Urinary Metabolic Biomarkers of Attentional Control in Children With Attention‐Deficit/Hyperactivity Disorder: A Dimensional Approach Through ¹H NMR‐Based Metabolomics

ABSTRACT

Enhancing the understanding of attention‐deficit/hyperactivity disorder (ADHD) by linking biological processes with behavioral manifestations is a primary objective of the Research Domain Criteria (RDoC) framework, which aims to transcend traditional diagnostic categories and enable a more precise understanding of mental disorders. This study aimed to replicate five data‐driven profiles of attentional control in school‐aged children and, for the first time, to explore associated metabolic biomarkers. Understanding these profiles and their biological underpinnings can become critical for improving ADHD diagnosis and developing new targeted interventions. A clinically well‐characterized sample of 83 children with (n = 37) and without (n = 46) diagnosed ADHD completed a virtual reality continuous performance test (VR‐CPT) and provided urine samples for analysis. Clustering analyses of VR‐CPT data identified and replicated five distinct attentional control subgroups, two of which—ADHD‐IMP and ADHD‐SP—exhibited clinically significant impairments in attention and hyperactivity but opposite performance profiles in response inhibition and latency of response. NMR‐based metabolomics further revealed that children in the ADHD‐IMP subgroup exhibited a distinct urinary metabolic signature, with alterations in metabolites such as 3‐indoxylsulfate, N‐phenylacetylglycine, 3‐methyl‐2‐oxovalerate, creatine, creatinine, pseudouridine, and trigonelline. These compounds are potentially linked to microbial activity, energy metabolism, and oxidative stress, biological pathways increasingly recognized in ADHD pathophysiology. Although no direct association emerged between these metabolites and behavioral clusters, combining both data types using machine learning, particularly Logistic Regression, substantially improved classification accuracy compared to using behavioral data alone. These findings highlight the potential of integrating behavioral and molecular markers to refine ADHD characterization and move toward more individualized approaches.

This study identified five attentional control subgroups in school‐aged children, with and without ADHD, using VR‐based behavioral tests. NMR metabolomics revealed subgroup‐specific urinary metabolic profiles linked to energy and amino acid metabolism. Combining behavioral and metabolic data with machine learning significantly improved ADHD classification, underscoring the potential of integrating molecular and behavioral markers for advancing ADHD diagnosis.

Abbreviations

AB

AdaBoost

ADHD

attention‐deficit/hyperactivity disorder

ADHD‐C

ADHD‐combined

ADHD‐IMP

ADHD‐impulsive

ADHD‐IN

ADHD‐inattentive

ADHD‐RS‐5

attention deficit/hyperactivity disorder rating scale

ADHD‐SP

ADHD‐slow processing

ASD

autism spectrum disorder

ATP

adenosine triphosphate

AUC

Area Under the Curve

BCAA

branched‐chain amino acid

BRIEF‐2

Behavior Rating Inventory of Executive Function‐2

CBCL/6–18

Child Behavior Checklist/6–18

DSM‐5‐TR

Diagnostic and Statistical Manual of Mental Disorders fifth edition

FC

fold changes

GB

gradient boosting

HPLC

high‐performance liquid chromatography

IQ

intelligence quotient

k‐NN

k‐nearest neighbors

LR

Logistic Regression

MPH

methylphenidate

MVDA

multivariate data analysis methods

NaN₃

sodium azide

ODD

Oppositional Defiant Disorder

PCA

principal component analysis

PLS‐DA

partial least squares discriminant analysis

PUFAs

polyunsaturated fatty acids

RF

random forest

RDoC

Research Domain Criteria

ROC

Receiver Operating Characteristic

SDRT

standard deviation of reaction time

SDQ

Strengths and Difficulties Questionnaire

SVM

Support Vector Machine

TD

typically developing

TSP

2,2,3,3‐d4‐(trimethylsilyl)propanoic acid

VR‐CPT

virtual reality continuous performance test

XGB

XGBoost

1. Introduction

Attention‐deficit/hyperactivity disorder (ADHD) is a common neurodevelopmental condition affecting around 5% of school‐aged children worldwide, characterized by persistent patterns of inattention, impulsivity, and hyperactivity [1]. As outlined in the Diagnostic and Statistical Manual of Mental Disorders (DSM‐5‐TR) [2], ADHD is traditionally classified into three presentations: predominantly inattentive (ADHD‐IN), predominantly hyperactive–impulsive (ADHD‐HI), and combined presentation (ADHD‐C), based on the predominance of specific symptoms. However, this classification system has faced increasing criticism for its limited clinical specificity [3], inability to capture the full spectrum of symptom variability [4], and contribution to high comorbidity rates [5] and overdiagnosis [6], limitations that complicate both accurate diagnosis and personalized treatment planning.

Given these limitations, there has been a growing shift toward more dimensional approaches in ADHD research, aiming to capture the disorder's complexity through cognitive, behavioral, and biological dimensions [7]. This paradigm shift aligns with initiatives such as the National Institute of Mental Health's Research Domain Criteria (RDoC) [8], which advocates for moving beyond traditional diagnostic categories and focusing on the continuous nature of ADHD traits (for a detailed review, see Nigg et al. [9]). Such approaches are designed to yield more clinically meaningful subgroups and to facilitate the identification of reliable biological markers that could improve diagnosis and treatment strategies [9]. Despite decades of research identifying multiple neurobiological pathways, especially within dopaminergic and noradrenergic circuits [10], ADHD still lacks clinically useful biomarkers [11, 12], a persistent gap largely attributed to the reliance on traditional diagnostic frameworks.

Among the various biological levels being explored, metabolomics has emerged as a particularly promising tool. This field uses untargeted profiling of small molecules in biological samples to detect metabolic alterations that may be linked to specific genetic, cellular, or behavioral phenotypes [13]. By examining metabolic profiles alongside clinical and neuroimaging data, metabolomics could refine ADHD classification, predict treatment response, and contribute to more personalized, biologically informed interventions [14, 15, 16].

Despite its potential, metabolomics research in ADHD remains in early stages, largely limited to case–control studies. To date, studies employing metabolomics techniques such as liquid chromatography–high‐resolution mass spectrometry (LC‐HRMS), liquid chromatography–tandem mass spectrometry (LC–MS/MS), and high‐performance liquid chromatography (HPLC) have produced varied results due to the heterogeneity within ADHD subtypes [9], variability in analytical methodologies, and the use of diverse biological matrices [17]. In serum, while some authors report positive associations between homocysteine levels and hyperactivity‐impulsivity [18], others show no significant correlations [19, 20]. Plasma studies show either no significant metabolite differences,‐ [21] or correlations between ADHD symptoms and metabolites like o‐phosphoethanolamine, 4‐aminohippuric acid, 5‐hydroxylysine, L‐cysteine, tryptophyl‐phenylalanine, and gentisic acid [22]. Urine, though less explored and analytically complex due to individual heterogeneity, has shown a similar trends, with some finding no significant associations [23, 24, 25] and others identifying potential biomarkers such as FAPy‐adenine, 3‐methylazelaic acid, and phenylacetylglutamine [26]. Notably, only one study has applied nuclear magnetic resonance (NMR) spectroscopy, identifying elevated hippurate levels in male ADHD twins [27].

The current study aims to address these gaps by integrating data‐driven behavioral phenotyping with untargeted metabolomic profiling in a pediatric ADHD sample. Building on prior research, where we attempted to improve ADHD classification by identifying attentional control profiles through clustering techniques [28], in this study, we expand upon that work by incorporating proton nuclear magnetic resonance (¹H NMR) spectroscopy to analyze urine samples. This allows us to explore how these behavioral profiles correlate with specific metabolic signatures, in line with the RDoC framework.

Specifically, this study aimed to (i) replicate attentional control profiles using clustering methods [28], (ii) examine metabolic differences among these data‐driven profiles, (iii) assess the capacity of these metabolic markers to predict profile categorization among children with ADHD, and, finally, (iv) compare the classification performance of various machine learning models using behavioral data alone and in combination with metabolic data. To this end, we applied ¹H NMR spectroscopy on urine samples, a non‐invasive, accessible method well‐suited for pediatric populations [29]. Despite the exploratory nature of this study, we hypothesized that: (i) attentional control profiles would be successfully replicated using clustering analyses [28], (ii) cluster profiles with greater clinical impairments would exhibit distinct metabolic signatures, (iii) metabolic markers would significantly predict behavioral subgroup classification, and (iv) models integrating statistically significant metabolites and behavioral data would yield improved diagnostic accuracy compared to behavioral data alone.

2. Materials and Methods

2.1. Procedure

Participants in this study were part of a larger research project investigating executive functions and neurobiological correlates in pediatric ADHD. Recruitment targeted families with and without ADHD‐diagnosed children via advertisements and mailing lists from private clinics, a public hospital's mental health and neurology departments, and educational centers (both public and private) in Almería, southern Spain. Families expressing interest completed a preliminary phone interview to verify eligibility, followed by an in‐person clinical assessment conducted by a licensed clinical psychologist with a doctoral degree.

During the assessment, parents completed a clinical interview and a set of behavioral rating scales to determine DSM‐5 criteria for ADHD [Attention Deficit/Hyperactivity Disorder Rating Scale (ADHD‐RS‐5) [30], Conners 3 ADHD Index [31], Strengths and Difficulties Questionnaire (SDQ) [32], Child Behavior Checklist/ 6–18 (CBCL/6–18) [33], and Behavior Rating Inventory of Executive Function‐2 (BRIEF‐2) [34]]. Whenever possible, clinical and school records were also reviewed to support the diagnosis.

Assessments were conducted individually, predominantly at the University of Almería's Psychology Laboratory, and occasionally in designated clinic and school spaces. Each assessment began with the assessment of attentional processes using a virtual continuous performance test (CPT). On the same day, urine samples were collected at the study facility.

Parents/legal guardians, and children over 12 years of age provided verbal and written informed consent. Ethical approval was obtained from the Bioethics Committee of Human Research of the University of Almería [UALBIO2017/018] and the Provincial Research Ethics Committee of the Torrecárdenas University Hospital [PSI2015‐70037‐R]. All procedures complied with the World Medical Association Declaration of Helsinki, the EU General Data Protection Regulation (2016/679) and Spain's Personal Data Protection and Digital Rights Law (3/2018). Participating families received a summary report of the assessment results.

2.2. Participants

Our sample consisted of 37 children with ADHD and 46 typically developing (TD) peers between 7 and 16 years old (M _age = 12.01, SD_age = 2.82, 40.96% girls). One experienced psychologist rated ADHD diagnosis as “present,” “subthreshold,” or “absent” according to DSM‐5 criteria following the abovementioned diagnostic procedure. As children with subthreshold ADHD experience incapacitating symptoms, although they do not reach the cut‐off criteria (they present three to five criteria) [35], they were included in the ADHD group. TD participants did not have a psychiatric history. We excluded children from both groups if they have a neurological illness, traumatic brain injury, or genetic disorders; a diagnosis of intellectual disability, autism spectrum disorder (ASD), or psychosis; sensory or motor impairments that prevent completion of the task; or intelligence quotient (IQ) < 70. Fifteen children (five ADHD and 10 TD) had also participated in Fernández‐Martín et al. [28]

Demographic and clinical characteristics are detailed in Table S1. ADHD and TD children did not differ in age (W = 907, p = 0.611) or sex distribution (χ ² [1] = 3.484, p = 0.062) but did in IQ (t(77.722) = 4.366, p < 0.001). The ADHD group scored significantly higher than the TD group on the SDQ and CBCL/6–18 scales. Seventeen ADHD children were on stimulant treatment with MPH. Ten children were treated with the prolonged‐release form of MPH (Concerta, Rubricrono; duration of action ~1–8 h), five children with the modified release (Medikinet; duration of action ~1–4 h), and two children with the immediate release (Rubifén, duration of action ~1–4 h). Medication was discontinued at least 24 h before the assessment as long as MPH improves CPT performance [36]. This timeframe also facilitated metabolic analyses since approximately 90% of the MPH dose is excreted in the urine within 16 h, primarily in the form of ritalinic acid [37].

2.3. Measures

2.3.1. Attentional Control

We used the virtual CPT “AULA” (“classroom” in English) [38], which assess attentional control, processing speed, impulsivity, and head‐motor activity in a classroom environment. A detailed description of this test can be found in Fernández‐Martín et al. [28].

2.3.2. Clinical Outcomes

2.3.2.1. SDQ

Parents completed the SDQ [32], an international and reliable scale to screen emotional and behavioral problems in children and adolescents aged 4 to 17 years. It contains 25 items divided between five scales: emotional symptoms, conduct problems, inattention/hyperactivity, peer relationship problems, and prosocial behavior. Normative data establishes a cutoff score of 17 (90th to 95th percentile) as indicative of significant impairments, whereas scores above 20 (≥ 95th percentile) are considered clinically significant [39].

2.3.2.2. Child Behavior Checklist (CBCL/6–18)

The CBCL/6–18 [33] is a widely used 113‐item questionnaire to assess the frequency and severity of emotional and behavioral problems in children and adolescents aged 6 to 18 years. The CBCL/6–18 covers a broad range of areas, including eight empirically based dimensions based on factor analyses (anxious/depressed, withdrawn/depressed, somatic complaints, social problems, thought problems, attention problems, rule‐breaking behavior, and aggressive behavior) that are grouped into three subscales of internalizing, externalizing, and total problems. CBCL/6–18 also includes six DSM‐oriented scales that comprise items consistent with DSM‐5 categories (depressive problems, anxiety problems, somatic problems, attention deficit/hyperactivity problems, oppositional defiant problems, and conduct problems). The CBCL/6–18 is standardized by sex and age [33]. T scores above 63–65 are considered clinically significant, while scores falling within the 60–63 range are considered borderline.

2.3.3. Urinary Sample Collection and Preparation for NMR

Children's urine samples were received and immediately frozen at −80°C to preserve their metabolome unaffected. Three samples from children with ADHD were missing, so urine preprocessing and analyses were applied to 80 participants (34 ADHD vs. 46 TD).

These urine samples were thawed for 30 min, and then centrifuged at 10,500 rpm at 4°C for 5 min. The 630 μL of the supernatant was then extracted in 1.5‐mL Eppendorf tubes using 70 μL of 1.5‐M phosphate buffer in D₂O (pH 7.4, containing the sodium salt of 2,2,3,3‐d ₄ ‐(trimethylsilyl)propanoic acid [TSP, 0.1% p/v], and sodium azide [NaN₃, 90 μM, an enzymatic inhibitor]). Finally, 600 μL of the mixture was transferred to 5‐mm NMR tubes (Eurisotop, Saint‐Aubin, France).

2.3.4. ¹H NMR Experiments

A Bruker Avance III 600‐MHz NMR spectrometer equipped with a 5‐mm QCI quadruple resonance pulse field gradient cryoprobe was employed for all ¹H NMR and two‐dimensional spectra. The experiment conditions were set at 300.0 ± 0.1 K without rotation, employing a temperature‐controlled SampleJet autosampler with up to 480 positions. The optimized acquisition parameters are those already reported in Tristán et al. [40] For structural elucidation of the metabolite set present in children's urine, two‐dimensional experiments including ¹H–¹H COSY, ¹H–¹H TOCSY, ¹H–¹³C HSQC, and ¹H–¹³C HMBC of a representative sample were recorded using standard Bruker sequences as described in Tristán et al. [40] Additionally, this identification of the metabolic profile was also supported by the assistance of public databases—The Human Metabolome Database (HMDB, http://www.hmdb.ca) and Complex Mixture Analysis Database (COLMAR, http://spin.ccic.ohio‐state.edu/index.php/colmar)—and private ones—Chenomx, Metabolite Reference Database (Bruker, bbiorefcode)—and literature searches.

2.4. Statistical Analyses

To illustrate the limitations of traditional ADHD diagnosis—the rationale for adopting the dimensional approach in the present study—we first compared groups on behavioral (Table S2 and Figure S1) and molecular measures (Table S3). As shown in the Supporting Information, traditional subtypes lacked specificity in distinguishing profiles within the ADHD group, supporting the need for data‐driven approaches.

2.4.1. Data‐Driven Attentional Control Profiles

Statistical analyses were conducted using R software [41]. We applied the same clustering procedure described in Fernández‐Martín et al. [28] to identify data‐driven attentional control profiles, conducting hybrid hierarchical k‐means clustering on the normalized t scores of the AULA test's main indices: Omissions, SDRT (standard deviation of reaction time), deviation of attentional focus, mean RT, commissions, and head movements. For further details on the selection and validation of the cluster cut‐off, see Fernández‐Martín et al. [28].

For group‐level comparisons, when data violated statistical assumptions, we applied robust models using 10% trimmed means and 2000 bootstrap samples [42, 43, 44]. Post hoc tests were adjusted for multiple comparisons, utilizing the Benjamini–Hochberg correction for robust models and Bonferroni correction for non‐robust ANOVA models. Statistical significance was set at p < 0.05.

2.4.2. Bucketing of ¹H NMR Spectral Data

The bucketing process of ¹H NMR spectra was performed using AMIX 3.9.12 software (Bruker BioSpin GmbH, Rheinstetten, Germany). Buckets were generated through a simple bucketing process (buckets of 0.04 ppm), and individual peak intensities were normalized to the total intensity in the δ _H 0.5 to 10 ppm region. The region containing the residual signal of water suppression (δ _H 4.64–4.80 ppm) was excluded from the bucket table employed in the analyses.

2.4.3. Multivariate and Univariate Metabolomics Data Analyses

Multivariate data analysis methods (MVDA) were applied through the SIMCA‐P software (v. 15.0, Umetrics, Sweden) to determine whether the urinary metabolic profile showed possible biomarkers of data‐driven profiles. Both unsupervised models, such as principal component analysis (PCA), and supervised models, such as partial least squares discriminant analysis (PLS‐DA), were applied to the urinary ¹H NMR spectral data through Unit Variance, along with some univariate analysis methods like Volcano Plot, which combines t test results and fold changes (FC).

Then, the same data matrices were statistically assessed through univariate data analysis methods, employing the online tool MetaboAnalyst [45]. FC and Log₂(FC) for discriminant metabolites among cluster profiles were estimated along with Wilcoxon rank‐sum tests to determine the significance of the metabolites (p < 0.05). All the metabolites that significantly changed between groups regardless of the FC value were considered. Instead of conducting all possible pairwise comparisons, we focused on those that align with the research objectives. We conducted t‐tests among the data‐driven profiles of clinical significance, that is, those with tscores above the clinical cut‐off (> 60), aiming to identify metabolic alterations that might distinctly contribute to distinct attentional control deficits. We also performed comparisons among each clinically significant cluster and normative profiles to identify metabolic changes associated with poor attentional control.

2.4.4. Assessment of Predictability Through Machine Learning

To assess whether clinical outcomes could be improved by the inclusion of the statistically significant metabolites identified, various machine learning models were evaluated using Python (v. 3.11.7) [46], and the following libraries: Pandas, Sklearn, XGBoost, NumPy, Seaborn, and Matplotlib. Specifically, random forest (RF), Support Vector Machine (SVM), logistic regression (LG), gradient boosting (GB), k‐nearest neighbors (k‐NN), Naïve Bayes (NB), XGBoost (XGB), and AdaBoost (AB) algorithms were implemented both with and without the presence of the metabolites. Performance metrics such as accuracy, precision, recall, and F1 score were computed for each model. Additionally, Receiver Operating Characteristic (ROC) curves were generated to assess model sensitivity, with the Area Under the Curve (AUC) values reported for each of the models. The model demonstrating the highest performance was further validated using cross‐validation techniques, incorporating confusion matrices to evaluate its robustness.

3. Results

3.1. Data‐Driven Attentional Control Profiles

We visually inspected AULA performance profiles across three‐ to five‐cluster solutions, confirming that each clustering structure—three, four, and five clusters—reflected a consistent distribution of ADHD and TD participants and replicated performance profiles across AULA indices (Figure S2). For k = 3, hybrid k‐means analyses identified one low‐performing subgroup constituted by 59.38% of ADHD participants; one subgroup with average scores (constituted by 75% TD participants); and one subgroup with intact performance but elevated SDRT and slow RT (formed by 52.17% TD participants). For k = 4, the low‐performing subgroup divided into two phenotypic ADHD subgroups, constituted by 83.33% and 45.83% ADHD participants, respectively. For k = 5, the average subgroup further split into average and high‐performance subgroups. In this cluster solution, 13 out of 15 participants (83.67%) who participated in the original and the present study were categorized in the same cluster.

Based on the majority rule among 30 clustering validation indices, the four‐cluster structure emerged as the optimal solution for explaining CPT performance among ADHD and TD participants (Figure S3), showing the highest internal consistency according to 13 well‐validated indices (Table S4). Although a smaller sample size in this study might have favored a four‐cluster solution to avoid small subgroups, the consistency of clustering structures across studies led us to retain the five‐cluster solution for deeper analysis of clinical and metabolic correlates.

The five data‐driven profiles were labeled using the same categorization as in Fernández‐Martín et al. [28], following the clinical cut‐off points provided by the validation study (Figure 1A). The clinically impaired subgroups ADHD‐slow processing (ADHD‐SP, n = 12) and ADHD‐impulsive (ADHD‐IMP, n = 15) were constituted by 83.33% and 46.67% of ADHD participants respectively. Average (n = 18) and high (n = 17) performers were constituted by 61.11% and 88.24% of TD participants, while sluggish performance (n = 21) was constituted by 52.68% of ADHD participants. Figure 1B illustrates the percentage distribution of participants from each cluster in ADHD and control groups. Robust one‐way ANOVA showed statistically significant differences in all AULA outcome measures among clusters, yielding large effect sizes. Test statistics and mean differences for each post hoc comparison are included in Table S5.

FIGURE 1.

Attentional control profiles measured by the virtual CPT AULA according to the five‐cluster solution. (A) 10% trimmed mean values of AULA main indices (t scores): omission errors, the standard deviation of reaction time (SDRT), time deviating the attentional focus from the blackboard, mean RT, commission errors, and total head movements. Error bars represent the 10% trimmed SEM. t scores ≥ 61 represent a clinically low performance. Dashed lines indicate cut‐offs for risk of attention problems (> 60 = at risk; > 70 = high risk). (B) Percentage distribution of each cluster in ADHD and control groups.

3.2. Clinical Correlates

Cluster's demographic and clinical features are presented in Table S6. We found that DSM‐5 subtypes for ADHD were similarly distributed across cluster profiles. Clusters significantly differed in age (F [4] = 4.50, p = 0.002, η ²   = 0.19) and IQ (F [4] = 4.97, p = 0.001, η ²   = 0.20) but did not in sex distribution (p = 0.37). Post hoc comparisons using Bonferroni correction showed that the ADHD‐IMP cluster was the youngest subgroup and high performers had the highest IQ.

Significant differences among cluster profiles were found in the inattention/hyperactivity subscale (F [4] = 2.56, p = 0.04, η ²   = 0.12) of the SDQ, and in the attention problems (F [4] = 3.04, p = 0.02, η ²   = 0.14), DSM ODD problems (F [4] = 3.43, p = 0.01, η ²   = 0.15), and DSM ADHD (F [4] = 3.39, p = 0.01, η ²   = 0.15) subscales of the CBCL/6–18. Adjusted post hoc tests showed that the ADHD‐SP subgroup had significantly higher scores on the inattention/hyperactivity subscale (p = 0.03) of the SDQ, and the attention problems (p = 0.01) and DSM ADHD (p = 0.005) subscales of the CBCL/6–18, than participants with high performance. ADHD‐IMP participants showed a significantly lower score on the DSM ODD problems (p = 0.03) subscale of the CBCL/6–18 than average performers. From a clinical point of view, the ADHD‐SP subgroup obtained borderline scores on the inattention/hyperactivity, emotional, peer problems, and total scales of the SDQ. This subgroup also reached > 98 scores on the attention problems subscale and borderline scores on the DSM ADHD subscales of the CBCL/6–18.

3.3. Urinary Biomarkers

3.3.1. Metabolic Profiling of Urine Samples

Urine samples were prepared as detailed in Section 2 of the present study. Figure 2 provides a comprehensive spectrum of a typical urine sample, offering a visual representation of all identified compounds.

FIGURE 2.

Representative ¹H NMR spectra of a urine sample. Assignments—(1) 3‐methyl‐2‐oxovalerate; (2) leucine; (3) isoleucine; (4) valine; (5) 3‐hydroxyisobutyrate; (6) isopropanol; (7) 3‐hydroxybutyrate; (8) fucose; (9) methylmalonate; (10) 3‐hydroxyisovalerate; (11) threonine; (12) lactate; (13) 2‐hydroxyisobutyrate; (14) 2‐phenylpropionate; (15) alanine; (16) lysine; (17) arginine; (18) acetate; (19) glutamate; (20) methionine; (21) glutamine; (22) acetone; (23) succinate; (24) pyruvate; (25) citrate; (26) DMA (dimethylamine); (27) TMA (trimethylamine); (28) creatinine; (29) creatine; (30) carnitine; (31) TMAO (trimethylamine‐N‐oxide); (32) taurine; (33) betaine; (34) phenylalanine; (35) methanol; (36) 4‐hydroxyphenylacetate; (37) glycine; (38) N‐phenylacetylglicine; (39) 3‐methylhistidine; (40) arabinose; (41) pseudouridine; (42) trigonelline; (43) cis‐aconitate; (44) urea; (45) tyrosine; (46) histidine; (47) 3‐indoxylsulfate; (48) tryptophan, (49) hippurate; (50) hypoxanthine; (51) formate.

Detailed information on the chemical shifts, coupling constants, and multiplicities for each metabolite is summarized in Table 1.

TABLE 1.

Metabolite assignment in urine sample extracts by ¹H NMR and diagnostic signals. Abbreviations: s = singlet; d = doublet; t = triplet; dd = doublet of doublets; m = multiplet; q = quadruplet; bs = broad signal.

Metabolites	Chemical shift (δ H, ppm) and J (Hz)
Amino acids
Valine	0.99 (d, J = 7.2 Hz), 1.04 (d, J = 7.2 Hz)
Isoleucine	0.94 (t), 1.00 (d, J = 7.2 Hz), 1.31 (m)
Leucine	0.95 (d), 0.96 (d), 1.72 (m)
Threonine	1.34 (d)
Lysine	1.45 (m), 1.53 (m), 1.74 (m), 1.91 (m), 3.04 (m)
Alanine	1.49 (d, J = 7.4 Hz)
Arginine	1.66 (m), 1.73 (m), 1.92 (m), 3.23 (t)
Glutamate	2.04 (m), 2.28 (m)
Glutamine	2.14 (m), 2.46 (m)
Tyrosine	6.87 (m), 7.17 (m)
Phenylalanine	7.32 (m), 7.37 (m), 7.42 (m)
Creatine	3.04 (s), 3.94 (s)
Trigonelline	4.52 (s), 8.86 (m) 9.00 (m), 9.30 (s)
Glycine	3.58 (s)
Taurine	3.27 (t), 3.43 (t)
Histidine	3.16 (dd), 3.26 (dd), 4.00 (dd), 7.15 (s), 8.02 (d)
Tryptophan	7.34 (m), 7.51 (d, J = 8 Hz), 7.74 (d, J = 8 Hz)
Organic acids
Acetate	1.93 (s)
Lactate	1.33 (d, J = 7.0 Hz), 4.13 (q, J = 7.0 Hz)
Succinate	2.34 (s)
Pyruvate	2.41 (s)
Citrate	2.54 (d, J = 15.8 Hz), 2.69 (d, J = 15.8 Hz)
Formate	8.46 (s)
Sugars
Fucose	1.16 (d), 1.22 (d), 4.57 (d), 5.25 (d)
Arabinose	4.53 (d), 5.25 (d)
Others
3‐methyl‐2‐oxovalerate	0.89 (t), 1.10 (d)
3‐hydroxyisobutyrate	0.92 (d)
Isopropanol	1.15 (d)
Methylmalonate	1.25 (d)
3‐hydroxybutyrate	1.16 (d), 2.31 (dd), 2.37 (dd), 4.14 (m)
3‐hydroxyisovalerate	1.27 (s)
2‐phenylpropionate	1.32 (d), 7.29 (m), 7.35 (m), 7.38 (m)
2‐hydroxyisobutyrate	1.36 (s)
Acetone	2.21 (s)
Methionine	2.13 (s), 2.61 (m)
Carnitine	2.45 (m), 3.16 (s), 3.41 (m)
DMA	2.73 (s)
TMA	2.87 (s)
Creatinine	3.07 (s), 4.06 (s)
Cis‐aconitate	3.13 (d), 5.69 (m)
3‐methylhistidine	3.94 (s), 7.07 (s), 7.91 (s)
Pseudouridine	4.18 (t), 4.30 (t), 7.70 (d)
4‐hydroxyphenylacetate	3.48 (s), 7.00 (m), 7.21 (m)
3‐indoxylsulfate	7.20 (m), 7.27 (m), 7.36 (s), 7.49 (m), 7.70 (m)
Hippurate	7.55 (m), 7.64 (m), 7.84 (m)
Betaine	3.27 (s), 3.95 (s)
TMAO	3.28 (s)
Methanol	3.36 (s)
N‐phenylacetylglycine	3.66 (s), 3.74 (d), 7.35 (m), 7.42 (m)
Hypoxanthine	8.20 (s), 8.23(s)
Urea	5.80 (bs)

3.3.2. Multivariate Data Analysis

A PCA model was applied to the ¹H NMR data set to determine possible underlying trends or groupings among the urine samples. Specifically, we aimed to visualize possible differences between traditional ADHD diagnoses (ADHD or control groups), and across the five data‐driven profile groups (sluggish, ADHD‐IMP, high, average, and ADHD‐SP groups). Figure 3 shows the obtained PC1 versus PC2 scores plots for urine samples, explaining 15.95% and 8.34% of the total variance, respectively, with an unclear separation for either the ADHD diagnosis or the data‐driven profile groups.

FIGURE 3.

PCA scores plot applied to the urine ¹H NMR dataset, scaled to Unit Variance. Different data‐driven profile groups were illustrated (sluggish: green; ADHD‐IMP: dark blue; high: red; average: yellow; ADHD‐SP: light blue), as well as the traditional ADHD diagnosis (ADHD: Δ, control: Ο).

To explore possible metabolic differences and similarities among the data‐driven profiles across urine samples, supervised methods were also conducted, specifically PLS‐DA models. As can be observed in Figure S4A, the samples did not group depending on the data‐driven profiles, which was supported by the model‐associated quality parameters. Then, another PLS‐DA model was also applied to other variables, including the traditional ADHD diagnosis (Figure S4B) and the traditional ADHD subtypes (Figure S4C), which also showed a lack of gathering among the samples. In all three cases—data‐driven profiles, traditional ADHD diagnosis, and traditional ADHD subtypes—the PLS‐DA models demonstrated signs of overfitting (CV‐ANOVA = 1), indicating a lack of predictive capacity.

3.3.3. Univariate Data Analysis

Wilcoxon rank‐sum tests were applied to explore possible differences among the cluster profile groups of the present study from a univariate data analysis perspective. When focusing on ADHD‐SP in comparison to ADHD‐IMP, some metabolites seemed to be significantly downregulated (lower) in the second group, namely, N‐phenylacetylglycine (p < 0.01), 3‐indoxylsulfate (p < 0.01), and 3‐methyl‐2‐oxovalerate (p < 0.05). On the other hand, these analyses also showed that the ADHD‐IMP group had significantly upregulated the urinary levels of creatine (p < 0.01) and downregulated those of creatinine (p < 0.05) in comparison to average and high performers. Besides, the ADHD‐IMP cluster exhibited significantly upregulated urinary levels of pseudouridine (p < 0.01), N‐phenylacetylglycine (p < 0.01), and 3‐indoxylsulfate (p < 0.05), while also showing downregulated urinary levels of trigonelline (p < 0.05) compared to average performers. Wilcoxon rank‐sum tests did not reveal significant differences between the ADHD‐SP cluster profile and both groups of good performers (average and high). All these univariate analysis results are detailed in Table 2, accompanied by their respective FC values. Additionally, box plots illustrating all group comparisons are provided in Figure S5.

TABLE 2.

Significant metabolic changes between attentional control data‐driven profiles in children's urine. The ¹H NMR chemical shift of each discriminant metabolite (p < 0.05), FC, and Log2(FC) is provided.

Metabolites	δ H (ppm)	FC	Log2(FC)	p
ADHD‐IMP vs. ADHD‐SP
N‐Phenylacetylglycine	7.42	0.669	−0.580	0.004
3‐Indoxylsulfate	7.50	0.711	−0.491	0.005
3‐Methyl‐2‐oxovalerate	1.10	0.817	−0.292	0.015
ADHD‐IMP vs. average performers
Pseudouridine	4.30	1.138	0.185	0.003
N‐phenylacetylglycine	3.66	1.136	0.185	0.004
Creatine	4.06	1.227	0.296	0.006
Creatinine	3.94	0.728	−0.458	0.030
3‐indoxylsulfate	7.50	1.342	0.424	0.027
Trigonelline	8.86	0.690	−0.536	0.038
ADHD‐IMP vs. high performers
Creatine	4.06	1.215	0.282	0.007
Creatinine	3.94	0.691	−0.534	0.007
ADHD‐SP vs. average performers
—	—	—	—	—
ADHD‐SP vs. high performers
—	—	—	—	—

3.3.4. Assessment of Predictability Through Machine Learning

Following the RDoC framework, we combined those significant metabolites derived from the univariate analyses with ADHD data‐driven profiles defined by specific points (Figure 1A). The goal was to assess how this integration impacts machine learning model performance, exploring the potential of metabolomics in ADHD diagnosis.

For this purpose, four performance metrics—accuracy, precision, recall, and F1 score were—calculated for eight different ML models. Each model was evaluated using two distinct datasets: a dataset (1) with behavioral data from the AULA virtual test (omissions, SDRT, deviation of attentional focus, mean RT, commissions, and head movements) and a dataset (2) combining these behavioral parameters with the significant metabolites earlier identified in the univariate study and given in Table 2 (trigonelline, 3‐indoxylsulfate, pseudouridine, creatine, creatinine, N‐phenylacetylglycine, and 3‐methyl‐2‐oxovalerate).

The combination of metabolites with behavioral parameters generally improved model performance, particularly for Logistic Regression, which achieved the highest overall results. When metabolites were added, as illustrated in Table 3, LR improved from an accuracy of 87.5%, precision of 90.6%, recall of 87.5%, and F1 score of 88.1% (clinical markers only) to an accuracy of 93.7%, precision of 94.4%, recall of 93.7%, and F1 score of 93.4%. This remarkable increase underscores the potential of integrating metabolites with behavior indicators to enhance the predictive accuracy of attentional control deficit profiles associated with ADHD.

TABLE 3.

Associated quality statistical parameters in % (accuracy, precision, recall, and F1 score) for each machine learning model applied to Datasets (1) and (2). The model with higher performance values is highlighted in bold.

	Behavioral parameters (1)				Behavioral parameters + metabolites (2)
ML models	Accuracy	Precision	Recall	F1	Accuracy	Precision	Recall	F1
Logistic regression	87.5	90.6	87.5	88.1	93.7	94.4	93.7	93.4
Random forest	75.0	80.3	75.0	76.3	68.7	78.1	68.7	70.9
Support Vector Machine	75.0	86.9	75.0	78.3	81.2	88.5	81.2	81.3
Gradient boosting	81.3	89.3	81.2	81.6	81.2	89.5	81.2	83.4
k‐nearest neighbors	68.8	80.2	68.7	71.0	87.5	92.5	87.5	88.4
Naïve Bayes	81.3	90.8	81.2	83.2	81.2	86.9	81.2	80.7
XGBoost	62.5	75.9	62.5	65.4	75.0	86.4	75.0	77.3
AdaBoost	68.8	78.1	68.7	70.9	68.7	78.1	68.7	70.9

The k‐NN model showed significant improvement with the addition of metabolites: accuracy increased from 68.8% to 87.5%, precision from 80.2% to 92.5%, recall from 68.7% to 87.5%, and F1‐score from 71.0% to 88.4%. Similarly, the XGBoost and SVM models improved across all metrics when moving from Dataset (1) to Dataset (2), with accuracy increases of 12.5% and 6.2%, respectively, and precision increases of 10.5% and 1.6%, respectively. These results support the value of incorporating metabolites for more accurate prediction of ADHD children's performance in the VR‐CPT test. However, not all models responded uniformly to the addition of metabolites. Random forest and Naïve Bayes saw a decline in performance, while GB and AdaBoost showed only marginal changes across the datasets.

All model performances were also evaluated using ROC curves. Figure 4 displays the ROC curves and AUC values for each of the eight ML models. LR achieved the highest AUC (0.99), indicating excellent class discrimination, with its ROC curve positioned near the upper left corner—an ideal outcome. GB also performed exceptionally well, with an AUC of 0.98, slightly lower than LR but still highly effective. Both the SVM and GB models demonstrated reliable classification, with AUCs of 0.97 and 0.98, and ROC curves close to optimal. By contrast, k‐NN, Naïve Bayes, and AdaBoost had lower AUC values, making them less effective for predicting ADHD children's performances in the VR‐CPT test.

FIGURE 4.

Multiclass ROC curves for the eight machine learning models generated using combined metabolite and behavioral markers as input data. The respective AUC values for each model are also shown, illustrating their effectiveness in class discrimination.

Given that LR demonstrated the best performance among models, we aimed to evaluate its predictive ability for attentional control deficits associated with ADHD through cross‐validation. Figure S6 presents the corresponding confusion matrices, comparing LR results across Datasets (1) and (2) (one with only behavioral variables and the other including both behavioral variables and metabolites, respectively), as well as metrics related to cross‐validation. For the first dataset (behavioral variables only), the LR confusion matrix highlights strong accuracy in predicting the sluggish, ADHD‐IMP, average, and ADHD‐SP classes, with no misclassifications in ADHD‐IMP, average, or ADHD‐SP. However, in the High class, two instances were misclassified as sluggish, indicating some overlap or difficulty differentiating these classes. Furthermore, the lower frequency of ADHD‐SP instances could affect the model's reliability in cross‐validation due to sample imbalance. The mean cross‐validation score was 91.25% (SD = 6.37%), with notable recall values: 87.5% for sluggish, 100% for ADHD‐IMP, high, and ADHD‐SP, and 75% for high. When metabolites were added in dataset (2), the LR model showed improved precision across most classes, particularly in sluggish, high, average, and ADHD‐SP. Some minor misclassifications were observed between sluggish and average, as well as high and average. Additionally, ADHD‐IMP had only one correctly classified instance, suggesting a potential issue due to sample size. The mean cross‐validation score increased to 92.5% (SD = 6.12%), with recall values reaching 100% for sluggish, ADHD‐IMP, average, and ADHD‐SP, and 75% for high.

4. Discussion

In the current study, we investigated whether data‐driven profiles of attentional control, identified through a virtual reality continuous performance task (“AULA”) [38], could reveal distinct urinary metabolomic signatures among children diagnosed with ADHD. While the traditional diagnosis was effective in differentiating children with and without ADHD, it failed to capture meaningful cognitive and metabolic variations within the ADHD group (see Supporting Information). In contrast, clustering analyses replicated five data‐driven attentional profiles [28], two of which—ADHD‐IMP and ADHD‐SP—were primarily composed of children with ADHD and exhibited distinct behavioral and metabolic features.

Consistent with our prior study [28], both ADHD‐IMP and ADHD‐SP children showed impairments in focused attention (omissions), inconsistent response times (SDRT), high distractibility (deviation from attention focus), and increased motor activity—patterns typically observed in ADHD [47, 48]. However, while ADHD‐IMP children were characterized by a high rate of impulsive responses, ADHD‐SP children showed notably slower reaction times. These findings suggest that inhibitory control and latency of response may be useful cognitive markers for distinguishing ADHD subgroups [28, 49, 50], whereas hyperactivity appears to be a shared feature across profiles [51]. Notably, both subgroups also displayed high ADHD symptom severity on the SDQ, with ADHD‐SP showing additional elevated scores in depressive and ODD symptoms—suggesting a potentially more severe, mixed internalizing/externalizing profile [52]. However, further replication in larger samples is needed to clarify their distinct psychopathological profiles.

Metabolomics analyses revealed that the ADHD‐IMP group—exhibiting the prototypical ADHD phenotype of inattention, impulsivity, and hyperactivity—displayed a distinct urinary metabolomic profile, marked by significantly downregulated levels of 3‐indoxylsulfate, N‐phenylacetylglycine, and 3‐methyl‐2‐oxovalerate compared to ADHD‐SP.

The 3‐indoxylsulfate—a tryptophan‐derived microbial metabolite produced via the PLP‐dependent (pyridoxal‐5′‐phosphate) activity of tryptophanase—has been previously found to be reduced in both untreated and Ritalin‐treated children with ADHD [53]. These reductions, along with low 3‐indoxylsulfate/tryptophan and 3‐indoxylsulfate/kynurenine ratios, may suggest a stable disruption in vitamin B6–dependent tryptophan metabolism, raising its potential as a trait‐like biomarker of the disorder. Beyond its microbial origin, 3‐indoxylsulfate acts as a neuroactive molecule capable of crossing the blood–brain barrier and modulating neural and behavioral processes. Dysregulated levels have been implicated in anxiety [54], depression [55], ASD [56], cognitive decline in chronic kidney disease [57], and Parkinson's disease [58], possibly due to its ability to trigger neuroinflammation [59] and oxidative stress [60], both mechanisms implicated in ADHD pathophysiology [15, 61].

Downregulated levels of N‐phenylacetylglycine—another gut microbial metabolite—have been linked to altered microbial metabolism [62], metabolic disturbances in liver injury [63], and probiotic‐induced shifts in ASD [64]. Similarly, 3‐methyl‐2‐oxovalerate, a product of isoleucine catabolism, crosses astroglial membranes via monocarboxylate transporter 1 and contributes to the glutamate–glutamine cycle or is further metabolized in glial cells [65]. While elevated levels are neurotoxic [66], the reductions observed in the ADHD‐IMP group may signal disrupted branched‐chain amino acid (BCAA) metabolism, potentially affecting mitochondrial function, neurotransmitter cycling, and brain energy homeostasis [65, 67]. Since BCAA metabolism is also modulated by gut microbial activity [68], these changes may reflect microbiota‐driven mechanisms linked to impulsive–hyperactive phenotypes.

Overall, our findings suggest that altered gut microbial metabolism may underlie core features of the impulsive–hyperactive ADHD profile. While we did not directly assess microbial composition, the consistent downregulation of key microbial‐derived metabolites (e.g., 3‐indoxylsulfate, N‐phenylacetylglycine, and 3‐methyl‐2‐oxovalerate) may reflect functional consequences of microbial dysbiosis—previously described in ADHD [69, 70]. These metabolic patterns may therefore serve as accessible biomarkers of gut–brain axis disruption. Although we did not control for diet or probiotic intake, prior evidence suggests that certain microbiota‐related alterations—including co‐metabolites and inflammatory markers—persist even after adjusting for nutritional factors [61, 71], supporting the notion that they reflect stable host–microbiome interactions rather than short‐term environmental influences. Furthermore, although stimulant medication can alter gut microbiota [72], the balanced distribution of medicated participants across ADHD profiles in our sample suggests that the observed metabolic differences are unlikely to be driven by treatment and may instead reflect trait‐like biological signatures of ADHD subgroups.

Additional differences between the ADHD‐IMP group and children with good performance were also identified. Compared to the latter, the ADHD‐IMP group showed increased urinary levels of 3‐indoxylsulfate and N‐phenylacetylglycine, along with elevated creatine and pseudouridine, and reduced creatinine and trigonelline. Creatine and creatinine are key regulators of cellular energy metabolism and influence GABAergic signaling and neuronal homeostasis [73, 74]. Similar imbalances have been observed in ASD, pointing to shared mechanisms involving oxidative stress and gut microbial pathways [29]. Neuroimaging studies have also reported elevated striatal creatine in treatment‐naive children with ADHD‐C, possibly reflecting an adaptive response to impaired dopaminergic control of glutamatergic transmission [72]. Dopamine deficits may result in glutamatergic spillover and excitatory stress, increasing the energy demands at synapses [74]. In this context, creatine accumulation may act as a buffer to sustain ATP availability under heightened synaptic load, aligning with the neuroenergetic hypothesis of ADHD, which proposes that catecholaminergic dysregulation compromises synaptic efficiency and increases energetic demands—particularly in fronto‐striatal circuits—driving symptoms such as hyperactivity and impulsivity [75, 76].

Notably, this interpretation appears specific to the ADHD‐IMP subgroup, as no comparable metabolic alterations were found in ADHD‐SP. Interestingly, the symptom profile characterizing ADHD‐IMP is also commonly reported in neurometabolic disorders [77], suggesting potential shared vulnerabilities or overlapping molecular mechanisms. In contrast, the ADHD‐SP profile—characterized by slower response latencies and increased cognitive effort—may reflect a different mechanism, in which elevated motor activity serves as a compensatory strategy to counteract cognitive inefficiencies rather than indicating excess synaptic energy [51]. While preliminary, these findings suggest divergent metabolic trajectories across ADHD subgroups and warrant further investigation in future studies.

Finally, dysregulated levels of pseudouridine—a modified nucleoside reflecting RNA degradation—have been associated with several metabolic and neurodevelopmental disorders [78], possibly via the alteration of neuronal function and gene expression [79]. Similarly, reduced trigonelline—a vitamin B3‐related compound with antioxidant, anti‐inflammatory, and neuroprotective effects [80]—may signal oxidative stress and neuroinflammatory processes, both implicated in cognitive dysfunction cognitive [80] and ADHD pathophysiology [61].

Despite the above‐mentioned metabolic differences, in our study, we found that urinary metabolites did not show predictive relationships with data‐driven profiles as a grouping variable when performing multivariate data analyses. Extensive research has explored biochemical alterations in individuals with ADHD [11, 12, 15] but research using significant metabolites as multivariate predictors, as is pursued in our study, is scarce. Specifically, the study by Tian et al. [26] explored the predictive ability of urinary metabolites determined through LC‐HRMS in ADHD. These authors found a good discriminant ability of amino acid metabolism and fatty acid metabolism to identify children with and without ADHD—the same metabolic routes in which we obtained group‐level differences.

The inclusion of metabolites significantly enhanced the performance of machine learning models, with LR achieving the highest accuracy, precision, recall, and F1 score. This underscores the value of integrating metabolic markers with behavioral data, as further supported by the robust performance observed in the confusion matrices and cross‐validation tests. While some models, such as k‐NN and SVM, exhibited larger improvements in accuracy and precision when transitioning from Dataset (1) to Dataset (2), for instance, k‐NN showed increases of 18.7% and 12.3%, respectively, and SVM improved by 6.2% and 1.6%, but LR still outperformed these models overall, particularly in terms of accuracy and cross‐validation consistency.

In summary, although no single metabolite emerged as a robust predictor of subgroup classification, the integration of metabolic and behavioral features substantially improved the accuracy and stability of machine learning models. These findings highlight the potential of multidimensional approaches to capture ADHD heterogeneity and support ongoing efforts toward biologically informed subtyping.

5. Study Limitations

Despite the promising findings, several limitations are worth mentioning: (i) The reduced sample size was due to clinical and logistical constraints, including challenges in obtaining consent for urine samples and completing the MPH washout. Despite this, the sample size was sufficient to identify replicable attentional control profiles and explore preliminary metabolic differences. (ii) Participants' CPT scores were generally attenuated compared to the original study [28], likely due to the inclusion of individuals undergoing MPH treatment. Although a wash‐out period (M = 57.12, SD = 55.61) was applied, prolonged medication could still induce neurobiological changes resulting in behavioral improvements [81], thereby affecting performance comparability with medication‐naïve children [36]. Nonetheless, the inclusion of medicated participants did not significantly alter cluster analysis results on the CPT, as we replicated the same profiles among both medication‐naïve [28] and wash‐out participants, suggesting that the VR‐CPT AULA profiles are reliable across populations. (iii) Although approximately 90% of MPH is excreted in urine within 16 h [37], it remains possible that MPH may have influenced the metabolic analysis. Despite no observed differences in excreted metabolites in a multivariate approach, prolonged MPH exposure could potentially alter metabolic pathways, particularly in catecholamine, glutamatergic, or bioenergetic systems [16, 81]. (iv) While alternative biological samples (e.g., blood) could reduce medication interference, they are more invasive for children. Urine, in contrast, still offers substantial overlap, capturing around 80% of metabolic changes detectable in blood [82]. (v) Urine samples were not collected following a fasting period, limiting our ability to control for potential dietary intake effects on the metabolic analysis. Taken together, the clinical relevance of these findings should be further evaluated in larger medication‐naïve samples with controls for age, sex, diet, and physical activity.

6. Conclusions

Our findings emphasize the clinical importance of using dimensional analyses to better characterize attentional behaviors in ADHD. We identified specific metabolic alterations in the impulsive–hyperactive subgroup (ADHD‐IMP) involving microbial‐derived metabolites, BCAA metabolism, energy regulation, and oxidative stress markers. These converging metabolic pathways suggest a coherent neurobiological substrate underlying impulsivity, possibly indicating increased vulnerability in this subgroup. Conversely, the ADHD‐SP subgroup showed no significant metabolic differences despite similar symptom severity, highlighting the biological heterogeneity of ADHD and suggesting distinct neurobiological mechanisms or the need for more sensitive detection methods.

The lack of linear relationships between urinary metabolites and ADHD symptoms, as observed by Bulut et al. [83], highlights the complexity of brain‐cognition relationships. Thus, further research aimed at identifying molecular data‐driven profiles in ADHD is crucial for advancing our understanding of metabolic alterations in the disorder. While this study aims to lay the groundwork for the development of a software application that integrates NMR‐based metabolomic data with behavioral markers to improve ADHD diagnosis, the relatively small sample size (N = 83), class imbalance, and urine sample complexity could impact model robustness and generalizability. Current work in undergoing in our laboratories to prioritize larger sample sizes, balanced data, and controls for medication effects to enhance clinical applicability and reliability.

Author Contributions

A.M. Salmerón: investigation, formal analysis, data curation, writing – original draft, writing – review and editing, and visualization. P. Fernández‐Martín: conceptualization, methodology, investigation, formal analysis, data curation, writing – original draft, writing – review and editing, and visualization. R. Rodríguez‐Herrera: methodology, investigation, and data curation; F.M. Arrabal‐Campos: investigation, formal analysis, and data curation; A.C. Abreu: investigation, formal analysis, data curation, and supervision; I. Fernández: conceptualization, methodology, resources, funding acquisition, supervision, and writing – review and editing; and P. Flores: conceptualization, methodology, resources, funding acquisition, supervision, and writing – review and editing.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

The supplementary material includes detailed demographic and clinical characteristics of the ADHD and TD samples, results from one‐way ANOVA models, attentional control profiles assessed using the virtual CPT AULA, and Wilcoxon rank‐sum tests comparing traditional ADHD subtypes with control participants. Additionally, it presents the results of clustering validation, with 14 indices supporting a four‐cluster solution as the most optimal. PLS‐DA score plots, boxplots of statistically significant metabolites, and confusion matrices are also provided.

Table S1 Demographic and clinical characteristics of the samples from ADHD and TD children. MPH = Methylphenidate. ODD = Oppositional defiant disorder. Clinically significant scores on the SDQ (≥ 95th percentile) and CBCL/6–18 (t scores ≥ 70) scales are in bold. Borderline clinical scores on the SDQ (90th to 95th percentile) and CBCL/6–18 (T scores between 65 and 70) scales are in italics.

Table S2. Robust one‐way ANOVA models revealed significant differences in AULA’s main outcomes between children with and without ADHD, although no clinically significant differences were observed between the ADHD‐C and ADHD‐IN subtypes. 10% trimmed mean differences (t‐scores) and associated bootstrap confidence interval are presented. ^‡ p < 0.001, ^† p < 0.01 and *p < 0.05. ¹The explanatory measure of effect size (ξ), a robust generalization of Pearson’s correlation allowing assumptions deviations and unequal sample sizes. Analogue to Cohen’s d, ξ = 0.15, 0.35 and 0.50 roughly correspond to small, medium and large effect sizes.^{64, 65}

Figure S1. Attentional control profiles measured by the virtual CPT AULA of ADHD‐C, ADHD‐IN and TD participants. 10% trimmed mean and SEM values of AULA main indices (t‐scores). T‐scores ≥ 61 represent a clinically low performance. Dashed lines indicate cut‐offs for risk of attention problems (>60 = at risk; >70 = high risk).

Table S3. Wilcoxon rank‐sum tests among traditional ADHD subtypes and control participants. Metabolic pathways in parentheses.

Figure S2. Attentional control profiles measured by the virtual CPT AULA according to three‐ and four‐ cluster structures. On the left, 10% trimmed mean and SEM values of AULA main indices. T‐scores ≥ 61 represent a clinically low performance. Dashed lines indicate cut‐offs for risk of attention problems (>60 = at risk; >70 = high risk). On the right, percentage distribution of each cluster in ADHD and TD groups. Graphical visualization for k = 5 is included in the main text.

Table S4. Fourteen clustering validation indices chose a four‐cluster structure as the most optimal cluster solution. Best fitting values for each index are boldfaced.

Figure S3. Graphical visualization of the best cluster solution among thirty validation indices.

Table S5. Robust One‐way ANOVA in AULA’s main outcome measures according to the five‐cluster solution. 10% trimmed mean differences (t‐scores) and associated bootstrap confidence interval are presented.

Table S6. Clusters’ demographic characteristics. Group‐level comparisons were assessed via one‐way ANOVAs. MPH = Methylphenidate. ODD = Oppositional defiant disorder. Clinically significant scores on the SDQ (≥ 95th percentile) and CBCL/6–18 (T scores ≥ 70) scales are boldfaced. Borderline clinical scores on the SDQ (90th to 95th percentile) and CBCL/6–18 (T scores between 65 and 70) scales are in italics.

Figure S4. PLS‐DA scores plots depending on the A) data‐driven cluster profile (groups are colored as follows—Sluggish: green, ADHD‐IMP: dark blue, High: red, Average: yellow, ADHD‐SP: light blue), also highlighting samples depending on the traditional diagnosis, B) ADHD diagnosis and C) ADHD subtype, applied to the urine ¹H NMR dataset to examine possible metabolic differences. PLS‐DA metrics are the following: A) R²X = 0.205, R²Y = 0.144, Q² = −0.15, CV‐ANOVA = 1; B) R²X = 0.182, R²Y = 0.578, Q² = 0.030, CV‐ANOVA = 1; and C) R²X = 0.201, R²Y = 0.272, Q² = −0.126, CV‐ANOVA = 1.

Figure S5. Box plots comparing urinary levels of metabolites that showed statistically significant differences between ADHD‐IMP and other performance‐based groups (ADHD‐SP, Average performers, and High performers).

Figure S6. Confusion matrices showing a comparison of the obtained results for the LR model using A) the first dataset (behavior variables) and B) the second dataset (metabolites and behavior variables). Abbreviations of the axis labels—0: Sluggish, 1: ADHD‐IMP, 2: High, 3: Average; 4: ADHD‐SP.

Salmerón A., Fernández‐Martín P., Rodríguez‐Herrera R., et al., “Urinary Metabolic Biomarkers of Attentional Control in Children With Attention‐Deficit/Hyperactivity Disorder: A Dimensional Approach Through ¹H NMR‐Based Metabolomics,” NMR in Biomedicine 38, no. 8 (2025): e70088, 10.1002/nbm.70088.

Data Availability Statement

The supporting data for this study is available upon request.