Aberrant functional connectivity for diagnosis of major depressive disorder: A discriminant analysis


  • Longlong Cao MD,

    1. Mental Health Institute of The Second Xiangya Hospital, Hunan Province Technology Institute of Psychiatry, Key Laboratory of Psychiatry and Mental Health of Hunan Province, Central South University, Hunan, China
    Search for more papers by this author
  • Shuixia Guo PhD,

    1. Mathematics and Computer Science College, Hunan Normal University, Hunan, China
    Search for more papers by this author
  • Zhimin Xue MD, PhD,

    1. Mental Health Institute of The Second Xiangya Hospital, Hunan Province Technology Institute of Psychiatry, Key Laboratory of Psychiatry and Mental Health of Hunan Province, Central South University, Hunan, China
    Search for more papers by this author
  • Yong Hu MSc,

    1. Mathematics and Computer Science College, Hunan Normal University, Hunan, China
    Search for more papers by this author
  • Haihong Liu MD,

    1. Mental Health Center, Xiangya Hospital, Central South University, Hunan, China
    Search for more papers by this author
  • Tumbwene E. Mwansisya MSc,

    1. Mental Health Institute of The Second Xiangya Hospital, Hunan Province Technology Institute of Psychiatry, Key Laboratory of Psychiatry and Mental Health of Hunan Province, Central South University, Hunan, China
    Search for more papers by this author
  • Weidan Pu MD,

    1. Mental Health Institute of The Second Xiangya Hospital, Hunan Province Technology Institute of Psychiatry, Key Laboratory of Psychiatry and Mental Health of Hunan Province, Central South University, Hunan, China
    Search for more papers by this author
  • Bo Yang PhD,

    1. Mental Health Institute of The Second Xiangya Hospital, Hunan Province Technology Institute of Psychiatry, Key Laboratory of Psychiatry and Mental Health of Hunan Province, Central South University, Hunan, China
    Search for more papers by this author
  • Chang Liu MM,

    1. Mental Health Institute of The Second Xiangya Hospital, Hunan Province Technology Institute of Psychiatry, Key Laboratory of Psychiatry and Mental Health of Hunan Province, Central South University, Hunan, China
    Search for more papers by this author
  • Jianfeng Feng PhD,

    1. Centre for Computational Systems Biology, School of Mathematical Sciences, Fudan University, Shanghai, China
    2. Department of Computer Science, University of Warwick, Coventry, UK
    Search for more papers by this author
  • Eric Y. H. Chen MD,

    1. Department of Psychiatry, LKS Faculty of Medicine, The University of Hong Kong, Hong Kong, China
    Search for more papers by this author
  • Zhening Liu MD, PhD

    Corresponding author
    1. Mental Health Institute of The Second Xiangya Hospital, Hunan Province Technology Institute of Psychiatry, Key Laboratory of Psychiatry and Mental Health of Hunan Province, Central South University, Hunan, China
    • Correspondence: Zhening Liu, MD, PhD, Mental Health Institute of The Second Xiangya Hospital, Hunan Province Technology Institute of Psychiatry, Key Laboratory of Psychiatry and Mental Health of Hunan Province, Central South University, Changsha, Hunan 410011, China. Email: zningl@163.com

    Search for more papers by this author



Aberrant brain functional connectivity patterns have been reported in major depressive disorder (MDD). It is unknown whether they can be used in discriminant analysis for diagnosis of MDD. In the present study we examined the efficiency of discriminant analysis of MDD by individualized computer-assisted diagnosis.


Based on resting-state functional magnetic resonance imaging data, a new approach was adopted to investigate functional connectivity changes in 39 MDD patients and 37 well-matched healthy controls. By using the proposed feature selection method, we identified significant altered functional connections in patients. They were subsequently applied to our analysis as discriminant features using a support vector machine classification method. Furthermore, the relative contribution of functional connectivity was estimated.


After subset selection of high-dimension features, the support vector machine classifier reachedup to approximately 84% with leave-one-out training during the discrimination process. Through summarizing the classification contribution of functional connectivities, we obtained four obvious contribution modules: inferior orbitofrontal module, supramarginal gyrus module, inferior parietal lobule-posterior cingulated gyrus module and middle temporal gyrus-inferior temporal gyrus module.


The experimental results demonstrated that the proposed method is effective in discriminating MDD patients from healthy controls. Functional connectivities might be useful as new biomarkers to assist clinicians in computer auxiliary diagnosis of MDD.

Evidence shows that major depressive disorder (MDD) patients at resting-state brain connectivities are aberrant compared with healthy controls (HC). Abnormal resting-state functional connectivities of distributed brain networks are believed to contribute to the MDD illness process.[1, 2] Abnormal resting-state functional connectivity might relate to affect dysregulations and cognitive deficits, and both are associated with the symptomatology of this disorder.[3] Antidepressant treatment has been reported to be associated with a normalization of abnormal functional connectivity mediated reward and emotional processing.[4-6] Briefly, abnormal functional connectivities observed in depression are closely associated with symptomatology and treatment response.

Although brain functional connectivity has been demonstrated, there is a drastic discrimination performance between depression patients and HC at the group level.[3, 7] To date, few studies used resting-state functional magnetic resonance imaging (fMRI) functional connectivity to discriminate MDD at an individual level. Additionally, previous research focused on selected regions of interest rather than examining functional connectivity patterns on a whole-brain scale. Furthermore, previous researches concentrated mostly on classification methods instead of features selection. Indeed, the qualities of selected features directly determine the accuracy of the classifier, so it is imminent to elevate feature selection qualities.

Traditionally, a depression diagnosis is based on a categorical taxonomy derived from responses to questionnaires about subjective experiences. Functional connectivity analysis has been used to study the neural substrates of depression, symptomatic changes and antidepressant responses.[4, 8, 9] This study further investigated the potential use of functional connectivity measures as objective biomarkers to provide critically needed assistance to clinicians.

In this research, we propose an effective classification method which automatically identifies discrepant functional connectivities and combines them into classification processes. The key steps include the following: (i) feature selection is conducted among all functional connectivities and significant discrepancy functional connectivities are picked out as discriminant features; (ii) the support vector machine (SVM) method of pattern recognition is adopted to discriminate MDD patients from HC; (iii) the performance of the new method is estimated by cross-validation measures; and (iv) we summarize and analyze the contribution of functional connectivities network on a modular perspective.

This study tried to optimally combine discriminative resting-state functional connectivities as biomarkers[10] in order to achieve more accurate diagnosis of MDD.[11] This may provide new perspectives to understanding the pathophysiology of MDD.


Subjects and preprocessing

Thirty-nine MDD patients (23 women, 16 men, mean age 27.99 ± 7.49 years old; mean education 12.00 ± 3.53 years) were recruited from the Second Xiangya Hospital of Central South University. The mean illness duration for MDD patients was 23.27 ± 37.82 months, and the mean score on the 17-item version of the Hamilton Rating Scale for Depression (HAMD) was 24.97 ± 4.99. Thirty-seven healthy controls (14 women, 23 men; mean age 28.22 ± 6.47 years old; mean education 13.32 ± 3.29 years) were also recruited. The two groups are well matched by sex (χ2-test = 4.873, P = 0.087), age (t = 0.266, P = 0.791) and education level (t = 1.458, P = 0.150).

All MDD patients were 18–45 years old, right-handed Han Chinese, with HAMD scores of at least 17. Patients and HC were excluded if they had any of the following criteria: a history of electroconvulsive therapy, substance abuse, neurological or other serious physical diseases and any contraindications for MRI. This study was approved by the Ethics Committee of the Second Xiangya Hospital.

All image data were acquired using a 1.5T Siemens GE Signa Twinspeed Scanner (General Electric Medical System, Milwaukee, WI, USA). A total of 180 volumes of echo planar images were obtained axially (repetition time = 2000 ms, echo time = 40 ms, slices = 20, thickness = 5 mm, gap = 1 mm, field of view = 24 × 24 cm2, resolution = 64 × 64, flip angle = 90°). Subjects kept their eyes closed until the end of the scanning process.

All the preprocessing was conducted using Statistical Parametric Mapping version 8 (http://www.fil.ion.ucl.ac.uk/spm/) and Data Processing Assistant for Resting-State fMRI. Images of the first 10 volumes were discarded to enable stabilization of the scanner and subjects to adapt to the environment. In brief, the remaining 170 functional scans were first corrected for within-scan acquisition time differences between slices, then realigned to the middle volume to correct for inter-scan head motions. Subsequently, the fMRI images were further spatially normalized to a standard template (Montreal Neurological Institute) and re-sampled to 3 × 3 × 3 mm3. After normalization, the blood oxygenation level-dependent (BOLD) signal of each voxel was first detrended to abandon linear trend and then temporal band-pass filtering (0.01 < f < 0.08 Hz) was performed in order to reduce low-frequency drift and high-frequency physiological noise. At last, nuisance covariates, including six head motion parameters, global mean signaling, cerebrospinal fluid signals and white matter signals, were regressed out from the BOLD signals.

To construct whole-brain functional networks, an automated anatomical labeling (AAL) atlas was employed to parcellate the brain into 90 regions of interest (ROI). The names of the ROI and their corresponding abbreviations are listed in Table 1. For each subject, the representative time series of each individual region was then obtained by simply averaging the fMRI time series over all voxels in this region. Finally, a 170 × 90 × 76 3D-times series matrix was acquired.

Table 1. Names and abbreviations of the regions of interest
Prefrontal lobe (PreF)Temporal lobe & medial temporal system (Tem)
Superior frontal gyrus, dorsolateral, SFGdorSuperior temporal gyrus, STG
Superior frontal gyrus, orbital, ORBsupTemporal pole: superior, TPOsup
Superior frontal gyrus, medial, SFGmedMiddle temporal gyrus, MTG
Superior frontal gyrus, medial orbital, ORBsupmedTemporal pole: middle, TPOmid
Middle frontal gyrus, MFGInferior temporal gyrus, ITG
Middle frontal gyrus, orbital, ORBmidHeschl gyrus, HES
Inferior frontal gyrus, opercular, IFGopercFusiform gyrus, FFG
Inferior frontal gyrus, triangular, IFGtriangHippocampus, HIP
Inferior frontal gyrus, orbital, ORBinfParahippocampal gyrus, PHG
Olfactory cortex, OLFAmygdala, AMYG
Gyrus rectus, RECParietal lobe (Par)
Anterior cingulated, ACGPostcentral gyrus, PoCG
Other parts of frontal lobe (OthF)Superior parietal lobule, SPG
Precentral gyrus, PreCGInferior parietal lobule, IPL
Supplementary motor area, SMASupramarginal gyrus, SMG
Median cingulated, DCGAngular gyrus, ANG
Rolandic operculum, ROLPrecuneus, PCUN
Insula (Ins)Paracentral lobule, PCL
Occipital lobe (Occ)Posterior cingulate gyrus, PCG
Calcarine fissure, CALCorpus striatum (CS)
Cuneus, CUNCaudate nucleus, CAU
Lingual gyrus, LINGPutamen, PUT
Superior occipital gyrus, SOGPallidum, PAL
Middle occipital gyrus, MOGThalamus (Tha)
Inferior occipital gyrus, IOG 

Features and feature selection

Most researches about functional connectivity simply focus on describing the group differences,[12] but this does not help classification at an individual level. Because functional connectivity can directly indicate interaction within brain areas and is easy to calculate, we adopted it as a classification feature. The consequent problem is a high-dimensional issue. Accordingly, it is necessary to reduce feature dimension (up to 4005). The selected features are not rendered robust by simply thresholding a small significance level of P-value,[13] because functional connectivity is very sensitive to noise, which is ineluctable even after filtering. In this paper, a holistic method was proposed for feature selection mainly based on probability density function (PDF-FS).

The basic flow of PDF-FS is as follows. First, the most significantly different functional connectivities between two groups are primitively selected by using t-test to ensure that major redundancy features are rejected. Second, for each primary selected functional connectivity, the probability density function of depressions and HC are estimated, respectively. We carry that out by using the non-parametric estimation method named Kernel density estimation based on the Gauss kernel,[14] which does not require any knowledge about the density of the data. Finally, we sort these functional connectivities in ascending order of overlap ratio (OR) and take the preceding K functional connectivities as the final feature subset. OR value of functional connectivity is calculated according to the following formula:

display math

where fi(x)(i = 1,2) are the corresponding PDF of functional connectivity of two groups. OR measures the area-two intersecting PDF curve. The smaller the OR is, the greater the difference is between the two groups. As Figure 1a,b shows, the PDF of two links with different OR are plotted (inferior frontal gyrus, orbital, right – pallidum, left [ORBinf.R-PAL.L], OR = 0.6299 vs caudate nucleus, right – inferior temporal gyrus, right [CAU.R-ITG.R], OR = 0.9854), and it is easy to see that the link between ORBinf.R-PAL.L is more distinguishing than the link between CAU.R-ITG.R.

Figure 1.

(a,b) Overlap rate of probability density function. Gray plus marks represent major depressive disorder (MDD) patients' functional connectivity values and the gray curve represents corresponding probability density function; on the converse, the black styles represent healthy controls. The larger the two intersecting curves surrounding the area, the smaller the difference between the two groups and the more difficult it is to distinguish them. The larger the differences among the different types of probability density function, the more easily they are distinguished. (image) Depression; (image) normal. (c,d) Classification accuracy, specificity and sensitivity of support vector machine (SVM). (c) T-test feature selection method, accuracy, specificity and sensitivity fluctuated with significance levels. (image) Accuracy; (image) Sensitivity; (image) Specificity. (d) Probability density function (PDF-FS) feature selection method, accuracy, specificity and sensitivity associated with selected numbers of features K. Accuracy is highest when K = 31 or 35. (image) Accuracy; (image) Sensitivity; (image) Specificity.


SVM is extensively applied into functional imaging data[13, 15-17] by constructing the maximum margin hyperplane, which finds the single hyperplane with the maximum distance between the plane and the points closest to the plane (see supporting information). Generally, SVM involves two stages: training and predication. During the training phase, the classifier is trained from the training data. Once the classifier is determined, it can be used to predict the class label (patient/health) when a new test sample is fed into the classifier.

The SVM procedure in this paper uses a toolkit named libsvm, written by Dr Lin Chih-Jen from Taiwan University (http://www.csie.ntu.edu.tw/~cjlin/libsvm/). Using a radial basis function as a kernel function (t = 2), parameter C was fixed to 10, which was used to trade off learning and extend ability, and other parameters were kept as default values.

Because of the small sample, we assessed classification machine efficiency using the leave-one-out cross validation (LOOCV) test. In each leave-one-out run, we used data from all but one of the 76 subjects to train the classifier. In the LOOCV, we selected one subject from all in turn but not at random.

Subsequently, the class assignment of the remaining subjects (one from each group), who so far had not seen the algorithm, was calculated during the test or application phase. All features of the training samples were sequentially screened by t-test and PDF-FS methods. The classifier, which was trained on the feature subset of the training set, was applied to test set corresponding features to predict label of test samples. All the prediction-accuracy rates (S = 76 times) on average were the actual estimated accuracy of the classification method.

Contribution network

Features in SVM which possess high weights make a great contribution during the classification process. Based on this principle, we normalized the square of each weight vector in each leave-one-out process (see supporting information); for example, the total contribution value of each classification is 1. All the feature weights in the classification were summed up and averaged over loop times. The mean value was used to show the contribution of each functional connectivity to the classification (contribution degree). At last, a connected network of functional connectivity was obtained, which consisted of the functional connectivities that appeared in the process of classification and their weights. In order to display universal existence of selected features, only the functional connectivities which appeared in the leave-one-out process more than half of the times were chosen, and their weights (e.g. average contribution degree) could indicate the significant different features between MDD subjects and HC.


Overlap ratio

When we performed t-test analysis on all features under significance level α = 0.002, there were 50 links left (see Table 2). OR value determined the sequence of the feature selected, but that did not mean those features used in classification would make a great contribution in the whole subset. In fact, the preceding 29 functional connectivities in Table 2 were identified coincidentally exactly with our contribution network, although not exactly in the sequence of contribution degree. Two functional connectivities could reveal this phenomenon well: (amygdala, left – precentral gyrus, right [AMYG.L- PreCG.R]), OR value was 0.6182, but the whole contribution degree was only 0.04. On the converse (supramarginal gyrus, left – fusiform gyrus, left [SMG.L-FFG.L]) OR value was only 0.7125, but the final contribution degree was 7.31.

Table 2. 50 links that owned the superior discriminative ability
AMYG.L – PreCG.R0.61820.0003SMG.L – FFG.L0.71250.0001
PAL.L – ORBinf.R0.62990.0000IPL.L – PCG.L0.71790.0002
SMG.R – FFG.L0.63720.0007ITG.R – MTG.L0.71980.0017
PUT.L – ORBinf.R0.63770.0000THA.L – PoCG.L0.72140.0011
CAU.L – MOG.R0.64530.0003AMYG.L – PreCG.L0.72710.0009
THA.R – SMG.L0.65130.0010INS.L – ORBinf.R0.72790.0001
THA.L – SMG.L0.65570.0001PAL.R – PoCG.L0.73000.0018
PUT.R – ORBinf.R0.66550.0000PoCG.L – INS.L0.73140.0006
HIP.L – PreCG.L0.66900.0002AMYG.L – SMA.L0.73640.0012
HIP.L – ORBinf.R0.67010.0006THA.L – PreCG.R0.73660.0016
IPL.R – PCG.R0.67110.0004SPG.L – PCG.R0.74120.0016
IPL.L – PCG.R0.67260.0000THA.R – PoCG.L0.74280.0010
THA.R – SMA.L0.68090.0009CAU.L – PoCG.L0.74310.0018
PUT.L – INS.R0.68230.0014CAL.R – MFG.R0.74390.0012
TPOmid.L – MFG.L0.68240.0014THA.R – PreCG.R0.74580.0011
CAU.R – PreCG.R0.68430.0015HES.L – MFG.R0.74790.0008
CAU.L – FFG.R0.69180.0010MTG.L – CAL.L0.74950.0010
INS.R – ORBinf.R0.69360.0002ROL.R – ORBinf.R0.74990.0011
PAL.R – ORBinf.R0.69470.0002THA.L – IPL.L0.75510.0016
PCUN.L – SFGdor.L0.69900.0015ROL.L – ORBinf.R0.75630.0009
STG.L – ORBinf.R0.70160.0007THA.L – SMA.R0.75750.0017
ITG.L – MTG.L0.70360.0001THA.R – ORBinf.R0.76820.0009
THA.L – ORBinf.R0.70540.0001PoCG.L – ROL.L0.76930.0008
SPG.L – HIP.L0.70660.0002THA.L – SMA.L0.77060.0014
PAL.L – SMA.L0.70700.0020CAU.L – SMA.L0.77810.0017

Classification of diagnosis

When we used the t-test feature selection method (see Fig. 1c, P ≤ 0.010), under significance level α = 0.003, the highest classification accuracy of SVM was 78.95% in discriminating results of MDD patients and HC. As the chart revealed: different significance levels could lead to different selected features and discrepancy classification accuracy. As long as feature numbers K surpassed limits, the accuracy declined sharply.

As shown in Figure 1c, the best preliminary feature significance level was α = 0.003. So we fixed α = 0.003 and then reduced dimension through PDF-FS. Additionally, selected feature number K was another influencing factor on classification accuracy. As shown in Figure 1d, The highest SVM classification accuracy reached 84.21% (P ≤ 0.009) when K = 31 or 35. Furthermore, by using the PDF-FS method, average accuracy was kept around 76.60%. It is noted that when t-test reached its highest accuracy, the feature numbers were 66.26, meanwhile the PDF-FS method halved the features number k.

Contribution module

We calculated the contribution of all the discriminant functional connectivity; finally we got a contribution network (see Fig. 2b). This generally consisted of four modules: centering on inferior orbitofrontal (ORBinf.R) module, supramarginal gyrus (SMG) module, inferior parietal lobule (IPL) – posterior cingulated gyrus (PCG) module and middle temporal gyrus (MTG) – inferior temporal gyrus (ITG) module. These connections played an important role in displaying the discrepancy connectivity between MDD patients and HC. In this paper, we studied these discrepancies on the perspective of modules instead of single functional connectivity.

Figure 2.

Contribution network analysis composed of feature selection. (a) Brain areas play important roles in discrimination process, the size of circles positively correlated with contribution degree. (b) Contribution network composed of functional connectivity during classification process, solid lines indicate major depressive disorder (MDD) patients' functional connectivity was higher than healthy controls' functional connectivity; on the contrary, dotted lines indicate healthy controls' functional connectivity was higher than MDD patients' functional connectivity.

Additionally, through the equal portion contribution value of each functional connectivity to both brain areas in each end, we could obtain contribution degree of each brain area (see Fig. 2a). Brain areas which played a major role in classification process included SMG.L, ORBinf.R, fusiform gyrus (FFG).L, caudate (CAU).L, and thalamus (THA).L.

These obviously discrepant connectivities and brain areas may enhance our knowledge of the pathophysiological mechanisms of MDD.


Previous studies indicated discrepant functional connectivities between MDD patients and HC. However, it is unknown whether these connectivities can be used as diagnostic biomarkers of MDD.[18] Indeed, whether the future diagnostic models built on the functional connectivity values can improve treatment prediction and clinical outcome depend on its accuracy performance.[19, 20] This study is according with our hypothesis.

The superiority of our discriminant classification method

This study has many advantages compared with previous research on methodology. Through calculation of Pearson's correlation coefficient of pairwise brain regions, each participant gets a 4005 dimension feature. Then it is necessary to reduce dimension before classification. Reduced dimension has two ways: one is through transformation or mapping of high dimension feature into low dimension space, for example the principal component analysis (PCA),[15] and the independent component analysis,[21] which usually transforms features into new vectors by weighting or decomposition. Although it can retain major classification information, conclusive analysis and the interpretation process will be ambiguous. The other way is feature selection, which by removing uncorrelated and redundant components in feature space, keeps a small part of features to achieve dimension reduction purpose.

In the present discriminant-analysis process, feature selection plays an important role. Many studies only simply pick out discrepant features by two-sample t-test or the recursive feature elimination (RFE) method or conduct linear transformation of features through the PCA method;[13, 21] meanwhile other approaches only select features depending on previous knowledge.[17] In fact, the quality of selected features can directly influence classification accuracy, and it turns out that more features in the classification process will not necessarily improve, but can even worsen the results.

T-test or T-Score is the most simple feature selection method.[13] It can remove tremendous indistinctive features. However, more efficient classification feature selection subsets still exist. Thus, we propose a supervised PDF-FS, which can ensure that high-quality features will be preferentially selected to carry out a further selection. But it is important to realize that computation of non-parametric estimation of probability density function is extremely complex. Consequently, we first carried out a forward selection by t-test, and only a few significant features participated in sorting by PDF-FS that greatly reduced the computational cost.

The reliability of the PDF-FS method outperformed previously utilized methods. Indeed, the classification results of the leave-one-out t-test (LOOTT) based on our database are as follows: the maximum accuracy is only 80.26%, the t-test is only 78.95%; thus, our LOOTT method is already superior to the t-test. The maximum accuracy of RFE is only 77.63%, still less than leave-one-out cross validation (LOOCT). In fact, LOOTT is an improved version of the t-test. In feature selection, the LOOTT method was used for pre-selection of features based on LOOCT training set; this was aimed at ensuring that the set of selected features was as concise as possible. By using the PDF-FS method, the highest accuracy of SVM can be improved by 5% (increase from 78.95% to 84.21%). The discriminant analysis accuracy performance is sufficient to assist clinical diagnosis or even build a diagnostic model. The number of selected features used was only half of the original numbers, this greatly streamlined the features subset and improved the quality of characteristics by making them contain more classification information. Thus, we constructed the contribution network by all the selected functional connectivities used during the cross-validation process. This can enhance information by extending significant difference from single connectivity to group level, and thus we can infer important brain regions and discrepant modules, which play a major role in the classification process.

OR can reflect distinguishing ability of single functional connectivity; meanwhile the ultimate contribution degree can reflect the classification contribution degree of single functional connectivity as an integral part of the overall population. The two most typically inconsistent functional connectivities regarding these two aspects are as follows:

  1. AMYG.L-PreCG.R, overlap rate was small, but the contribution was only 0.04. It reveals that this functional connectivity was always prior selected to the classification of feature subset, but virtually it made a small contributions degree every time. Clinically, the psychopathological role of this functional connectivity was ambiguous; and
  2. SMG.L-FFG.L, a single functional connectivity. Theoretically its distinguishing ability was not high, but the final value of its contribution degree was tremendous. This reveals that it did not participate in each classification process, but once participated in, it made a great contribution. Just as described below, the psychopathological significance of this functional connectivity is virtually very significant.

Finally, we chose the functional connectivities which appeared in the leave-one-out process more than half the times and weight network of their contribution degree to construct a contribution network. These discrepant functional connectivities and brain areas played an important role in revealing the discrepancy between MDD patients and HC. These findings indicate the identified components had disease-related significance. The contributions of identified components were positively correlated with their psychopathological significance.

The psychopathological significance of cardinal modules

Depressed mood and markedly diminished interest in previously enjoyed activities (i.e., anhedonia) are key characteristics of MDD.[22] The ORB cortex plays a crucial role in emotion-processing circuits.[23] Brain regions (hippocampus, insular and putamen), which are associated with emotion regulation, were involved in the ORBinf module and made a greater contribution in the classification process. The ORB gyrus is implicated in reward-directed learning.[24] The ORB cortex links reward to hedonic experience.[25] This may explain the patients' all-encompassing low mood and anhedonia.

IPL is involved in multisensory perception and integrating information.[26] Studies indicate a prominent role for the posterior cingulate cortex in pain.[27] IPL and PCG are both involved in pain perception.[28] We presume this module is associated with interpretation of sensory information, especially pain sense. This neural circuit hyperactivity may lead to human body sensory hypersensitivity; this may also be associated with cenesthopathy and somatic symptoms.

IPL is included in one network engaging working memory regions.[29] The posterior cingulate has been linked to thought suppression and cognitive control.[30] The PCG and IPL both belong to default mode network.[29] Several studies have shown a pattern in depression of balky transitions from introspective thought to work that requires conscious effort and frequent slippage into the default mode during cognitive tasks.[31] This is closely associated with poor concentration and distractability in depressed patients.[31]

Available studies demonstrated that dysfunction of SMG and FFG best predicted impairment in reading words.[32] A recent study has found that left SMG and middle FFG are closely related to reading ability.[33] FFG responded to pictorial stimuli, meanwhile SMG correlated with the phonological processing.[34] Based on the above view and our results, the authors infer that this neural circuit may be associated with abnormal effective connectivity across visuo-attentional and audito-attentional network, which is likely to contribute to deficiency in attention filtering of information.[35]

The right MTG is associated with processing of angry and disgusted facial expressions. Inferior temporal gyri is also involved in emotional processing. Compared with HC, patients with depression demonstrated greater activation in right middle/inferior temporal gyrus to expression of strong disgust.[36] So we infer that this module may be related to emotional processing abnormalities in MDD.

Our study has several limitations. First, we did not check the head motion difference between the two subject groups. Thus the findings of this study might be partially confounded by head motion. Second, AAL atlas ROI specifically and structural ROI generally are rather coarse and tend to combine functionally distinct ROI, but in this study we did not combine with functionally ROI. Thus these findings might have been confounded by AAL atlas.

In summary, a more effective classification method was proposed in this research, which could automatically identify discrepant functional connectivities for distinguishing MDD patients from HC. The results demonstrated that the proposed method not only achieved a promising classification performance so as to be used as biomarker for clinical diagnosis of MDD patients, but also identified discriminative functional connectivities that are informative for MDD diagnosis. In future studies, it will also be of great interest to examine resting-state functional connectivity alteration in unmediated depressed patients before and after antidepressant treatment in unraveling antidepressant drug action at the neural level.[4, 19]


Our researches were supported by grants from the National 973 Program of China (2011CB707800 to Dr Z.N. Liu), and the National Natural Science Foundation of China (81271485, 81071092 to Dr Z.N. Liu, 81000587 to Dr Haihong Liu, and 11271121 to Dr Shuixia Guo), and Specialized Research Fund for the Doctoral Program of Higher Education (20110162110017 to Dr Z.N. Liu, 20100162120048 to Dr Haihong Liu). None of the authors has anything to disclose.