Brain morphological changes across behaviour spectrums in attention-deficit/hyperactivity disorder Authors: Tianzheng Zhong, Feng Wang, Jianfeng Qiu, Weizhao Lu Published: 2025 ABSTRACT AbstractBackgroundAttention-deficit/hyperactivity disorder (ADHD) is a common neurodevelopmental disorder with behavioural symptoms and grey matter volume (GMV) changes. However, previous studies have not fully elucidated the progressive and causal GMV changes associated with behavioural symptoms in ADHD.AimsThis study aimed to explore the causal relationship between GMV alterations and behavioural symptoms in children and adolescents with ADHD using behaviour-causal structural covariance network (BCaSCN) analysis.MethodsStructural magnetic resonance imaging (sMRI) data from 135 children and adolescents with ADHD and 182 neurotypical controls (NCs) were analysed. ADHD subtypes were identified based on GMV using a clustering algorithm to address the neuroanatomical heterogeneity. To investigate the causal relationships of GMV changes related to behavioural symptoms, sMRI data were sequentially ordered by ADHD index, inattentive index and hyperactive/impulsive index values to generate pseudo-time series data. These data were then analysed using region-of-interest-based BCaSCN analysis to explore potential progressive patterns of GMV change.ResultsNeuroanatomical subtyping revealed two ADHD subtypes with distinct GMV patterns compared with NCs. BCaSCN analysis showed that ADHD subtype 1 was closely associated with inattentiveness, involving prominent nodes in the frontal regions and cerebellum. In contrast, ADHD subtype 2 was more strongly linked to overall disease severity, with the cerebellum and hippocampus as primary hubs.ConclusionsADHD is associated with heterogeneous changes in GMV corresponding to distinct behavioural domains, highlighting the need for subtype-specific diagnostic and therapeutic strategies.Keywords: Attention Deficit Disorder with Hyperactivity, Behaviour, Brain, Causality, Neuropsychiatry WHAT IS ALREADY KNOWN ON THIS TOPIC Attention-deficit/hyperactivity disorder (ADHD) is a common neurodevelopmental disorder with behavioural symptoms and grey matter volume (GMV) changes. However, previous studies have not fully elucidated the causal relationship of GMV changes associated with behavioural symptoms in ADHD. Attention-deficit/hyperactivity disorder (ADHD) is a common neurodevelopmental disorder with behavioural symptoms and grey matter volume (GMV) changes. However, previous studies have not fully elucidated the causal relationship of GMV changes associated with behavioural symptoms in ADHD. WHAT THIS STUDY ADDS This study identified two neuroanatomical subtypes of ADHD, each showing distinct and progressive GMV alteration patterns that may correspond to different behavioural domains. This study identified two neuroanatomical subtypes of ADHD, each showing distinct and progressive GMV alteration patterns that may correspond to different behavioural domains. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY This study revealed structural heterogeneity in ADHD with distinct causal patterns that may underlie different behavioural aspects of the disorder, emphasising the need for subtype-specific diagnostic and therapeutic strategies. This study revealed structural heterogeneity in ADHD with distinct causal patterns that may underlie different behavioural aspects of the disorder, emphasising the need for subtype-specific diagnostic and therapeutic strategies. Introduction Attention-deficit/hyperactivity disorder (ADHD) is one of the most prevalent neurodevelopmental disorders in children and adolescents, with patients typically manifesting symptoms of inattention, hyperactivity and impulsivity.1 2 It is estimated that approximately 7.6% of the global population is affected by ADHD.3 ADHD is a heterogeneous condition; the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5), classifies ADHD into three clinical presentations: predominantly inattentive, predominantly hyperactive/impulsive and combined.1 Advances in neuroimaging have led to a rapid expansion of ADHD research.4 5 Structural magnetic resonance imaging (sMRI) technique has been widely used to investigate the brain morphology and structural characteristics in individuals with ADHD,6 revealing significant grey matter volume (GMV) changes in the cerebellum, prefrontal cortex, accumbens, amygdala, caudate, hippocampus and putamen compared with neurotypical controls (NCs).7 8 Previous studies have highlighted the crucial roles of these regions in mediating ADHD symptoms.9,15 Specifically, disruptions in the cerebellar-thalamo-frontal circuit have been implicated in attention deficits and hyperactive symptoms.9 10 The prefrontal cortex in individuals with ADHD matures at a slower rate, leading to impairments in executive function, attention and motor planning.11 In terms of subcortical nuclei, the accumbens, a key region in reward processing, has been linked to motivational and emotional impairments in ADHD.12 Moreover, amygdala volume has been associated with hyperactivity and general emotional regulation, and changes in this region may impair emotional recognition.13 The caudate nucleus, responsible for activating attention, shows asymmetries that contribute to attentional impulsivity in ADHD.14 15 However, findings from sMRI studies of ADHD have been inconsistent,5 possibly because the three presentations defined in the DSM-5 do not fully capture the heterogeneity of brain structure and morphology in ADHD. Several studies have suggested that these clinical presentations may not correspond to homogenous changes in brain structure but rather reflect overlapping behavioural characteristics.16,18 Several approaches have been proposed to better understand the impact of structural heterogeneity on the pathogenesis of ADHD. For example, Li et al identified three subgroups of children with ADHD based on subcortical volumes using a community detection clustering algorithm.19 Similarly, Wang et al applied the heterogeneity through discriminative analysis (HYDRA) algorithm to classify ADHD into two subtypes based on GMV, white matter volume and functional measures.20 Although these studies have advanced our understanding of neuroanatomical heterogeneity in ADHD, the causal relationships between brain structural changes and behavioural symptoms in ADHD and ADHD subtypes remain unknown. Causal structural covariance network (CaSCN) analysis based on pseudo-time series sMRI data can provide insights into the progressive and causal characteristics of GMV changes21 and has been used to elucidate the pathological processes underlying progressive brain structural changes in disorders such as epilepsy and schizophrenia.21,23 This study hypothesised that ADHD neuroanatomical subtypes are associated with heterogeneity in progressive and causal brain structural changes, which may correspond to different behavioural domains. To test this hypothesis, we first applied a clustering algorithm to resolve the heterogeneity of ADHD and then introduced a novel behaviour-CaSCN (BCaSCN) analysis based on CaSCN to characterise causal relationships between brain structural alterations and behavioural domains in ADHD. Materials and methods Subjects This study was approved by the institutional review board of our university and conducted in accordance with the Declaration of Helsinki. T1-weighted sMRI data were obtained from the ADHD-200 dataset (http://fcon_1000.projects.nitrc.org/indi/adhd200). The following exclusion criteria were applied for data cleaning (figure 1): (1) left-handedness or mixed-handedness or missing handedness information; (2) missing data for ADHD index, inattentive index or hyperactive/impulsive index; (3) age >18 years and (4) poor image quality, defined as a 5-point Likert scale rating below three by an experienced radiologist. To minimise subjectivity in manual evaluation, we also used the Computational Anatomy Toolbox (CAT12, http://dbm.neuro.unijena.de/cat12/) software for image quality rating. sMRI data with CAT12 quality ratings below B+were also excluded. Finally, sMRI data from 317 children and adolescents (182 NCs and 135 ADHD) from five sites were included. Demographic and clinical characteristics of both groups are presented in online supplemental table S1. Flowchart of participant selection in this study. ADHD, attention-deficit/hyperactivity disorder. Voxel-wise morphology analysis The following processing steps were carried out on the T1-weighted sMRI data using CAT12: (1) skull stripping, denoising and bias correction; (2) segmentation of the images into grey matter, white matter and cerebrospinal fluid regions; (3) registration of the images into the Montreal Neurological Institute standard template; (4) modulation of grey values into voxel-wise volumes and estimation of the total intracranial volume (TIV) for all participants; (5) spatial smoothing of the GMV images using a Gaussian kernel with a 6 mm full width at half-maximum; (6) generation of an optimal threshold mask from all sample data to eliminate non-grey-matter voxels. To estimate the differences in GMV between individuals with ADHD and NCs, two-sample t-tests were conducted on the GMV maps. Site, TIV, age and sex were treated as nuisance covariates in the generalised linear model. A false discovery rate (FDR) correction at p<0.05 was used to control false positives, with a cluster-extent threshold of 50 voxels. Subtyping of ADHD To address brain structural heterogeneity in children and adolescents with ADHD, regional volumes were extracted for all participants according to the automated anatomical labelling (AAL) brain atlas. The effect of age, sex, site number and TIV was eliminated using a generalised linear model. HYDRA, an advanced semisupervised learning algorithm, was applied to identify the ADHD neuroanatomical subtypes with brain regional volumes as features (online supplemental figure S1a).24 HYDRA involved the following steps: first, participants with ADHD were labelled as positive and NCs as negative. Second, a convex polyhedron with K planes (equal to the number of clusters) was constructed to separate ADHD from NCs. Then, an extended standard linear maximum-edge classifier was used to calculate the distance between each participant with ADHD and each hyperplane. Finally, each participant with ADHD was assigned to the nearest hyperplane, resulting in K clusters (ADHD subtypes). In the clustering process, we used a 10-fold cross-validation strategy. In each cross-validation step, ninefold data were used for clustering. After 10-fold cross-validation, each participant with ADHD had nine clustering labels. A total of 20 clustering consensus steps were then performed to determine the final clustering label for each participant with ADHD using a co-occurrence matrix generated from the nine labels. The number of clusters (K) was set from 2 to 10, and the optimal cluster number was determined using the Adjusted Rand Index (ARI).25 To test the effect of feature selection on subtyping, the following steps were performed for the 116 features: first, we used Levene’s test to examine the homogeneity of variance for each feature between the NC and ADHD groups, resulting in the exclusion of 20 features. Next, we employed the intraclass correlation coefficient (ICC) to assess the consistency of the remaining features across sites with an ICC threshold of 0.15, excluding 68 features. Finally, we removed features with an absolute correlation coefficient with the ADHD index <0.15, leaving 11 features. We then reperformed HYDRA with the remaining features. Furthermore, to evaluate the impact of different site harmonisation methods on the resulting subtypes, we applied ComBat to harmonise the variables across sites prior to HYDRA subtyping.24 Agreement between subtypes derived using ComBat and those using the original approach was assessed using the ARI and Jaccard index. Voxel-wise comparisons of GMV between ADHD subtypes and NCs After dividing participants with ADHD into subtypes, two-sample t-tests were performed on the GMV maps to compare each ADHD subtype and NCs, controlling for age, sex, site number and TIV. An FDR correction at p<0.05 was used to control false positives. To investigate progressive changes in brain GMV associated with disease severity, participants within each ADHD subtype were further divided into three stages according to ADHD index: stage 1 (ADHD index 30–50), stage 2 (51–70) and stage 3 (71–90). Comparisons of GMV maps between each subgroup and NCs were performed, controlling for age, sex, site number and TIV. FDR correction at p<0.05 was used to control false positives, with a cluster-extent threshold of 50 voxels. CaSCN and BCaSCN analysis To model the progression of ADHD from mild to severe stages, we ordered the GMV maps of each ADHD subtype by increasing ADHD index values. When two patients shared the same ADHD index value, their GMV maps were further ordered by the corresponding inattentive and hyperactive/impulsive indexes. The resulting ordered sequence of GMV maps for each ADHD subtype formed a pseudo-time series representing disease progression. To investigate the causal relationships for GMV changes in ADHD subtypes, CaSCN analysis was performed by applying region-of-interest (ROI)-wise bivariate signed-path Granger causality (GC) analysis (lag order=1) based on the REST (http://www.restfmri.net) toolbox on pseudotime series (online supplemental figure S1b). We defined ROIs using the AAL brain atlas, which divides the brain into 116 regions. We calculated the weighted out-degree and in-degree values for each ROI, where the out-degree value was the sum of GC values projected by the corresponding ROI to all other ROIs, and the in-degree value was the sum of GC values received by the corresponding ROI.22 We applied a threshold to GC values with an absolute value >0.16, corresponding to an FDR correction at p<0.05.22 Nodes with out-degree values or in-degree values >18.4 (115×0.16=18.4) were defined as the prominent nodes. To further evaluate the reliability of prominent nodes, we used block bootstrap sampling of the pseudo-time series (block size=4) and reported the occurrence rate of each prominent node across 1000 bootstrap iterations. However, as the above process sequenced sMRI data according to a single ADHD index, it could not fully capture the two major behavioural domains of ADHD: inattention and hyperactivity-impulsivity. To better characterise causal GMV changes across these behavioural domains, we performed BCaSCN analysis. Specifically, GMV maps within each ADHD subtype were ordered according to inattentive and hyperactive/impulsive index values from low to high. Pseudo-time series of GMV maps were then generated, and ROI-wise GC analyses were performed (online supplemental figure S1c). We also calculated the out-degree value and in-degree value of ROIs and identified the prominent nodes associated with the two behavioural domains. Statistical analysis Statistical analyses were performed using SPSS software (V.24.0). Differences in demographic and clinical indicators between NCs and participants with ADHD were examined using the χ2 test for sex and t-test for age, ADHD index, inattentive index, hyperactive/impulsive index, full-scale IQ (FIQ), verbal IQ (VIQ) and performance IQ (PIQ). Demographic and clinical differences among NC and ADHD subtypes were further evaluated using the Bonferroni method for post-hoc analysis. P<0.05 was considered statistically significant. Results Demographic and clinical data of participants with ADHD and NCs Demographic and clinical characteristics of participants with ADHD and NCs are presented in online supplemental table S2. Significant group differences were found in age, sex, FIQ, VIQ, PIQ, ADHD index, inattentive index and hyperactive/impulsive index. GMV differences between participants with ADHD and NCs Voxel-based morphology (VBM) analysis was conducted to evaluate the differences in GMV maps between participants with ADHD and NCs. No significant GMV differences were observed after FDR correction, which may have resulted from the structural heterogeneity of ADHD. Two neuroanatomical ADHD subtypes HYDRA estimated the consistency of the clusters by assigning the number of clusters (K) from 2 to 10. The clustering results showed that ARI reached its maximum when K=2, indicating that dividing ADHD into two neuroanatomical subtypes based on the volume of ROIs was optimal (online supplemental figure S2).1,14 The resulting ARI graph (online supplemental figure S3) again peaked at K=2, confirming that the initial feature set produced stable clustering results.515,26 Following HYDRA classification, demographical and clinical characteristics were compared among the ADHD subtypes and NCs (online supplemental table S3). Significant differences were observed in sex, FIQ, VIQ, PIQ, ADHD index, inattentive index and hyperactive/impulsive index between each ADHD subtype and NCs. There were no significant differences between subtypes 1 and 2 on these demographical and clinical scales. Subtype 1 exhibited slightly lower ADHD, inattentive and hyperactive/impulsive indexes compared with subtype 2. Group comparisons in GMV maps between each ADHD subtype and NCs revealed distinct volumetric patterns. Specifically, subtype 1 exhibited an increase in GMV across the whole brain (T>2.450, p<0.05, FDR-corrected), whereas subtype 2 showed global GMV reductions (T<–2.450, p<0.05, FDR-corrected) (figure 2). MR images showing distinct patterns of GMV in the two ADHD subtypes compared with NCs (p<0.05, FDR-corrected, cluster size threshold ≥50 voxels). Specifically, ADHD subtype 1 exhibits increased GMV, while ADHD subtype 2 shows decreased GMV. Increased GMV is shown in red and decreased GMV in blue, overlaid on an axial template. The colour bar represents t-values from the two-sample t-test. ADHD, attention-deficit/hyperactivity disorder; FDR, false discovery rate; GMV, grey matter volume; NC, neurotypical control. Progressive GMV changes in each ADHD subtype To investigate the progressive patterns of GMV, each ADHD subtype was further divided into three stages based on the ADHD index: stage 1 (ADHD index 30–50), stage 2 (51–70) and stage 3 (71–90). Comparisons of GMV between each subtype and NCs are shown in figure 3, and demographical and clinical data are provided in online supplemental table S4. MR images showing progressive alterations in GMV in the two subtypes of ADHD compared with NCs. Increased GMV is shown in red and decreased GMV in blue, overlaid on an axial template (p<0.05, FDR-corrected, cluster size threshold ≥50 voxels). Based on the ADHD index, participants in each subtype were further divided into three stages: stage 1 (ADHD index 30–50), stage 2 (51–70) and stage 3 (71–90). ADHD, attention-deficit/hyperactivity disorder; FDR, false discovery rate; GMV, grey matter volume; NC, neurotypical control. For ADHD subtype 1, both stage 1 (n=30) and stage 2 (n=26) showed increased GMV mainly in the frontal, parietal and temporal regions (T>2.953, p<0.05, FDR-corrected) compared with NCs. In stage 3 (n=11), GMV increases further extended to the cerebellum in addition to the aforementioned regions (T>2.953, p<0.05, FDR-corrected). For ADHD subtype 2, stage 1 (n=22) showed no significant GMV differences compared with NCs, whereas stages 2 (n=29) and 3 (n=17) both exhibited GM reductions mainly in the bilateral cerebellum, insula, limbic systems, frontal, temporal and occipital regions (T<–2.953, p<0.05, FDR-corrected). CaSCN and BCaSCN results To explore the causal relationships between GMV changes in ADHD subtypes and behavioural symptoms, we performed CaSCN analysis on pseudo-time series ordered by ADHD index values (figure 4). Subtype 1 showed no out-degree prominent nodes and one in-degree prominent node, namely, the cerebellar vermis 9 (occurrence rate=10.50%). For ADHD subtype 2, three out-degree prominent nodes were identified: the right cerebellar region 10 (occurrence rate=10.70%), left hippocampus (occurrence rate=10.90%) and cerebellar vermis 10 (occurrence rate=9.80%). No in-degree prominent nodes were identified for subtype 2 (online supplemental table S5 and S6). GC analysis based on ROI illustrating the causal relationships among ROIs in different subtypes of ADHD. (A, C) Bivariate signed-path GC analysis results showing ROI-wise causal networks for subtypes 1 (A) and 2 (C). (B, D) Weighted out- and in-degree sums for the top 30 ROIs in subtypes 1 (B) and 2 (D). Weighted out-degree/in-degree values for all 116 ROIs across both subtypes are provided in online supplemental figures S5 and S6. ADHD, attention-deficit/hyperactivity disorder; Angular, angular gyrus; Calcarine, calcarine cortex; Caudate, caudate nucleus; Cerebellum_3, cerebellum area 3; Cerebellum_4_5, cerebellum areas 4 and 5; Cerebellum_6, cerebellum area 6; Cerebellum_7b, cerebellum area 7b; Cerebellum_8, cerebellum area 8; Cerebellum_9, cerebellum area 9; Cerebellum_10, cerebellum area 10; Cerebellum_Crus1, cerebellum crus I; Cingulum_Ant, anterior cingulum; Cingulum_Mid, middle cingulum; Cingulum_Post, posterior cingulum; CRBL10, cerebellum area 10; Frontal_Inf_Oper, opercular part of inferior frontal gyrus; Frontal_Inf_Orb, orbital part of inferior frontal gyrus; Frontal_Inf_Tri, triangular part of inferior frontal gyrus; Frontal_Med_Orb, medial orbital superior frontal gyrus; Frontal_Mid, middle frontal gyrus; Frontal_Sup_Medial, medial part of superior frontal gyrus; Fusiform, fusiform gyrus; GC, Granger causality; Heschl, Heschl gyrus; HIP, hippocampus; L, left; Lingual, lingual gyrus; Occipital_Mid, middle occipital gyrus; Occipital_Sup, superior occipital gyrus; Parietal_Sup, superior parietal gyrus; ParaHippocampal, parahippocampal gyrus; Postcentral, postcentral gyrus; Precentral, precentral gyrus; R, right; Rectus, gyrus rectus; ROI, region-of-interest; Rolandic_Oper, Rolandic operculum; Temporal_Mid, middle temporal gyrus; Temporal_Pole_Mid, middle temporal pole; Temporal_Pole_Sup, superior temporal pole; T emporal_Sup, superior temporal gyrus; Vermis_1_2, Vermis area 1 and 2; Vermis_8, Vermis area 8; Vermis_9, Vermis area 9; Vermis10, Vermis area 10. In addition to sequencing sMRI data according to the ADHD index, we ordered the data by inattentive index and hyperactive/impulsive index and performed ROI-wise BCaSCN analyses (figure 5). We found that for ADHD subtype 1, the inattentive domain was associated with a greater number of prominent nodes, indicating stronger links between inattentive symptoms and brain structural changes in this subtype. In contrast, hyperactive/impulsive symptoms were more closely linked to brain structural changes in ADHD subtype 2. ROI-wise BCaSCN analysis based on behavioural-domain scales revealing prominent nodes driving changes in the brain region associated with behavioural tendency during disease progression. Prominent nodes with out-degree or in-degree values >18.4 in the ROI-wise BCaSCN results for ADHD subtype 1 (A) and ADHD subtype 2 (B). The sum of weighted out-degree and in-degree values for all 116 ROIs in subtypes 1 and 2 is presented in online supplemental figures S7–S10. Bars represent the sum of weighted out-degree/in-degree values of the prominent nodes. ADHD, attention-deficit/hyperactivity disorder; BCaSCN, behaviour-causal structural covariance network; Cau, caudate nucleus; CRBL10, cerebellum area 10; CRBL45, cerebellum areas 4 and 5; CRBL6, cerebellum area 6; HES, Heschl gyrus; L, left; ORBinf, orbital part of inferior frontal gyrus; ORBsup, orbital part of superior frontal gyrus; PoCG, postcentral gyrus; R, right; ROI, region-of-interest; SFG, superior frontal gyrus; TPOsup, superior temporal pole. Discussion Main findings In this study, we analysed sMRI and clinical data from 135 children and adolescents with ADHD and 182 NCs. Using the HYDRA algorithm, participants with ADHD were first divided into two neuroanatomical subtypes. To identify subtype-specific changes in GMV, VBM analysis was performed to compare each subtype with NCs. To explore the progressive and causal patterns of GMV changes in brain regions in relation to clinical symptoms, sMRI data within each subtype were ranked according to ADHD index, inattentive index and hyperactive/impulsive index, generating pseudo-time series data that represented disease progression. ROI-wise GC analysis was then performed on these pseudo-time series data. The results indicated that the use of different symptom indices as severity references yielded distinct sets of prominent nodes across the whole brain between the two subtypes, further revealing the complexity and diversity of brain structural changes in ADHD. Voxel-wise comparisons of GMV between participants with ADHD and NCs revealed no significant differences, which contrasts with current understanding of the pathophysiological mechanisms underlying ADHD.7 8 Given that ADHD is a heterogeneous disorder with diverse manifestations, the changes in GMV within the patient population are likely to vary substantially. This heterogeneity may hinder the detection of region-specific GMV changes when conducting direct group-level comparisons between individuals with ADHD and NCs.26 To address the impact of heterogeneity on VBM results, we employed the HYDRA algorithm to further subdivide participants with ADHD into two neuroanatomical subtypes. The primary advantage of HYDRA is its ability to leverage known group distinctions to guide the discovery of unrecognised, clinically relevant subtypes within a heterogeneous patient population like ADHD. Specifically, HYDRA integrates existing biological information distinguishing ADHD from NCs to reduce neuroanatomical heterogeneity. Subsequent group analysis revealed that compared with NCs, ADHD subtype 1 exhibited increased global GMV, whereas ADHD subtype 2 exhibited decreased global GMV, which was consistent with our previous results.20 For ADHD subtype 1, we observed a trend of increased GMV in specific brain regions with increasing disease severity, primarily involving the superior frontal gyrus (SFG) and inferior parietal gyrus (IPG). Consistent with our findings, Seidman et al reported a significant increase in GMV in the dorsolateral prefrontal cortex, which is a part of the SFG, in individuals with ADHD.27 Similarly, Sowell et al reported an increase in GMV in the bilateral IPG in individuals with ADHD.28 The prefrontal cortex, which includes the SFG, is implicated in attention, emotional regulation and higher-level functions such as goal management, working memory and inhibitory control.29 The IPG is implicated in spatial working memory processes and is considered an important hub for spatial attention transfer.30 31 Taken together, the progressive GMV increases in the SFG and IPG observed in ADHD subtype 1 may be associated with the inattentive symptom domain of ADHD. For ADHD subtype 2, we observed atrophy of grey matter in the bilateral cerebellum, SFG, middle frontal gyrus (MFG) and left hippocampus with increasing disease severity. Consistent with our findings, previous studies have also reported grey-matter reductions in several cortical regions in individuals with ADHD, including the bilateral cerebellum and SFG,32 the MFG33 and the left insula.34 The cerebellum is considered a key brain region in the pathogenesis of ADHD, playing a pivotal role in the attentional process, impulse control and motor coordination.35,38 The MFG has been associated with sensory integration and cognitive information processing in ADHD.39 The hippocampus is implicated in the impaired regulation of motivation and emotion in ADHD.40 Moreover, the anterior hippocampus encodes spatial and temporal relationships among sensory experiences,41 and structural changes in this region may underlie some of the characteristic behaviours observed in ADHD.42 43 The progressive GMV reductions in these regions might contribute to the more severe symptoms in ADHD subtype 2. To further investigate the causal relationships underlying GMV changes in specific brain regions in relation to clinical symptoms, we conducted an ROI-wise GC analysis on pseudo-temporal series of sMRI data for each ADHD subtype. Using various behavioural domains as reference indicators of disease severity enabled the identification of the potential clinical focus of each ADHD subtype. BCaSCN analysis demonstrated that ADHD subtype 1 exhibited stronger causal relationships among brain regions associated with inattention, suggesting that attentional dysfunction may represent the primary impairment in this subtype. In contrast, subtype 2 demonstrated a more widespread pattern of causality, involving not only the ADHD index but also multiple domains such as inattention and hyperactivity/impulsivity. These findings highlight the complexity of the pathophysiological mechanisms underlying ADHD516,18 and are consistent with the differences observed in clinical scales between the two ADHD subtypes. Limitations This study has several limitations. First, the BCaSCN approach, which is based on pseudo-temporal series sMRI data, may not accurately capture the true temporal progression of the disorder. Additionally, the progressive patterns of GMV changes across ADHD subgroups were derived using the ADHD index, which may not comprehensively reflect the severity of the condition. Second, this study employed cross-sectional data, providing only an initial snapshot of structural changes in brain regions in individuals with ADHD. Future longitudinal and follow-up investigations are necessary to gain a more comprehensive understanding of the dynamic process of brain structural changes associated with clinical symptoms in ADHD. Implications ADHD is a neurodevelopmental disorder associated with structural changes throughout the brain. However, the causal relationships between GMV changes and clinical symptoms in ADHD remain unclear. In this study, by applying HYDRA, we first categorised ADHD into two subtypes, thereby addressing the structural heterogeneity of ADHD. We then constructed pseudo-time series sMRI data to model disease progression for each subtype and used BCaSCN analysis to explore the causal relationships in GMV changes in different ADHD subtypes. The results revealed distinct patterns of GMV alterations: subtype 1 showed causal changes more strongly associated with inattention, whereas subtype 2 was linked to overall disease severity. These findings reveal structural heterogeneity in ADHD, suggesting distinct causal patterns linked to different behavioural domains and emphasising the need for subtype-specific interventions tailored to their specific neural mechanisms. Translating these findings into clinical practice could offer tentative but targeted guidance for future clinical strategies. Individuals with ADHD subtype 1, characterised by prominent frontal–cerebellar involvement in inattention, may benefit most from cognitive training regimens specifically targeting attention networks. In contrast, subtype 2, with its widespread GMV atrophy and greater overall severity, may require more intensive, multimodal interventions combining pharmacological and comprehensive behavioural therapies. Future research validating these subtypes prospectively could pave the way for such personalised treatment approaches for ADHD. Supplementary material Biography Tianzheng Zhong is a fifth-year undergraduate student in the School of Radiology at Shandong First Medical University in China, expecting to graduate with a Bachelor of Medicine in 2026. He enrolled at Shandong First Medical University in 2021, where he has undertaken comprehensive studies in core disciplines such as Medical Imaging Technology, Radiological Diagnostics, and Radiation Therapy. His academic performance has been consistently outstanding, demonstrating a strong foundation in both theoretical knowledge and practical skills. Driven by a keen interest in interdisciplinary medical innovation, he has actively engaged in scientific research training. His academic focus lies in the application of artificial intelligence in medical image analysis, neuroimaging, and causal modelling. Additionally, he has gained valuable clinical experience through practicums in the imaging departments of several affiliated hospitals, where he effectively translated classroom learning into patient care. He has also participated in volunteer medical service activities, further solidifying his commitment to the field. As an aspiring radiology professional, he is determined to pursue a career in medical imaging or radiation therapy, aiming to contribute to advancing diagnostic and therapeutic technologies for patient benefit. Footnotes Funding: This study was funded by the Department of Science and Technology of Shandong Province, Natural Science Foundation of Shandong Province (ZR2023QH109) and Taishan Scholars Program of Shandong Province (TS201712065). Provenance and peer review: Not commissioned; externally peer reviewed. Patient consent for publication: Consent obtained from parent(s)/guardian(s). Ethics approval: This study involves human participants and this study was approved by the Institutional Review Board of Shandong First Medical University in accordance with the Declaration of Helsinki (No. 2025-H034). Participants gave informed consent to participate in the study before taking part. Data availability statement Data were provided by the ADHD-200 dataset (http://fcon_1000.projects.nitrc.org/indi/adhd200). References Associated Data Supplementary Materials Data Availability Statement Data were provided by the ADHD-200 dataset (http://fcon_1000.projects.nitrc.org/indi/adhd200).