Attention-deficit/hyperactivity disorder (ADHD) is a common neurodevelopmental disorder in childhood and adolescence. The reported incidence of ADHD differs between both studies and countries. The estimated worldwide prevalence of ADHD is up to 7.1% in children and adolescents (Thomas et al., 2015). The Diagnostic and Statistical Manual of Mental Disorders (DSM), Fifth Edition (DSM-5) characterizes ADHD by its three cardinal symptoms: inattention, hyperactivity, and impulsivity. These symptoms affect healthy psychological functioning and are associated with mental health problems as well as impaired academic and social functioning (Missiuna et al., 2014; Asherson et al., 2016). The disorder is not limited to children and adolescents; for 40–60% of children with ADHD, the disorder persists into adulthood and results in lifelong impairment (Faraone et al., 2000; Volkow and Swanson, 2013).

Currently, the standard diagnostic procedure for ADHD is based on clinical interviews and symptom questionnaires. Although these questionnaires serve as objective and quantitative measures in conjunction with the subjective criteria listed in DSM fourth edition (DSM-4) and DSM-5, diagnosis is still time-consuming. Clinically, diagnosis requires more retrospective reports from parents and teachers than it does patient self-reports (Sibley et al., 2012). Integrating clinical questionnaires with behavioral and cognition test results is complex and challenging and requires an experienced expert evaluator (Morrill, 2009). Therefore, this process must be carefully performed by trained professionals following repeated observations and by using reports obtained from parents, teachers, or other caregivers. Moreover, the evaluator must confirm the absence of any other underlying disorders that could be mislabeled as ADHD (Rader et al., 2009). According to a survey by Morrill (2009), more than 85% of clinical staff (primary care physicians) expect auxiliary diagnostic tools for ADHD to be available. The selective use of neuropsychological tests is one such auxiliary diagnostic approach. ADHD is typically associated with substantial cognitive deficits in various neuropsychological studies, such as reaction time variability, intelligence and achievement, vigilance, working memory, and response inhibition (Pievsky and McGrath, 2018). Thus, ADHD may interfere with several brain functional structures, and this interference is reflected in behavior.

Although physiological characteristics are not included in the current diagnostic criteria, the existing body of neuroscientific evidence from structural and functional neuroimaging studies indicates a correspondence between the abnormal behaviors used for ADHD diagnosis and brain features; for example, some brain areas may have a smaller volume and be prone toward hyperactivity or hypoactivity (Konrad and Eickhoff, 2010) in patients with ADHD compared with healthy patients, such as the frontal lobe (Durston et al., 2004; Paloyelis et al., 2007), cerebellum (Semrud-Clikeman et al., 2000; Hart et al., 2012), motor cortex (Mostofsky et al., 2006; McLeod et al., 2014), and deep brain area (Schneider et al., 2006; Konrad and Eickhoff, 2010; Castellanos and Proal, 2011). To accelerate the rate at which neuroscience informs clinical practice, The ADHD-200 Consortium (2012) constructed a large-scale ADHD data set (sample size over 1,000) and a global competition to integrate information technology with brain image analysis. A survey of follow-up studies (The ADHD-200 Consortium, 2012) demonstrated that neuroimaging data, especially data from resting state functional magnetic resonance imaging (rs-fMRI), are highly promising for the diagnosis, assessment, and treatment evaluation of patients with ADHD. These objective biological tools can be used by an individual or auxiliary clinician to inform an ADHD diagnosis during their decision-making regarding treatment.

Resting state is defined as spontaneous brain activity in the absence of any specific cognitive task during the acquisition of a brain signal or image; this definition reduces the difficulty of clinical diagnosis of ADHD. Rs-fMRI has been widely applied in the diagnosis of neurological or psychiatric diseases such as ADHD, autism spectrum disorder, Alzheimer’s disease, depression, etc (Uddin et al., 2010; Lee et al., 2013; Fox et al., 2014). A variety of neurological evidence of brain network abnormalities in ADHD revealed through rs-fMRI have been reported. For example, abnormalities in default mode networks (DMN), including increases (Tian et al., 2008) or decreases in network homogeneity (Uddin et al., 2008) or decreases in connectivity (Castellanos et al., 2008; Uddin et al., 2008) in patients with ADHD were observed by using seed-based correlation analysis (SCA). Regional homogeneity analyses of rs-fMRI results have also revealed that patients with ADHD have similarly high activation in the sensory areas of the brain (Cao et al., 2006; Krain and Castellanos, 2006).

Seed-based correlation analysis (SCA) is the oldest and most popular method for investigating abnormal brain activity in rs-fMRI research on ADHD and other psychiatric disorders (Biswal et al., 1995). Due to the lack of a task design, referring the literature to selected regions of interest (ROIs) as seeds to estimate functional connectivity (FC) is common in rs-fMRI research. Because this selection usually stands with strong evidence, the results of abnormal brain investigations are more accepted and understood in clinical research. However, some seed details, such as shape, size, and boundary may induce bias in the results, and the results may be difficult to replicate or apply. By contrast, the use of general whole-brain anatomical or functional maps as seeds can increase the reproducibility of SCA results. Craddock et al. (2013) used the automated anatomical labeling (AAL) (Tzourio-Mazoyer et al., 2002), CC200, and CC400 (Craddock et al., 2012) atlases to divide the whole brain into several seeds and construct the resting networks. Regardless of the seed selection method, the auxiliary diagnostic performance reported in the ADHD-200 consortium competition (The ADHD-200 Consortium, 2012) was an accuracy of only 60 ± 5%. Thus, the model-driven seed selection method may be unsuitable for diagnosing multisymptomatic neurodevelopmental disorders. If a more flexible, data-driven seed selection method that is responsive to abnormalities in ADHD could be identified, providing auxiliary diagnostic information with SCA should become more effective.

The aim of this study is to evaluate the feasibility of diagnosing ADHD by analyzing rs-fMRI data. To achieve this goal, we propose a smart seed selection method for diagnosing ADHD in children with rs-fMRI data. We refer regional homogeneity (ReHo) analysis results. This data-driven method can determine specific seeds with within-group homogeneity and between-group differences. Furthermore, principal component analysis was used to reduce the SCA feature dimensions, and linear discriminant analysis (LDA) was used to distinguish healthy children and those with ADHD. The classification accuracies of different feature extraction approaches were also compared.

At present, no ideal biological test for ADHD diagnosis exists. Currently, ADHD is primarily evaluated by observing the symptoms listed in the diagnostic criteria. The availability of more physiological information (e.g., brain images or signals) is key to improving the diagnosis or treatment of ADHD by medical teams. Such physiological information can provide objective information about cognitive deficits and increase the confidence of the medical team in diagnosis. Rs-fMRI is an excellent acquisition method for physiological brain information for people with psychiatric disorders. However, analyzing rs-fMRI data and confirming observations is challenging due to the lack of a preexisting model; in this study, we propose a procedure for automatically determining effective seeds for investigating brain abnormalities in ADHD (ReHo-FC). The procedure reduces the difficulty of seed selection (location, shape, and size) in estimations of brain interconnections, improving the search for an effective seed; Our proposed feature combination of ReHo-FC and LFC can achieve a success rate of 84.87% (LDA) for identifying patients with ADHD by 10 times random sampling of the data from 149 individuals in the NYU data set. The identified ReHo seeds have been noted in studies on ADHD. The present results are sufficient for a medical center to construct an auxiliary diagnostic information system in accordance with its needs and resources (e.g., individuals, instruments, and technicians).

As noted in the previous sections, the challenge of SCA is the difficulty of selecting or defining an effective ROI as a seed. The use of anatomical and functional atlases to select seeds and then estimating brain networks is a straightforward method of analyzing abnormal FC and resting state networks (RSNs). In this study, we compared the overlap between the identified ReHo seeds and the atlas maps. The overlapping area was only approximately 10–40% of the atlas map size. Asynchronous voxels may affect the representativeness of the extracted signal. Yao et al. (2015) demonstrated bias in RSNs from different atlas and suggested that the quantification of brain networks might be substantially affected by the selection of atlas. In an example of a use of ADHD-200 with a functional atlas in practice, Chang et al. (2012) used the CC400 atlas to estimate the whole-brain connectivity network to detect ADHD. The accuracy rates were 67% (46.9% selectivity and 82.27% specificity) and 62% (41.8% selectivity and 73.13% specificity); Dey et al. (2014) reached 64.4% (selectivity 30.66% and 84.1% specificity) by using the CC200 atlas; The proposed method, in which four precise ReHo seeds were used to estimate FC, an accuracy of 81% and balanced selectivity and specificity were achieved. Moreover, the accuracy of each seed approached or exceeded 70%—higher than whole-brain atlas FC. This result supports the introduction of a data-driven method (e.g., ReHo) to improve selection efficiency in FC analysis.

Moreover, the difficulty of using multisite data is an obstacle to identifying effective auxiliary FC patterns for diagnosis. The scanning device, parameters, and protocol may differ between sites participating in the ADHD-200. For example, individuals in the NYU data set were required to close their eyes, but this is optional for individuals in the Peking data set. Moreover, individuals are insufficiently categorized into subgroups or uncategorized in some data sets. The aggregation of data sets recorded from different institutions generates a clustered structure within the data that may cause a batch effect (Olivetti et al., 2012). The presence of batch effects conflicts with one of the basic assumptions of statistical learning theory: that data are independent and identically distributed (iid). The use of non-iid data may result in suboptimal training. During the ADHD-200 competition, most teams (including our team) used non-iid information to train a general classifier model. The final scores of most teams were approximately 60 ± 5% and biased to detecting healthy individuals. Even the winners (Eloyan et al., 2012) had a substantial bias for healthy individuals (accuracy 89%, sensitivity 21%, and specificity 94%). Beyond the competition, the consistency of abnormal functional connectivity in the ADHD-200 dataset across different data centers was investigated by Zhou et al. (2019). The results show that the abnormality of seed-based connectivity in children with ADHD was inconsistent among all datasets, including those from the same research site (but not the same equipment). This inconsistency was observed even in areas frequently reported to be associated with the default mode network (DMN). Therefore, Shao et al. (2019), Riaz et al. (2020) and Ke et al. (2022) used this independent site training method. Although the proposed classifiers and the selected sites are different, they are better than the results in the above competition in discriminating ADHD patients. Itani et al. (2018) not only developed separate classification models for different sites but also proposed a method for automatically selecting an appropriate site-specific classifier based on the similarity of the testing data for prediction. The results are also better than most research results in the above competition. In this study, we focused on the NYU data set and applied the proposed feature (ReHo-FC) with LDA model. The result could reach 83.89% in total accuracy and was balanced in terms of sensitivity (82.63%) and specificity (83.24%). Moreover, the same procedure was applied to the Peking-1 data set (other data sets have the problems of overly small or excessively unbalanced samples), and our method reached 94.71% total accuracy, 76.67% sensitivity, and 98.57% specificity. Both results were superior to above results in ADHD-200. Therefore, using multiple data sets to establish multiple independent classification models may be more suitable than establishing a general model. Alternatively, as suggested by Olivetti et al. (2012), multiple data sets could be connected by using a statistical model such as Combat (Johnson et al., 2007) and Seurat Integration (Stuart et al., 2019) to correct the batch effect.

Another approach, deep learning, is usually presented as an end-to-end learning process. Deep learning can alleviate feature engineering requirements and reduce domain knowledge requirements. Its multilayer, nonlinear learning strategy provide more flexibility in features extraction and classification. Hence, deep learning is a promising approach for informaticists in abnormal brain investigation. In the practical application part of the ADHD-200, most studies have used general atlases to generate whole-brain FC features for the deep learning model. The approximate classification accuracies of the general deep learning architecture were approximately 60–75% (Riaz et al., 2017, 2018; Zou et al., 2017; Chen et al., 2020). Mao et al. (2019) skipped feature extraction and directly imported overall rs-fMRI images to construct customized long short-term memory (LSTM) and four-dimensional convolutional neural network (CNN) classification models. However, the accuracy rates (68.8 and 71.3%) were not substantially better than the aforementioned FC-based models. Thus, we propose using an FC feature calculation method with effective guidance (ReHo seed) and deemphasizing the customization of classifiers. As presented in Table 3, the proposed method is superior for distinguishing patients with ADHD. An accuracy of 81.8% can be achieved with simple linear classification. With the application of a two-dimensional CNN architecture, the accuracy rate can be increased to 82.5%. We believe that this large performance improvement compared with previous models is primarily due to reducing the size of the input feature space by identifying meaningful features. Although deep learning is a powerful tool in processing high-dimensional data, its use with MRI images is typically hindered by the relatively small size of available samples (Mao et al., 2019). The sample size required to successfully train a deep learning model depends on multiple factors, such as the type of data, network size and architecture, type of stochasticity in the data, dimensionality of the feature space, regularization schemes, and the actual target function that the deep learning neural network is intended to learn, to name only a few (Koppe et al., 2021). A systematic review even notes that, for sample sizes less than 200, deep learning models may have heterogeneous performance results (Quaak et al., 2021). The advantages of deep learning in terms of dimensionality reduction instead resulted in overfitting in a small sample data set. Because deep learning relies on multiple hidden layers, clinical staff would have difficulty adjusting the architecture to avoid overfitting and to fit their data set. Moreover, the deep learning outcome typically lacks interpretable auxiliary information. The decision process of deep learning models is also difficult to understand and visualize for human beings. Therefore, we do not recommend directly importing deep learning architectures for auxiliary diagnosis and treatment of ADHD at this stage. We are more inclined to use meaningful and well-known features in the clinic, and the comparatively simple and easily adjustable machine learning models could provide auxiliary diagnostic information.

Although the reasons for abnormal RSNs in ADHD are unclear, recent studies in neuroscience have identified a number of abnormal RSNs, including the DMN, ventral attention, dorsal attention, and the frontal and motor networks in patients with ADHD (Castellanos and Proal, 2011). We used a number of ReHo-FC features that belong to a DMN subnetwork (e.g., from the mPFC to the precuneus or anterior temporal lobe network or from the motor area to the mPFC). We also observed some potential connections, such as in the frontal-striatal and frontal-striatal-cerebellar circuits of the putamen, through the ReHo-FC analysis. These circuits have been frequently reported as being associated with ADHD in recent years (Krain and Castellanos, 2006; Vaidya and Stollstorff, 2008; Durston et al., 2011; Cubillo et al., 2012; Kucyi et al., 2015). These circuits enables changes in behavior according to environmental demands and can be considered to affect cognitive or behavioral control (Durston et al., 2011). Circuit hyperactivation causes children with ADHD to be unable to avoid distractions while performing a task, efficiently manipulate information in working memory, and switch between tasks or priorities (Castellanos et al., 2006). However, the result that almost 70% accuracy in ADHD classification was achieved when the circuit was observed in the task-free resting state is intriguing. We surmise that structural changes in myelin and gray matter revealed by diffusion tensor imaging caused this abnormal functional connectivity (Liston et al., 2011).

This study may have a limitation in the generalization for classification of ADHD because we excluded subjects of a small sample size (e.g., The Supplementary Appendix Table 3 presents small population of non-ADHD psychiatric symptoms in the control group and neurologic or psychiatric comorbidity in the ADHD group). We need more data with large sample size to help classification of various ADHD subtypes. Structural and functional MRI researches have revealed that various comorbidities may be involved in the pathology of ADHD (Schneider et al., 2006; Singh et al., 2006; Wolosin et al., 2009). The contribution of comorbidity on classification of ADHD subtypes remains to further investigation.

In the future, clinical auxiliary diagnosis information could be adapted to various ADHD conditions such as subtypes, comorbidities, and even age (a critical improvement in the DSM-V standard) to enhance evaluator confidence in diagnosis. Second, although our proposed features can achieve good classification rates with the NYU (83%) and Peking (94%) data sets, the deleterious effects of using multisite data sets cannot be ignored. At present, each data set is used independently, which enables the extraction of numerous meaningful features and results in an improved the accuracy rate. We expect the system to undergo more substantial development at various hospitals and pediatric centers. Finally, by integrating the models of multiple independent centers with methods such as reinforcement learning, transfer learning, bagging, and bootstrapping, developers can increase the clinical confidence in the auxiliary medical information provided by the method.

Publicly available datasets were analyzed in this study. This data can be found here: ADHD-200.

Ethical review and approval was not required for the study on human participants in accordance with the local legislation and institutional requirements. Written informed consent from the participants’ legal guardian/next of kin was not required to participate in this study in accordance with the national legislation and the institutional requirements.

T-HH completed the writing of the manuscript and the summary of all the data. C-CK and F-ZS provided rs-fMRI analysis. S-FL provided the machine learning analysis. All authors contributed to the article and approved the submitted version.