Fifty-eight drug-naïve ADHD patients were recruited from the Out-patient Clinic of the First Affiliated Hospital of Wenzhou Medical University. Meanwhile, 31 healthy controls matched in gender, age and IQ were recruited from local schools. The diagnosis of ADHD was performed by three experienced psychiatrists according to the Fifth Edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-V). For all the participants, the exclusion criteria were: (1) conduct disorder, oppositional defiant disorder, or any other psychotropic disorder besides ADHD; (2) history of major neurological or physical illness; (3) current or history use of psychotropic drugs; (4) intelligence quotient (IQ) lower than 80; (5) MR scanning contraindications. This study was approved by the Ethics Committee of the First Affiliated Hospital of Wenzhou Medical University. All participants were well informed and written informed consent was obtained before the test.

The revised version of Conners’ Parent Rating Scale (CPRS) (Conners et al., 1998) was used to assess children’s behavioral problems, including conduct problems, study problems, psychosomatic disorders, impulse hyperactivity, anxiety, and hyperactivity.

The computerized Wisconsin Card Sorting Test (WCST) (Feldstein et al., 1999) was used to assess the executive function, including cognitive flexibility, working memory, continuous attention, and response inhibition. In specific, patients were required to match two cards according to the color, shape or quantity. After 10 trials of correct matching, one completed session was counted and the rule would be changed. The task was ended if the participants complete six sessions or 128 trials. The perseverative errors (at least two successive errors) and non-perseverative errors were counted. Note that non-perseverative errors were counted as the difference between the total number of errors and the number of perseverative errors.

Images were acquired in a 3T GE Sigma HDX 3.0 scanner (GE Healthcare, United States) with the following parameters: repetition time (TR) = 2,000 ms, echo time (TE) = 30 ms, 31 axial slices, slice thickness = 4 mm, slice gap = 0.2 mm, flip angle = 90°, matrix size = 64 × 64, field of view (FOV) = 192 × 192 mm, and 240 volumes. High-spatial-resolution T1-weighted 3D images were obtained axially using the following parameters:162 axial slices, repetition time (TR) = 7,200 ms, echo time (TE) = 2.2 ms, flip angle = 7°, section thickness = 1 mm, gap = 0, field of view (FOV) = 256 × 256 mm, data matrix = 256 × 256. During the scanning, sponge earplugs were used to reduce scanner noise, and foam cushions were used to minimize head movement. The subjects were required to be relax, stay still, close their eyes, and stay awake.

Data Processing and Analysis for Brain Imaging (DPABI) (Yan et al., 2016) was used for data preprocessing: (1) the first 10 volumes were eliminated to enable subjects to adapt to the scanning environment and the magnetic field to be stable; (2) slice timing correction; (3) head motion correction (subjects were excluded from further analysis if their maximal head motion exceeded 3 mm displacement or 3°of rotation, 8 ADHD patients and 3 HCs were therefore excluded, leaving 50 ADHD patients and 28 HCs); (4) structural image alignment, aligning the individual T1-weighted image to the average functional image through six-degree-of-freedom linear transformation; (5) structural image segmentation, which divides the structural image into gray matter, white matter and cerebrospinal fluid; (6) nuisance covariates regression, including the Friston-24 motion parameters, white matter signals, cerebrospinal fluid signals and linear drift. Note that the multiple direction motion parameters and the mean frame-wise displacement (FD) of each subject were calculated to evaluate and compare the head motion, results were listed in Table 1 and Supplementary Table 1; (7) based on the segmented images, the individual space was transformed to the MNI space using the DARTEL (diffeomorphic anatomical registration through exponentiated lie algebra) tool, and the resampled voxel size is 3 mm × 3 mm × 3 mm.

R-fMRI indices were calculated with following procedures:

ALFF (Zang et al., 2007; Zou et al., 2008). After Fourier transform of time series of a voxel, ALFF takes the mean amplitude of a specific frequency range (0.01–0.08 Hz) as the intensity of local activity of the voxel.

ReHo (Zang et al., 2004). The Kendall’s coefficient of concordance between the time series of a given voxel and the time series of its closest voxel.

DC (Buckner et al., 2009; Zuo et al., 2012). In this study, we calculate the Pearson correlation coefficient between the time course of each voxel and the time course of every other voxel in the whole brain. For each voxel, DC denotes the sum of connections that showed a correlation coefficient exceeding 0.25 (Buckner et al., 2009; Yan et al., 2017). Correlations below this threshold were considered mainly caused by signal noise. Meanwhile, it has been proven that different threshold selection from 0.1 to 0.3 did not significantly change the network structure (Buckner et al., 2009). The obtained DC values were converted to Z-scores by subtracting the global mean DC and dividing by the SD of the whole brain DC, the obtained voxel-wise DC map.

VMHC (Zuo et al., 2010). Function images were transformed into a symmetric space. Then, Pearson’s correlation coefficients between time series from homologous brain regions were calculated (VMHC). Finally, the VMHC maps were transformed by Fisher Z-transformation for subsequent analysis.

GSC (Hahamy et al., 2014; Power et al., 2017). BOLD signals from all the voxels were averaged as the global signal. Pearson correlation coefficients between each voxel time course and the global signal were calculated to get the GSC maps, finally The GSC maps included in the subsequent analysis were Fisher Z-transformed.

In particular, dynamics of these R-fMRI indices were analyzed using Temporal Dynamic Analysis (TDA) toolkits based on DPABI (Yan et al., 2016). Specifically, R-fMRI indices mentioned above were computed with the hamming windows (window length = 32 TR, window step = 4 TR, generating 50 windows). Previous studies have revealed that even a window length ranged from 30 to 60 s is able to capture dynamic functional connectivity fluctuations and different window lengths do not yield significantly different results (Preti et al., 2017). Also, previous studies have demonstrated that the step size does not dramatically affect the variance of dynamic properties of R-fMRI (Liao et al., 2019). Therefore, we used a window length of 32 TR (64 s) and a step size of 4 TR (8 s) (Yang et al., 2020). More importantly, different window length (48 TR/64 TR) were used to validate our results (see Supplementary Materials). The R-fMRI indices mentioned above were prefixed with a d to indicate the dynamic indices. For example, dALFF represents dynamic ALFF. Further, the standard deviation (SD) of each index across the time series from each voxel was calculated and normalized across the voxels to measure the variability in these dynamic R-fMRI indices. Finally, the normalized SD maps of each dynamic R-fMRI index were smoothed (FWHM = 4 mm) to improve the signal-to-noise ratio.

The concordance of five dynamic R-fMRI indices (ALFF, ReHo, DC, VMHC, and GSC) were calculated. In specific, volume-wise and voxel-wise concordance was measured by Kendall’s W among these indices (Yan et al., 2017). The volume-wise concordance (spatial concordance) was calculated in two steps. First, compute the Kendall’s W of the five indices across voxels for each window. Second, the dynamic volume-wise concordance was computed as the averaged Kendall’s W across all time windows for each participant. The voxel-wise concordance (temporal concordance) was measured by the Kendall’s W of the five indices across all time windows. By doing this, each voxel (voxel-wise) was assigned a concordance value. These voxel-wise concordance maps were then smoothed (FWHM = 4 mm) for subsequent analyses. Thus, for each participant, they got one volume-wise concordance value and N (N = number of voxels) voxel-wise concordance values.

Importantly, we validated the results with the window length of 48 and 64 TR.

We compared age, IQ, ADHD-relevant clinical variables between ADHD patients and healthy controls using two-sample t-tests. Meanwhile, gender distributions were compared using chi-square test. These statistical analyses were performed with SPSS 23.0.

The dynamic R-fMRI indices and voxel-wise concordance were compared using two-sample t-test, and the whole-brain gray mask was applied, and the results were corrected by multiple comparisons using Gaussian random field theory (GRF, voxel-wise p < 0.001, cluster-wise p < 0.05, two-tailed). Meanwhile, the volume-wise concordance was compared using analysis of covariance (ANCOVA). Further, Pearson or Spearman partial correlation was used to quantify the associations between dynamic indices and ADHD-relevant clinical test scores. Note that the age, gender, IQ, and head motions are controlled as covariates.

The characteristics of samples were summarized in Table 1. In particular, the ADHD group and the HCs group showed no significant differences in age, gender, head motion, or IQ. As measured by CPRS, ADHD patients exhibited significantly higher hyperactive and impulsive, and had more serious behavior and learning problems. In Wisconsin Card Sorting Test, the number of perseverative errors in ADHD group was higher than that in HC group (P = 0.014).

Compared with HCs, the ADHD group exhibited increased dALFF variability in right middle occipital gyrus and decreased dALFF variability in left middle frontal gyrus (GRF, voxel-wise p < 0.001, cluster-wise p < 0.05, two-tailed; see Table 2 and Figure 1). No significant between-group difference was found for the dynamics of other R-fMRI indices.

The ANCOVA results indicated no significant differences in volume-wise concordance between the ADHD group and HC group. However, ADHD patients showed decreased voxel-wise concordance in left middle frontal gyrus (GRF, voxel-wise p < 0.001, cluster-wise p < 0.05, two-tailed; see Table 3 and Figure 2).

No significant correlations were found between the CPRS scores and altered dALFF variability for the ADHD group. However, the abnormal voxel-wise concordance of the left middle frontal was negatively correlated with the non-perseverative errors in WCST of patients with ADHD (Pearson partial correlation, r = −0.367, p = 0.012; see Figure 3). Meanwhile, we found no correlation between left middle frontal voxel-wise concordance and non-persistent errors in HC group (Pearson partial correlation, r = 0.099, p = 0.646; see Figure 3).

The validation results with the window length of 48 and 64 TR showed good agreement with the results mentioned above (Supplementary Tables 2–5 and Supplementary Figures 1–6).

Specifically, for the variability, ADHD patients exhibited increased dALFF variability in right middle occipital gyrus and decreased one in left middle frontal gyrus, which may reflect the hyper-activation in visual cortex and hypo-activation in frontal cortex. This is consistent with a previous study in which decreased ReHo was found in frontal and increased ReHo and ALFF was found in visual cortex (An et al., 2013).

The development of the frontal lobe is essential for the acquisition, execution and control of a wide range of functions (Rosch and Mostofsky, 2019). And frontal lobe is a component of the brain region of the cognitive-attention network (cingulate-frontal lobe-parietal lobe), these areas are typically involved in cognitive, attentional, executive function, response inhibition, motor control, and working memory (Bush, 2011). Deficits in frontal lobe can lead to a series of developmental disorders like ADHD, and increase the psychopathological vulnerability. Many studies have revealed that the structural and functional abnormalities in the frontal lobe of patients with ADHD (Albajara Sáenz et al., 2019). For example, the volume of gray matter (Shaw et al., 2011) and the ALFF (Yang et al., 2011) in the middle frontal gyrus was decreased in ADHD patients. One previous study also exhibited decreased functional connectivity between right middle frontal gyrus and whole brain, and the decreased was associated with lower IQ (Shaw and Sudre, 2021). Consistent with the previous studies, we also observed decreased dALFF variability in middle frontal gyrus. The decreased dALFF variability indicated abnormal temporal fluctuations of brain activity in this region. Such an abnormal pattern may underlie the deficit in cognitive control of patients with ADHD, and the failure in cognitive control plays a key role in the main symptoms of ADHD.

Structural and functional alterations in occipital were also reported in previous studies. For example, reduced gray matter volume in the right middle occipital gyrus (Wang L. J. et al., 2020) and hyper-activation of occipital lobe function was found in ADHD (Tian et al., 2008; Icer et al., 2019), indicating that atypical visual processing may play a key role. Previous study also reported that the decreased connectivity between the frontal cortex and visual cortex was associated with increased severity in ADHD symptoms (Wang M. et al., 2020). Meanwhile, the interaction between the dorsal attention network and visual cortex is important for maintaining attention (Shulman et al., 2009). These findings suggest that brain activities in frontal and occipital regions and their interactions may play a key role in ADHD, which awaits further investigation.

Alterations in concordance among the dynamic R-fMRI indices were also identified in patients with ADHD. A recent study found that R-fMRI regional indices was high concordant in brain cortical and subcortical areas across the time window (Yan et al., 2017), and the voxel-wise concordance reflects homogeneity between these indices. Several studies on schizophrenia or MDD have found abnormal voxel-wise concordance in multiple brain regions (Zhu et al., 2018, 2019). These findings suggest that the concordance of dynamic indices may be an important biomarker for neuropsychiatric disorders. In our study, ADHD patients exhibited decreased voxel-wise concordance in the left middle frontal, which was consistent with the decreased dALFF variability in middle frontal gyrus, indicating maladaptive in frontal lobe.

In addition, we found that the voxel-wise concordance change in left middle frontal gyrus was negatively correlated with non-perseverative errors in WCST of ADHD group. Whereas, we did not find a correlation between them in the HCs. WCST is widely used to assess executive function, including cognitive flexibility, working memory, continuous attention and response inhibition (Godinez et al., 2012). Non-perseverative error measures the difficulties in working memory, sustained attention, and response inhibition (Barceló, 1999; Barceló and Knight, 2002; Godinez et al., 2012). Furthermore, non-perseverative errors have been found to be associated with frontal lobe function in patients with ADHD (Qian et al., 2018). However, no significant difference in non-perseverative errors was found between ADHD patients and healthy controls. This may due to the relatively small sample size, which might induce more variance and undermine the difference between these two groups. Nevertheless, this implies that the concordance in left middle frontal gyrus can be a sensitive biomarker that can capture the changes in non-preservative errors in ADHD patients.

Our findings highlighted the role of frontal lobe in the neuropathological mechanism of ADHD, and demonstrated that concordance of R-fMRI indices can serve as a biomarker to indicate the deficit in frontal lobe. To date, the diagnosis of ADHD is still based on symptom scores from clinical interviews, which can be subjective and can be influenced by many factors, such as the experience, stereotype, and cultural differences. Neuropsychiatric biomarkers can help to build a standard and objective diagnosis system, which can benefit the patients. R-fMRI is among the neuropsychiatric tools that have the potential to find the biomarkers (Abi-Dargham and Horga, 2016). Such biomarkers, combined with the rapidly developed machine learning algorithms, researchers are able to precisely classify the ADHD from healthy control (Du et al., 2018). R-fMRI indices such as ReHo (Wang et al., 2013) and functional connectivity (Fair et al., 2012) have been frequently used by these researchers. Functional concordance incorporates features from multiple metrics, which are fully explored in the current study, can also be used to train the machine learning models, and thereby can promote the accuracy of classification and prediction of ADHD.

Further, the behavioral dysfunction may be too subtle to be measured, just like the non-perseverative errors that failed to be capture by the current study. However, as mentioned above, the R-fMRI indices as proposed by the current study are able to evaluated the maladaptive of brain functions.

This study has several limitations. First, no significant differences were found among some dynamic indices, which may be due to the relatively small sample size. A larger sample size is needed to validate the alterations in dynamic brain function in ADHD and examine the relationship between dynamic variables and psychological data. Second, the current study only focus on the BOLD signal from gray matter. However, previous studies have demonstrated the white matter signals also provide valuable information (Peer et al., 2017; Ding et al., 2018; Li et al., 2020a,b). Hence, further studies are need to explore the altered variability and concordance of white matter signals in patients with ADHD. Third, this study used a variety of indices to reflect different patterns of dynamic spontaneous brain activity, but the seed-based functional connectivity analysis (FC) and independent component analysis (ICA) were not included in the analysis, this is because ICA may be biased by component selection and FC may be biased by seed definition. In future studies, more supplementary R-fMRI indices should be considered, which may help to describe the spontaneous brain activity more thoroughly and understand its underlying physiological processes more deeply.