Humans exhibit diverse characteristics of emotion, cognition, and behavior that describe individuals. Many different psychometric parameters from questionnaires or behavioral tasks have been developed to evaluate human characteristics of cognition/behavior (Diener et al., 1985; Raine, 1991; Carver and White, 1994; Yamauchi et al., 2009). Brain function underlies cognition/behaviors, and thus, it is possible that the characteristics of an individual can be evaluated by measuring brain function.

Functional MRI (fMRI) is the most widely used non-invasive method of measuring human brain function (Ogawa et al., 1992; Kim and Ugurbil, 1997). Measurements of human cognition/behavior by fMRI require psychometric parameters describing cognition/behavior that are embodied as tasks to induce neuronal processing in the brain (brain activation). Some psychometric parameters are easy to formalize, whereas others are not (Rupp and Zumbo, 2006). For the former type of parameters, a task can be established to evoke brain activation, and the corresponding fMRI signals of brain activation can be detected from the relevant brain areas.

Brain imaging by task-based fMRI (tb-fMRI) has provided information about the brain areas and brain networks that represent given tasks (Poldrack et al., 2013). However, most brain imaging studies have focused on specific psychometric parameters (Gauthier et al., 2000; Grill-Spector et al., 2004; Fernández-Alcántara et al., 2016; Krendl and Kensinger, 2016; Kogler et al., 2017) rather than considering a comprehensive set of psychometric parameters that can describe the diverse human characteristics of cognition/behavior. This is because it is very difficult to identify brain areas/networks responsible for the many diverse human characteristics due to limitations in task designs and time to perform tests.

In contrast to tb-fMRI, resting-state fMRI (rs-fMRI) signals come from intrinsic brain activities not designated to explicit tasks (Fransson, 2005; Fox and Raichle, 2007), and thus, there is no information to link a measured rs-fMRI signal with a specific brain function. However, functional networks produced from the correlation of rs-fMRI signals with certain brain areas, such as the default mode network, are known to represent various aspects of brain function (Greicius et al., 2004; Mason et al., 2007). Recent studies have further shown that task performance can be estimated by brain networks identified by rs-fMRI (Tavor et al., 2016; Craig et al., 2017; Song et al., 2017). In addition, some studies have proven that brain areas or functional connectivity in a rs-fMRI network are correlated with the scores of an explicit task used to identify functional areas by tb-fMRI (Finn et al., 2015; Meskaldji et al., 2016). These previous studies suggested that rs-fMRI signals can be used to identify brain networks that represent sensory and higher-order cognitive function or higher-order social function (Finn et al., 2015; Lei et al., 2017; Yang et al., 2017). That is, rs-fMRI signals may represent functional infrastructures for processing information in the brain that is highly involved in the brain function related to cognition/behavior. These characteristics of rs-fMRI may allow diverse information about the brain to be obtained from rs-fMRI signals acquired during a single MRI scan if appropriate supporting information is provided that can explain the characteristics of rs-fMRI signals or the correlation between brain areas based on the signals.

Using behavioral data about psychometric parameters as the supporting information, we attempted to test whether functional networks of the brain could be identified by rs-fMRI signals and psychological parameters (Figure 1). Recently some studies have shown that it is possible to decode brain states from fMRI responses using machine learning algorithms, such as a support vector machine (Guo et al., 2015; Altmann et al., 2016; Hu et al., 2017; Zafar et al., 2017). To verify that the functional networks identified represent the corresponding psychological parameters, and also to evaluate human characteristics from the rs-fMRI signals, we used a support vector machine (SVM).

We attempted to derive a specific SVM classifier related to each identified functional network. Second, we aimed to identify a set of functional networks of the brain to comprehensively evaluate human characteristics (Figure 1). For this purpose, we prepared a set of 130 psychometric parameters (Table 1; Supplementary References), including those related to intelligence quotient (IQ), emotional quotient (EQ), and social ability/skill that can explain the most conceivable aspects of human cognition/behavior by which an individual can be described.

The experiments in this study comprised a primary experiment and a supplementary experiment. The primary experiment was designed to identify functional networks and to design the SVM classifiers. The supplementary experiment was performed to test the SVM classifiers developed in the primary experiment.

The primary aims of this study were to prove that functional networks identified by rs-fMRI signals and psychometric parameters represent brain functions to which corresponding cognition/behavior are related and to determine whether the functional networks can be interpreted in a similar way to functional areas identified by tb-fMRI signals upon task stimulation. Another aim was to identify a set of functional networks for comprehensively evaluating human characteristics.

We found functional networks corresponding to 111 out of 130 psychometric parameters and derived a multiclass SVM classifier for each psychological parameter. These findings demonstrate that each rs-fMRI functional network can represent a corresponding cognition/behavior and that the set of functional networks reported here, which correspond to 111 psychometric parameters, can be used to comprehensively evaluate human characteristics.

Previous studies have attempted to identify functional networks from rs-fMRI signals. Some have identified functional networks by correlation between rs-fMRI signals and known ROIs (Wang et al., 2009; Tian et al., 2011), and others have derived functional networks by correlation between a known seed area or image voxel of the brain and other brain areas/voxels (Greicius et al., 2003; Fox et al., 2009) or by independent component analysis (ICA) in which a group of brain areas sharing a common component of an independent signal are identified to constitute a functional network (Damoiseaux et al., 2006; De Luca et al., 2006). However, the functional roles of the networks were interpreted based on various cognition/behavior related to the tasks in tb-fMRI (van den Heuvel and Hulshoff Pol, 2010), as a combination of functional areas previously identified by tb-fMRI and the tasks performed during tb-fMRI were used in the identification procedure of the functional networks. Therefore, it remains unclear whether functional networks can be identified by psychological indices (psychometric parameters consisting of self-reported questionnaire scores) and whether rs-fMRI signals can represent the corresponding cognition/behavior. By identifying brain networks related to psychometric parameters and the deriving multiclass SVM classifiers corresponding to those psychological parameters, our results demonstrate that rs-fMRI functional networks can represent cognition/behavior. In tb-fMRI, SVM classifiers have been used to prove functional specificity of a brain area by testing whether the brain area can discriminate the related stimulus exemplar from other stimuli (MacEvoy and Epstein, 2009). Similarly, each of our identified classifiers could significantly discriminate the scores of the corresponding psychological parameter, showing that each functional network represented the corresponding cognition/behavior. The results indicate that the identified functional networks can be used in a similar way to the functional areas identified by tb-fMRI for brain imaging. Recent studies on rs-fMRI support our results that the functional brain networks identified by rs-fMRI signals contain intrinsic information of the brain system and have shown that similar brain maps obtained by tb-fMRI can also be acquired by rs-fMRI (Finn et al., 2015; Tavor et al., 2016).

The identification of 111 functional networks and the derivation of SVM classifiers for each of the 111 functional networks suggest that the variance of rs-fMRI signals between subjects reflects differences in cognition/behavior, which also indicates that variations in neural systems develop differently in each individual. Therefore, it is inferred that brain plasticity with different genetic and environmental conditions (Bouchard, 2004) varies among participants and that such variation in plasticity could appear in the functional network representation.

Several studies of brain plasticity at the systems level have been performed. Long-term plasticity has been shown in primary sensory areas, such as in visual areas of blind subjects and auditory areas of deaf subjects (Karni and Sagi, 1991; Gaser and Schlaug, 2003), which could be considered as modifications to brain “hardware.” Mid-term plasticity was also shown in high-level areas like the hippocampus among people who have extensively trained for many years, such as in taxi drivers (Maguire et al., 2000). In addition, short-term plasticity was shown in high-level areas such as the hippocampus in those undergoing intense cognitive function training, such as in those studying for a difficult medical exam for a few months or during physical juggling training for a few weeks (Draganski et al., 2006). Such plastic changes have been observed by rs-fMRI and VBM. Considering the brain plasticity observed by rs-fMRI or VBM, we can infer that the 163 functional areas and 111 functional networks identified by the same modalities used in the previous studies also reflect brain plasticity.

We performed the leave-one-out cross validation in this study because we thought it gives less biased predictions. Recently a study has suggested conservative evaluation for reliability of cross-validation methods in applying machine learning algorithms for small sample sizes (Varoquaux, 2017). In this study, we completely separated data sets for the training and the test and could acquire similar results for the training and the test. In addition, the average accuracy 23.6% of the test data is still above a corrected chance level (about 17% for 150 samples), corrected by the number of samples, which was proposed by a previous study (Combrisson and Jerbi, 2015). These support that our classifiers are reliable.

In terms of the general use of classifiers, accuracy is an important factor. The accuracy of the SVM classifiers is low, although it is above the chance level. One reason is that the number of classes - 8 - could be too high. For 3-class SVM classifiers we could get about 53% accuracy; similarly, we can expect about 80% accuracy for a binary classifier. In this study, we used the same number of classes for all psychometric parameters when designing SVM classifiers, although there was an optimal number of clusters. We used the same algorithm and kernel for all psychometric parameters. However, other classification algorithms or kernels may be better for some psychometric parameters depending on their features and data structures. In future studies, it will be necessary to optimize the classification algorithm to each parameter and to optimize the number of classes to suit the evaluation of human characteristics, especially transient changes in characteristics by brain plasticity resulting from education, training, or diseases. However, even with the 8-class SVM classifier we derived, we can significantly evaluate human characteristics. For example, in a supplementary experiment (not published), we obtained rs-fMRI data from one participant 8 times across 2 weeks. We found that 66 out of 70 SVM classifiers gave the same score (level) more than 4 times, which means that repeated measurements enable our 8-class SVM classifiers to be applied to estimate a wide variety of human characteristics with high specificity (more than 87%; Table 5).

For training, the SVM classifier that we used here only has information regarding the edges of the functional networks with which the correlations of the psychometric scores were comparatively low, 0.51 ± 0.07 (mean ± SD), as shown in Figure S2, which might be a limitation of the performance of the SVM classifiers. But other factors related to the topology of a functional network as a graph, such as centrality and mean path, can be used to improve the accuracy. Additional data from other modalities, such as fractional anisotropy or mean diffusivity from diffusion-weighted imaging, may also be useful for further improving the accuracy. In the future, we plan to use these types of data to further improve the accuracy for generalizing our classifiers.

The tb-fMRI signal is known to be induced by stimulation through neuro-vascular coupling, and the site at which an fMRI signal is measured is known to contain neurons that are processing the information related to a given stimulus. Therefore, tb-fMRI is known to be a direct method to identify functional areas. However, resting-state fMRI is considered to be indirect because additional supporting information is needed to characterize the connectivity, such as psychological parameters or behavioral information, and it remains unclear whether neurons in brain areas of a rs-fMRI functional network are directly related to the processing of the information required for the expected cognitive or behavioral brain function. Therefore, even in the case in which the data acquired by the two modalities can yield the same results in evaluating cognition/behavior, the functional mechanism represented by brain areas or networks may be different. This may explain why ROIs for a psychometric parameter were not identified but a functional network was identified. This type of problem related to intrinsic functional characteristics should be further examined to elucidate the mechanism of brain function.

To understand the details of the functional role of the functional network of each psychological parameter, tb-fMRI may be needed. However, in this study, the aim was to prove that functional networks can be identified from rs-fMRI signals and psychometric parameters and to identify biomarkers of cognition/behavior to evaluate a wide variety of human characteristics that can be used to describe individuals. Therefore, although the details of identified functional networks should be investigated further, we believe that our aim of identifying functional networks/classifiers that characterize most of the intended psychometric parameters has been achieved.

Regarding the size of the subject population for significantly obtaining brain information from rs-fMRI and psychological parameters, the population size of 153 subjects for this study is considered to be appropriate because a previous study suggested that more than 100 subjects can provide reliable variation for estimating brain plasticity (Termenon et al., 2016).

To the best of our knowledge, this study derived the largest number of functional networks/classifiers (or identified functional brain networks) reflecting cognition/behavior among fMRI studies performed to date. Although there are several large databases around the world that are represented by the “human connectome project” (www.humanconnectome.org), none of the databases include as many psychometric parameters as included in this study. This is the first study of its kind in the field of brain imaging that reveals the possibility to describe an individual based on a comprehensive set of diverse human characteristics. However, other psychometric parameters may need to be added to obtain more classifiers so that an individual can be described as completely as possible. In the present study, 19 psychological parameters were not significantly associated with a functional network, for which we may need to devise new tasks and use tb-fMRI to elucidate the corresponding functional networks. In addition, it may be possible that abilities or emotions related to the psychometric parameters are not reflected in resting state fMRI signals, which may be an interesting topic to investigate in a future study.

Taken together, our results demonstrate that; (i) rs-fMRI signals include intrinsic information of brain function related to cognition/behavior, (ii) functional networks identified by psychometric parameters can represent corresponding cognition/behavior, and (iii) the set of functional networks/classifiers identified here can be used to comprehensively evaluate human characteristics.

We identified a set of 128 functional networks of cognition/behavior by rs-fMRI that span a variety of human characteristics and psychometric parameters, and we derived 123 multiple-class SVM classifiers corresponding to 111 psychometric parameters. This demonstrates that we can identify functional areas or networks of the brain not only by tb-fMRI but also by rs-fMRI. It also demonstrates that we can evaluate cognition/behavior and develop biomarkers for a wide variety of human characteristics using the 111 dimensions of the data obtained from a single rs-fMRI scan. The data and classifiers may also be applied to longitudinal studies or studies evaluating educational, training, or career development programs.

Y-WS, YK, and SO designed this study. The psychological data for this study were acquired by YK, CA, and YO. The anatomical and functional magnetic resonance images (MRI) data were acquired by Y-WS and DK. The VBM analysis of anatomical data was conducted by Y-WS and DK. Brain network analysis and SVM analysis of functional data was conducted by U-SC. Y-WS wrote the first draft of the manuscript with SO, YK, and U-SC contributed to revise this manuscript. All authors reviewed this manuscript.