understanding psychophysiological interaction and its relations to beta series correlation. Brain Imaging and Behavior 15, 958-973.

Dougherty, D.M., Mathias, C.W., Dawes, M.A., Furr, R.M., Charles, N.E., Liguori, A., Shannon, E.E., Acheson, A., 2013. Impulsivity, attention, memory, and decision-making among adolescent marijuana users. Psychopharmacology (Berl) 226, 307-319.

Esterman, M., Rothlein, D., 2019. Models of sustained attention. Curr Opin Psychol 29, 174-180.

Feng, Q., Ren, Z., Wei, D., Liu, C., Wang, X., Li, X., Tie, B., Tang, S., Qiu, J., 2024. Connectome-based predictive modeling of Internet addiction symptomatology. Soc Cogn Affect Neurosci 19.

Greene, A.S., Gao, S., Scheinost, D., Constable, R.T., 2018. Task-induced brain state manipulation improves prediction of individual traits. Nature Communications 9, 2807.

Harakeh, Z., de Sonneville, L., van den Eijnden, R.J., Huizink, A.C., Reijneveld, S.A., Ormel, J., Verhulst, F.C., Monshouwer, K., Vollebergh, W.A., 2012. The association between neurocognitive functioning and smoking in adolescence: the TRAILS study. Neuropsychology 26, 541-550.

Hart, C.L., van Gorp, W., Haney, M., Foltin, R.W., Fischman, M.W., 2001. = . Neuropsychopharmacology 25, 757-765.

Lawrence, N.S., Ross, T.J., Stein, E.A., 2002. Cognitive mechanisms of nicotine on visual attention. Neuron 36, 539-548.

Lisdahl, K.M., Price, J.S., 2012. Increased marijuana use and gender predict poorer cognitive functioning in adolescents and emerging adults. J Int Neuropsychol Soc 18, 678-688.

O'Halloran, L., Cao, Z.P., Ruddy, K., Jollans, L., Albaugh, M.D., Aleni, A., Potter, A.S., Vahey, N., Banaschewski, T., Hohmann, S., Bokde, A.L.W., Bromberg, U., Buchel, C., Quinlan, E.B., Desrivieres, S., Flor, H., Frouin, V., Gowland, P., Heinz, A., Ittermann, B., Nees, F., Orfanos, D.P., Paus, T., Smolka, M.N., Walter, H., Schumann, G., Garavan, H., Kelly, C., Whelan, R., 2018. Neural circuitry underlying sustained attention in healthy adolescents and in ADHD symptomatology. Neuroimage 169, 395-406.

Potter, A.S., Newhouse, P.A., 2008. Acute nicotine improves cognitive deficits in young adults with attention-deficit/hyperactivity disorder. Pharmacol Biochem Behav 88, 407-417.

Ren, Z., Daker, R.J., Shi, L., Sun, J., Beaty, R.E., Wu, X., Chen, Q., Yang, W., Lyons, I.M., Green, A.E., Qiu, J., 2021. Connectome-Based Predictive Modeling of Creativity Anxiety. Neuroimage 225, 117469.

Rosenberg, M.D., Finn, E.S., Scheinost, D., Papademetris, X., Shen, X., Constable, R.T., Chun, M.M., 2016. A neuromarker of sustained attention from whole-brain functional connectivity. Nat Neurosci 19, 165-171.

Rosenberg, M.D., Scheinost, D., Greene, A.S., Avery, E.W., Kwon, Y.H., Finn, E.S., Ramani, R., Qiu, M., Constable, R.T., Chun, M.M., 2020. Functional connectivity predicts changes in attention observed across minutes, days, and months. Proc Natl Acad Sci U S A 117, 3797-3807.

Shen, X., Finn, E.S., Scheinost, D., Rosenberg, M.D., Chun, M.M., Papademetris, X., Constable, R.T., 2017. Using connectome-based predictive modeling to predict individual behavior from brain connectivity. Nat Protoc 12, 506-518.

Valentine, G., Sofuoglu, M., 2018. Cognitive Effects of Nicotine: Recent Progress. Curr Neuropharmacol 16, 403-414.

Verbruggen, F., Aron, A.R., Band, G.P.H., Beste, C., Bissett, P.G., Brockett, A.T., Brown, J.W., Chamberlain, S.R., Chambers, C.D., Colonius, H., Colzato, L.S., Corneil, B.D., Coxon, J.P., Dupuis, A., Eagle, D.M., Garavan, H., Greenhouse, I., Heathcote, A., Huster, R.J., Jahfari, S., Kenemans, J.L., Leunissen, I., Li, C.S.R., Logan, G.D., Matzke, D., Morein-Zamir, S., Murthy, A., Pare, M., Poldrack, R.A., Ridderinkhof, K.R., Robbins, T.W., Roesch, M.R., Rubia, K., Schachar, R.J., Schall, J.D., Stock, A.K., Swann, N.C., Thakkar, K.N., van der Molen, M.W., Vermeylen, L., Vink, M., Wessel, J.R., Whelan, R., Zandbelt, B.B., Boehler, C.N., 2019. A consensus guide to capturing the ability to inhibit actions and impulsive behaviors in the stop-signal task. Elife 8.

Whelan, R., Conrod, P.J., Poline, J.B., Lourdusamy, A., Banaschewski, T., Barker, G.J., Bellgrove, M.A., Buchel, C., Byrne, M., Cummins, T.D., Fauth-Buhler, M., Flor, H., Gallinat, J., Heinz, A., Ittermann, B., Mann, K., Martinot, J.L., Lalor, E.C., Lathrop, M., Loth, E., Nees, F., Paus, T., Rietschel, M., Smolka, M.N., Spanagel, R., Stephens, D.N., Struve, M., Thyreau, B., Vollstaedt-Klein, S., Robbins, T.W., Schumann, G., Garavan, H., Consortium, I., 2012. Adolescent impulsivity phenotypes characterized by distinct brain networks. Nat Neurosci 15, 920-925.

Yoo, K., Rosenberg, M.D., Hsu, W.T., Zhang, S., Li, C.R., Scheinost, D., Constable, R.T., Chun, M.M., 2018. Connectome-based predictive modeling of attention: Comparing different functional connectivity features and prediction methods across datasets. Neuroimage 167, 11-22.

Young, J.W., Finlayson, K., Spratt, C., Marston, H.M., Crawford, N., Kelly, J.S., Sharkey, J., 2004. Nicotine improves sustained attention in mice: evidence for involvement of the alpha7 nicotinic acetylcholine receptor. Neuropsychopharmacology 29, 891-900.

Zhao, W., Makowski, C., Hagler, D.J., Garavan, H.P., Thompson, W.K., Greene, D.J., Jernigan, T.L., Dale, A.M., 2023. Task fMRI paradigms may capture more behaviorally relevant information than resting-state functional connectivity. Neuroimage, 119946.

Zhao, W., Palmer, C.E., Thompson, W.K., Chaarani, B., Garavan, H.P., Casey, B.J., Jernigan, T.L., Dale, A.M., Fan, C.C., 2021. Individual Differences in Cognitive Performance Are Better Predicted by Global Rather Than Localized BOLD Activity Patterns Across the Cortex. Cereb Cortex 31, 1478-1488.