Using Machine Learning to Derive Neurobiological Subtypes of General Psychopathology in Late Childhood

Abstract

Traditional mental health diagnoses rely on symptom-based classifications. Yet this approach can oversimplify clinical presentations as diagnoses often do not adequately map onto neurobiological features. Alternatively, our study used structural imaging data and a semi-supervised machine learning technique, heterogeneity through discriminative analysis (HYDRA), to identify neurobiological subtypes in 9- to 10-year-olds with high psychopathology endorsements (n = 9,027). Our model revealed two stable neurobiological subtypes (adjusted Rand index = 0.38). Subtype 1 showed smaller structural properties, elevated conduct problems and ADHD symptoms, and impaired cognitive performance compared to Subtype 2 and typically developing youth. Subtype 2 had larger structural properties, cognitive abilities comparable to typically developing youth, and elevated internalizing symptoms relative to Subtype 1 and typically developing youth. These subtypes remained stable in their neurobiological characteristics, cognitive ability, and associated psychopathology traits over time. Taken together, our data-driven approach uncovered evidence of neural heterogeneity as demonstrated by structural patterns that map onto divergent profiles of psychopathology symptoms and cognitive performance in youth.

General Scientific Summary

This study identified neurobiological subtypes underlying high levels of general psychopathology. Smaller/thinner brains were associated with specific externalizing symptoms and cognitive deficits. Larger/thicker brains were associated with elevated internalizing symptoms and intact cognition. Neural characteristics, cognition, and psychopathology among the subtypes remained largely stable across a two-year period.

Introduction

While the Diagnostic and Statistical Manual of Mental Disorders (DSM-5-TR) serves as the gold standard for mental health criteria (American Psychiatric Association, 2013), individuals with the same diagnosis often exhibit significant heterogeneity in clinical outcomes, treatment response, neurobiological attributes, and symptom characteristics (Barzilay et al., 2020; Martin et al., 2007; Monroe et al., 2019; van Loo et al., 2014).

Many have explored subtyping within diagnostic categories to better understand the heterogeneity of psychological disorders. Clustering and machine learning techniques enable data-driven psychological subtyping, using characteristics of neurobiology (e.g., gray matter volume, functional connectivity) instead of categorical diagnoses. Uses have included examining heterogeneity in conditions like internalizing disorders, attention-deficit/hyperactivity disorder (ADHD), and psychosis across the lifespan from late childhood to adulthood (Barth et al., 2018; Chew et al., 2022; Itahashi et al., 2020; Kaczkurkin et al., 2020; Liang et al., 2020). Neurobiological-driven subtypes may elucidate individual mechanistic differences despite a shared diagnosis. When examining psychological profiles associated with neurobiological subtypes, many studies find discrete subtypes associated with comparable overall symptom scores, such as a depressive symptom total score (Drysdale et al., 2017; Kaczkurkin et al., 2020; Wang et al., 2021). This demonstrates similar outward symptoms reflecting different neurobiological underpinnings, indicating multiple pathways to the same clinical profile defined as by the standard diagnostic system. For this reason, subtyping on symptoms and behaviors alone may not be sufficient for elucidating heterogeneity. Rather, subtyping on the basis of neurobiological features may be required for pathophysiological classifications.

Despite progress, subtyping studies often define samples based on existing diagnoses, recruiting only those who meet criteria for specific disorders before exploring neurobiological subtypes (Drysdale et al., 2017; Liang et al., 2020). For example, a study may use a major depressive disorder diagnosis as a criterion for participant inclusion before clustering within this group. This approach lets diagnoses guide the research, despite evidence of heterogeneity within disorders, and raises concerns that the clustering sample is already influenced by expected characteristics of a particular disorder.

The present study models continuous dimensions of psychopathology coupled with the machine learning approach known as heterogeneity through discriminative analysis (HYDRA) in a large sample of 9- to 10-year-old children (N = 9,027) from the Adolescent Brain Cognitive Development (ABCD) Study. Psychopathology is operationalized using a bifactor model; this proposes a general psychopathology factor (p factor; on which all CBCL items load) and three orthogonal specific dimensions (internalizing, conduct problems, ADHD; Moore et al., 2020). While two of the factors reflect DSM constructs, we did not categorize participants with binary diagnoses. Instead, our model assigns continuous scores to each participant across the four factors, capturing individual differences across different aspects of psychopathology. This approach accommodates dimensionality, subthreshold symptoms, and comorbidity in our diverse sample, separating variance specific to each dimension.

Additionally, HYDRA allows neurobiological features to define subtypes in children with higher general psychopathology scores in the community sample. HYDRA has parsed heterogeneity in various neuropsychiatric disorders (e.g., Alzheimer’s disease, internalizing disorders, psychosis) and across ages (e.g., pre-adolescent through late adulthood), revealing subtypes with distinct neurostructural, functional, cognitive, and psychopathological profiles (Chand et al., 2020; Kaczkurkin et al., 2020; Varol et al., 2017). HYDRA’s effectiveness is critical given that we study these phenotypes across various psychiatric dimensions and ages (late childhood through early adolescence) within a single sample.

To our knowledge, prior research has not used HYDRA methods to identify data-driven neurobiological subtypes of cortical thickness and volume associated with general psychopathology or examined associated features over time. This study’s subtyping aspect is exploratory due to its data-driven nature. Additionally, our pre-adolescent sample is a strength since many subtyping studies focus on adults (Chand et al., 2020; Chew et al., 2022; Drysdale et al., 2017; Varol et al., 2017). Many psychopathology symptoms as well as cognitive control and executive functioning skills undergo substantial changes during early adolescence (Kessler et al., 2012; Luna, 2009). This coincides with profound brain development, with certain regions peaking in structural development around ages 11–12 (Gennatas et al., 2017; Giedd et al., 1999). Our study captures this structural growth period (up to age 12) and given prior cross-sectional findings suggesting HYDRA subtype stability across age 8 to 23 years of age, we expect the resulting neurobiological subtypes to remain stable during this time (Kaczkurkin et al., 2020).

Methods

Participants

We used data from the baseline, first-year follow-up, and second-year follow-up of the ABCD Study (release 4.0). As per ABCD Study protocols, all participants provided assent/parental consent, and Vanderbilt University Institutional Review Board approved access to the deidentified dataset. The sample comprised 11,876 children aged 9 and 10 years recruited from 21 locations in the United States. After excluding participants with missing data or failed neuroimaging quality assurance measures and allowing for participant per household, the final sample included 9,027 participants. Additional details about the ABCD Study can be found elsewhere (Casey et al., 2018; Hagler et al., 2018) and in the supplement. Prior to analysis, we randomly selected one member from each family in the ABCD Study dataset, which includes twins, triplets, and siblings.

Measure of Psychopathology

Psychopathology was defined by the parent-rated Child Behavior Checklist (CBCL) of behaviors and emotions; multiple response options on items yielded 119 variables for modeling (Achenbach, 2009). To capture the dimensional and hierarchical aspects of psychopathology, we used Moore et al. (2020)’s bifactor model, featuring a general psychopathology factor and three specific factors (internalizing symptoms, conduct problems, and ADHD). This was developed and validated on ABCD Study data, highlighting its applicability in the present study. An exploratory structural equation model identified specific latent dimensions, while a confirmatory bifactor analysis modeled a general psychopathology factor capturing shared symptoms across disorders. All factors are orthogonal, with each CBCL item loading on both the general and a specific factor, exemplified by items like “can’t sit still” loading on both general psychopathology and ADHD factors. Each psychopathology dimension was separately modeled at each timepoint. For detailed modeling, validity, and reliability information, refer to Moore et al. (2020) and the supplement. Primary caregivers reported on child’s biological mother’s mental illness history (see supplement).

Measure of Cognitive Ability

Baseline and second-year follow-up cognitive abilities were assessed with tasks from the National Institute of Health (NIH) Toolbox including the Dimensional Change Card Sort Test (set-shifting), Pattern Comparison Processing Speed Test (processing speed), and Flanker Inhibitory Control and Attention Test (attention and inhibition). Card Sort at the second-year follow-up was not examined due to small sample size (n = 52). Details about these tasks are available in previous work and the supplement (Weintraub et al., 2013).

Image Acquisition, Processing, and Quality Assurance

Using a whole brain approach, we analyzed 68 cortical regions from Desikan et al.’s (2006) surface-based atlas procedure and 19 subcortical regions from Fischl et al.’s (2002) automated labeling procedure using FreeSurfer (Desikan et al., 2006; Fischl et al., 2002). Procedures for image acquisition, processing, and quality assurance evaluation have been detailed elsewhere (Casey et al., 2018; Hagler et al., 2018) and in the supplement.

Clustering Procedures

HYDRA is a semi-supervised machine learning algorithm that distinguishes cases (i.e., those with significant psychopathology) from controls (i.e., typically developing) and then identifies subtypes within the cases that are maximally different from controls by defining a polytope and then clustering cases based on their association with faces of the polytope (Varol et al., 2017). Thus, HYDRA requires labels denoting participant cases and controls. Gaussian mixture modeling (GMM), a probabilistic model, labeled participants before HYDRA clustering.

The results of GMM and its metrics (silhouette coefficient, Jensen-Shannon divergence metrics) were used to separate the sample into cases (high psychopathology endorsement) and controls (low psychopathology endorsement). As shown in the results section, GMM identified a segmentation point within the sample, assigning participants to either ‘high’ or ‘low’ psychopathology endorsement based on their probabilities of belonging to these components.

Following GMM labeling, we identified neurostructural subtypes within cases (high endorsement) based on their differences from controls (low endorsement) using HYDRA. Subtyping features included gray matter volume and cortical thickness of 68 cortical regions, and gray matter volume from 19 subcortical regions while accounting for age and sex. HYDRA was executed once using baseline neural properties. We derived one-, two-, and three-cluster solutions; the clustering solution with the highest adjusted Rand index (ARI) was chosen for subsequent analyses.

Statistical Analysis

After neurostructural subtyping with HYDRA, we conducted multiple linear regressions and analyses of covariance (ANCOVA) to examine group differences in structural features, cognitive ability, and psychopathology factor scores while controlling for age, sex, and site/scanner model. We also assessed the stability of these psychological, cognitive, and neural associations by employing the maximum available timespan of two years. Omnibus ANCOVAs and pairwise post-hoc tests were adjusted for multiple comparisons using the false discovery rate (q < .05) with the stats package in R version 3.6.1. Sensitivity analyses adjusted for total intracranial volume and average cortical thickness, as well as maternal education and income.

Transparency and Openness

We follow the APA Style Journal Article Reporting Standards (JARS; Kazak, 2018). The ABCD Study is publicly available and can be accessed through the National Institute of Mental Health Data Archive (https://nda.nih.gov/abcd; DOI: 10.15154/1523041). The R code (version 4.3.1; R Core Team, 2023; https://www.R-project.org/) and corresponding wiki for the analytic procedures can be found on GitHub at https://github.com/VU-BRAINS-lab/Reimann_HYDRA_psychopathology and https://github.com/evarol/HYDRA. The study design and analyses were not preregistered.

Results

Gaussian Mixture Modeling

We used GMM to evaluate the optimal segmentation point of bifactor-defined general psychopathology factor scores into cases and controls. Two components exhibited the optimal solution (silhouette coefficient = 0.63; Jensen-Shannon divergence = 0.01; see supplement for further detail). These two components represented children with high (N_Case = 5,047; General_average = 0.66) and low (N_Control = 3,980; General_average = −0.70) psychopathology symptom endorsement, and were used to define HYDRA group membership.

HYDRA Clustering

The stability of the cluster solutions produced by HYDRA was assessed via 3-fold cross-validation, showing a peak at k = 2 (ARI = .38) compared to a three-cluster solution (ARI = .33). This suggests the presence of two neurobiological subtypes within the GMM-defined high general psychopathology cases, henceforth referred to as Subtype 1 (N = 2,668) and Subtype 2 (N = 2,379). We then examined whether these subtypes differed from GMM-defined controls (N = 3,980; henceforth referred to as typically developing) on the demographic, cognitive, and psychopathology measures in Table 1.

Table 1.

Sample demographic and behavioral descriptives across neurobiologically-derived subtypes and typically developing youth.

	TD Youth (N = 3980)		Subtype 1 (N = 2668)		Subtype 2 (N = 2379)		Full Sample (N = 9027)
Female (Male), n	2116 (1864)		1191 (1477)		1011 (1368)		4318 (4709)
	Mean	SD	Mean	SD	Mean	SD	Mean	SD
Age at baseline, years	9.92	0.62	9.91	0.62	9.90	0.62	9.90	0.62
Maternal Education, years	15.17	2.72	14.37	2.79	15.29	2.51	14.97	2.71
Psychopathology
Baseline
General	−0.70	0.37	0.67	0.6	0.64	0.57	0.06	0.83
Externalizing	−0.10	0.49	0.14	0.77	0.06	0.73	0.02	0.66
ADHD	−0.02	0.54	0.14	0.81	0.02	0.82	0.04	0.71
Internalizing	−0.05	0.64	0.08	0.75	0.14	0.77	0.04	0.72
1st Year Followup
General	−0.47	0.56	0.53	0.76	0.54	0.72	0.10	0.83
Externalizing	−0.11	0.52	0.16	0.76	0.04	0.71	0.01	0.66
ADHD	−0.03	0.56	0.16	0.77	0.06	0.81	0.05	0.70
Internalizing	−0.002	0.67	0.08	0.74	0.16	0.74	0.06	0.72
2nd Year Followup
General	−0.42	0.59	0.5	0.78	0.49	0.74	0.10	0.82
Externalizing	−0.09	0.51	0.14	0.73	0.02	0.67	0.01	0.63
ADHD	−0.03	0.57	0.13	0.74	0.07	0.77	0.04	0.68
Internalizing	−0.01	0.65	0.04	0.73	0.17	0.74	0.05	0.70
Cognitive Tasks
Baseline
Flanker Task	94.66	8.64	92.93	9.71	94.41	8.64	94.10	9.00
Pattern Accuracy	89.03	14.44	87.04	14.7	87.91	14.43	88.14	14.54
Card Sorting	93.50	8.96	91.14	9.98	93.02	8.93	92.67	9.31
2nd Year Followup
Flanker Task	100.84	7.09	99.14	8.00	100.57	7.35	100.28	7.47
Pattern Accuracy	104.73	14.79	102.05	15.11	103.44	15.02	103.61	14.98

Note: Externalizing, ADHD, and Internalizing each refer to specific factors derived from bifactor model

Demographic Analyses

There were no significant age differences among Subtype 1, Subtype 2, and TD youth (Table 2). Both Subtype 1 and Subtype 2 had significantly fewer females compared to TD youth; no other sex differences were significant. Subtype 1 had lower maternal education compared to both Subtype 2 and TD youth, while no significant difference in maternal education was found between Subtype 2 and TD youth.

Table 2.

Demographics, psychopathology, and cognitive performance group differences.

			Subtype 1 vs. TD				Subtype 2 vs. TD				Subtype 1 vs. Subtype 2				R2
	F/ X 2	p	B	SE	t	p fdr	B	SE	t	p fdr	B	SE	t	p fdr	Est	L CI	U CI
Descriptives
Maternal Education 1	79.50	<.01	−0.13	0.07	−10.90	<.01	0.02	0.08	1.66	.10	−0.15	0.08	−11.21	<.01	.02	.01	.03
Age 2	0.83	0.44	−0.01	0.19	−0.87	0.58	−0.01	0.19	−1.21	0.58	0.00	0.21	0.34	0.73	.00	.00	.00
Gender 2	83.52	<.01	−0.34	0.05	−6.81	<.01	−0.42	.05	−8.22	<.01	0.09	0.06	1.53	.12	AIC: 12420
Psychopathology
Baseline 3
General	7984.43	<.01	0.74	0.01	107.11	<.01	0.69	0.01	99.67	<.01	0.02	0.01	2.85	<.01	.65	.63	.66
Externalizing	104.16	<.01	0.15	0.02	13.56	<.01	0.10	0.02	8.39	<.01	0.05	0.02	4.26	<.01	.05	.04	.06
ADHD	35.95	<.01	0.09	0.02	7.53	<.01	0.02	0.02	1.43	.15	0.07	0.02	5.34	<.01	.04	.03	.04
Internalizing	72.83	<.01	0.10	0.02	8.44	<.01	0.12	0.02	10.23	<.01	−0.03	0.02	−1.96	.04	.04	.03	.05
1st Year Followup 4
General	2098.20	<.01	0.55	0.02	54.33	<.01	0.53	0.02	52.14	<.01	0.00	0.02	0.17	.86	.36	.35	.38
Externalizing	106.98	<.01	0.17	0.02	13.71	<.01	0.09	0.02	7.07	<.01	0.08	0.02	5.67	<.01	.05	.05	.06
ADHD	41.35	<.01	0.11	0.02	8.72	<.01	0.05	0.02	4.31	<.01	0.05	0.02	3.78	<.01	.04	.03	.05
Internalizing	41.64	<.01	0.06	0.02	5.19	<.01	0.10	0.02	8.22	<.01	−0.04	0.02	−2.98	<.01	.04	.03	.05
2nd Year Followup 5
General	1513.82	<.01	0.50	0.02	45.94	<.01	0.48	0.02	44.23	<.01	0.00	0.02	0.20	0.84	.31	.29	.32
Externalizing	77.08	<.01	0.15	0.02	11.87	<.01	0.07	0.02	5.42	<.01	0.08	0.02	5.57	<.01	.05	.04	.06
ADHD	21.49	<.01	0.07	0.02	5.94	<.01	0.02	0.02	1.49	.14	0.05	0.02	3.91	<.01	.03	.03	.04
Internalizing	50.86	<.01	0.05	0.02	3.89	<.01	0.12	0.02	9.17	<.01	−0.07	0.02	−4.97	<.01	.05	.04	.06
Cognitive Tasks
Baseline
Flanker Task 6	32.53	<.01	−0.08	0.22	−7.10	<.01	−0.01	0.23	−1.26	.21	−0.07	0.25	−5.11	<.01	.06	.05	.07
Pattern Accuracy 8	12.34	<.01	−0.05	0.36	−4.31	<.01	−0.02	0.37	−1.85	.06	−0.03	0.40	−2.10	.05	.08	.07	.09
Card Sorting 7	51.71	<.01	−0.10	0.23	−8.88	<.01	−0.02	0.24	−1.51	.13	−0.08	0.26	−6.45	<.01	.07	.06	.09
2nd Year Followup
Flanker Task 9	30.21	<.01	−0.10	0.23	−6.91	<.01	−0.02	0.24	−1.10	.27	−0.08	0.25	−5.14	<.01	.05	.04	.06
Pattern Accuracy 10	13.84	<.01	−0.06	0.45	−4.60	<.01	−0.02	0.46	−1.65	.09	−0.04	0.51	−2.59	.02	.08	.07	.09

^1.df = 7698;

^2.df = 9024;

^3.df = 8994;

^4.df = 7655;

^5.df = 7116;

^6.df = 8874;

^7.df = 8875;

^8.df = 8858;

^9.df = 5935;

^10.df = 5907

Note: Second-year follow-up card sorting data n = 52; data not used. R-squared calculated for the entire model. Externalizing, ADHD, and Internalizing each refer to specific factors derived from bifactor model

Cognition Analyses

Baseline

Subtype 1 performed worse than TD youth across all three tasks (Table 2). Compared to Subtype 2, Subtype 1 performed worse on the inhibitory control and card sorting tasks but did not differ in processing speed. Subtype 2 and TD youth did not differ across any cognitive task (Figure 1).

Figure 1. Group Differences in Psychopathology Symptoms and Cognitive Performance.

Note. The Y-axis displays estimates derived from the fitted model to assess group differences. Each vertical line corresponds to a 95% confidence interval (CI), with the mean line representing the comparison group, typically developing youth (TD). The subtype (S1 or S2) is significantly different from TD when its CI does not contain 0, the mean of TD.

Second-year Follow-up

Subtype 1 continued to perform worse than TD youth across all examined tasks (Table 2). Compared to Subtype 2, Subtype 1 performed worse on the inhibitory control and processing speed tasks. Subtype 2 and TD youth did not differ across any cognitive task.

Psychopathology Analyses

Baseline

Subtype 1 showed higher psychopathology across all dimensions compared to TD youth. Subtype 2 also had elevated endorsement across all dimensions compared to TD youth, except Subtype 2 and TD youth did not differ on ADHD symptoms. When comparing the subtypes, Subtype 1 had greater endorsement of general psychopathology, ADHD symptoms, and conduct problems, while Subtype 2 had higher endorsement of internalizing symptoms (Figure 1; Table 2). Results were consistent when adjusting for maternal history of psychopathology.

First-year Follow-up

Both Subtype 1 and Subtype 2 showed significantly elevated psychopathology across all four dimensions when compared to TD youth. When comparing the two subtypes to one another, Subtype 1 continued to show greater endorsement of ADHD and conduct problems and Subtype 2 still showed greater endorsement of internalizing symptoms. However, Subtype 1 and Subtype 2 no longer differed on general psychopathology at the first-year follow-up (Table 2). Results were consistent when adjusting for maternal history of psychopathology.

Second-year Follow-up

Subtype 1 showed higher psychopathology across all four dimensions compared to TD youth. Subtype 2 showed elevated endorsement in general psychopathology, internalizing symptoms, and conduct problems compared to TD youth at the second-year follow-up; however, there was no difference in ADHD endorsement. When comparing the two subtypes, Subtype 1 showed greater endorsement in ADHD and conduct problems, while Subtype 2 showed higher endorsement in internalizing symptoms. At the second-year follow-up, there was no difference in general psychopathology endorsement between Subtype 1 and Subtype 2 (Table 2). Results were consistent when adjusting for maternal history of psychopathology.

Structural Analyses

Of note, because we are clustering on brain volume and cortical thickness, it is expected that the subtypes will differ from controls and from each other on these structural measures. The following neural analyses are provided not to illustrate significance (they are already expected to be significant), but to show instead how the groups compare relative to each other and in what direction (positive or negative effects).

Main Analyses at Baseline

See Tables S1–3 for volume and cortical thickness analyses. Subtype 1 had globally smaller gray matter volume in all cortical and subcortical regions compared to the other groups (Subtype 1 < TD < Subtype 2; Figure 2). Subtype 1 also showed thinner cortices in 66 out of 68 regions compared to TD youth and all regions compared to Subtype 2. In contrast, Subtype 2 had globally thicker cortices and larger gray matter volumes in all cortical and subcortical regions compared to Subtype 1 and TD youth (Subtype 1 < TD < Subtype 2). These findings were consistent when examining group differences in total gray matter volume, total intracranial volume, and average cortical thickness instead of examining individual regions (Subtype 2 < TD < Subtype 1; Table S4).

Figure 2. Baseline Cortical and Subcortical Gray Matter Volume and Cortical Thickness Group Differences.

Note. Brain images plot the t values by region for each contrast, with blue indicating smaller volumes/thinner cortices and red indicating larger volumes/thicker cortices. Subtype 1 (S1) showed smaller cortical and subcortical volumes and thinner cortices than subtype 2 (S2) or typically developing youth (TD). S2 showed larger volumes and thicker cortices than TD.

Sensitivity Analyses at Baseline

Sensitivity findings were consistent with primary analyses while adjusting for maternal education, income, and maternal history of psychopathology (see supplement). Multiple region associations were not significant after adjusting for total intracranial volume and average cortical thickness, though many still remained (Tables S5–6).

Main Analyses at Second-year Follow-up

Continuing to use subtypes derived at baseline, we investigated the stability of gray matter volume and cortical thickness associations at the two-year follow-up. The results largely mirrored those at baseline (Tables S7–9). Additionally, these patterns were consistent for total gray matter volume, total intracranial volume, and average cortical thickness (Table S10).

Sensitivity Analyses at Second-year Follow-up

Sensitivity findings at the second-year follow-up were consistent with analyses at baseline, when adjusting for maternal education and income. Also similar to the baseline sensitivity analyses, a number of regional associations were no longer significant after adjusting for total intracranial volume and average cortical thickness at the second-year follow-up, though many remained (Tables S11–12).

Discussion

We identified two distinct neurostructural subtypes in youth with high general psychopathology using semi-supervised heterogeneity through discriminative analysis. Subtype 1, defined by smaller brain volumes and thinner cortices, displayed impaired cognitive deficits and elevated ADHD and conduct problems relative to the other two groups. Conversely, Subtype 2, marked by greater brain volume and thicker cortices, showed intact cognitive ability as well as elevated internalizing symptoms relative to the other two groups. These findings remained stable over two years.

Our results are consistent with prior studies identifying distinct neurobiological subtypes related to general psychopathology, internalizing disorders, and ADHD (Barth et al., 2018; Itahashi et al., 2020; Liang et al., 2020; Wang et al., 2021). Previous work studying children through young adulthood also found two neurobiological subtypes with globally larger and smaller neurostructural properties in an independent sample of 1,141 youth ages eight to 23 years with internalizing symptoms (Kaczkurkin et al., 2020). Unlike prior studies which examine adults and/or wide age ranges, our study features a sample of children of a narrowly defined age range to examine neurobiological features of late childhood transitioning to early adolescence.

Many subtyping studies reveal differences in specific symptoms, but largely do not differ across total clinical scores of psychological symptoms (Drysdale et al., 2017; Kaczkurkin et al., 2020; Wang et al., 2021). Our study also show neurobiologically-distinct subtypes with diverse specific symptom and cognitive profiles, despite similar general psychopathology endorsement,. This underscores neurobiological features to understand psychopathology, especially since our subtypes deviate in distinct directions from typical development. Furthermore, our findings indicate that our neurobiological subtypes remained largely stable in terms of brain volume, cortical thickness, cognition, and psychopathology symptoms. Longitudinal analysis did not show convergence of brain volume and thickness from either subtype toward typical development. Future studies may investigate the ongoing stability of these subtypes as the ABCD Study releases data across adolescence and young adulthood.

Subtype 2 exhibited larger brain volume and thicker cortices than typically developing youth. Subtype 2 also showed higher endorsement of the internalizing dimension, but intact executive functioning. Intact neurostructural properties may contribute to better cognitive performance compared to Subtype 1, which had global structural deficits. However, these unimpaired structural characteristics did not offer protection against some forms of psychopathology, with Subtype 2 showing the highest internalizing endorsements among all groups. This is consistent with previous work identifying a subtype with distinct neurostructural properties and elevated psychological symptoms but intact cognitive performance across a wide age range from late childhood to early adulthood (Kaczkurkin et al., 2020).

Our study adopts a continuous approach to general psychopathology, using hierarchical modeling rather traditional diagnostic categories. This is significant for several reasons. First, traditional diagnostic criteria may miss individuals with symptoms that do not meet diagnostic threshold for a specific disorder. For example, Barzilay and colleagues (2020) found that 41% of 6,487 adolescents met DSM criteria for at least one lifetime anxiety disorder, while 92% endorsed some level of anxiety symptoms. Traditional diagnostic criteria might underestimate subclinical symptoms severity, potentially leading to inadequate support for these individuals. Secondly, many studies, including subtyping research, often define samples based on pre-established diagnoses. This approach can introduce bias into the sample selection process, as the chosen diagnoses may influence the expected characteristics of the study population. In contrast, our sample includes minimal to severe symptom presentations, embracing the transdiagnostic and heterogeneous nature of psychopathology rather than solely utilizing diagnostic thresholds.

Our methods align with the broader clinical science frameworks, such as Research Domain Criteria (RDoC), which aim to redefine psychopathology beyond symptoms alone (Insel et al., 2010). In our results, we observed two subtypes with opposing neurostructural characteristics both exhibiting the high endorsement of general psychopathology presentation, suggesting different underlying etiologies. Relying solely on symptoms would not allow us to distinguish between these subtypes based on differences in neurodevelopment.

In line with RDoC, our hybrid dimensional-categorical approach reconceptualizes psychopathology as brain disorders which map onto clinically relevant variations. Our dimensional approach, combined with HYDRA subtyping, deconstructs traditional diagnostic groups to model brain metrics predicting individual differences in emotion, behavior, and cognition. Incorporating pathophysiology holds promise for yielding psychopathology classifications with a strong biological basis and informing clinical decisions for specific interventions, monitoring, and levels of care (Drysdale et al., 2017; Wang et al., 2021). Of note, this is not to suggest that imaging will be necessary in clinical diagnosis – rather, understanding the pathophysiological differences between individuals can help us refine our symptom-based approaches which will then be able to better detect the conditions we want to diagnose and treat.

Study limitations may include use of parent reports to assess child psychopathology, reflecting parental interpretations of their child’s behavior and emotions. While self-reports may better reflect a child’s internal state, youth self-reports are not yet available at this age in the ABCD Study. Furthermore, we used maternal education and income as a proxy for socioeconomic status. Future studies may improve upon this by considering different socioeconomic measures and additional environmental or genetic factors. Despite limitations, the strengths of this study advance our understanding of neurostructural differences in children with psychopathology symptoms, potentially supporting biomarkers to identify at-risk children and enhance early intervention strategies.