Refined measure of functional connectomes for improved identifiability and prediction

Abstract

Brain functional connectome analysis is commonly based on population‐wise inference. However, in this way precious information provided at the individual subject level may be overlooked. Recently, several studies have shown that individual differences contribute strongly to the functional connectivity patterns. In particular, functional connectomes have been proven to offer a fingerprint measure, which can reliably identify a given individual from a pool of participants. In this work, we propose to refine the standard measure of individual functional connectomes using dictionary learning. More specifically, we rely on the assumption that each functional connectivity is dominated by stable group and individual factors. By subtracting population‐wise contributions from connectivity patterns facilitated by dictionary representation, intersubject variability should be increased within the group. We validate our approach using several types of analyses. For example, we observe that refined connectivity profiles significantly increase subject‐specific identifiability across functional magnetic resonance imaging (fMRI) session combinations. Besides, refined connectomes can also improve the prediction power for cognitive behaviors. In accordance with results from the literature, we find that individual distinctiveness is closely linked with differences in neurocognitive activity within the brain. In summary, our results indicate that individual connectivity analysis benefits from the group‐wise inferences and refined connectomes are indeed desirable for brain mapping.

1. INTRODUCTION

Functional connectivity profiles, or functional connectomes (FCs) calculated from functional magnetic resonance imaging (fMRI) time series have provided a promising measure to study connectivity patterns in human brains (Anderson, Ferguson, Lopez‐Larson, & Yurgelun‐Todd, 2011; Hutchison & Morton, 2015; Khalili‐Mahani et al., 2012; Weber, Soreni, & Noseworthy, 2014). Traditionally, patterns of brain activity are generally inferred at a population level. Often, we refer to such type of analyses as group‐wise or population‐wise approaches. A large number of studies evaluate functional connectomes at a population‐wise level. For instance, many studies have linked functional connectivity with various disease states or developmental stages (Allen et al., 2011; Khadka et al., 2013; Konrad & Eickhoff, 2010; Lynall et al., 2010; Qin et al., 2015; Sheline & Raichle, 2013; Yuhui et al., 2015). Simultaneously, dynamic FC analyses can help identify reoccurring patterns within human brains (Abrol, Chaze, Damaraju, & Calhoun, 2016; Allen et al., 2014; Cai, Zhang, et al., 2018; Cai, Zille, et al., 2018; Damaraju et al., 2014; Hutchison, Womelsdorf, Gati, Everling, & Menon, 2013; Rashid, Damaraju, Pearlson, & Calhoun, 2014). However, population‐wise studies generally overlook the existing connectivity heterogeneity within each group.

Each individual brain is unique. During the development process of the human brain, a lot of elements such as genes and environment may affect connectivity patterns. As a consequence, neural development and even cognition can be altered (Frost & Goebel, 2012; Mueller et al., 2013). Recently, several fMRI studies considering the information provided by subject‐level variability, have paved the way to the new promising avenue of detecting individual differences (Chamberland et al., 2017; Mueller et al., 2013). For example, Miranda‐Dominguez et al. proposed a model‐based approach that was capable of predicting the time series of each subject (Miranda‐Dominguez et al., 2014). Finn et al. demonstrated that FC could distinguish subjects from a large population, and predict individual fluid intelligence (Finn et al., 2015). Kaufmann et al. reported that functional connectomes developed into a more stable, individual wiring pattern during adolescence. Besides, they also showed that a delay in this network‐tuning process was associated with mental diseases (Kaufmann et al., 2017). Inspired by these works, we aim to refine the standard measure of individual functional connectivity. Our motivation is that, by reducing group‐wise contributions, intersubject variability across FC profiles should be improved.

Nowadays, various computational approaches have been proposed to identify subnetworks at a population level. Methods based on matrix factorization probably are most commonly used in the fMRI research community. Therein, independent component analysis, as one of the most representative methods, is applied to evaluate the hidden spatio/temporal structure from fMRI time series (Calhoun, Adali, Pearlson, & Pekar, 2001; Calhoun, Liu, & Adalı, 2009). Alternatively, relying on sparse priors, dictionary learning method has attracted increasing attention (Batmanghelich, Taskar, & Davatzikos, 2012; Eavani et al., 2012; Varoquaux, Gramfort, Pedregosa, Michel, & Thirion, 2011). For example, Eavani et al. utilized a sparse dictionary learning (SDL) model to determine distinct functional subnetworks (Eavani et al., 2012). Instead of finding patterns using fMRI series as inputs, their approach evaluates a group‐wise low‐rank decomposition for FC. While the resulting dictionaries are common to subjects, they are differentially expressed for each individual (i.e., representation matrix). In this work, we show that we may rely on results from the SDL model to refine standard measures of individual functional connectomes.

Here, we include data obtained from the publicly available Philadelphia Neurodevelopmental Cohort (PNC) (Satterthwaite et al., 2016). PNC is a large‐scale study of child development that incorporates rich multimodal neuroimaging, which is well suited to test the ability of the brain FC as a fingerprint across different fMRI protocols (Kaufmann et al., 2017). Through assessing the identification ability across rest and task sessions, we hope to manifest that an individual's underlying intrinsic functional architecture is reliable enough to detect him or her from a pool of participants regardless of how the brain is engaged during imaging. More importantly the transformative brain changes occurring from childhood to adolescence are critical for the shaping of individual developmental trajectories in cognitive and social function, personality, and mental health (Paus, Keshavan, & Giedd, 2008). Because participants are ranging from 8 to 22 years in PNC data sets, additional intersubject variability may be captured thereby revealing the individual FC transition within this period of development. All imaging data from the PNC is acquired in a short period of time. Due to the short study timeline and the lack of development phase, product sequences were used. The MRI protocol is comprised of scans designed to obtain information on brain structure, perfusion, structural connectivity, resting state functional connectivity (from 44 min, 14 s to 50 min, 32 s), working memory function (from 20 min, 40 s to 32 min, 19 s) and emotion identification (from 10 min, 4 s to 20 min, 40 s). The total scanning time of the entire protocol is 50 min, 32 s. More details about the MRI protocol can be found in Satterthwaite et al. (2016).

In the following, we will focus on two types of experiments. First, we evaluate the effect of refined FC profiles on the individual identifiability and demonstrate that modified FC values will lead to a higher distinctiveness in functional connectomes relative to raw connectivity. In the second experiment, we show that the proposed measure can also improve the prediction power of cognition, and disclose the link between the individual connectivity profile and neurocognitive behavior. More specifically, we introduce the refined measure of individual FC by applying the SDL model in Section 2. We then analyze the performance of the proposed approach using two experiments mentioned above in Section 3. Finally, some discussions and concluding remarks are given in Sections 4 and 5, respectively.

2. METHODS

As in previous work, functional brain networks were considered to be dominated by stable group and individual factors (Gratton et al., 2018; Xie et al., 2018). In this section, we utilize a data‐driven SDL model to estimate group factors from functional connectomes. Then, group contributions are removed to improve the individual identifiability.

2.1. Data acquisition and preprocessing

Data used in this work were collected through the PNC project, which was a large‐scale collaboration between the Brain Behavior Laboratory at the University of Pennsylvania and the Children's Hospital of Philadelphia (Satterthwaite et al., 2016). It is available via the dbGap database with nearly 900 participants ranging in age from 8 to 21 years. In this article, all subject data were acquired on the same scanner (Siemens Tim Trio 3 Tesla, Erlangen, Germany; 32 channel head coil) using the same imaging sequences. Blood oxygen level dependent (BOLD) fMRI was acquired using a whole‐brain, single‐shot, multislice, gradient‐echo (GE) echoplanar (EPI) sequence of 124 volumes with the following parameters: TR/TE = 3000/32 ms, flip = 90°, FOV = 192 × 192 mm, matrix = 64 × 64, slice thickness/gap = 3 mm/0 mm. The resulting normal voxel size was 3.0 × 3.0 × 3.0 mm. Prior to time series acquisition, a 5‐minute magnetization‐prepared, rapid acquisition gradient‐echo T1‐weighted (MPRAGE) image (TR = 1810 ms, TE = 3.51 ms, FOV = 180 × 240 mm, matrix = 256 × 192, effective voxel resolution of 1.0 × 1.0 × 1.0 mm) was acquired to aid spatial normalization to standard atlas space. Prior to scanning, in order to acclimate subjects to the MRI environment, a mock scanning session was conducted for each individual using a decommissioned MRI scanner and head coil. Mock‐scanning was accompanied by acoustic recordings of the noise produced by gradient coils for each scanning pules sequence. During these sessions, feedback regarding head movement was provided using the MoTrack (Psychology Software Tools, Inc, Sharpsburg, PA) motion tracking system. In order to further minimize motion, subjects' heads were stabilized in the head coil using one foam pad over each ear and a third over the top of the head. In addition, a stringent inclusion criterion was incorporated to limit the influence of motion (subject data with a maximum translation of >1.5 mm were excluded).

In the experiments, three paradigms of fMRI sessions including the resting state, emotion identification task and working memory task (fractal n‐back) were involved. Note that not all participants have these three paradigms of data. Thus we filtered out participants lacking one or two protocols of fMRI scanning sessions. Following data selection, a cohort of 623 participants (age 15.09 ± 7.49 years, 280 male and 343 female) was included in our analyses. In addition, all subjects underwent a one‐hour long computerized assessment adapted from tasks applied in functional neuroimaging studies to evaluate a broad range of cognitive domains. In this study, we relied on the performance scores (ratio of total correct responses) from the wide range achievement test (WRAT), which assessed the general learning ability (such as reading recognition, spelling, and arithmetic computation). While some reservations remain, the WRAT is still an effective method to estimate intelligence quotient (IQ) values (Griffin, Mindt, Rankin, Ritchie, & Scott, 2002).

Functional images were preprocessed using an automatic pipeline based on SPM12 (http://www.fil.ion.ucl.ac.uk/spm/). Note that all paradigms of fMRI data were preprocessed applying the same strategy. Standard preprocessing steps contained motion correction, spatial normalization to standard Montreal Neurological Institute (MNI) space and spatial smoothing with a Gaussian kernel (FWHM = 3 mm). The effect of head motion (six parameters) was further regressed out and BOLD time series went through a band‐pass filter (0.01 Hz to 0.1 Hz) to suppress physiological artifacts. Finally, we reduced the dimensionality of the data by employing the standard 264 region of interests (ROIs) template proposed by Power et al. with a 5 mm sphere radius parameter (Power et al., 2011). In addition, the AAL 116 parcellation template was also involved to evaluate our proposed framework (Figures S1–S4).

To facilitate the understanding of the behavior for different brain regions, we rely on the study from Smith et al. to assign 264 nodes into 10 functional networks corresponding to the primary resting state networks (RSNs) (Smith et al., 2011). These functional networks include: medial visual (“Med Vis,” RSN1), occipital pole visual (“OP Vis,” RSN2), lateral visual (“Lat Vis,” RSN3), default mode (“DMN,” RSN4), cerebellum (“CB,” RSN5), sensorimotor (“SM,” RSN6), auditory (“Aud,” RSN7), executive control (“EC,” RSN8), and two strongly lateralized frontoparietal maps (“FPR,” RSN9 and “FPL,” RSN10). For each of d = 264 nodes in our network, we calculate the membership relative to each of 10 RSNs. More specifically, for a specific RSN, we extract values at the location of nodes with a 4 mm radius sphere. Then, for each node, we transform these values into z‐scores. If the absolute maximum value of z‐scores for a certain node is above the threshold (∣z∣  > 3), we assign it to the given RSN. Eventually, a total of 181 nodes were left as displayed in Figure 1.

Figure 1.

Axial and sagittal view for 10 resting state functional module maps proposed by Smith et al. Med Vis, Medial Visual (RSN1); OP Vis, Occipital Pole Visual (RSN2); Lat Vis, Lateral Visual (RSN3); DMN, Default Mode Network (RSN4); CB, Cerebellum (RSN5); SM, Sensorimotor (RSN6); Aud, Auditory (RSN7); EC, Executive Control (RSN8); FPR, Frontoparietal Right (RSN9); FPL, Frontoparietal Left (RSN10) [Color figure can be viewed at http://wileyonlinelibrary.com]

2.2. Group‐wise functional connectivity patterns

Sparse dictionary learning was used to identify functional subnetworks within the human brain (Eavani et al., 2012). Let us assume we have n ∈ ℕ subjects. For each subject, BOLD time series with n_t time points and p ROIs (n_t, p ∈ ℕ) are available. C_i ∈ ℝ^p × p(i ∈ 1, 2, …, n) denotes a correlation matrix for each subject. Here, C_i(b₁, b₂) is the Pearson correlation between ROIs b₁ and b₂ across the entire BOLD time series. Due to the symmetric form of the correlation matrix, we discard the upper triangular part of C_i. This leads to the edge weight vector e_i = vec(C_i) ∈ ℝ^{p(p − 1)/2} for every subject. Then, we concatenate edge weight vectors from all subjects to generate the data Y = [e₁, e₂, …, e_n] with the size of m × n, where m = p(p − 1)/2. Here, the task of identifying the sparse representation of the functional connectivity across subjects (Y) can be modeled as an SDL problem. We would like to approximate the given data Y by solving the following formulation:

$$\begin{array}{ll}& \underset{D,X}{\min }\parallel Y-\mathit{DX}{\parallel }_F^2\\ & \mathit{\text{subject to}}\parallel x_i{\parallel }_0\le L,i=1,2,\dots ,K.\end{array}$$ (1)

where L is a non‐negative model parameter to control the sparsity level of dictionaries. D ∈ ℝ^m × K denotes the dictionaries, and K is the size of dictionaries. X = [x₁, x₂, …, x_K] ∈ ℝ^K × n is the representation matrix, and ∥ · ∥₀, ∥  · ∥_F define the 0 and Frobenius norm, respectively. The problem in Equation (1) can be reformulated as a sparse, group‐wise approximation of whole connectivity patterns across subjects.

To solve the formulation proposed by Equation (1), we can rely on an algorithm named K‐SVD (Aharon, Elad, Bruckstein, et al., 2006). K‐SVD is an efficient algorithm and designed to be a direct generalization of the K‐means clustering method. In the algorithm, the expression in Equation (1) is minimized iteratively. First, we fix the atoms D and try to find the best coefficient matrix X. In this way, we rewrite the penalty term as:

$$\parallel Y-\mathit{DX}{\parallel }_F^2=\sum _{i=1}^n\parallel y_i-D{x}_i{\parallel }_2^2$$ (2)

Hence, the problem in Equation (1) can be rearranged in the following form.

$$\begin{array}{ll}& \underset{x_i}{\min }\parallel y_i-D{x}_i{\parallel }_2^2\\ & \mathit{\text{subject to}}\parallel x_i{\parallel }_0\le L,i=1,2,\dots ,K.\end{array}$$ (3)

This problem can be solved by pursuit algorithms, and the orthogonal matching pursuit (OMP) algorithm is used here. In the second stage, we search for a better dictionary. At each time, we fix all columns in D except one, d_k, and find a new column d_k and its corresponding coefficients x^k (the kth column in X) to reduce the mean squared error (MSE). Here, the SVD is suggested to find the alternative d_k and x^k.

2.3. Increased individual identifiability by removing group factors

Inspired by the work proposed by Gratton et al. (2018), we believe that each functional connectivity is comprised of group and individual factors. Group components are expected to be more expressed by some subjects. By contrast, individual components are considered to carry the intersubject variability information for each subject. To improve individual identifiability, we first estimate group factors from the functional connectivity across individuals. Then, we remove their contributions from each subject's connectivity pattern to achieve that goal.

As discussed in the previous section, by solving Equation (1), we obtain a set of group‐wise dictionaries encoded by the column of D. For each subject i = 1, 2, …, n, dictionaries are performed to approximate the associated correlation matrix C_i using the corresponding representation vector x_i. In this view, dictionaries are expressed in all subjects. Because we want to improve the intersubject variability, contributions of these dictionaries can be excluded from each correlation matrix C_i to obtain a new refined FC ${\widehat{C}}_i$. The refined functional pattern is defined as follows:

$${\widehat{C}}_i=C_i-\mathit{mat}\left(D{x}_i\right)$$ (4)

where mat(Dx_i) ∈ ℝ^p × p is the correlation matrix reconstructed through the lower triangular information Dx_i. An illustration of the workflow can be seen in Figure 2. As a result, we believe that refined FC ${\widehat{C}}_i,i=1,2,\dots ,n$ can be used to better describe the subject‐specific pattern in contrast to the raw FC by removing common or group‐wise networks. Note that during data processing, the SDL framework discussed earlier was performed on data from each fMRI paradigm individually.

Figure 2.

An illustration of the workflow to decompose the raw FC into both the group‐wise and subject‐specific patterns using a SDL model. Note that the main assumption is that each functional connectivity can be decomposed into the population‐wise and individual FCs. Meanwhile, the subject‐specific FC carries the most of the identifiability information for each subject [Color figure can be viewed at http://wileyonlinelibrary.com]

2.4. Analysis of estimated group contribution

The population contribution extracted by the SDL approach may depend on the group selection. To validate this, for each kind of fMRI session, we randomly divide it into two sections with equal number of subjects. Then, we perform the algorithm on these two sections individually. When we obtain their group factors, a two‐sample t test with false discovery control (significant level α = .01) is implemented on them to detect the significant connections that are different. This whole procedure is repeated 500 times. Finally, the different percentage is estimated as:

$$\mathrm{D}\mathit{\text{ifferent percentage}}=\frac{\mathrm{S}\mathit{\text{ignificant different connections}}}{\mathrm{W}\mathit{\text{hole connections}}}\times 100$$ (5)

We also test how the number of subjects affects the performance of the SDL model. First, we randomly choose 20 people (the least sample value is equal to the dictionary size) and perform the SDL approach on them to extract the group contribution of each subject. We set the resultant group factors as a start point. Then, we gradually add new subjects into the data set and repeat the same pipeline to analyze the group contributions of the selected 20 subjects. Frobenius norm is then used to calculate the distance between the new group factors and start point. After reducing the group contribution to each subject, the intersubject variability is enhanced. To check the main difference in the spatial pattern of FC before and after depressing the group factors (C_i and ${\widehat{C}}_i$), a two‐sample t test with false discovery control (significant level α = .01) is implemented.

2.5. Individual identifiability analysis

For the target FC with a given subject and a given protocol (e.g., emotion identification), we want to identify that the connectivity pattern from another protocol (e.g., working memory) belongs to the same subject. First, we created a database consisting of connectivity matrices from all subjects, Y_m = [vec(C₁), vec(C₂), …, vec(C_n)], where vec(C_i) is a lower triangular part of the correlation matrix, m denotes protocol and i represents subject. To predict the subject identity, we calculated the similarity between target FCs and the ones from another protocol. Similarity scores are simply estimated by the Pearson correlation of connectivity patterns in various modalities. We then selected the FC that had the highest similarity score for the target connectivity pattern. If both the FCs are from the same subject, the identification accuracy is considered to be 100%. Otherwise, it is set to be 0%. Because we have three modalities (resting state [RS], emotion identification [EI], and working memory [WM]) in total, this results in three possible combinations (e.g., RS to EI, RS to WM, and WM to EI).

2.6. Prediction ability for cognitive behaviors

To test whether refined FC profile can better represent cognitive behaviors, we describe it from two scenarios: continuous phenotype and classification analysis. Note that cognitive behavior or cognitive ability in the following experiments refers to reading recognition, spelling, and arithmetic computation skills represented by the WRAT scores.

2.6.1. Continuous phenotype

To determine how refined connectomes can describe learning ability for each subject, we rely on the WRAT scores discussed earlier. Note that to avoid age effects on the WRAT scores, subjects whose age was below 16 years were removed, eventually leaving 267 subjects in total retained (Kaufmann et al., 2017). We use the leave‐one‐subject‐out cross‐validation to demonstrate whether refined FC profiles can better predict the individual WRAT score relative to raw FCs. During this procedure, the connectivity profile from one subject is selected as the test data, and FCs from remaining n − 1 subjects are considered as the training data. For each fold, we first perform a feature selection step, which calculates the correlation between each edge of FC profiles (Pearson correlation between two ROIs) and phenotype observations on the training set. If the correlation is significant (threshold α = .01) and positive, the corresponding feature is retained for the following steps. We then feed the training data to a simple linear regression model to fit the WRAT scores. Finally, we input the test data from the excluded subject into the model to generate a WRAT score. The predictive power is assessed by the correlation values between the predictive and observed WRAT scores across all iterations.

While assessing the predictive ability for each functional network, we repeat the leave‐one‐subject‐out cross‐validated procedure discussed earlier for three fMRI data paradigms. Under this condition, the feature selection step is restricted to each of 6 functional networks (merge the RSNs 1, 2, 3 into the visual network, the RSNs 9, 10 into the frontoparietal network, and remove the cerebellum network [only 4 nodes]). Due to limited nodes left for a certain network, we modify the statistical threshold from α = .01 to α = .1 to ensure that more features pass the selection step. Six sets of predicted WRAT scores are generated for each fMRI paradigm. The correlation between predicted and observed scores is defined as a measure of predictive power.

2.6.2. Classification analysis

We investigate whether we can classify two subsets according to cognitive ability. To perform the experiment, we define low and high cognitive groups according to WRAT scores as follows. First, we convert WRAT scores into z‐scores. Next, we extract subjects that are in either the upper or lower δ‐th percentile of the distribution of z‐scores. That means we only keep subjects that have the highest or lowest δ% WRAT scores. Here, we consider cases of δ ∈ {10,20,30}. Then we perform the feature selection procedure as described in continuous phenotype part. Note that correlation coefficients are retained at both the significant level α = .01 and α = .1 in this case. We use a support vector machine approach with Gaussian kernel for the classification (Cortes & Vapnik, 1995). Additionally, we intend to figure out whether applying refined FCs can improve the ability of discrimination relative to using raw FC profiles.

Through two subset extractions, a relatively low number of subjects are retained (54 subjects for δ = 10, 108 subjects for δ = 20, and 162 subjects for δ = 30). Thus, the experiment was replicated 100 times. For each run, we divided subjects into a training set 75%) and a testing set (25%). A standard SVM function built in Matlab with Gaussian kernel was applied, and a grid search was performed to optimize parameters within the SVM (e.g., the radius of Gaussian kernel, the weight of the soft margin cost function).

2.7. Brief description of experiments

Before performing our experiments, we first preprocess the PNC data, optimize the key parameters involved in our proposed model, investigate the contribution from the estimated group elements, and assign 264 ROIs to 10 functional networks. In order to verify the performance of our proposed method, several experiments are repeated under the automated anatomical labeling 116 (AAL 116) atlas. The results can be found in Data S1. Then, we perform our experiments that mainly include two sections. For the first section, we evaluate the effect of refined FC profiles on the individual identifiability. Further, we design an experiment to assess how refined connectivity influence the intersubject variability along the development of the human brain. Next, we explore how refined connectomes affect the intersubject variability of each functional network.

For the second section, we determine whether refined FC profiles can better describe cognitive behavior. To further validate that refined FCs have a better ability to describe cognitive behavior, we perform a classification analysis on the subsets generated based on the WRAT scores.

3. RESULTS

In this section, refined functional connectivity patterns defined in Equation (4) are estimated from the fMRI data set. By comparing with raw FCs, we assess the performance of refined FC profiles to identify individuality and predict cognitive behavior from a large population. The details are described later.

3.1. Tuning parameter selection

Parameter optimization has a significant effect on the performance of the SDL model. The model described in Equation (1) included two key parameters: the dictionary size K and the sparsity level for the atoms L. For the SDL model, we consider that each fMRI paradigm possesses a set of parameters with the best performance. Through applying the SDL framework to refine the FC profiles, we hope to improve the ability of individual identification. Thus, we use the identification accuracy for different combinations of fMRI paradigms to select the key parameters.

To figure out the effect of these parameters on the identification accuracy, we run the algorithm with varying values of K and L (K and L changed from 2 to 15 and 1 to 15, respectively). This range was referred in Eavani et al.'s study (Eavani et al., 2012). Then we repeated the identification procedures mentioned in the previous section to extract refined functional patterns. The best pair of parameters, K and L, were chosen to be the combination that provided high accuracy for each situation (RS to EI, RS to WM, and WM to EI). The optimization results are described in Figure 3.

Figure 3.

Model selection: choice of dictionary size K and sparsity L based on the identification accuracy. (a) resting state to emotion; (b)resting state to working memory; (c) working memory to emotion [Color figure can be viewed at http://wileyonlinelibrary.com]

In Figure 3, the hotter the color, the higher the accuracy. By checking the results, we obtained that the identification accuracy changed as the K and L varied in the K‐SVD algorithm. Meanwhile, three different cases (Figure 3a–c) had the similar results. That is, for a fixed K and incrementally varying sparsity level L, the accuracy increased gradually. Simultaneously, variation in K alone also displayed the same trend. By balancing the performance in these three situations, the best pair of parameters for the SDL was chosen for each protocol. Specifically, K = 15, L = 13 was chosen for the resting state fMRI data, and K = 15, L = 14 was selected for both the emotion identification and working memory fMRI data.

3.2. Investigation of the group contribution

We first checked if the population contribution depended on the group selection. It should be noted that for the key parameters involved in all experiments, we used the results generated from the tuning parameter section. As a consequence, the choice of group members does not have a crucial influence on the population contribution extracted by the SDL model. Even for the resting state fMRI that is unconstrained during the collecting process, the different percentage is around 4% (only 2% for task fMRI sessions in Figure 4a).

Figure 4.

Analysis of group contribution extracted by the SDL model. A) Effect of group selection on the resultant population contribution. Box plot denotes the percentage of the significant difference between two parts for all fMRI protocols (red: resting state; blue: working memory; green: emotion identification) across 500 times run. B) Effect of the subject number on the resultant group contribution. Here, frobenius norm is applied to calculated the distance between two group factors. C) Main differences in the spatial pattern of FC before and after reducing the group contribution. Significant connections are estimated through a two‐sample t‐test with false discovery control (significant level α = 0.01). It should be noted that colorful connections in the figure represent the connectivity within a certain functional network. Otherwise, gray lines denote the inter‐network connections [Color figure can be viewed at http://wileyonlinelibrary.com]

Then, we described how the number of subjects affected the performance of the algorithm. As shown in Figure 4b, we observe that the distance for all three fMRI protocols first increases and then remains steady. The elbow point for this trend is around 75. It demonstrates that as the sample size reaches a critical point, the group factor of a specific subject is hardly affected by the newly added data. Because the subject number is large enough in the following experiments, the resulting group contribution is robust and repeatable. In addition, combining the results in Figure 4c and 5a, we observe that the significant difference for the FC largely lies in the intranetwork connections. Through applying the refining procedure, the intranetwork activity is principally weakened relative to the internetwork connections.

Figure 5.

Evaluation of the influence of SDL on identification accuracy. A1‐A3) From the left to right: the group averaged FC of the original data for resting state, emotion identification and working memory fMRI; the group averaged FC of the reconstructed data applying refined functional patterns for three modalities. Refined FCs are extracted using the SDL method as described in the methods section. B1‐B3) For the top line of the figure, from left to right: identifiability matrix (i.e. Pearson correlation coefficient between functional connectivity across subjects and modalities) of the original data for all three combinations (rest to emotion, rest to working memory and working memory to emotion); identifiability matrix of the refined FC data. Note that row and column subject order is symmetric. Thus, diagonal elements are correlation scores from the matched subjects, while off diagonal coefficients are from the unmatched subjects. Mean correlation coefficients for both diagonal (matched) and off‐diagonal (unmatched) elements are also displayed (bottom, error bars presents ± s.d.). ** indicates P < 10‐5 for two‐tailed t‐test. C) Identification accuracy for both original and refined FCs in all three situations [Color figure can be viewed at http://wileyonlinelibrary.com]

3.3. Refined FCs affecting the individual identifiability

We first explore how refined FCs affect the individual distinctiveness in contrast with original functional patterns. Next, we perform the age‐wise and subnetwork‐wise analyses on the identification accuracy applying the SDL method.

3.3.1. Effects of refined FCs

To investigate the sensitivity of identifiability to the refined FC, we tested the group average functional connectomes before and after the processing of the fMRI data with three paradigms. Original FC from various paradigms shared the similar format (left in Figure 5a1–a3). However, for the refined FC, we can see obvious differences across paradigms, especially between the resting state and task (EI, WM) fMRI (right in Figure 5a1–a3). We speculate that they are caused by the increase in the intersubject variability.

To further analyze the difference in identifiability, we calculated correlations between connectivity matrices of all 623 subjects across three possible combinations for both raw and refined FCs. The results are shown in the top row of Figure 5b1–b3 (left: original FCs; right: refined FCs). Note that the row and column is symmetric by subject. Hence, diagonal elements were correlation scores from the matched subjects, while off‐diagonal elements are the ones from the unmatched subjects. In a comparison of raw cross‐subject correlation coefficients, we found that with refined FCs, the difference of coefficients between the matched and unmatched subjects were significantly higher. For the pair of emotion identification and working memory fMRI, this condition was the most pronounced. Besides, the mean and SD of both the diagonal and off‐diagonal elements were also displayed in the bottom line of Figure 5b1–b3. By applying refined FC profiles, the difference between the diagonal and off‐diagonal elements is largely enhanced. It demonstrates that refined FCs indeed benefit for individual discrimination, and the pair of EI to WM results in the best performance.

We then estimated the identification accuracy for three fMRI session combinations to validate that refined FCs can increase the subject‐specific identifiability. The procedures have already been discussed in the section of individual identifiability analysis, and the results are shown in Figure 5c. When checking the identification results in Figure 5c, we observe that for the pairs of RS to EI and RS to WM, the identification accuracy rose from below 60% to 80% by implementing the refined procedure. By contrast, the pair of WM to EI gained around 15% by applying refined connectomes. Meanwhile, with refined connectivity, the accuracy of the WM to EI was highest among three combinations (about 93%). As a consequence, refined FCs can significantly enhance the intersubject variability (accuracy significantly increases for all situations), and the combination WM to EI possesses the greatest identification power. This confirms the results in Figure 5b1–b3.

3.3.2. Age‐wise analysis for refined FCs

As a second analysis, we performed the identification accuracy, as age increased, using both raw and refined FCs. First, we averaged identification accuracies across all the FC combinations to indicate the overall performance of connectome distinctiveness with advancing age. Note that we refer to the measure of identification accuracy as connectome distinctiveness, as it resembles how well an individual's connectome discriminates that subject from the rest of the cohort. The results are shown in Figure 6a. During this step, we repeatedly fit a smooth function with automatic estimation of the smoothness parameter between age and connectome distinctiveness, computing mean and SD across 10,000 bootstraps (Fjell et al., 2010). Similarly, the change of connectome distinctiveness with age for each FC combination was considered as well (Figure 6b). Further, to assess whether the obtained accuracies were significantly above chance, we performed 10,000 nonparametric permutation tests (two‐sided) on the subject identification. It illustrates that identification accuracies are highly above chance for both raw and refined FCs (Figure 6a, all p < .0001, none of 10,000 permutation tests exceeding the true values).

Figure 6.

Age‐wise identification accuracy analysis. Red lines correspond to the accuracy when applying the refined functional connectome measure (SDL). Blue lines refer to the accuracy applying raw FCs. A) Association between connectome distinctiveness and age, yielding robust differences between raw and refined FCs. Here, we average the identification accuracies across all combinations to form the connectome distinctiveness. The curves represent mean (solid line) and s.d. (shaded areas) of smooth function fitting obtained based on 10,000 permutations. It should be noted that p‐values are obtained across 10,000 permutation testing. B) Relationship between the connectome distinctiveness of each identification combination and age. (a) Identification accuracy using for ”RS to EI” prediction combination. (b) Identification accuracy using for ”RS to WM” prediction combination. (c) Identification accuracy using for ”WM to EI” combination. Similarly, all the results are generated across 10,000 bootstraps [Color figure can be viewed at http://wileyonlinelibrary.com]

The resulting identification rates, for both raw and refined FCs, can be seen in Figure 6. Overall, identification rates estimated by refined FCs performed much better than raw FCs' accuracy except for a certain period (from 8 to 8.5 years for the pair of WM to EI). This verifies that refined FCs indeed benefit discriminating the individual identity from a pool. Besides, connectome distinctiveness based on raw or refined FCs showed a clear increase with age, especially for the typical puberty period (from 11 to 15 years old, period obtained from bootstrapping). The outcome perfectly agrees with the previous study proposed by Kaufmann et al. (2017). Interestingly, during this period, refined FCs provided an obvious increase on the individual identifiability for all three fMRI paradigm combinations, especially for the combination of WM to EI. In comparison with the rest to task combinations (RS to EI and RS to WM), refined connectomes extracted from the pair of task to task (WM to EI) gave a fairly large accuracy escalation. We hypothesize that group factors removed by our approach are dominated by the resting state fMRI. Within the WM to EI identification task, the resting state fMRI signal can be considered as the nuisance. Results indicate that our proposed measure using the SDL model can greatly help to reveal the subject‐specific features.

3.3.3. Subnetwork‐wise analysis for refined FCs

To explore how refined connectomes affect the discriminative power within the brain, we investigated identification rates for each functional network mentioned in data acquisition and preprocessing section. Note that in this experiment, we merged the RSNs 1, 2, 3 into the visual network and the RSNs 9, 10 into the frontoparietal network. Due to the small number of nodes (only four), the cerebellum was not involved. It led to six functional networks in total, which were considered in the following analysis.

In Figure 7, we display the histogram of identification rates for each imaging session combination and each functional network estimated by both raw and refined FCs. Similar to what has been observed using the full set of ROIs in Figures 5c and 6, using refined FCs produced a significant increase in the identification accuracy for every case. Hence, the refinement procedure applying the SDL is pretty robust in this work. Compared with the task to task combination, the gain in the accuracy was more significant with the rest to task pairs, especially for the visual, sensorimotor, and frontoparietal networks. Besides, we would like to point out that the frontoparietal network possessed the most discriminative power for all session combinations. This is in accordance with the related studies (Finn et al., 2015; Kaufmann et al., 2017). Interestingly, the visual network did not display strong identifiability across all combinations with raw FCs. However, after the refinement, the highest increase was obtained in the visual brain areas for all situations. Previous studies have suggested that the visual system was highly associated with group factors (Cai, Zhang, et al., 2018; Cai, Zille, et al., 2018; Jolles, van Buchem, Crone, & Rombouts, 2010; Zille, Calhoun, Stephen, Wilson, & Wang, 2017). We believe that the subject‐specific features in the visual network are concealed by the group influence. Hence, discriminative power in the visual area is significantly improved by removing group contributions.

Figure 7.

Subnetwork‐wise analyses (6 functional networks) identification accuracy for different fMRI session combinations. Blue bars correspond to the accuracy when using the refined function connectome measure (SDL). Brown bars refer to the accuracy applying original FCs. (a) Identification accuracy for ”RS to EI” prediction combination. (b) Identification accuracy for ”RS to WM” prediction combination. (c) Identification accuracy for ”WM to EI” combination [Color figure can be viewed at http://wileyonlinelibrary.com]

Overall, observations made in these experiments confirm the assumption that the functional connectivity can be divided into group and subject‐specific factors. Applying the SDL approach, we can effectively remove the group‐wise contribution from each individual connectivity. To do so, the discriminative power for each subject is significantly enhanced. In the next section, we further consider whether similar gains can be made for cognition prediction by applying refined FCs.

3.4. Prediction of connectivity for cognitive behavior

As described by Finn et al. (2015), individual differences in functional connectivity are relevant to individual differences in behavior. In this section, we explore whether refined FC profiles can better describe cognitive behaviors. Note that cognitive behavior or cognitive ability in the following experiments refers to reading recognition, spelling, and arithmetic computation skills represented by the WRAT scores. More specifically, we first test refined FC profiles' prediction ability for cognitive behaviors. Further, applying refined FCs, we investigate whether we could train a classifier to better discriminate between two subsets based on differing cognitive ability.

3.4.1. Continuous phenotype

Similar to the individual identification analysis, we first tested whether refined FC profiles within the whole brain could better predict cognitive behavior for novel subjects. Prediction scores applying both raw and refined FCs for each fMRI scanning session are displayed in Figure 8a. By looking at the results in Figure 8a, we observed the correlation between predicted and observed scores estimated by refined FCs was significantly better than raw FCs for all combinations of fMRI data. In contrast with raw FCs, the range of predicted scores extracted by refined FC profiles was much narrower, especially for resting state and working memory fMRI data. To discriminate these predictions from random guess, 100 nonparametric permutation tests were performed for every scenario and the results are shown in Figure 8a. The results illustrate that the prediction of cognitive behavior (correlation between predicted WRAT and observed scores) is above chance for both raw and refined FCs (all p < .01, none of 100 permutation tests exceeding the true values). In addition, more points in Figure 8b,d,f were located at the 95% confidence interval relative to Figure 8a,c,e. That indicates refined FCs.

Figure 8.

Connectivity profiles predict cognitive behavior. A) Results from a leave‐one‐subject‐out cross‐validation (LOOCV) analysis comparing predicted and observed WRAT scores. Scatter plot shows prediction results based on the positive features in the whole brain with a threshold α = 0.01. Each dot represents one subject, and the area between dashed lines reflects 95% confidence interval for the best‐fit line which used to assess the predictive power of the model. R values are the correlation coefficients between predicted and observed WRAT scores. (a‐b) resting state fMRI, (c‐d) emotion identification fMRI and (e‐f) working memory fMRI. Note that 100 nonparametric permutation tests are performed and p‐value < 0.01 for all prediction conditions listed above. B) Results from a LOOCV analysis which feature selection is restricted to within‐network edges for six functional networks with a threshold α = 0.1. Similarly, r value in y axis means the correlation between predicted and observed WRAT scores, x axis indicates the network label. Note that for a specific network, if no features pass the statistical threshold step, a missing bar is used to represent it. (a) resting state fMRI, (b) emotion identification fMRI and (c) working memory fMRI [Color figure can be viewed at http://wileyonlinelibrary.com]

In the second stage, we compared the predictive ability for each functional network between raw and refined FC profiles and explored which network contributed the most predictive power. The results can be seen in Figure 8b. As a general observation, there were significant gains in predictive power when using refined FC profiles compared to raw ones. Across all networks, default mode (r = .29 for refined FCs vs. r = .19 for raw FCs) and sensorimotor (r = .48 for refined FCs vs. r = 6.54 × 10⁻¹⁹ for raw FCs) appeared to show the best predictive power in the resting state fMRI. Besides, the best predictive performance was provided for auditory (r = .48 for refined FCs vs. r = .11 for raw FCs) and executive control (r = .51 for refined FCs vs. r = .12 for raw FCs) when using the emotion identification profiles as predictors. Finally, FC values related to the working memory task were the most predictive in the visual (r = .64 for refined FCs vs. r = .35 for raw FCs) and frontoparietal (r = .69 for refined FCs vs. r = .51 for raw FCs) regions. Note that within all fMRI scanning sessions, frontoparietal network possessed a considerably strong predictive ability (r = .66, . 58, and .69 for resting state, emotion identification, and working memory fMRI, respectively). Interestingly, by comparing with the results in the previous section, we observed that the networks that were significantly discriminative between individuals could also better explain cognitive behavior (such as visual, executive control, and frontoparietal networks). Applying refined FC profiles, the individual identifiability variation is more closely related to the change of predictive power for cognitive performance.

3.4.2. Classification analysis

For further analysis, we investigated whether we could discriminate between two subsets based on cognitive ability. As we can see from Figure 9, regardless of the significance level α and percentile values δ, applying refined connectomes led to an improved classification accuracy compared with raw FCs for all fMRI scanning sessions. Meanwhile, SD values estimated by refined FCs are significantly reduced at the relatively low percentiles (δ = 10, 20). When using raw FC profiles, classification accuracy was similar among situations with different percentile values δ. By contrast, with refined FC profiles, the great disparity (relatively low value of δ) between high and low cognitive groups induced high classification accuracy. According to our understanding, the great distinction between groups should make us classify them easily. This tendency was well expressed by refined FC profiles instead of raw ones. Interestingly, refined FC profiles from emotion identification task provide the greatest improvement in classifying WRAT related subpopulations. More specifically, at the significance level α = .01, the accuracy increased from 64.3% to 92.8% for δ = 10, from 62.9% to 85.2% for δ = 20 and from 68.3% to 81.7% for δ = 30. When using the significance level of δ = 0.1, the improvement in classification was from 64.3% to 92.8% for δ = 10, from 66.7% to 81.5% for δ = 20 and from 70.7% to 82.9% for δ = 30.

Figure 9.

Classification results between low and high cognitive groups based on WRAT scores. Within the feature selection step, two significant level values are involved (A: α = 0.01, B: α = 0.1). For each significant level, three percentile values are considered ((a): δ = 10, (b): δ = 20, (c): δ = 30). Blue boxes represent the classification accuracy applying refined FC profiles for all fMRI scanning sessions (resting state, working memory and emotion identification). Black boxes provide the classification accuracy using raw FC profiles for all fMRI scanning sessions [Color figure can be viewed at http://wileyonlinelibrary.com]

4. DISCUSSION

Recently, some studies pointed out that functional connectivity profiles could be considered as a fingerprint‐like measure to identify subjects from a large population (Finn et al., 2015; Kaufmann et al., 2017). Meanwhile, Gratton et al. (2018) proved that functional brain networks were dominated by stable group and individual factors. Inspired by these findings, we believe that subtracting population contributions from connectivity patterns can enhance the identifiability of FC profiles. In this study, we used a SDL method to extract group factors of FCs. Then, we removed them from connectivity patterns to generate refined FC profiles. Applying three fMRI scanning sessions (resting state, emotion identification, and working memory fMRI), we tested the effect of refined FCs on the prediction power following similar procedures proposed by Finn et al. (2015). Through both the age‐wise and subnetwork analyses, we further estimated refined FCs' identifiability relative to raw FC profiles. Furthermore, we investigated whether refined FCs benefited the description of cognitive behavior. In particular, we considered two scenarios containing both the prediction of continuous phenotypes and the discrimination of two groups with different cognitive ability.

When checking the group average connectomes, we found that original connectivity from different modalities shared the similar pattern. However, refined FC profiles demonstrated a significant difference between various fMRI paradigms, especially between the resting state and task (EI, WM) fMRI signals. To our knowledge, after the use of SDL, group contributions are excluded from functional connectomes. Hence, the intersubject variability is enhanced. In this way, the similarity in functional connectivity across various fMRI sessions is largely suppressed. To further analyze the individual identifiability affected by refined FCs, we evaluated the power by calculating correlations of connectivity matrices across all subjects. Regardless of protocol combinations, the similarity among unmatched subjects estimated by refined FCs was significantly reduced relative to raw FC profiles. The gain of individual discrimination was obtained when using refined connectomes. Meanwhile, the pair of EI to WM possessed the best identifiability. Compared with raw connectomes, the accuracy using refined FC profiles increased from below 60% to 80% for the combinations of RS to EI and RS to WM. Specifically, the identification rate for the pair of EI to WM reached the highest level of 93%.

To further assess the performance of refined FC profiles, we performed both the age‐wise and subnetwork analyses. As the age increased, both raw and refined FCs provided a significant improvement in connectomes distinctiveness for each combination, which was highly consistent with the previous study (Kaufmann et al., 2017). This inclination was the most pronounced for the typical puberty period. However, for each combination, better identification rates were gained by refined connectomes. Specifically, during the age period from 11 to 15 years, the improvement of individual identification was higher than other ages. It demonstrates that as the brain matures, intersubject variability becomes larger and larger. Compared with the pairs of rest to task (RS to EI and RS to WM), refined connectomes estimated by the task to task combination (WM to EI) got a larger increase in identification rates. Meanwhile, we received the highest accuracy using the fMRI session combination of EI to WM. We believe that without any stimuli, group factors play a more dominant role in the brain activity. A fairly large proportion of group factors are expressed in the resting state fMRI. When applying SDL method to remove population contributions, intersubject variability remaining in the task related fMRI is enhanced relative to rest session. Meanwhile, due to more individualized cognitive processes involved in the task fMRI, it is reasonable that connectomes from the pair of EI to WM give the best performance in individual distinctiveness. Through the subnetwork analysis, we investigated the contribution of each FC network. Similarly, a significant increase in identification rates was achieved using the refined FC profiles, and this phenomenon was more salient for rest to task session combinations. In addition, we observed that the frontoparietal network made an enormous contribution to the intersubject variability as reported in previous studies (Finn et al., 2015; Kaufmann et al., 2017). Interestingly, for all fMRI session combinations, the strong identifiability only appeared in the visual network when group influence was excluded. We hypothesize that individual features in the visual connectome are dominated by group factors. Through applying our refinement procedure to reduce the group contribution, this intersubject variability can be enhanced.

Furthermore, we explored whether refined connectomes had a positive effect on individual differences in behaviors. First, we tested the predictive power of refined FCs for cognitive behavior. Regardless of fMRI session combinations, refined FC profiles significantly improved the predictive ability based on cognitive behaviors. Therein, the frontoparietal network as a brain region of last development and maturity, showed a strong prediction power on cognitive behavior. Taking into account the results in the previous section, we determined that the networks with a stronger individuality can also better predict cognitive ability. It indicates that individual distinctiveness is positively correlated with subject‐specific variations in cognitive activity. Similarly, refined connectomes could also enhance classification rates on discriminating high and low cognitive groups according to WRAT scores. Refined FC profiles from fMRI emotion identification task gained the highest increase in accuracy. We believe that the fMRI data collected by emotion identification task can better reflect the individualized cognitive processes in human brains.

During our experiments, the task fMRI sessions produced the best fingerprinting results. This suggests that both individual distinctiveness and cognitive prediction using the resting state fMRI are proven to be the most challenging for identifying individual differences. In this work, we relied on FCs computed over the entire scanning session. However, many works indicated that during the resting state, significant variations could be observed in a short time, and dynamic FC analysis has drawn more and more attention (Cai, Zhang, et al., 2018; Cai, Zille, et al., 2018; Damaraju et al., 2014; Hutchison et al., 2013; Marusak et al., 2017; Rashid et al., 2014). If we incorporate dynamic measurements, it may further improve the results. This static analysis might be one limitation of our work. However, it is interesting to notice again that the highest improvements for some experiments made by refined FCs are always related to resting state fMRI (including rest to task combinations), showing the importance of resting state fMRI.

5. CONCLUSION

In this work, we proposed to apply the group‐wise analysis to improve subject‐specific distinctiveness using an SDL approach. Results for individual recognition validated that refined FC profiles could significantly increase subject‐specific identifiability. In particular, the combination of EI to WM showed the best performance for connectomes distinctiveness. We also showed that intersubject variability increases as the brain matures, consistent with existing studies. Likewise, we observed that the frontoparietal network contributed a lot to this power. The strong identifiability of the visual network was disclosed by applying refined connectomes. In addition, we found that refined FC profiles also helped improve prediction of cognitive behaviors across fMRI scanning sessions. Individual distinctiveness was more closely related to differences in cognition by applying refined connectomes. In summary, the findings in this study indicate that analyses of individual fMRI data can benefit from group‐wise methods. In turn, taking the individual variation into account may also help population‐wise studies.

Supporting information

Fig. S1 (AAL116 template) Evaluation of the influence of sparse dictionary learning on identification accuracy. A1‐A3) From the left to right: the group averaged FC of the original data for resting state, emotion identification and working memory fMRI; the group averaged FC of the reconstructed data applying refined functional patterns for three modalities. Refined FCs are extracted using the sparse dictionary learning method as described in the method section. B1‐B3) For the top line of the figure, from left to right: identifiability matrix (i.e., Pearson correlation coefficient between functional connectivity across subjects and modalities) of the original data for all three combinations (rest to emotion, rest to working memory and working memory to emotion); identifiability matrix of the refined FC data. Note that row and column subject order is symmetric. Thus, diagonal elements are correlation scores from the matched subjects, while off‐diagonal coefficients are from the unmatched subjects. Mean correlation coefficients for both diagonal (matched) and off‐diagonal (unmatched) elements are also displayed (bottom, error bars presents ± s.d.). ** indicates p < 10–5 for two‐tailed t‐test. C) Identification accuracy for both original and refined FCs in all three situations.

Figure S2. (AAL116 template) Age‐wise identification accuracy analysis. Red lines correspond to the accuracy when applying the refined function connectome measure (sparse dictionary learning). Blue lines refer to the accuracy applying raw FCs. A) Association between connectome distinctiveness and age, yielding robust differences between raw and refined FCs. Here, we average the identification accuracy across all identification combination to form the connectome distinctiveness. The curves represent mean (solid line) and s.d. (shaded areas) of smooth function fitting across 10,000 bootstraps. It should be noted that P‐values are obtained through permutation testing across 10,000 permutations. B) Relationship between the connectome distinctiveness of each task combination and age. (a) Identification accuracy using for “RS to EI” prediction combination. (b) Identification accuracy using for “RS to WM” prediction combination. (c) Identification accuracy using for “WM to EI” combination. Similarly, all the results are generated across 10,000 bootstraps.

Figure S3. (AAL116 template) Connectivity profiles predict cognitive behaviour. A) Results from a leave‐one‐subject‐out cross‐validation (LOOCV) analysis comparing predicted and observed WRAT scores. Scatter plot shows prediction results based on the positive features in the whole brain with a threshold α = .01. Each dot represents one subject, and the area between dashed lines reflects 95% confidence interval for the best‐fit line which used to assess the predictive power of the model. (a‐b) resting state fMRI, (c‐d) emotion identification fMRI and (e‐f) working memory fMRI.

Figure S4. (AAL116 template) Classification results between low and high cognitive groups based on WRAT scores. Within the feature selection step, two significant level values are involved (A:α = .01, B:α = .1). For each significant level, three percentile values are considered ((a):δ = 10, (b):δ = 20, (c):δ = 30). Blue boxes represent the classification accuracy applying refined FC profiles for all fMRI scanning sessions (resting state, working memory and emotion identification). Black boxes provide the classification accuracy using raw FC profiles for all fMRI scanning sessions.

Cai B, Zhang G, Hu W, et al. Refined measure of functional connectomes for improved identifiability and prediction. Hum Brain Mapp. 2019;40:4843–4858. 10.1002/hbm.24741

DATA ACCESSIBILITY

The data that support the findings of this study are openly available in dbGap at same https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000607.v1.p1.