Brain iron deficiency in children with attention deficit hyperactivity disorder using iterAtive magnetic suscePtibility sources sepARaTion (APART-QSM)

Abstract

Background

IterAtive magnetic suscePtibility sources sepARaTion (APART-QSM), a recently proposed susceptibility source separation method, can differentiate paramagnetic and diamagnetic susceptibility distributions related to iron and myelin, respectively. This study aimed to investigate whether paramagnetic susceptibility values of deep gray matter structures combined with machine learning algorithms could be used to identify individuals with attention-deficit/hyperactivity disorder (ADHD) and to further explore ADHD-related pathogenesis.

Methods

Thirty-six ADHD and 35 age, sex-matched healthy controls (HCs) were recruited. The paramagnetic susceptibility mapping obtained by using APART-QSM method was normalized and the positive susceptibility values of deep gray matter structures, including the bilateral caudate nucleus, putamen, pallidum, and thalamus, were extracted. Random forest (RF) and support vector machine (SVM) were adopted to build machine learning models based on regional positive susceptibility values. The accuracy, sensitivity, specificity and the area under the curve (AUC) were used to evaluate the classification performance.

Results

Lower positive susceptibility values of the left caudate nucleus and bilateral pallidum were found (Caudate_L: 0.0231±0.0045 vs. 0.0261±0.0051, Pallidum_L: 0.0431±0.0114 vs. 0.0503±0.0141, Pallidum_R: 0.0426±0.0119 vs. 0.0488±0.0120, P<0.05, uncorrected). However, no significant correlations were found between decreased iron levels and attention performance. Both classifiers achieved good performance, particularly the RF model with an AUC of 0.756, sensitivity of 77.8%, specificity of 68.6% and accuracy of 73.2%.

Conclusions

Our findings revealed iron deficiency of deep gray matter nuclei in children with ADHD, and machine learning models combined with APART-QSM could be used to distinguish ADHD from HCs, providing a potential biomarker for further understanding of ADHD pathophysiology and facilitating early diagnosis.

Introduction

Attention-deficit/hyperactivity disorder (ADHD) is a child-onset neurodevelopmental disorder characterized by deficits in attention, impulsivity and hyperactivity (1). The estimated prevalence of ADHD in children is approximately to 7.6%, with a higher prevalence for boys (2). While the resulting behavior and cognitive symptoms actually have an impact on the academic, social and occupational functioning, the pathophysiological mechanisms underlying ADHD remain to be investigated. Along with the findings that multiple risk factors contribute to the development of ADHD, including genetic, early environment factors, and the interactions between them, recent findings of dopaminergic system deficits are suspected to play a central role in the pathophysiology of ADHD (1,3,4). Importantly, the proposed dopamine-related mechanism is supported by the fact that the psychostimulants are the first-line treatment for ADHD, which primarily act on the striatal dopaminergic system (5).

Notably, iron, as an essential element, plays a vital role in many neurological processes, including oxygen transportation, neuronal development, myelin formation, as well as neurotransmitter synthesis (6). Early reduction in serum iron levels has been found to be associated with cognitive and behavioral symptoms, and increased risk for a variety of disorders, such as restless legs syndrome, autism spectrum disorder, Tourette’s syndrome and ADHD (7,8). Although previous studies have found peripheral serum iron decreases in ADHD (9,10), considering the fundamental relationship between brain iron and the dopaminergic system, which plays a central role in the pathophysiology of ADHD, the investigation of brain iron has gained increasing attention in recent years. Based on different magnetic resonance imaging (MRI) techniques [transverse relaxation rate T2*, magnetic field correlation (MFC) and quantitative susceptibility mapping (QSM)], previous studies (11-14) using regions of interest (ROIs) and whole-brain voxel-size analysis consistently reveal a significant reduction in brain iron content in medication-naïve ADHD children and that psychostimulants use may normalize the brain iron in pediatric ADHD (15). Additionally, those studies have shown that the deep gray matter nuclei are not only iron-rich structures but also involved in the pathogenesis circuits in ADHD, thus, the abnormalities in iron content within these deep gray matter nuclei warrant thorough investigation. QSM analysis enables quantitative detection of source susceptibilities. However, because paramagnetic and diamagnetic substances often coexist in most regions of the brain, traditional QSM analysis can only provide voxel-averaged susceptibility mapping, derived from the mixture of paramagnetic and diamagnetic susceptibility sources (16). This averaging effect can lead to the cancellation of signals from paramagnetic and diamagnetic components.

Recently, the proposed susceptibility separation method, IterAtive magnetic suscePtibility sources sepARaTion (APART-QSM), employs a comprehensive complex data model and iterative voxel-size magnitude decay kernel estimating algorithm to separately obtain the paramagnetic and diamagnetic mappings (17). While paramagnetic susceptibility mapping is mainly for the brain iron estimation and diamagnetic susceptibility mapping is often influenced by myelination, the capability to separate paramagnetic susceptibility portion from diamagnetic susceptibility is hypothesized to be more reliable and specific to estimate brain iron distribution than traditional QSM analysis in which the effects of paramagnetic and diamagnetic susceptibility would counteract to some extent. Moreover, the APART-QSM method has shown good application in investigating the changes in brain iron during age-related brain development due to the high-quality separate susceptibility mapping that it generates (18). Meanwhile, the susceptibility separation method has demonstrated the ability to provide a more comprehensive understanding of the evolution of paramagnetic and diamagnetic lesions in multiple sclerosis (19). However, up till now, the use of APART-QSM to quantify the brain iron content in ADHD has not been previously reported.

Hence, in the study, we employed the APART-QSM method to investigate paramagnetic iron alterations in iron-rich deep gray matter structures in ADHD. Further, machine learning algorithms combined with the paramagnetic iron values derived from APART-QSM analysis were adopted to explore whether abnormal iron distribution in deep gray matter structures could serve as a potential imaging biomarker to distinguish ADHD from healthy controls (HCs). We expected that our proposed method could advance our understanding of the pathological mechanism of ADHD and provide a potential biomarker for early diagnosis. We present this article in accordance with the CLEAR reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2800/rc).

Methods

Subjects

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the institutional review board of The First Affiliated Hospital, Sun Yat-sen University (No. [2019]328), and has been registered online (https://clinicaltrials.gov/; Identifier: ChiCTR2100048109). All individuals with ADHD included in our study were recruited from the pediatrics department of The First Affiliated Hospital of Sun Yat-sen University from March 2019 and February 2021. Each subject or their legal guardians signed an informed consent form approved by the ethics committee of The First Affiliated Hospital of Sun Yat-sen University. For ADHD patients, the inclusion criteria were as follows: (I) all patients with ADHD were diagnosed by two pediatricians with >20 years of experience according to the Diagnostic and Statistical Manual of Mental Disorders 5^th edition (DSM-V) criteria (20), as well as the parent and teacher reports on Conners Symptom Questionnaire (21) (a partially simplified version of Conner scale were used in practice for clinical feasibility). Individuals were diagnosed as ADHD positive if their parent-rated and teacher-rated ADHD indices ≥75^th percentile. Namely, the subjects had a diagnosis of ADHD by two experienced pediatricians at the beginning of the study. (II) Patients ranged in age from 6 to 14 years and were right-handed. The exclusion criteria were as follows: (I) with a history of head injury or any other neurological dysfunction; (II) other psychiatric conditions, such as tic disorder, conduct disorder, oppositional defiant disorder, autism spectrum disorder, bipolar disorder and obsessive-compulsive disorder; (III) a history of medication therapy that may affect the development of central nervous system; (IV) MRI contraindication; (V) MRI data with obvious artifacts and head motion which were evaluated by two experienced radiologists. All HCs were assessed by the same questionnaire, with parent’s and teacher’s ADHD indices <75^th percentile to exclude ADHD diagnosis. An experienced pediatrician conducted a clinical interview with HCs to rule out any other brain injury, neurological and psychiatric disorders. The exclusion criteria for the ADHD group were also applied to the HCs group.

The digital cancellation test (DCT) (22-24) was used to evaluate the attention performance of ADHD patients, which could be a measure of executive function. Participants were asked to cancel the specified digit in a series of trials, and then cancellation test scores and concentration index were calculated according to the correct number, omission errors and commission errors. Higher scores and an index mean a better attention level.

MRI acquisition

All images were acquired on a 3.0 T scanner (SIGNA Pioneer GE Healthcare, WI, USA) equipped with a 32-channel head coil. During scanning, all participants fixed their heads with foam padding to minimize head motion and wore earplugs to reduce scanning noise. A 3D multi-echo gradient echo (GRE) sequence with 8 echoes was obtained using the following parameters: repetition time (TR) = 46.7 ms, echo time₁ (TE₁) =4.72 ms, TE₈ =41.68 ms, ΔTE =5.28 ms, flip angle =20°, thickness/gap =2.0 mm/no gap, slice number =76, field of view (FOV) = 256 mm × 256 mm, and matrix = 256×256, scanning time =5 min 19 s. Three-dimensional high-resolution structural T1-weighted data were acquired using the fast spoiled gradient recalled echo (FSPGR) sequence with the following parameters: TR =8.6 ms, TE =3.3 ms, flip angle =12°, slice thickness/gap = 1.0 mm/no gap, slice number =192, FOV = 256 mm × 256 mm, and matrix =256×256, scanning time =5 min 28 s. In addition, T2-weighted fluid attenuation inversion recovery (FLAIR) images were acquired to exclude any brain abnormalities.

Paramagnetic susceptibility mapping reconstruction and normalization

The APART-QSM postprocessing is mainly composed of two procedures: paramagnetic susceptibility mapping reconstruction and spatial normalization. The procedures of paramagnetic susceptibility mapping reconstruction were as follows: the brain extraction tool for tissue mask extraction, Rapid Opensource Minimum spanning treE algorithm (ROMEO) total field calculation for phase unwrapping (25), variable kernel sophisticated harmonic artifact reduction for phase data (VSHARP) method for the background field removal (26), morphology enabled dipole inversion (MEDI) algorithm for calculating the susceptibility mapping (27-29), and finally APART-QSM algorithm for reconstructing the paramagnetic and diamagnetic susceptibility mapping (17,30). The detailed information of APART-QSM algorithm could be referred to the previous study (17). All the mentioned procedures were performed using MATLAB (R2021b). In addition, to control the potential effect of excessive head motion, the exclusion of 3D T1w data was mainly based on visual inspection to check for artifacts associated with blurring, ghosting and head motion. For QSM data acquired with a GRE sequence with 8 echoes, the multi-volume GRE magnitude data were realigned to obtain the volume-to-volume head motion, namely, framewise displacement (FD). Using FD >0.3 mm as a threshold, subjects with more than 50% of volumes exceeding the threshold were excluded.

The spatial normalization pipeline of paramagnetic susceptibility mapping to Montreal Neurological Institute (MNI) space was established using FMRIB Software Library (FSL, https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/) (31). In brief, the single-subject brain mask was applied to the magnitude image of the first echo to obtain skull-stripped images. The skull-stripping of the individual 3D T1w was performed based on the T1w segmentation procedure to retain the brain tissue only. Then, the linear registration from the magnitude image of the first echo to the individual 3D T1w image, followed by nonlinear registration from the individual 3D T1w to the MNI template, was combined to generate the composite transforms using FSL. After that, the obtained concatenated transforms were applied to the paramagnetic susceptibility mapping to achieve the normalization to the MNI space. Eight ROIs in the deep gray matter structures were defined according to the Neuromorphometrics Atlas (http://www.neuromorphometrics.com/), including the bilateral caudate nucleus, putamen, pallidum, and thalamus. The averaged paramagnetic susceptibility value of each ROI was extracted and used as the input features for the subsequent machine learning analysis.

Classification model with machine learning

We employed two machine learning algorithms, support vector machine (SVM) and random forest (RF), to distinguish ADHD from HCs with the paramagnetic susceptibility values of the defined ROIs as input features. The SVM classifier is a representative supervised algorithm to perform classification by finding a hyperplane to maximize the margin between classes in a high-dimensional space (32). Based on the ensemble learning method, the RF classifier integrates the prediction results of multiple independent decision trees, and the final classification decision is determined by the prediction with the most votes from all trees (33). The performance of these classifiers was estimated by a nested leave-one-out cross-validation (LOOCV) method, which had the most unbiased estimate of test error and was more suitable for small sample cases. The outer LOOCV loop was used to split the training data and testing data, in which one subject was considered as testing data to perform validation and the remaining subjects were defined as training data to train the model, while the inner loop was conducted on the training data for hypermeter optimization (i.e., number of decision trees for RF, penalty coefficient C and kernel width parameter γ for SVM) via grid-search strategy with 5-fold cross-validation. In fact, this process was repeated N times (for N subjects) for the outer-loop, and all subjects had a predictive value for the final classification performance. Finally, we took the sensitivity, specificity, accuracy and area under the curve (AUC) into account to evaluate the classification performance. The above-mentioned workflow was implemented in Python 3.8 using the scikit-learn package (34).

Statistical analysis

The demographic and clinical data between ADHD and HCs were compared using a two-sample t-test or Chi-squared test as appropriate. A two-tailed P<0.05 was considered to be significant. The mean positive susceptibility values of ROIs were also compared between groups, with false discovery rate (FDR) correction for multiple comparisons. To investigate whether the regions with significant abnormality in positive susceptibility value were related to the attention performance, partial correlation analyses were also conducted in the ADHD group with age and sex as covariates.

Results

Demographic and clinical characteristics

In our study, 36 ADHD and 35 HCs were included. There were no significant differences in sex (male/female: 32/4 vs. 31/4) or age (mean ± SD: 8.74±2.17 vs. 8.74±1.74 years). Demographic and clinical information are shown in Table 1.

Table 1. Demographic and clinical features of ADHD and HCs.

Characteristics	ADHD (n=36)	HCs (n=35)	Statistics	P value
Age, years, mean ± SD [range]	8.74±2.17 [6–14]	8.74±1.74 [6–13]	−0.03	0.998
Sex (male/female), n	32/4	31/4	0	>0.99
Concentration index, mean ± SD	15.62±16.89	NA	NA	NA
Cancellation test score, mean ± SD	5.76±30.56	NA	NA	NA

ADHD, attention-deficit hyperactivity disorder; HCs, health controls; SD, standard deviation.

ROI-based comparison between groups

The representative images of ROIs segmentation overlaid with spatially normalized paramagnetic susceptibility mapping are shown in Figure 1. The mean paramagnetic susceptibility values of the ROIs are extracted and summarized in Table 2 and Figure 2. Compared with HCs, individuals with ADHD showed significantly lower positive susceptibility values in the left caudate nucleus and bilateral pallidum (P<0.05, uncorrected), and no significant differences were found in other regions. Nevertheless, the group differences did not reach statistical significance after FDR correction.

Figure 1.

The representative images of the defined deep gray matter nuclei overlaid in paramagnetic susceptibility mapping. The paramagnetic susceptibility mapping was derived from quantitative susceptibility data acquired with multi-echo gradient echo sequence. The deep gray matter nuclei (caudate, pallidum, putamen and thalamus) were defined and labeled on paramagnetic susceptibility mapping. The above shaded gray bar indicated the quantification of positive susceptibility values range in ppm, with darker regions corresponding to lower iron content.

Table 2. The mean, standard deviation and P value of the positive susceptibility values in the deep gray matter nuclei between groups.

ROIs	ADHD	HCs	P value	Cohen’s d
Caudate_L	0.0231±0.0045	0.0261±0.0051	0.011	−0.63
Caudate_R	0.0235±0.0046	0.0252±0.0044	0.122	−0.38
Putamen_L	0.0218±0.0040	0.0217±0.0046	0.893	0.02
Putamen_R	0.0220±0.0035	0.0209±0.0042	0.271	0.28
Pallidum_L	0.0431±0.0114	0.0503±0.0141	0.022	−0.56
Pallidum_R	0.0426±0.0119	0.0488±0.0120	0.033	−0.52
Thalamus_L	0.0218±0.0028	0.0218±0.0045	0.950	0.00
Thalamus_R	0.0216±0.0030	0.0217±0.0041	0.871	−0.03

Data are presented as mean ± SD. Cohen’s d, the effect sizes indicated the magnitude of differences between groups, with negative values suggesting lower values in the ADHD compared to HCs. ADHD, attention-deficit/hyperactivity disorder; HCs, healthy controls; L, left; R, right; ROIs, regions of interest; SD, standard deviation.

Figure 2.

Comparisons of ROIs-based positive susceptibility values between ADHD and HCs. Caudate_L: 0.0231±0.0045 vs. 0.0261±0.0051, Pallidum_L: 0.0431±0.0114 vs. 0.0503±0.0141, Pallidum_R: 0.0426±0.0119 vs. 0.0488±0.0120. Asterisks indicate statistically significant group differences (*, P<0.05, uncorrected). ADHD, attention-deficit/hyperactivity disorder; HCs, healthy controls; L, left; R, right; ROIs, regions of interest.

Classification results of machine learning models

The classification performance of the classifiers is shown in Figure 3. In our study, the included cohort exhibited a nearly equivalent distribution of ADHD and HCs. The SVM classifier achieved an AUC of 0.718, sensitivity of 75.0%, specificity of 65.7% and accuracy of 70.44%. The performance of RF classifier was better than the SVM classifier, with an AUC of 0.756, sensitivity of 77.8%, specificity of 68.6%, and accuracy of 73.2%. Since RF was an ensemble of decision trees, the feature importance could be obtained directly by averaging the importance index calculated based on the mean decrease in impurity of the RF classifier over all LOOCV loops. Thus, the feature importance is also ranked and shown in Figure 4.

Figure 3.

The classification performance of machine learning models. (A) ROC curves of RF and SVM model. (B) The confusion matrix of the RF model. ADHD, attention-deficit/hyperactivity disorder; AUC, area under the curve; FPR, false positive rate; HCs, health controls; RF, random forest; ROC, receiver operating characteristic; SVM, support vector machine; TPR, true positive rate.

Figure 4.

The ranked feature importance across all LOOCV loops. L, left; LOOCV, leave-one-out cross-validation; R, right.

Correlation analysis

No significant correlations were found between positive susceptibility values and cancellation test scores, as well as the concentration index (Table S1).

Discussion

In this study, we used APART-QSM method to evaluate the potential iron-related pathophysiology of ADHD, and combined machine learning algorithms with paramagnetic-based susceptibility values of deep gray matter structures to identify individuals with ADHD, and three main findings were obtained: (I) the positive susceptibility values obtained from paramagnetic component mapping combined with machine learning algorithms (both SVM and RF models) achieved promising classification performance; (II) the proposed APART-QSM method for separating substances with paramagnetic and diamagnetic susceptibility was able to reveal brain iron deficiency which was consistent with previous findings of ADHD; (III) the significantly decreased positive susceptibility values were found in caudate nucleus and pallidum, but there was no significant correlation with attention performance.

In the past few years, machine learning algorithms combined with morphological, structural and functional imaging measurements have made substantial progress in revealing the pathogenesis mechanisms in ADHD (35). Nevertheless, these studies showed spatial inconsistency in the involved regions across studies, with cortical thickness variations, gray matter volumetric changes, white matter integrity disruptions and network-level dyshomeostasis. Notably, pathological changes such as brain iron content were supposed to precede macroscopic changes such as decreased brain volume in ADHD. Thus, it is very rewarding to evaluate brain iron homeostasis during pathological conditions, in which QSM analysis is often regarded as an efficient way for iron evaluation. Particularly, APART-QSM employed in our study could separate opposing sources to minimize the effect of diamagnetic content on the paramagnetic property of brain iron at the voxel level, which was limited by conventional QSM analysis (17). Compared to a similar study using the conventional QSM analysis to assess the total susceptibility of the basal ganglia nuclei in ADHD (36), our study had more positive results with statistically significant differences in the iron content of nuclei. This difference may be because the new method excludes the interference of diamagnetic substances such as myelin, accentuating the iron content difference. Consistent with previous studies using other methods in data analysis (11,12,37), our study found lower positive susceptibility values in the deep nuclei areas, indicating a reduction in brain iron levels in ADHD. Additionally, with the combination of machine learning models, the positive susceptibility values of deep gray matter nuclei showed good accuracy in differentiating ADHD from HCs, providing a potential biomarker for diagnosing ADHD. Furthermore, the diagnostic model based on the overall brain iron levels in deep gray matter nuclei, instead of a particular region, indicated that the spatial distribution pattern of iron within deep gray matter nuclei contains diagnostically significant information. Therefore, the findings of brain iron deficiency in ADHD confirmed the feasibility of APART-QSM for accurate estimation of brain iron and further highlighted the important role of iron in the pathogenesis of ADHD.

As an essential element for brain development, iron homeostasis is usually regulated by various factors to ensure the normal function of many neurophysiological processes. Critically, brain iron concentration is preferentially rich in deep gray matter structures, including the basal ganglia and thalamus, to maintain myelin synthesis, neurotransmitter synthesis and metabolism, and abnormal iron levels in these regions are associated with many diseases. Excessive brain iron accumulation was pathologically implicated in neurodegenerative diseases (38). In Parkinson disease, brain iron deposition induced oxidative damage and promoted Lewy body aggregation (39). Similarly, Alzheimer disease progression involved the interaction between brain iron overload, amyloid-beta plaques and tau proteins (40,41). In contrast to this, iron deficiency was often involved in neurodevelopmental diseases, including ADHD. Regarding the pathophysiological mechanisms between iron deficiency and ADHD, we proposed the following possible interpretations. First, lower iron levels were related to the disrupted dopamine receptors, dopamine transporters and dopamine synthesis (42,43). The resulting hypoactive dopaminergic system could contribute to the core symptoms related to impulsivity, inattention and hyperactivity in ADHD, and its action on the core frontostriatal circuits may provide a biological basis for neurocircuitry abnormalities in ADHD (44). Second, as iron was required for axon myelination, decreased iron levels were supposed to lead to demyelination. This aligned with neuroimaging evidence of white matter microstructure alterations (45) and further lent support to a long-standing hypothesis that dysregulated myelination may contribute to the brain developmental delay in ADHD (46). However, considering that brain iron deficiency was also associated with other neurodevelopmental disorders such as autism spectrum disorder and tic disorder (47,48), whether there was overlap or specificity in the pathogenesis needs to be explored in combination with genetic, cellular and molecular studies.

Of note, we found that the positive susceptibility values of the caudate nucleus and pallidum were significantly lower than those in HCs, and the caudate and pallidum also appeared as the top features among the defined deep nuclei in the RF algorithm. The caudate and pallidum are important components of the frontostriatal circuits associated with cognitive control and reward processing in ADHD (49). Neurotransmitters such as dopaminergic activity could exert neuromodulatory influence over behavior and cognition via the frontostriatal circuitry. In addition, studies using MFC imaging and diffusion tensor imaging (DTI) as proxy for measuring brain iron found reduced striatal iron levels in children with ADHD (13,50). With the use of QSM, Tang et al. and our previous study also revealed widespread decreased brain iron levels in frontal areas, substantia nigra, anterior cingulate gyrus, olfactory cortex as well as striatal structures (i.e., caudate nucleus, putamen and pallidum) in ADHD (11,12). Moreover, the brain iron levels decreased in the striatal structures in individuals with medication-naïve ADHD, but were comparable to HCs in individuals with psychostimulant-medication ADHD, providing preliminary evidence for the positive effect of psychostimulants on decreased striatal iron levels in ADHD (15). Thus, combined with existing research, our findings further emphasize the importance of decreased striatal iron levels in the pathogenesis of ADHD. However, the finding that there were no correlations between significantly decreased iron levels and attention concentration levels in ADHD seemed beyond expectation. As an etiologically, pathologically and clinically heterogeneous disease, ADHD was traditionally divided into three subtypes, namely inattentive, hyperactive and combined type. The DCT employed in our study was mainly used to evaluate the attention concentration level and could not fully characterize the symptoms of ADHD. Hence, we speculated that the heterogeneity of ADHD included in our study and the one-sidedness of the scale may be a reasonable explanation for this.

Limitations

Some limitations in this study need to be noted. First, the sample size in our study was relatively small, despite the use of nested cross-validation strategies to reduce the risk of overfitting, larger and multi-center cohorts would be necessary to confirm the generalizability. Second, the findings regarding the relationship between iron deficiency and clinical characteristics in previous studies remained inconsistent in ADHD. According to our findings, no significant associations between decreased brain iron levels and attention performance were found, thus, other clinical assessment scales need to be used in larger samples to verify the associations with decreased brain iron levels. Third, previous studies have also examined reduced system iron levels in children with ADHD. However, it remained unclear whether peripheral iron levels have an effect on brain iron levels, therefore, more comprehensive clinical information, such as serum ferritin, should be collected in the future. Fourth, it was worth noting that the ROI-based comparisons between groups showed significant differences only in the uncorrected analysis, future studies with a larger sample size may be more able to increase the statistical power. Fifth, while the relationships between body mass index (BMI), intelligence quotient (IQ) and brain iron level remained poorly understood, future studies should incorporate BMI assessment and IQ evaluation via Wechsler Intelligence Scale for Children (WISC) to address these issues. Additionally, future studies might need to recruit a more balanced sample to elucidate the effect of sex. Finally, our study was cross-sectional, future longitudinal studies may provide insight into elucidating the dynamic changes of brain iron during disease progression and verifying the effect of medication treatment on brain iron in children with ADHD.

Conclusions

In conclusion, the positive susceptibility values of striatal regions combined with machine learning algorithm could differentiate children with ADHD from HCs. The decreased iron levels in the caudate nucleus and pallidum might be associated with the disrupted deep gray matter dopamine system and demyelination hypothesis. These findings provide insight into the role of iron in the pathogenesis of ADHD and highlight the utility of APART-QSM in assessing iron-related pathophysiology in neurological diseases.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the ethics committee of The First Affiliated Hospital of Sun Yat-sen University (No. [2019]328). Informed consent was obtained from all subjects or their legal guardians who participated in the study.