Attention-deficit/hyperactivity disorder (ADHD) is an early-onset, functionally impairing and prevalent disorder that is associated with greater risk for a host of negative and impairing outcomes [1], including alcohol problems [2]. Adolescents and adults with ADHD are at greater risk for developing alcohol use disorders and problems, and in individuals with ADHD, the lifetime prevalence of alcohol dependence is ~ 3–11% and of any alcohol use disorder, it is ~ 43% [2]. The comorbidity of ADHD with alcohol use disorder is associated with additional comorbidities and worse response to treatment [2]. Adolescent alcohol use is associated with attenuated grey matter and deficits in cognitive processes affected by ADHD, including attentional and executive functions, leading to worse cognitive and developmental outcomes for adolescents with ADHD who frequently consume alcohol. In turn, worse adolescent cognitive outcomes are associated with greater risk for adulthood alcohol binge drinking and, over time, worse health and socioeconomic status [2]. Better understanding the causes of ADHD and identifying predictors in adolescence of prognosis in young adulthood is key to advancing the effectiveness of early identification of at-risk individuals and to the individualization of prevention and treatment, i.e. precision psychiatry.

A clinical phenotype is the clinically observable and relevant characteristics of a disorder, i.e. manifest symptoms. Data indicate efforts to determine etiology and predict prognosis relying on the clinical phenotype have been largely unsuccessful, arguably because the ADHD clinical phenotype is heterogeneous in terms of causes, manifestation, and course [3, 4]. Specifically, a multitude of developmental pathways can lead to a clinical phenotype that is multifaceted both with regard to core symptoms (i.e., difficulties with regulating activity, attention, and impulses [1]), and with regard to associated features (e.g. emotional features [5], executive functioning [6], and reward processing [7]). This multifaceted clinical phenotype, in turn, leads to diverse outcomes.

Intermediate phenotypes are biological markers that are heritable and, considering an etiological framework, located between genetic predisposition and manifest symptoms [8, 9]. Relative to the clinical phenotype, intermediate phenotypes, by virtue of their homogeneity, are hypothesized to have better explanatory and prognostic power [7]. Evidence indicates reward processing may be an ADHD intermediate phenotype, as findings show associations between differences in reward processing and ADHD [3, 4, 7]. Research on biomarkers of reward processing in ADHD is comprised almost exclusively of case–control and diagnostic biomarker studies. A diagnostic biomarker (1) confirms or detects the absence/presence of a condition and/or (2) differentiates across presentations (subtypes) of that disorder [10]. In many cases, between-group differences across ADHD and control groups were not detectable/ replicated [11,12,13,14], or biomarkers did not differentiate diagnostic groups [7], leading to the conclusion that the biomarker is clinically irrelevant or uninformative. Yet, findings of case–control studies may be misleading as even in the absence of between-group differences in the biomarker, there may be a difference in the extent to which (or whether) the biomarker is associated with functional outcomes. In case of reward processing, even in the absence of between-group differences in neural reward response, there is a between-groups difference in how neural reward response is associated with affective and alcohol outcomes. For example, in adolescents at-risk for ADHD, a negative association was observed between neural reward response and depression and a positive association was observed between neural reward response and hazardous alcohol use. In adolescents not at-risk for ADHD, neural reward response was not associated with depression and it was negatively associated with hazardous alcohol use [14]. By definition, diagnostic biomarker studies assess the extent to which a given biomarker of an intermediate phenotype converges with the categorical clinical phenotype even though the very utility of the biomarker lies in being an improvement upon and thus nonredundant with the clinical phenotype. Both case–control and diagnostic biomarker studies, albeit informative about differences at the group level, are by nature uninformative about causes and course.

Taken together, the test of clinical utility of a biomarker of an intermediate phenotype is not whether it differs or differentiates between groups, i.e. whether it is a diagnostic biomarker. Rather, the apt test of such utility is whether it explains the causes of or the course of the disorder, i.e. whether it is a risk or a prognostic biomarker [7]. A biomarker that indicates the potential for developing a disorder or medical condition in an individual who does not currently have clinically apparent disorder or the medical condition is classified as a susceptibility/risk biomarker. The concept is similar to prognostic biomarkers, except that the key issue is the association with the development of a disease rather than prognosis after one already has the diagnosis [10]. A prognostic biomarker indicates the likelihood of a clinical event or outcome, or the progression or recurrence of the disorder in individuals with the condition [10].

Event-related potentials (ERPs) are changes in the electroencephalogram (EEG) as a result of specific events (i.e. stimuli) that reflect, physiologically, the synchronous activity of neuronal populations and psychologically, different cognitive functions, e.g. affective, cognitive, motor, of perceptual processes that are experimentally probed by stimuli or a task [15]. ERPs are appropriate and ideal for assessing aspects of reward processing defined and differentiated based on their occurrence in time [16]. Moreover, given their acceptance by participants, cost effectiveness, and relatively high movement tolerance, ERPs are also suitable for collecting data from large clinical samples longitudinally [17, 18]. Case–control studies indicate between-group differences in ERPs to reward across ADHD and control groups, e.g. adolescents and children with ADHD exhibited enhanced ERPs to escaping delay [19] and to salience of reward [20] as well as greater improvements in behavioral inhibition as a result of social rewards [21].

In some cases, between-group differences in ERPs to reward across ADHD and control groups were not detected, e.g. between adults and children with and without ADHD to error and to inhibition [22,23,24], to probabilistic reward learning [25], or with regard to improved performance as a result of reward [26]. Diagnostic biomarker studies indicate in adolescents, ERPs of reward do not differentiate adolescents with and without ADHD [7].

Here, we examine whether electrophysiological indices of reward processing are ADHD risk and prognostic biomarkers. We index ADHD risk by ADHD polygenic risk scores (PRS), which reflect the cumulative effect of frequent genetic variants [5], and index ADHD prognosis via alcohol use. Specifically, our aims were to examine whether (Aim 1) in a sample of adolescents, ERP measures of reward anticipation are associated with ADHD PRSs, and whether (Aim 2) in a sample of adolescents with the ADHD clinical phenotype, ERPs of reward anticipation are associated, longitudinally, with alcohol use. We hypothesized that ERP measures of reward anticipation are associated with ADHD PRSs and longitudinally, with alcohol use.

Across analyses, we account for the effects of age, sex and depression, given an established link between reward processing and these variables [27, 28]. We also account for the effects of ADHD severity; first, to ensure that its shared variance with ADHD PRSs does not account for findings and second, to ensure that any findings obtained reflect effects of the intermediate phenotype beyond the clinical phenotype.

Data analyzed in the current study were collected at the first two assessment sessions of the second (Wave 1) and fourth (18-month follow-up, i.e. Wave 2) years of a longitudinal study, the Budapest Longitudinal Study of ADHD and Externalizing Disorders.

Participants were excluded if they exhibited cognitive ability at or below the percentile rank that corresponds to a full-scale IQ score of 80 on the Wechsler intelligence scale for children–Fourth Edition (WISC-IV) or the Wechsler adult intelligence scale–Fourth Edition (WAIS–IV) [29, 30]; met diagnostic criteria for bipolar, obsessive–compulsive or psychotic disorder on the Structured Clinical Interview for DSM-5 Disorders, Clinical Version (SCID-5-CV); had a prior autism spectrum disorder (severity ≥ 2) diagnosis; reported a neurological illness; and had visual impairment (uncorrected, impaired vision < 50 cm).

Following written informed assent (adolescents) and written informed assent (parents), adolescents completed a series of tests. At Wave 1, the first assessment session comprised clinical interview and cognitive testing, genetic sampling, and completion of questionnaires. The second assessment session comprised an EEG measurement and completion of questionnaires. At Wave 2, the first assessment session comprised completion of questionnaires. The second assessment session comprised an EEG measurement. Questionnaires were completed by parents via Psytoolkit [31, 32] and Qualtrics (Version June 2020–May 2023) (Qualtrics, Provo, UT). The longitudinal study was approved by the National Institute of Pharmacy and Nutrition (OGYÉI/17089-8/2019). The study has been performed in adherence to the ethical standards of the 1964 Declaration of Helsinki and its later amendments.

ADHD classification was determined using parent-report on the ADHD Rating Scale-5 (ARS-5) [33]. To be classified as at-risk for ADHD, adolescents had to meet a total of ≥ 4 of the Diagnostic and Statistical Manual of Mental Disorders (5th ed.; DSM-5) ADHD symptoms (from either domain). To be classified as diagnosed with ADHD (for purposes of research), adolescents had to meet a total of ≥ 6 (youth < 17 years old) or 5 (youth ≥ 17 years old) of the DSM-5 ADHD inattentive (IA) or hyperactive/impulsive (H/I) symptoms and exhibit impairment (i.e., rating of ≥ 2 = moderate impairment) in ≥ 3 areas of functioning.

Addressing different questions, findings with samples drawn from the larger longitudinal study, have been previously published [7, 14, 34,35,36].

Participants were N = 304 adolescents oversampled for ADHD, i.e. recruited from the community and hospitals, as detailed in previous publications [34, 35]; at baseline, adolescents were between 14 and 17 years (Mage = 15.78 years, SD = 1.08; 39.5% female); n = 132 (43.4%) met criteria for at-risk for ADHD. At Wave 2, data were available for n = 233 adolescents (23% attrition), of whom n = 99 (42.5%) were classified at baseline as at-risk for ADHD. At Wave 2, adolescents at-risk for ADHD were between 15 and 19 years (Mage = 17.08 years, SD = 1.07; 29.3% female).

Participants’ average cognitive ability was in the 61st percentile (SD = 20.86). Based on net household income per person, compared to the 2020 Hungarian average of ~ 147 000 HUF [37], with a sample average of 156 374 HUF (SD = 77 685), participating adolescents and their families were from a somewhat above-average socioeconomic background t(303) = 2.104, p = 0.036, Cohen's d = 0.121 (95% CI[0.008, 0.233]). For details on medication washout, see Supplement.

Items from the self-reported European School Survey Project on Alcohol and Other Drugs (ESPAD) master questionnaire [38] were used to assess alcohol consumption, binge drinking, and drunkenness across lifetime, during the last 12 months, and during the last 30 days. The parent-reported ARS 5 [33] was used to assess ADHD. Prior findings indicate acceptable psychometric properties for both the ESPAD [34, 38,39,40,41] and the ARS 5 [7, 14, 33]. In the current sample, the Binge Drinking (ωbaseline = 0.940; ωT2 = 0.938), the Consumption (ωbaseline = 0.916; ωT2 = 0.934), and the Drunkenness (ωbaseline = 0.944; ωT2 = 0.938) subscales of the ESPAD and the ARS-5 Total (ωbaseline = 0.954) exhibited acceptable internal consistency and were used in analyses. For details, see Supplement.

The MID task [42, 43] is the recommended task for probing reward anticipation [44] and its electrophysiological version, the e-MID task is appropriate for differentiating electrophysiological response to anticipation and receipt of reward [45]. Evidence indicates reliability of e-MID ERPs [16] as well as convergent validity between e-MID ERPs and self-report reward processing [46]. For description of the employed MID parameters and version, see Supplement. For analyzed ERP variables, see Analytic Plan.

Details and procedures for EEG data recording and processing have been described previously [7]. Electrodes and time windows were selected based on the literature [7, 16, 27, 45, 47] based on when and where ERPs were maximal during our pilot studies: Cue P3 at Pz, POz, P1, and P2, for the 450–650 ms time window; Target P3 at CPz, Pz, P1, and P2, for the 200–375 ms time window; SPN at CPz, Pz, CP1, CP2, P1, and P2, for the -200–0 ms time window; and RewP at CPz, Cz, FCz, CP1, CP2, FC1, and FC2, for the 225–325 ms time window [7].

Genomic DNA was isolated from saliva samples. Samples were processed following manufacturer guidelines and recommendations [48] and genotyped using the Illumina Infinium Global Screening Array-24 v3.0 BeadChip by LIFE & BRAIN GmbH (Bonn, Germany).

All analyses were conducted in RStudio (version 2023.09.1. Build 494, R version 4.3.2.). For packages used, see Table S1.

Data preparation involved imputation of missing data. Missing alcohol use data were substituted using multiple imputation with a state-of-the-art deep learning method, for details see [49]. One of five generated datasets was used.

ADHD PRSs were calculated based on a discovery dataset involving 38,691 individuals with ADHD and 186,843 controls [50]. Using SNP cutoff of p < 0.50, the number of ADHD PRS SNPs was 99,330 with an associated R2 of ≈3.7%. For details, see Supplement.

Exploratory factor analysis (EFA) EFA was conducted with the aim of dimension reduction, on 48 ERP variables: indices of amplitude and latency for Cue P3, Target P3, SPN, and RewP to conditions of win, lose, neutral win and neutral lose; indices of amplitude for Cue P3, Target P3, SPN, and RewP to win-lose, win-neutral win, lose-neutral lose difference scores.

EFA was conducted applying promax rotation (based on correlations between ERP variables) and 15 factors (based on parallel analysis). Items were first eliminated if they loaded poorly (< 0.40 on any factor) [51, 52] and then if they loaded on more than one factor (> 0.40 on ≥ two factors) [51, 52]. Dual loading items were eliminated starting with the item whose second highest loading (absolute value) was the highest. After each elimination, parallel analysis was rerun until no additional items were indicated for removal.

Considering eigenvalues > 1 [53] and factors with > two variables [51, 52], two factors were retained (factor-item loadings ≥ 0.820). The first factor (ERPf1TargetP3) included Target P3 amplitude variables, to win (0.939), lose (0.919), neutral win (0.911), and neutral lose (0.912). The second factor (ERPf2SPN) included SPN amplitude variables to win (0.823), lose (0.899), neutral win 0.850), and neutral lose (0.820). Target P3 to win and to lose trials achieved acceptable internal consistency by the ~ tenth trial, MID SPN to win and to lose trials by the ~ 20th trial, Target P3 to neutral win and to neutral lose by the ~ 14th trial, and SPN to neutral win and to neutral lose by the ~ 26th trial (Figure S1).

Regression analysis Across Aim 1 and 2 models, linear regression analyses were conducted. Across Aim 1 and Aim 2 models, covariates were baseline age, sex, ADHD severity, and Depressive Problems T scores. For Aim 1 models, independent variables were ADHD PRSs; dependent variables were ERPf1TargetP3 and ERPf2SPN. For Aim 2 models, independent variables were ERPf1TargetP3 and ERPf2SPN; dependent variables were Wave 2 values of alcohol use (ESPAD binge drinking, consumption, and drunkenness subscales). For Aim 1 models, additional covariates were the first four genetic principal components and for Aim 2 models, additional covariates were baseline values of the outcome variable. p-values corresponding to the effect of interest were adjusted for false discovery rate (FDR) [54].

Across models, distribution of residuals was checked using normality tests (Anderson–Darling, Lilliefors-corrected Kolmogorov–Smirnov as well as visual inspection of diagnostic plots (histograms, density and Q-Q plots); homoscedasticity using the studentized