Introduction
Major depressive disorder (MDD) is a common mental disease that severely limits psychosocial functioning and diminishes quality of life (Malhi & Mann, Reference Malhi and Mann2018), and tends to emerge during adolescence (Kessler et al., Reference Kessler, Avenevoli, Costello, Georgiades, Green, Gruber and Merikangas2012). Depressive symptoms in adolescence are associated with an increased risk of long-term adverse outcomes, such as suicide, psychiatric comorbidities, and chronic course and recurrent episodes of illness (Aalto-Setälä, Marttunen, Tuulio-Henriksson, Poikolainen, & Lönnqvist, Reference Aalto-Setälä, Marttunen, Tuulio-Henriksson, Poikolainen and Lönnqvist2002; Kiviruusu, Strandholm, Karlsson, & Marttunen, Reference Kiviruusu, Strandholm, Karlsson and Marttunen2020; Pine, Cohen, Gurley, Brook, & Ma, Reference Pine, Cohen, Gurley, Brook and Ma1998). The rapid development of psychoradiology has advanced our understanding of the brain mechanisms of MDD. However, neurobiological research in adolescents has lagged behind that in adults. The brain abnormalities in adolescent MDD can be different from those in adults (Cullen et al., Reference Cullen, Westlund, Klimes-Dougan, Mueller, Houri, Eberly and Lim2014), due to significant maturational changes in the adolescent brain (Giedd et al., Reference Giedd, Blumenthal, Jeffries, Castellanos, Liu, Zijdenbos and Rapoport1999). Thus, the neurobiology of adolescent MDD needs further study.
Resting-state functional magnetic resonance imaging (rs-fMRI) has been a valuable and noninvasive tool to detect functional abnormalities in patients with MDD. Functional connectivity (FC) is measured with rs-fMRI by quantifying intrinsic functional brain organization (Biswal, Yetkin, Haughton, & Hyde, Reference Biswal, Yetkin, Haughton and Hyde1995; Friston, Reference Friston2002). Previous studies have demonstrated functional impairments in the limbic-prefrontal circuits (Connolly et al., Reference Connolly, Ho, Blom, LeWinn, Sacchet, Tymofiyeva and Yang2017; Geng et al., Reference Geng, Wu, Kong, Tang, Zhou, Chang and Wang2016; Pannekoek et al., Reference Pannekoek, van der Werff, Meens, van den Bulk, Jolles, Veer and Vermeiren2014; Wu et al., Reference Wu, Tu, Sun, Geng, Zhou, Jiang and Kong2019a), abnormalities in interactions among large-scale brain networks (Sacchet et al., Reference Sacchet, Ho, Connolly, Tymofiyeva, Lewinn, Han and Yang2016), and disruptions in the topological organization of brain functional networks (Wu, Li, Zhou, Zhang, & Long, Reference Wu, Li, Zhou, Zhang and Long2020b) in adolescent MDD. However, most of previous studies performed FC analysis based on the assumption that connectivity was static within the entire time courses, and did not consider the important dynamic aspect over time. Emerging evidence suggests that the human brain is a large, interacting dynamic network, and its architecture of coupling among brain regions varies across time (Allen et al., Reference Allen, Damaraju, Plis, Erhardt, Eichele and Calhoun2014; Calhoun, Miller, Pearlson, & Adalı, Reference Calhoun, Miller, Pearlson and Adalı2014; Chang & Glover, Reference Chang and Glover2010; Kang et al., Reference Kang, Wang, Yan, Wang, Liang and He2011; Preti, Bolton, & Van De Ville, Reference Preti, Bolton and Van De Ville2017). In this context, a new concept of ‘chronnectome’ has been proposed to emphasize the dynamic characteristics of brain connectivity (Calhoun et al., Reference Calhoun, Miller, Pearlson and Adalı2014), implying that traditional static FC analysis based on the whole scan may be too simple to capture important time-varying information of brain activity. The spatial functional chronnectome is an innovative mathematical model designed to capture dynamic features in the organization of brain networks derived from rs-fMRI data (Calhoun et al., Reference Calhoun, Miller, Pearlson and Adalı2014; Preti et al., Reference Preti, Bolton and Van De Ville2017).
Abnormal time-varying FC has been found in MDD patients. For example, altered temporal variability of FC related to the medial prefrontal cortex was observed in adults with MDD and correlated with symptom severity (Kaiser et al., Reference Kaiser, Whitfield-Gabrieli, Dillon, Goer, Beltzer, Minkel and Pizzagalli2016). Additionally, temporal dynamic patterns of the ventromedial prefrontal cortex underlie the association between rumination and depression (Gao et al., Reference Gao, Biswal, Yang, Li, Wang, Chen and Yuan2023). Some other studies found that adult MDD patients spent more time in the weakly-connected state (Yao et al., Reference Yao, Shi, Zhang, Zheng, Hu, Li and Hu2019; Zhi et al., Reference Zhi, Calhoun, Lv, Ma, Ke, Fu and Sui2018), and MDD-caused FC disruptions mainly occurred in this state (Yao et al., Reference Yao, Shi, Zhang, Zheng, Hu, Li and Hu2019). However, previous dynamic FC studies mainly focused on adult MDD, only few studies have been conducted in adolescents with MDD. A recent preliminary study identified aberrant frontoinsular-default network dynamics that associated with illness severity and rumination in adolescent depression based on co-activation pattern analysis (Kaiser et al., Reference Kaiser, Kang, Lew, Van Der Feen, Aguirre, Clegg and Pizzagalli2019). Subsequent study has shown that brain functional network dynamics underlie symptom heterogeneity in late adolescence-onset MDD (Marchitelli et al., Reference Marchitelli, Paillère-Martinot, Bourvis, Guerin-Langlois, Kipman, Trichard and Artiges2022). Although these studies have preliminarily explored the adolescent MDD-related dynamic FC alterations, patterns of change in whole-brain FC and network topological properties in the context of dynamic FC have not been fully understood in adolescent MDD.
The main purpose of the present study was to investigate the dynamic FC in adolescents with MDD, with a focus on the temporal properties of dynamic FC states and the variability of network topological organization. We hypothesized that (1) the dynamic FC state analysis would reveal abnormal temporal properties in first-episode drug-naïve adolescent MDD patients; and (2) the dynamic graph theory analysis would reveal altered network topology in first-episode drug-naïve adolescent MDD patients.
Methods
Participants
This study was approved by the Research Ethics Committee of West China Hospital, Sichuan University. Written informed consent was obtained from all participants and their parents/legal guardians. One hundred and ten adolescent MDD patients aged 12 to 18 years were recruited at West China Hospital of Sichuan University from January 2018 to January 2021. Eighty-seven healthy adolescents aged 12 to 18 years were recruited from the local community via advertisement. Diagnosis of depression was determined by two experienced clinical psychiatrists according to the criteria of the DSM-V. The severity of depression was rated using the 17-item Hamilton Rating Scale for Depression (HAMD). Adolescents with MDD were eligible if they were having their first episode of depression with HAMD total score ⩾ 18, and had not received any psychotropic medication treatment. Healthy adolescents were eligible if they and their first-degree relatives had no current or past MDD diagnoses. Exclusion criteria for both groups included: (1) history of neurological diseases or head injury; (2) history of chronic medical conditions; (3) history of developmental disorders or other psychiatric disorders; and (4) the presence of contraindications for MRI scans. Ten adolescent MDD patients were excluded due to a history of head injury (n = 1), a history of anxiety disorder (n = 7), a history of substance use disorder (n = 1), and the presence of MRI contraindications (n = 1). Two healthy adolescents were excluded due to MRI contraindications.
MRI data acquisition
MR imaging was performed on a 3.0-T Discovery MR750 scanner (GE Healthcare, Milwaukee, WI). Functional images and high-resolution T 1-weighted anatomical images were obtained from all subjects (see online Supplement for details).
Image preprocessing and head motion control
Data preprocessing was conducted using the DPABI v4.1 toolbox (http://rfmri.org/dpabi) (Yan, Wang, Zuo, & Zang, Reference Yan, Wang, Zuo and Zang2016). Preprocessing steps mainly included: (1) removal of the first 10 volumes; (2) slice-timing and realignment; (3) spatial normalization; and (4) spatial smoothing. Stringent inclusion criteria were adopted to minimize the potential effects of head motion on dynamic FC (see online Supplement for details).
Group ICA and identification of RSNs
A detailed overview of the framework is summarized in online Supplementary Fig. S1. To decompose the rs-fMRI data into different independent components (ICs), spatial group independent component analysis (ICA) implemented in the GIFT v4.0b software (http://mialab.mrn.org/software/gift) was performed. ICA included the following steps: (1) a two-step principal component analysis was used to decompose the data into 100 ICs; and (2) ICA decomposition was performed using the Infomax algorithm (Bell & Sejnowski, Reference Bell and Sejnowski1995), and this step was repeated 20 times in ICASSO (Himberg, Hyvärinen, & Esposito, Reference Himberg, Hyvärinen and Esposito2004) to obtain a stable and reliable set of components. Subsequently, the spatial maps and corresponding time courses were generated for each IC; and (3) a ICA back reconstruction algorithm (Calhoun, Adali, Pearlson, & Pekar, Reference Calhoun, Adali, Pearlson and Pekar2001) was used to back-project the group ICs. This step can reconstruct subject-specific spatial maps and time courses for each IC.
Of the decomposed 100 ICs, 53 were identified as meaningful, based on the following criteria (Allen et al., Reference Allen, Damaraju, Plis, Erhardt, Eichele and Calhoun2014; Cordes et al., Reference Cordes, Haughton, Arfanakis, Wendt, Turski, Moritz and Meyerand2000; Kim et al., Reference Kim, Criaud, Cho, Díez-Cirarda, Mihaescu, Coakeley and Strafella2017): (1) peak activations of spatial maps located in gray matter; (2) low spatial overlap with known vascular, ventricular, motion, and susceptibility artifacts; (3) time courses dominated by low frequency fluctuations and characterized by a high dynamic range. Based on a sorting function of spatial regression, as well as the similarities to the intrinsic connectivity networks described in previous studies (Allen et al., Reference Allen, Damaraju, Plis, Erhardt, Eichele and Calhoun2014; Tu et al., Reference Tu, Fu, Zeng, Maleki, Lan, Li and Kong2019, Reference Tu, Fu, Mao, Falahpour, Gollub, Park and Kong2020; Zhi et al., Reference Zhi, Calhoun, Lv, Ma, Ke, Fu and Sui2018), the 53 ICs were sorted into seven functional networks: sensorimotor network (SMN), visual network (VN), auditory network (AN), default mode network (DMN), cognitive-control network (CCN), cerebellar network (CB), and subcortical network (SC) (Fig. 1).
The 53 independent components (ICs) identified by group independent component analysis. (a) ICs were sorted into seven functional domains based on their anatomical and functional properties: sensorimotor (SMN), visual (VN), auditory (AN), default mode (DMN), cognitive control (CCN), cerebellar (CB), and subcortical (SC) networks. The spatial maps of each domain were overlaid on a standard template. (b) Group averaged static functional connectivity (FC) between IC pairs was computed using the entire resting state data. Color bar of the static FC matrix represents the correlation z values. Each of the 53 ICs was rearranged by network group based on the seven functional networks.
Additional postprocessing was performed on time courses of the selected 53 ICs to regress out the remaining noise sources. Postprocessing steps included: (1) detrending linear, quadratic, and cubic trends; (2) 3D-despiking; (3) regression of the six parameters of head movement; and (4) low-pass filtering with a high frequency cut-off of 0.15 Hz.
Computation for dynamic FC
A sliding window approach was used to detect time-varying changes of FC within the 53 IC networks during the whole rs-fMRI scans. The entire time courses of rs-fMRI data were segmented into 148 rectangle windows with a window length of 22 TRs (i.e. 44 s) and a step size of 1 TR. Then, these rectangle windows were convolved with a Gaussian (σ = 3 TRs) function. This step generated a series of tapered windows, which were employed to calculate covariance values. The window length of 22 TRs was selected based on a previous study (Preti et al., Reference Preti, Bolton and Van De Ville2017), which suggested that window sizes between 30 and 60 s were able to successfully capture resting-state dynamic FC fluctuations. In each sliding window, we calculated the covariance value of each pair of ICs, resulting in a 53 × 53 pairwise covariance matrix. Furthermore, to promote sparsity in estimation, the L1 norm penalty was implemented in the graphical LASSO framework (repeated 100 times) (Friedman, Hastie, & Tibshirani, Reference Friedman, Hastie and Tibshirani2008). A Fisher's r-to-z transformation was performed to convert values in the resulting 53 × 53 pairwise FC matrices into z scores. To control for the effect of possible covariates, the converted z scores were residualized with age, gender, and mean framewise displacement (FD) using multiple linear regression.
Dynamic FC state analysis
To assess reoccurring FC states, k means clustering with Manhattan distance function was performed on all windowed 53 × 53 FC matrices for all subjects and was repeated 500 times. The optimal number of clusters was determined to be three (k = 3) using the Dunn's index (online Supplementary Fig. S2), with each cluster representing a dynamic FC state. To visualize and compare dynamic FC patterns of the two groups, the averaged values of all subject-specific centroids assigned to each state were calculated for both adolescent MDD patients and healthy controls (HCs), respectively, thus generating group-specific centroids of all states (online Supplementary Fig. S3).
Three state transition metrics were calculated to quantify the temporal properties of dynamic FC states (Kim et al., Reference Kim, Criaud, Cho, Díez-Cirarda, Mihaescu, Coakeley and Strafella2017): (1) fractional window, which represents the proportion of time spent in each state; (2) mean dwell time, which represents how long the participant stayed in a certain state; and (3) number of transitions, which represents how many times either state changed from one to the other. Additional validation analyses were performed to test the consistency and reliability of our results using different window sizes (see online Supplement for details).
We also described the meanings of the three dynamic FC states from two aspects (Wu et al., Reference Wu, He, Shi, Xia, Zuang, Feng and Qiu2019b): the level of global integration and the connection characteristics of each functional network (see online Supplement for details).
Dynamic graph theoretical analysis
A graph theoretical analysis was used to evaluate the variance in topological organization of the FC, wherein we defined those selected 53 ICs as nodes and the correlation coefficients between each pair of ICs as edges. For each subject, all 148 windowed 53 × 53 FC matrices were binarized under a series of sparsity (S) thresholds (0.12 ⩽ S ⩽ 0.38, with an interval of 0.01) to generate adjacency matrices. This threshold range was selected to ensure that the thresholded networks were estimable for small-worldness (Zhang et al., Reference Zhang, Wang, Wu, Kuang, Huang, He and Gong2011) (online Supplementary Fig. S4).
At each sparsity threshold, we calculated both global and nodal topological metrics. The global metrics included: (1) network efficiency (global efficiency [Eglob] and local efficiency [Eloc]); and (2) small-world metrics (clustering coefficient [Cp], characteristic path length [Lp], normalized clustering coefficient [γ], normalized characteristic path length [λ], and small-worldness [σ]). The nodal metrics included nodal degree, nodal efficiency, and nodal betweenness. A brief interpretation of these graph metrics is shown in online Supplementary Table S1. We then calculated the area under the curve (AUC) for each network metric within the whole range of sparsity thresholds. This strategy has been widely used in previous studies (Wu et al., Reference Wu, Li, Zhang, Zhang, Long, Gong and Jia2020a, Reference Wu, Li, Zhou, Zhang and Long2020b; Yue, Wang, Li, Ren, & Wu, Reference Yue, Wang, Li, Ren and Wu2021). Finally, to examine dynamic graph properties of FC networks, we calculated the variance of AUC of both global and nodal metrics across all sliding windows as suggested by previous study (Yu et al., Reference Yu, Erhardt, Sui, Du, He, Hjelm and Calhoun2015).
Statistical analysis
Between-group differences in demographic and clinical data were analyzed using χ2 test and independent two-sample t test (two-tailed). Between-group differences in temporal properties and dynamic graph metrics were determined using nonparametric permutation tests (10 000 iterations). Partial correlation analyses were used to estimate relationships between altered dynamic FC metrics and clinical variables in adolescent MDD. The false discovery rate (FDR) was used to correct for multiple comparisons (p < 0.05, FDR corrected). Age, gender, and mean FD were set as nuisance covariates.
Exploratory multivariate pattern analysis
To assess whether dynamic FC can detect adolescent MDD at the individual level, exploratory multivariate pattern analysis was performed in accordance with a previous study (Liu et al., Reference Liu, Guo, Fouche, Wang, Wang, Ding and Chen2015) (see online Supplement for details).
Results
Demographic and clinical characteristics
Ninety-four adolescent MDD patients and 78 HCs were finally included. There were no significant differences in age, gender, education level and body max index between the two groups (all p > 0.05). Detailed demographic and clinical characteristics of the two groups are shown in Table 1.
Demographic and clinical characteristics of the participants
MDD, major depressive disorder; BMI, body mass index; HAMD, Hamilton Depression Rating Scale.
a The p value was calculated by using χ2 test.
b The p value was calculated by using independent two-samples t test.
Note: All quantitative data are expressed as mean ± standard deviation; numbers for gender data.
Intrinsic FC networks
Figure 1 shows spatial maps of all 53 ICs and group-averaged static FC matrix. The 53 ICs were organized into to seven networks: SMN, VN, AN, DMN, CCN, CB, and SC. The detailed information and spatial maps of all 53 ICs are provided in online Supplementary Table S2 and Figs S5–S11.
Dynamic FC state analysis
Dynamic FC states and characteristics
Three dynamic FC states were identified (Fig. 2a). State 1 occurred most frequently (17 728 times, 70%), and both State 2 (3887 times, 15%) and State 3 (3841 times, 15%) occurred at a low frequency. Figure 2b shows the top 10% of the strongest connections. In State 1, the strongest connections were located mainly within networks with positive correlations, and between the DMN and CCN with negative correlations. In State 2 and State 3, the strongest connections were located mainly within or across the SMN, VN, and AN.
Cluster centroids and characteristics of dynamic functional connectivity (FC) states across all subjects (window length = 22 TR). (a) Cluster centroids for each state and the corresponding total number of occurrences and percentage of total occurrences (shown at the top right of the centroid matrices). (b) Circular graphs show the top 10% strongest connections (i.e. the top 10% maximum correlation coefficients [absolute value]) of the FC matrix in each state. Each square color represents one of the seven networks. (c) Modular organization of the centroid matrix for each state. Each ball color represents a specific module. The results were visualized using the BrainNet Viewer package. (http://www.nitrc.org/projects/bnv). (d) Radar and line maps show mean FC strength within and between networks for all three states. SMN, sensorimotor network; VN, visual network; AN, auditory network; DMN, default mode network; CCN, cognitive control network; CB, cerebellar network; SC, subcortical network.
The three dynamic FC states showed different characteristics in modular organization (Fig. 2c). Specifically, the centroid matrix of State 3 has the highest modularity measure (Q = 0.552). In State 3, the 53 ICs were divided into two functional modules: module 1 mainly involved ICs in SMN, VN and AN, and module 2 mainly involved ICs in DMN, CCN, CB, and SC. The centroid matrix of State 2 has the lowest modularity measure (Q = 0.077) and showed a similar modular distribution to the centroid matrix of State 3. In State 2, all 53 ICs were also divided into two functional modules: a module 1 mainly involving ICs in SMN, VN, and AN, and a module 2 mainly involving ICs in DMN, CCN, CB and SC. The centroid matrix of State 1 showed a moderate modularity measure. In this state, all 53 ICs were divided into three modules: module 1 was mainly composed of ICs in DMN and CCN, module 2 was mainly composed of ICs in SMN, AN, and CCN, and module 3 was mainly composed of ICs in SMN and VN.
The centroid matrices of the three dynamic FC states also showed different patterns with respect to the mean FC strength within and between the seven intrinsic brain networks (Fig. 2d and online Supplementary Table S3). In each state, the mean FC strength within networks was higher than that between networks. In general, the mean within-network FC strength was the lowest in State 1, the highest in State 2, and the moderate in State 3. In addition, the mean FC strength between a network and all other networks was the strongest in State 2, but weaker in State 1 and State 3.
Temporal properties of dynamic FC states
Adolescent MDD patients, relative to controls, showed significantly higher fractional windows (p = 0.020) and longer mean dwell time (p = 0.026) in State 1, and exhibited significantly lower fractional windows (p = 0.022) and shorter mean dwell time (p = 0.020) in State 2 (Fig. 3). Additional validation analyses revealed that the main results remain unchanged (online Supplementary Tables S4–S6 and Figs S12, S13).
Group comparison of temporal properties of dynamic functional connectivity states. The fractional windows (a), mean dwell time (b) and number of transitions (c) are shown using violin plots for the adolescent major depressive disorder (MDD) group (orange) and healthy control (HC) group (green). Horizontal lines indicate group medians (solid lines) and upper and lower quartiles (dashed lines). Asterisks represent p < 0.05.
FC strength of dynamic FC states
Notably, not all subjects had the windowed matrix assigned to each state; thus, only the subject-specific matrices assigned to that state were used to determine between-group difference in FC strength. In State 1, significantly decreased FC strength were observed within the AN and VN and between the SMN-VN, SMN-DMN, SMN-CCN, and AN-CCN (p = 0.010, network-based statistic corrected) in adolescent MDD patients compared to controls (online Supplementary Fig. S14). No significant between-group difference in FC strength was observed in State 2 and State 3.
Dynamic topological metrics
Adolescent MDD patients showed significantly higher variance in Eglob (p = 0.002) and Eloc (p = 0.011) than HCs (Fig. 4). There were no significant differences in variance of any other network metrics between the two groups.
Group comparison of variance in global topological metrics. The variance of the global efficiency (Eglob) (a), local efficiency (Eloc) (b), clustering coefficient (Cp) (c), characteristic path length (Lp) (d), normalized clustering coefficient (γ) (e), normalized characteristic path length (λ) (f) and small-worldness (σ) (g) across functional connectivity matrices are shown using violin plots for the adolescent major depressive disorder (MDD) group (orange) and healthy control (HC) group (green). Horizontal lines indicate group medians (solid lines) and upper and lower quartiles (dashed lines). Asterisks represent p < 0.05.
Relationship with clinical variables
For depressed adolescents, no significant correlations were observed between altered dynamic FC properties and clinical variables.
Single-subject classification of adolescent MDD patients and HCs
Using LOOCV strategy, the linear SVM classifier achieved an accuracy of 68.18% (sensitivity = 71.77%, specificity = 63.77%, p = 0.006) for classification of adolescent MDD patients v. HCs by using the 175 highest ranked functional connections in State 1 (online Supplementary Table S7 and Fig. S15), and the ROC curve revealed an AUC value of 0.722 (Fig. 5). However, the classification accuracy was not significant using the optimal FC features in State 2 (p = 0.085) or State 3 (p = 0.229).
Receiver operating characteristic curve of the classifier. FC, functional connectivity.
In State 1, 135 consensus features were identified in the LOOCV, and these consensus functional connections were mainly located within or across the large-scale brain functional networks (online Supplementary Fig. S16).
Discussion
The present study investigated time-varying FC in adolescent MDD patients, with a focus on the temporal properties of dynamic FC states as well as the dynamic graph properties of FC networks. The main findings were: (1) adolescent MDD patients spent more time in the weakly-connected and relatively highly-modularized State 1, but less time in the strongly-connected and low-modularized State 2; (2) Disrupted FC among multiple networks were found in State 1, and FC in State 1 showed the best performance for adolescent MDD classification; and (3) a higher variability in the network efficiency (both global and local level) was observed in depressed adolescents, suggesting impaired local segregation and global integration of functional brain networks.
In our study, although the two groups shared similar dynamic FC states, adolescent MDD patients showed altered temporal properties compared to the HC group. Specifically, depressed adolescents, relative to controls, spent more time in State 1, in which all 53 ICs tended to be subdivided into three functional modules with a relatively high modularity index, representing a higher level of segregation. In contrast, for adolescent MDD patients, there was a significant lack of State 2, in which all 53 ICs tended to be subdivided into two functional modules, and the lowest modularity index in State 2 is suggestive of a higher level of global integration. The brain dynamically reconfigures its functional organization to support diverse cognitive task performances (Bassett & Bullmore, Reference Bassett and Bullmore2017). Successful reconfiguration underlying better task performance relies not only on sufficiently independent processing in specialized subsystems (i.e. segregation) but also on effective global cooperation between different subsystems (i.e. integration) (Bassett & Bullmore, Reference Bassett and Bullmore2017; Deco, Tononi, Boly, & Kringelbach, Reference Deco, Tononi, Boly and Kringelbach2015; Fair et al., Reference Fair, Dosenbach, Church, Cohen, Brahmbhatt, Miezin and Schlaggar2007; Shine, Reference Shine2019; Sporns, Reference Sporns2013). In healthy individuals, the resting brain's functional organization is configured to maintain a balance between network segregation and integration, and this segregation–integration balance empowers the brain to support diverse cognitive abilities (Wang et al., Reference Wang, Liu, Cheng, Wu, Hildebrandt and Zhou2021). Thus, our findings of increased time of functional segregation and decreased time of functional integration in adolescent MDD patients may reflect an abnormal dynamic network configuration. The optimal balance between network segregation and integration might be disrupted in depressed adolescents, which lead to impaired functional flexibility.
Additionally, for the weakly-connected State 1, disrupted FC predominantly involved the CCN and sensory networks (i.e. SMN, VN, and AN) was observed in adolescent MDD patients. The CCN consists mainly of fronto-parietal regions and is involved in top-down regulation of attention and working memory tasks (Cole & Schneider, Reference Cole and Schneider2007), while the sensory networks are involved in sensory perception and motor process which are responsible for information communication with external environment. Individual's perception can be affected by goal-directed decisions, which further results in the corresponding modulation of sensory cortical activity, and this process actually reflects the top-down control and allows individuals to flexibly navigate multiple streams of sensory information (Gazzaley, Cooney, McEvoy, Knight, & D'Esposito, Reference Gazzaley, Cooney, McEvoy, Knight and D'Esposito2005a; Gazzaley, Cooney, Rissman, & D'Esposito, Reference Gazzaley, Cooney, Rissman and D'Esposito2005b). Thus, abnormal interactions between the CCN and sensory networks might underlie the deficits of the high-order control over sensory process in depression. Although we found FC disruptions associated with the sensory-related regions, abnormal resting-state FC of sensory networks is not commonly reported in previous neuroimaging studies on both adolescent-onset and adult-onset depression. It is noteworthy that most of previous functional neuroimaging studies calculated FC using time courses within the entire rs-fMRI scan (i.e. static FC), and the FC abnormalities involved the sensory networks may occur over a relatively short timescale. Thus, the static FC analytical strategy may not be sensitive enough to detect FC alterations of the sensory regions. This also highlights the need for conducting dynamic FC analysis to discover additional important information that cannot be captured by static FC analysis. Interestingly, disrupted FC only occurred in State 1, and adolescent MDD patients spent more time in this state. Therefore, we speculate that the abnormal FC in adolescent MDD patients may be more related to the weakly-connected state, reflecting a state-specific FC dysfunction. Similarly, a recent study found that FC impairment caused by MDD in adults also mainly occurred in the weakly-connected state, and FC strength between the DMN and CCN and within the CCN in this state was correlated with symptom severity (Yao et al., Reference Yao, Shi, Zhang, Zheng, Hu, Li and Hu2019). Furthermore, additional SVM analysis revealed significant single-subject classification of adolescent MDD patients and controls with an accuracy of 68.18%. Additionally, some FC showing significant between-group differences overlapped with those identified consensus functional connections. Thus, our results suggest that dynamic FC, especially FC in the weakly-connected state, might have potential diagnostic value for adolescent MDD.
This study also revealed the instability of the dynamic FC in adolescent MDD, as characterized by the higher variability of both global and local efficiency. In general, dynamic FC based on rs-fMRI reflects fluctuations in the exchange of information between brain regions during resting state, which occurs in an organized sequential manner. Eglob and Eloc measure the ability of a network to transmit information at the global and local level, respectively. Thus, the increased variance in network efficiency in adolescent MDD patients means low and unstable information transmission efficiency within/between the functional networks, and further suggests impaired local segregation and global integration of functional brain networks in adolescents with MDD. Notably, previous studies that performed graph theoretical analysis on static whole-brain FC have demonstrated disrupted topological organization in first-episode drug-naïve adolescents (Wu et al., Reference Wu, Li, Zhou, Zhang and Long2020b) and adults (Zhang et al., Reference Zhang, Wang, Wu, Kuang, Huang, He and Gong2011) with MDD. Therefore, our findings are consistent with previous studies. However, a recent study found that compared with the HC group, the total number of transitions between states in MDD patients showed a decreasing trend, and the difference between groups was close to the level of significance (p = 0.070) (Yao et al., Reference Yao, Shi, Zhang, Zheng, Hu, Li and Hu2019). Their findings suggest that the stability of dynamic FC in MDD patients may not be impaired. Several factors may have contributed to these inconsistent results. First, most of the patients they included were treated with antidepressants, which may have an impact on the stability of FC. Second, all patient samples included in their study were adults other than adolescents; thus, our findings may reflect different pathophysiological mechanisms between adolescent-onset and adult-onset MDD. In fact, a previous study has observed distinct patterns of cortical abnormalities between adolescents and adults with MDD, suggesting that MDD may impact brain gray matter structure in a highly dynamic way (Schmaal et al., Reference Schmaal, Hibar, Sämann, Hall, Baune, Jahanshad and Veltman2017). Prior research has shown that resting-state functional network dynamics change over adolescent development (Hutchison & Morton, Reference Hutchison and Morton2015). Thus, developmental factors may contribute to distinct patterns of change in dynamic FC between adolescent and adult MDD.
A recent study also investigated dynamic FC properties in first-episode drug-naïve adolescents with MDD, and found similar results to our study (Zheng et al., Reference Zheng, Chen, Jiang, Zhou, Li, Wei and Cheng2022). Notably, they used a parcellation strategy (i.e. Dosenbach 160 atlas) to define the brain functional networks. The predefined template may affect the accurate delineation of functional network regions. Our study used ICA to extract the intrinsic brain functional networks, rather than a predefined atlas. The ICA is a data-driven approach to decompose rs-fMRI data into functionally homogeneous regions, and enables a whole-brain analysis without resorting to atlas-defined regions of interest that may merge or imprecisely delineate functionally distinct areas (Allen, Erhardt, Wei, Eichele, & Calhoun, Reference Allen, Erhardt, Wei, Eichele and Calhoun2012). Unlike their study, we did not find significant associations between clinical characteristics and altered dynamic FC properties after correction for multiple comparisons. These different findings might be attributed to sample size and methodology. Overall, our study further confirmed abnormal dynamic FC in adolescent MDD patients using a more rigorous methodological framework. However, the clinical relevance of abnormal dynamic FC properties needs to be further explored in future researches.
Limitations
Several limitations should be acknowledged in our study. First, this was a cross-sectional study with a relatively small sample size, which may affect the statistical power. Second, the MRI acquisition parameters need to be mentioned in dynamic FC analysis. Fine temporal resolution and a sufficient length of acquisition are both important factors for obtaining reliable results. Although previous study has successfully captured the dynamics of fluctuation using similar rs-fMRI parameters (i.e. TR = 2 s, total volumes = 180) (Tu et al., Reference Tu, Fu, Zeng, Maleki, Lan, Li and Kong2019), an optimal fast MRI acquisition, such as simultaneous multi-slice acquisition (Moeller et al., Reference Moeller, Yacoub, Olman, Auerbach, Strupp, Harel and Uğurbil2010), is needed in future studies. Third, we did not perform a longitudinal study to observe the adolescent MDD-related alterations of dynamic FC over time, thus restricting our direct understanding of dynamic FC changes as the disease progresses. Fourth, in machine learning analysis, the SVM model was only internally validated using a LOOCV strategy, the performance of the model should be validated externally in an independent dataset in the future. Fifth, we did not perform a priori sample size calculations and power analyses, which may affect the statistical results of our study. Finally, the study sample size was not large enough to examine developmental effects. Prior research has shown that resting network dynamics change over adolescent development (Hutchison & Morton, Reference Hutchison and Morton2015). Future longitudinal researches that target the developmental effects on brain network dynamics may provide new insight into network abnormalities in adolescent MDD.
Conclusions
In summary, our study suggests abnormal dynamic functional network configuration in adolescent MDD, as characterized by impaired local segregation and global integration of the functional networks as well as disturbed segregation-integration balance. Furthermore, dynamic FC, especially FC in the weakly-connected state, may act as potential and candidate biomarkers for the diagnosis of adolescent MDD. Thus, the dynamic FC abnormalities observed in this study may provide new perspectives for understanding the neurobiology of adolescent MDD.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291723003719
Author's contributions
QG and ZJ conceptualized the project. QG, ZJ, and BW designed the study. BW, XL, YC, HX, and XW contributed to literature search, data collection, data analysis and data interpretation. BW drafted the manuscript. QG, ZJ, and NR critically revised the manuscript. All authors approved the final version of the manuscript.
Funding statement
This study was supported by the National Natural Science Foundation of China (82271947, 81971595, 81820108018, and 81621003), the National Key R&D Program of China (2022YFC2009900), the Key Program of Natural Science Foundation of Sichuan Province (2022NSFSC0047), the 1.3.5 Project for Disciplines of Excellence–Clinical Research Incubation Project, West China Hospital, Sichuan University (2020HXFH005), and the CAMS Innovation Fund for Medical Sciences (CIFMS) (2022-I2M-C&T-B-104).
Competing interests
None.
