The framework of graph theory provides useful tools for investigating the neural substrates of neuropsychiatric disorders. Graph description measures may be useful as predictor variables in classification procedures. Here, we consider several centrality measures as predictor features in a classification algorithm to identify nodes of resting-state networks containing predictive information that can discriminate between typical developing children and patients with attention-deficit/hyperactivity disorder (ADHD). The prediction was based on a support vector machines classifier. The analyses were performed in a multisite and publicly available resting-state fMRI dataset of healthy children and ADHD patients: the ADHD-200 database. Network centrality measures contained little predictive information for the discrimination between ADHD patients and healthy subjects. However, the classification between inattentive and combined ADHD subtypes was more promising, achieving accuracies higher than 65% (balance between sensitivity and specificity) in some sites. Finally, brain regions were ranked according to the amount of discriminant information and the most relevant were mapped. As hypothesized, we found that brain regions in motor, frontoparietal, and default mode networks contained the most predictive information. We concluded that the functional connectivity estimations are strongly dependent on the sample characteristics. Thus different acquisition protocols and clinical heterogeneity decrease the predictive values of the graph descriptors.
Attention-deficit/hyperactive disorder (ADHD) is a neurodevelopmental disorder with a prevalence of around 5.3% in children and adolescents [
Graph theoretical analysis is an emerging component in the field of connectomics and brain network analysis based on neuroimaging data [
Pattern recognition methods based on machine learning techniques have shown to be a promising approach to the analysis of neuroimaging data [
Graph theory descriptors can be used as predictor variables (i.e., features) in a machine-learning framework. Merging graph theoretical approaches and machine learning techniques might provide a better-adjusted way to scrutinize the impairment of RSNs in ADHD as well as mapping predictions to a single individual case. In this study, we investigated the use of network centrality measures as predictive features to discriminate between typical developing children and ADHD patients with both inattentive and combined presentations. In addition, we investigated possible differences between inattentive and combined ADHD groups. The ADHD-200 dataset [
The publicly available resting-state fMRI data from the ADHD-200 Consortium were used in the present study. The images were acquired at five different sites: Peking University, Kennedy Krieger Institute, NeuroIMAGE sample, New York University Child Study Center, and Oregon Health & Science University (OHSU). The subject sample consisted of 609 subjects, 340 controls (mean age [standard deviation] − 11.59
All research protocols from institutes contributing to the ADHD-200 Consortium received local approval by their respective IRB. All the data distributed via the International Neuroimaging Data-sharing Initiative (INDI) are fully anonymized in accordance with HIPAA Privacy Rules. Further details concerning the sample and scanning parameters can be obtained by request to the ADHD-200 Consortium.
Step-wise data preprocessing was previously conducted by the NeuroBureau community using the Athena pipeline and consisted in the systematic and homogeneous processing of all resting-state fMRI data. The following steps were carried out: exclusion of the first four EPI volumes; slice time correction; deobliquity of the dataset; head motion correction using the first volume as a reference; exclusion of voxels at non-brain regions by masking the volumes; averaging the EPI volumes to obtain a mean functional image; coregistration of this mean functional image to the subjects’ correspondent anatomical image; spatial transformation of functional data into template space; extraction of BOLD time series from white matter and cerebrospinal fluid using masks obtained from segmenting the structural data; removing trend and motion effects through linear multiple regression; temporal band-pass filtering; spatial smoothing using a Gaussian filter.All preprocessed images are available at the website
A representative set of 400 brain-wide regions of interest (ROIs) was chosen for defining the network nodes used for connectivity analysis and the construction of the graphs. The ROIs were determined by using the method developed by Craddock et al. [
The mathematical definitions of these measures are described in Table
Measure | Definition |
---|---|
Degree |
|
Closeness |
|
Betweenness |
|
Eigenvector |
|
Burt’s constraint |
|
Degree is a straight and intuitive way to quantify nodes centrality, and it is defined as the number of edges connected to a particular node. The closeness centrality is the average distance between a given node and all other nodes of the network. Betweenness quantifies the influence of a node and is defined as the number of shortest paths passing through it. The basic rationale underlying eigenvector centrality is that connections with more central nodes increase the nodes influence in the network. Hence, different weights are attributed to a vertex depending on the centrality of the connected nodes. Finally, Burt’s constraint value is inversely proportional to the number of connections of a node and increases with the number of strong mutual connections [
The centrality measures of each graph’s nodes were used as features (i.e., predictor variables) in an independent classification analysis. Classification was performed using a linear support vector machine (SVM) algorithm [
Finally, in order to identify the most discriminative regions, we built brain maps highlighting the 5% brain regions with greater predictive values. We used the approach proposed by Mourão-Miranda et al. [
Table
Typical developing versus ADHD classification: sensitivity, specificity, and score for each centrality descriptor.
Typical Developing versus ADHD | ||||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Descriptor | Cut-off | Peking | Kennedy Krieger | NeuroIMAGE | New York | OHSU | All sites | |||||||||||||
Spec | Sens | Score | Spec | Sens | Score | Spec | Sens | Score | Spec | Sens | Score | Spec | Sens | Score | Spec | Sens | Score | |||
Unweighted graph | Closeness | 0.1 | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% |
0.15 | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | ||
0.25 | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | ||
Between | 0.1 | 61% | 37% | 49% | 80% | 24% | 52% | 52% | 47% | 50% | 59% | 56% | 58% | 45% | 37% | 41% | 62% | 45% | 54% | |
0.15 | 59% | 40% | 50% | 72% | 19% | 46% | 61% | 47% | 54% | 59% | 55% | 57% | 52% | 40% | 46% | 63% | 44% | 54% | ||
0.25 | 63% | 38% | 51% | 77% | 19% | 48% |
|
|
|
57% | 60% | 59% | 48% | 31% | 40% | 61% | 44% | 52% | ||
Degree | 0.1 | 66% | 44% | 55% | 82% | 19% | 51% | 43% | 42% | 43% | 59% | 59% | 59% | 50% | 43% | 46% | 63% | 43% | 53% | |
0.15 | 65% | 50% | 57% | 82% | 24% | 53% | 57% | 47% | 52% |
|
|
|
50% | 37% | 44% | 63% | 40% | 51% | ||
0.25 | 61% | 49% | 55% | 77% | 19% | 48% |
|
|
|
|
|
|
48% | 31% | 40% | 65% | 44% | 54% | ||
|
||||||||||||||||||||
Weighted graph | Closeness | — | 62% | 35% | 48% | 87% | 5% | 46% | 57% | 42% | 49% | 46% | 51% | 48% | 62% | 54% | 58% | 65% | 31% | 48% |
Between | — | 62% | 44% | 53% | 85% | 0% | 43% | 48% | 37% | 42% | 52% | 58% | 55% |
|
|
|
62% | 43% | 53% | |
Degree | — | 59% | 44% | 51% | 80% | 29% | 54% |
|
|
|
|
|
|
55% | 37% | 46% | 62% | 50% | 56% | |
EVC | — | 59% | 38% | 49% | 85% | 19% | 52% | 61% | 37% | 49% |
|
|
|
57% | 31% | 44% | 74% | 41% | 58% | |
Burt | — | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% |
Table
ADHD types classification: sensitivity, specificity, and score for each centrality descriptor. Note: the accuracy measures could not be obtained at the NeuroIMAGE site due to the small number of ADHD-combined subjects.
Inattentive ADHD versus Combined ADHD | ||||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Descriptor | Cut-off | Peking | Kennedy Krieger | NeuroIMAGE | New York | OHSU | All sites | |||||||||||||
Spec | Sens | Score | Spec | Sens | Score | Spec | Sens | Score | Spec | Sens | Score | Spec | Sens | Score | Spec | Sens | Score | |||
Unweighted graph | Closeness | 0.1 | 100% | 0% | 50% | 100% | 0% | 50% | — | — | — | 100% | 0% | 50% | 100% | 0% | 50% | 0% | 100% | 50% |
0.15 | 100% | 0% | 50% | 100% | 0% | 50% | — | — | — | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | ||
0.25 | 100% | 0% | 50% | 100% | 0% | 50% | — | — | — | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% | ||
Between | 0.1 | 34% | 69% | 52% |
|
|
|
— | — | — | 63% | 37% | 50% |
|
|
|
62% | 45% | 54% | |
0.15 | 31% | 71% | 51% |
|
|
|
— | — | — | 58% | 33% | 45% |
|
|
|
58% | 47% | 53% | ||
0.25 | 41% | 69% | 55% | 88% | 0% | 44% | — | — | — | 63% | 37% | 50% |
|
|
|
64% | 46% | 55% | ||
Degree | 0.1 | 45% | 61% | 53% |
|
|
|
— | — | — | 63% | 35% | 49% |
|
|
|
60% | 49% | 54% | |
0.15 | 28% | 71% | 50% | 94% | 20% | 57% | — | — | — | 66% | 33% | 49% |
|
|
|
60% | 55% | 58% | ||
0.25 | 34% | 71% | 53% | 88% | 20% | 54% | — | — | — | 64% | 42% | 53% |
|
|
|
|
|
|
||
|
||||||||||||||||||||
Weighted graph | Closeness | — | 34% | 71% | 53% |
|
|
|
— | — | — | 63% | 35% | 49% | 74% | 33% | 54% | 60% | 37% | 49% |
Between | — | 28% | 78% | 53% | 100% | 0% | 50% | — | — | — | 66% | 37% | 51% | 74% | 17% | 45% | 58% | 46% | 52% | |
Degree | — | 31% | 73% | 52% | 88% | 20% | 54% | — | — | — | 64% | 37% | 51% |
|
|
|
59% | 51% | 55% | |
EVC | — | 34% | 71% | 53% | 69% | 0% | 34% | — | — | — | 66% | 37% | 51% |
|
|
|
65% | 46% | 56% | |
Burt | — | 100% | 0% | 50% | 100% | 0% | 50% | — | — | — | 100% | 0% | 50% | 100% | 0% | 50% | 100% | 0% | 50% |
Interestingly, the mean score (across sites) and the score from whole-sample classification were very similar, except when using betweenness and degree in unweighted graphs (Figure
Classification scores ([specificity + sensitivity]/2) for each centrality measure.
Regarding the identification of the brain regions with greater contribution to prediction, we chose only the classifications with accuracy above 70%. Figure
Discriminant regions for betweenness centrality (weighted graph) in typical developing versus ADHD classification at OHSU.
Figure
Discriminant regions for unweighted betweenness, weighted degree, and weighted eigenvector centrality in the classification between ADHD types at OHSU.
At present, resting-state fMRI is a well-established tool for the assessment of spontaneous brain activity. Graph theoretical measures provide a suitable framework for the investigation of the structures of complex neural networks. In addition, the application of machine-learning algorithms has been of great impact on developing more advanced neuroimaging studies of psychiatric disorders [
When the whole sample was used, none of the centrality measures had a relevant predictive power beyond chance. However, significant prediction values were observed at the OHSU site. Thus both within- and between-site variability have a negative impact on the extraction of predictive information and consequently on classification. In the OHSU sample, betweenness centrality measures contained predictive information for the classification of ADHD and control subjects with a score of 73%. After an extensive analysis of sample characteristics and acquisition parameters, we hypothesize that the classification score at OHSU was higher than the other scores for two main reasons: (i) the sample was approximately balanced between typically developing controls (42 subjects) and ADHD patients (35 subjects), while the group sizes were very different at the other sites; (ii) OHSU EPI acquisition has the largest voxel size (3.8 mm) and the 3T system was equipped with a 12 channels head coil (as opposed to 8) which increases the signal-to-noise ratio.
When the 5% nodes with greater predictive values were mapped, a sparse pattern of brain regions was observed. In fact, widespread brain alterations in ADHD are supported by findings of impaired interregional connectivity between the nodes of large-scale functional networks (reviewed in [
A promising finding was observed for the degree centrality in the whole sample analysis on the classification of the disorder types. In the within-site analyses, relatively high scores were observed for degree, betweenness, and eigenvector centralities. However, as the sample size is smaller in these cases, variability is increased. Moreover, the mean scores of within-site analyses were almost identical to the ones from the whole sample analysis. Brain regions mapped for betweenness measures included nodes of the right frontoparietal network. This network has been implicated in attentional and executive processes and is thought to be impaired in ADHD. Cubillo et al. [
The measure of degree centrality, when applied to the separation between ADHD types, produced the highest classification scores in areas of the sensory-motor network and of the DMN, mainly in parietal cortex and the precuneus. These findings are in agreement with our hypothesis, based on consistent results in the literature [
In conclusion, a novel approach of applying graph theoretical measures was shown to be useful for testing our hypothesis regarding resting-state network impairment in ADHD disorder. In particular, distinct patterns of network dysfunction were evident for both inattentive and combined ADHD subtypes. The classification scores for discriminating between ADHD and healthy subjects were close to chance. Clearly, within-site analysis improves prediction levels when compared to whole sample analysis, suggesting that heterogeneity across the sites may strongly limit the application of the method as a potential clinical support. The functional connectivity estimation is strongly dependent on the samples’ characteristics. Thus, in order to advance the pathophysiological knowledge of ADHD, we emphasize the importance of further multicentric studies with more homogeneous acquisitions.
Dr. Luis Augusto Rohde has been a member of the speakers’ bureau/advisory board and/or acted as a consultant for Eli-Lilly, Janssen-Cilag, Novartis, and Shire in the last three years. He receives authorship royalties from Oxford Press and ArtMed. He has also received travel awards from Shire for his participation of the 2014 APA meeting. The ADHD and Juvenile Bipolar Disorder Outpatient Programs chaired by him received unrestricted educational and research support from the following pharmaceutical companies in the last three years: Eli-Lilly, Janssen-Cilag, Novartis, and Shire.
The authors declare that there is no conflict of interests regarding the publication of this paper.
The authors are grateful to the Sao Paulo Research Foundation—FAPESP (Grants 2012/13390-9, 2013/10498-6, and 2013/00506-1) and to CAPES Brazil. The authors would also like to thank the institutes funding the ADHD-200 Consortium: The Commonwealth Sciences Foundation, Ministry of Health, China (200802073); The National Foundation, Ministry of Science and Technology, China (2007BAI17B03); The National Natural Sciences Foundation, China (30970802); The Funds for International Cooperation of the National Natural Science Foundation of China (81020108022); The National Natural Science Foundation of China (8100059); the Open Research Fund of the State Key Laboratory of Cognitive Neuroscience and Learning; The Autism Speaks Foundation and the (National Institutes of Health) NIH (R01 NS048527, R01MH078160, and R01MH085328); the Johns Hopkins General Clinical Research Center (M01 RR00052); the National Center for Research Resources (P41 RR15241); the Intellectual and Developmental Disabilities Research Center (HD-24061); the (Netherlands Organisation for Scientific Research) NWO-Groot, the (National Institutes of Mental Health) NIMH (R01MH083246); Autism Speaks; The Stavros Niarchos Foundation; The Leon Levy Foundation; An endowment provided by Phyllis Green and Randolph Cōwen; K99/R00 MH091238 (Fair); R01 MH086654 (Nigg); the Oregon Clinical and Translational Research Institute (Fair); the Medical Research Foundation (Fair); UNCF/Merck (Fair); the Ford Foundation (Fair); Cognitive & Brain Systems Maturation (5R01 MH067924, Luna); Reward Processing in Adolescence (1R01 MH080243, Luna); Functional Anatomy of Adolescent ADHD: Defining markers of recovery (K01MH82123, Velanova); The Brooks Family Fund; R01 HD057076 (Schlaggar); R01 NS046424; NIH NINDS NRSA (Church); NIH NIMH R21 (Schlaggar) and TSA (Schlaggar); TSA (Church).