machine-learning classifier for patients with major depressive...

Research ArticleMachine-Learning Classifier for Patients withMajor Depressive Disorder Multifeature ApproachBased on a High-Order Minimum Spanning TreeFunctional Brain Network

Hao Guo12 Mengna Qin1 Junjie Chen1 Yong Xu3 and Jie Xiang1

1College of Computer Science and Technology Taiyuan University of Technology Taiyuan China2National Laboratory of Pattern Recognition Institute of Automation The Chinese Academy of Sciences Beijing China3Department of Psychiatry The First Hospital of Shanxi Medical University Taiyuan China

Correspondence should be addressed to Hao Guo feiyu guosinacom

Received 16 June 2017 Revised 10 October 2017 Accepted 9 November 2017 Published 14 December 2017

Academic Editor Marko Gosak

Copyright copy 2017 Hao Guo et alThis is an open access article distributed under the Creative Commons Attribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

High-order functional connectivity networks are rich in time information that can reflect dynamic changes in functionalconnectivity between brain regions Accordingly such networks are widely used to classify brain diseases However traditionalmethods for processing high-order functional connectivity networks generally include the clustering method which reduces datadimensionality As a result such networks cannot be effectively interpreted in the context of neurology Additionally due to thelarge scale of high-order functional connectivity networks it can be computationally very expensive to use complex network orgraph theory to calculate certain topological properties Here we propose a novel method of generating a high-order minimumspanning tree functional connectivity network This method increases the neurological significance of the high-order functionalconnectivity network reduces network computing consumption and produces a network scale that is conducive to subsequentnetwork analysis To ensure the quality of the topological information in the network structure we used frequent subgraph miningtechnology to capture the discriminative subnetworks as features and combined this with quantifiable local network featuresThenwe applied a multikernel learning technique to the corresponding selected features to obtain the final classification results Weevaluated our proposed method using a data set containing 38 patients with major depressive disorder and 28 healthy controlsTheexperimental results showed a classification accuracy of up to 9754

1 Introduction

Resting-state functional magnetic resonance imaging (rs-fMRI) using blood oxygenation level-dependent (BOLD)signals as neurophysiological indicators can detect sponta-neous low-frequency brain activity and has been successfullyapplied to the diagnosis of neuropsychiatric diseases suchas schizophrenia [1ndash4] Alzheimerrsquos disease [5ndash7] epilepsy[8ndash10] attention deficit hyperactivity disorder (ADHD) [11]and stroke [12 13] Resting functional brain network analysishelps clarify the mechanisms of neuropsychiatric disordersand has the potential to provide relevant imaging markersthat may offer new perspectives for the diagnosis and evalua-tion of clinical brain diseases [2] In traditional brain network

analysis it is assumed that the correlation between differentbrain regions does not change with time during rs-fMRIscanning Networks constructed usingmethods based on thisassumption are called low-order networks [14]

However this assumption may lead researchers to over-look the dynamic interaction patterns between brain regionsduring the entire scan which are essentially time-varyingIndeed several recent studies have indicated that functionalconnectivity analyses can be rich in dynamic temporal infor-mation [15 16] High-order functional connectivity networkscontain abundant dynamic time information so this methodhas been proposed and applied in the diagnosis of braindiseases [14 17]

HindawiComputational and Mathematical Methods in MedicineVolume 2017 Article ID 4820935 14 pageshttpsdoiorg10115520174820935

2 Computational and Mathematical Methods in Medicine

Themost commonmethod for constructing a high-orderfunctional network is the dynamic sliding window methodin which the whole rs-fMRI time series is divided into severaltime windows [18] A low-order functional connectivity net-work is built in each time window and then all the low-ordernetworks are stacked A clustering algorithm is performedto divide all relevant time series into several clusters Theaverage time series of each cluster is then taken as a newnodeand the Pearson correlation coefficient is calculated betweeneach node pair as the weight of connectivity [14]

In this method clustering is employed to decrease theassociated computational costs and classification accuracyis greatly influenced by the randomness of the selectionof initial clustering centers and the number of clustersHowever because the time series of all connectivities withineach cluster are averaged the network loses neurologicalinterpretability

In the present study we used the minimum spanningtree method [19] to reduce the computational cost while pre-serving the core framework of networks A classic approachin graph theory this unbiased method greatly simplifies thenetwork structure while preserving its core framework thusavoiding the influences of network sparseness and otherparameters on network structure It also guarantees thenetworkrsquos neurological interpretability and has been widelyused in previous studies [20ndash22]

Furthermore the traditional feature extraction methodin minimum spanning tree networks uses a quantifiablenetworkwith local features for classification of brain diseasessuch as degree clustering coefficient minimum path lengthand eccentricity [23 24] However a clear shortcomingof this method was the chance that some of the usefultopology information in the network (including connectionpatterns in the sample itself and the common connectionpatterns between the samples) would be lost resulting inreduced classifier performance Frequent subgraph miningtechnology was proposed to mine discriminative subgraphpattern features for machine-learning classification of braindiseases [25 26] Subgraph pattern features could account forthe connection pattern information between multiple brainregions but it was not sensitive to changes in single brainregions [27] Therefore both methods can lead to loss ofsample information

Here we propose a novel feature extraction method thatcombines quantifiable local network features with subgraphpattern features Specifically we computed degree eccentric-ity and betweenness centrality of each brain region as localnetwork features and extracted the subgraph features usinga frequent subgraph mining method for a group of healthycontrols (HC) and a group of people with major depressivedisorder (MDD)Then a kernel function for each type of fea-ture was constructed namely a vector kernel (local networkfeatures) and a graph kernel (subgraph features) Finallythe two kernel matrices were combined and a multikernelsupport vector machine was constructed as a classifier Theproposed method achieves better classification performancethan traditionalmethods that use only a single type of feature

Table 1 Subject demographics and clinical characteristics

HC MDD 119901 values

Age 17ndash51(266 plusmn 94)

17ndash49(284 plusmn 968) 044a

Sex (malefemale) 1315 1523 057b

Handedness (RL) 280 380

HAMD NA 15ndash42(228 plusmn 133)

Data are minimumndashmaximum (mean plusmn standard deviation) HAMD 24-item Hamilton scale atwo-sample two-tailed 119905-test btwo-tailed Pearsonrsquoschi-squared test

2 Materials and Methods

21 Proposed Framework Figure 1 shows the flowchart ofthe proposed method which includes four main steps (1)data acquisition andpreprocessing (2) network constructionin which a high-order functional connectivity network isconstructed first followed by construction of a minimumspanning tree network (3) feature extraction and selectionin which two types of feature are extracted and selected(the first is used to calculate quantifiable local networkfeatures (degree betweenness centrality and eccentricity)and uses the KolmogorovndashSmirnov test for feature selectionand the second is used to mine frequent subgraphs from theHC and MDD groups and selects the most discriminativesubnetworks as the subgraph patterns) (4) constructionof a classification model in which the kernel matrix ofthe two types of feature is calculated The multiple-kernelsupport vectormachine (SVM) is adopted to combine the twoheterogeneous kernels enabling the distinction of individualswith MDD from healthy controls

22 Data Acquisition and Preprocessing The study was car-ried out in accordance with the recommendations of themedical ethics committee of Shanxi Province (referencenumber 2012013) All subjects provided written informedconsent in accordance with the Declaration of HelsinkiTwenty-eight healthy subjects and thirty-eight people withMDD underwent rs-fMRI in a 3T scanner (Siemens Trio3-Tesla scanner Siemens Erlangen Germany) Participantdemographic information is shown in Table 1

Data collection was completed at the First Hospital ofShanxi Medical University Radiologists familiar with fMRIperformed all scans During each scan the participant wasasked to relax with their eyes closed and not think aboutanything in particular but to stay awake and avoid fallingasleep Each scan consisted of 248 contiguous echo-planarimaging (EPI) functional volumes (33 axial slices repetitiontime (TR) = 2000ms echo time (TE) = 30ms thicknessskip= 40mm field of view (FOV) = 192 times 192mm matrix =64 times 64mm and flip angle = 90∘) The first 10 volumes inthe time series were discarded to account for magnetizationstabilization See Supplemental Text S1 for detailed scanningparameters

Data preprocessing was performed in SPM8 (httpwwwfilionuclacukspm)with slice-timing and head-movementcorrections Two samples containing a translation of more

Computational and Mathematical Methods in Medicine 3

gSpan

gSpan

Local network features

Selectedfeatures

NC

MDD

fMRI data AAL template Regional meantime series High-order FC

network Minimum spanning tree

(2) Network construction

Frequent subgraphsDiscriminative

subgraphs

Subgraph features

(3) Features extraction and selection subgraph and MST(4) Classification with multikernel SVM

Graph kernel matrix

Vector kernel matrix

Multikernel SVM

Degree

Eccentricity

BetweennessKolmogorov-Smirnov Test

(1) Data acquisition and preprocessing

middot middot middot


Figure 1 Basic framework Illustration of the basic framework of the method used (1) Data acquisition and preprocessing (2) networkconstruction (the high-order functional connectivity network is constructed first followed by construction of the minimum spanning treenetwork) (3) feature extraction and selection (two types of feature are extracted and selected one is to calculate quantifiable local networkfeatures (degree betweenness centrality and eccentricity) with the KolmogorovndashSmirnov test used for feature selection and the other isto mine frequent subgraphs from the HC and MDD groups and select the most discriminative subnetworks as the subgraph patterns)(4) classification model construction (kernel matrix is calculated for two types of feature and then the multiple-kernel support vectormachine (SVM) is adopted to combine these heterogeneous kernels for distinguishing individuals with MDD from healthy controls) AALautomated anatomical labeling fMRI functional magnetic resonance imaging FC functional connectivity HC healthy controls MDDmajor depressive disorder

than 30mm and rotation of more than 30∘ were excludedfrom the final analysis of 66 samples Functional imageswere normalized using the 12 parameters from the affinetransformation and the cosine-based nonlinear transforma-tion from the normalization of the anatomic image to theMontreal Neurological Institute (MNI) space Additional

normalization of the functional data sets to the SPM8 EPItemplate was then performed and the data were resampledto a voxel size of 3 times 3 times 3mm using a sinc interpolation Nosmoothing kernel was applied to limit spurious local connec-tivity between voxels Finally we performed linear detrendingand band-pass filtering (001ndash010Hz) to reduce the effects of


Sliding windows Low-order FC networks Stacked low-order FC network

High-order FC network

Pearsonrsquos correlation

Figure 2 High-order functional connectivity network construction flowchart (1) Partition the entire rs-fMRI time series into multipleoverlapping segments of subseries by adopting a fixed-length sliding window (2) construct temporal low-order FC networks in each timewindow (3) stack all low-order FC networks for all subjects (4) construct a high-order FC network for each subject by taking the low-orderFC as a new vertex and the pairwise Pearsonrsquos correlation coefficient between each pair of these new vertices as the weight FC functionalconnectivity

low-frequency drift and high-frequency physiological noiseThen for each subject the brain space of the fMRI imageswas parcellated into 90 regions of interest (ROIs) (45 in eachhemisphere) based on the automated anatomical labeling(AAL) atlas [40] and each region was defined as a node inthe network Each regional mean time series was regressedagainst the average cerebral spinal fluid and white-mattersignals as well as the six parameters from motion correctionThe residuals of these regressions constituted the set ofregional mean time series used for undirected graph analysis

23 Construction of the High-Order MinimumSpanning Tree Network

231 High-Order Functional Connectivity Network A high-order functional connectivity network was constructed usinga flowchart with the following steps (Figure 2) (1) partitionthe entire rs-fMRI time series into multiple overlapping seg-ments of subseries by adopting a fixed-length slidingwindow(2) construct temporal low-order functional connectivitynetworks in each time window (3) stack together all low-order functional connectivity networks for all subjects (4)construct a high-order functional connectivity network foreach subject by taking the low-order functional connectivityas the new nodes and the pairwise Pearson correlationcoefficient between each pair of nodes as the path weight

To enable construction of a low-order functional connec-tivity network in each time window we divided the wholetime series 119909(119897)119894 into a number of overlapping subseries seg-ments using the sliding time-window method Specificallyif the length of the sliding window is 119873 and the step sizebetween two successive windows is 119878 let 119909(119897)119894 isin 119877119873 denotethe 119896th segment of the subseries extracted from 119909(119897)119894 The totalnumber of segments generated by this approach is given by

119870 = lfloor(119872 minus119873)119904 rfloor + 1 1 le 119896 le 119870 (1)

The length of our sliding window was 90 and the steplength was 1 For an illustration of the sliding window seeSupplemental Figure S1

For the 119871th subject the 119896th segment in the subseries forall ROIs can be expressed in matrix form as

119909(119897) (119896) = [119909(119897)1 (119896) 119909(119897)2 (119896) 119909(119897)119877 (119896)] isin 119877119873times119877 (2)

where 119877 is the total number of ROIs Then the entry forthe 119896th temporal functional connectivity matrix for the 119871thsubject 119862(119897)(119896) can be obtained by the Pearson correlationbetween the 119894th and 119895th ROIs The 119870 temporal functionalconnectivity networks for the 119871th subject can be establishedby taking 119910(119897)119894119895 as nodes and 119862(119897)119894119895 (119896) as the weights of newedges as per the following equation

119862(119897)119871 (119896) = (119909(119897)119894 (119896) 119862(119897)119894119895 (119896)) (119896 = 1 2 119896) (3)

In this way it is possible to construct119870 dynamic temporalfunctional connectivity networks for each subject For eachROI pair (119894 119895) for the 119871th subject we can concatenate 119862(119897)119894119895 (119896)to obtain a correlation time series

119910(119897)119894119895 = [119862(119897)119894119895 (1) 119862(119897)119894119895 (2) 119862(119897)119894119895 (119896)] isin 119877119870 (4)

We can then stack together all the dynamic temporal func-tional connectivity networks for each subject as per (4)

The main goal of this article is to reveal the intrinsicrelationship between the correlation time series 119910(119897)119894119895 andthe dynamic temporal information contained within it Wecalculated the Pearson correlation coefficient between eachpair of correlation time series for each subject as follows

119867(119897)119894119895119901119902 = corr (119910(119897)119894119895 119910(119897)119901119902) (5)

Thus the construction of a high-order functional connectiv-ity network is achieved by taking 119910(119897)119894119895 as new nodes and119867(119897)119894119895119901119902 as the weights of new edges and then connectingnodes 119910(119897)119894119895 and 119910(119897)119901119902 The new high-order functional connec-tivity network can be represented as

119866(119897)119867 = (119910(119897)119894119895 119867(119897)119894119895119901119902) (6)


Table 2 Definitions and formulae of minimum spanning tree network properties

Concept Explanation FormulaDegree Number of links for a given node 119896119894 = sum

119895isin119873

119886119894119895

EccentricityLongest shortest path from a referencenode to any other node in the minimum

spanning treeEcc(V) = max119889(119906 V)

Betweenness centrality Fraction of all shortest paths that passthrough a particular node BC119894 = 1

(119899 minus 1)(119899 minus 2) sumℎ119895isin119873ℎ =119895ℎ =119894

120588(119894)ℎ119895

120588ℎ119895119886119894119895 connection between 119894 and j 119889(119906 V) shortest path from u to v 120588ℎ119895 number of shortest paths between h and j 120588(119894)

ℎ119895 number of shortest paths between h and

119895 which passthrough i

Therefore 119867(119897)119894119895119901119902 can be said to represent the high-ordercorrelation and the corresponding network 119866(119897)119867 representsthe high-order functional connectivity network The high-order correlation indicates the linear correlation strengthbetween two correlation time series and reflects the interac-tion between up to four brain regions Compared with thetraditional network the high-order functional connectivitynetwork not only takes into account the time-varying char-acteristics of functional connectivity but also represents themore complex and abstract interaction patterns among brainregions

232 Minimum Spanning Tree To further reduce the com-plexity of the high-order functional connectivity network weconstructed a minimum spanning tree This is a weightedsubnetwork (fully connected network) that connects all thenodes in the network without forming loops and has theminimum total weight of all possible spanning trees [24]We constructed the minimum spanning tree based on theweighted network Since we were interested in determiningthe strongest connection in the network we used Kruskalrsquosalgorithm to obtain the strongest connection weights [41]This algorithm first sorts the edges into descending weightorder and then starts the construction of the minimumspanning tree from the largest-weight edge adding the nextlargest-weight edge until all nodes 119873 are connected in anacyclic subnetwork consisting of 119872 = 119873 minus 1 edges Whenthe addition of an edge forms a loop this edge is ignoredFor more information regarding Kruskalrsquos algorithm seeSupplemental Text S2

24 Feature Extraction and Selection After completion ofthe network we extracted features of two different typesquantifiable local network features of theminimum spanningtree and subgraph patterns from frequent subgraph miningWe selected quantifiable local network features of the mini-mum spanning tree using the Kolmogorov-Smirnov test Forthe connected patterns from frequent subgraph mining weused discriminative scores to select the most discriminativesubgraphs

241 Local Network Features and Selection Methods Weselected the local network properties of the minimum span-ning tree (degree betweenness centrality and eccentricity)as features We calculated the three properties of each node

in the high-order minimum spanning tree network Table 2gives the definition and formula of these three propertiesWeused multilinear regression analysis to assess the confound-ing effect of age sex and educational attainment on eachnetwork attribute The independent variable was the meanof each network attribute (except for the degree owing toits nature) and the dependent variables were age sex andeducational attainment The results showed no significantcorrelations between betweenness centrality eccentricityand corresponding variables (see Supplemental Table S1 forresults)

We used the Kolmogorov-Smirnov test [42] to select thequantifiable local network features of theminimum spanningtree (119901 lt 005) The results were then corrected using theBenjamini-Hochberg false positive rate (119902 = 005) [43]242 Frequent Subgraph and Discriminative Evaluation

(1) Frequent Subgraph In this paper subgraph pattern extrac-tion was mainly based on frequent subgraph mining Thefrequent subnetwork refers to the connected patterns thatappear most often in the network [44] The purpose offrequent subnetwork mining is to uncover the most frequentconnected patterns (ie subnetworks) in the whole network[25] We applied this algorithm to the HC and MDD groupsIn the field of data mining a large number of frequentsubgraph mining methods have been proposed [45 46]including a priori-based graph mining [47] and the frequentsubgraph discovery algorithm [48] Here we used the well-known gSpan algorithm [49] to extract the frequent subnet-works from the functional connectivity network Because ofits high efficiency in graph traversal and subgraphmining thegSpan algorithm has been widely applied in many researchfields including neural imaging [25ndash27]

The gSpan algorithm works as follows [50] First thegSpan constructs a new lexicographic order among graphsand maps each graph to a unique minimum depth-firstsearch (DFS) code as its canonical label Then based on thelexicographic order gSpan uses the DFS strategy to efficientlymine frequently connected subgraph patterns In the presentstudy we termed the hierarchical search space of frequentsubgraphs the ldquoDFS code treerdquo where each node in thetree represents a DFS code (ie subgraph) The 119896 + 1thlevel subgraph is generated from the 119896th level subgraph (ie


parent) by adding one frequent edge Finally all subgraphswith nonminimal DFS codes are pruned to avoid redundantcandidate generations In subgraph mining the number ofsubgraphs is mainly controlled by frequency Given a set ofgraphs 119866 the frequency of a subgraph 119892119904 is defined as

fq (119892119904 | 119866) =1003816100381610038161003816119892119904 is a subgraph of 119892 119892 isin 1198661003816100381610038161003816

119866 (7)

The DFS lexicographic order used in frequent subgraphmining and the gSpan algorithm are described in detail inSupplemental Text S3

(2) Discriminative Evaluation The discriminative subnet-work can be used as a feature for classification [51] but it isworth noting that gSpan is only used for mining the frequentsubgraph which by itself has no discriminative power Forinformation on the discriminative capabilities of differentsubgraphs see Supplemental Figure S2 However some of thefrequent subnetworks may have less discriminative informa-tion for classification To address this problem we selectedthe most discriminative subnetworks from the frequentsubnetworks using subgraph discriminative scores (whichexpress subgraph frequency differences) [27] This strategy iscalled frequent-scoring feature selection In the present studythe method involved choosing the same number of frequentsubgraphs from the HC and MDD groups calculating andsorting the discriminative scores of frequent subgraphs andselecting the top 119896 subnetworks with higher discriminativescores Thus 2 lowast 119896 discriminative subnetworks are selectedFor the given graphs 119866119875 and 119866119899 119866119875 = 1198921199011 1198921199012 119892119901119898refers to the set of frequent subgraphs for all positive samplesand 119866119899 = 1198921198991 1198921198992 119892119899119898 refers to the set of all frequentsubgraph features for negative samples The discriminativescores 119878(119892119904) of subgraph 119892119904 can be calculated as

119878 (119892119904) = 10038161003816100381610038161003816119891119902 (119892119904 | 119866119875) minus 119891119902 (119892119904 | 119866119899)10038161003816100381610038161003816 (8)

The discriminative score of a subgraph pattern 119892119904 is simplydefined as the difference between its positive frequency andnegative frequency A larger score reflects a larger differencebetween the patterns in the two groups 119878(119892119904) = 1 indicatesthat the subgraph 119892119904 exists in all graphs for the HC group andthat there is no such pattern in any graph for theMDD group119878(119892119904) = minus1 indicates that the subgraph 119892119904 exists in all graphsfor the MDD group and that there is no such pattern in anygraph for the HC group

25 Construction of Classification Model The classificationmodel chosen in this paper is a multikernel SVM Recentstudies on multikernel learning have shown that the integra-tion of multiple kernels can significantly improve classifica-tion and enhance interpretability of results [52] Generallythe integration of the kernel is achieved by linear combinationof multiple kernels

119896 (119909 119910) =119872

sum119894=1

119886119894119896119894 (119909 119910) st119872

sum119894=1

119886119894 = 1 (9)

where 119896119894(119909 119910) is a basic kernel built for subjects 119909 and 119910119872 is the number of kernel matrices required and 119886119894 is thenonnegative weighting parameter

A graph kernel can be regarded as a group of similaritiesbetween a pair of subjectsThe brain network data is mappedfrom the original network space to the feature space andthe similarity between the two brain networks is furthermeasured by comparing their topology In this study weused theWeisfeiler-Lehman subtree based on theWeisfeiler-Lehman isomorphism test [53] to measure the topologicalsimilarity between paired connectivity networksThis type ofgraph kernel can effectively capture topological informationfromgraphs and improve performanceGiven two graphs thebasic process of theWeisfeiler-Lehman test is as follows if thetwo graphs are unlabeled (ie the nodes of the graph havenot been assigned labels) each node is first labeled with thenumber of edges that are connected to that node Then ateach iteration step the label of each node is updated based onits previous label and the labels of its neighbors That is thesorted set of updated node labels for each node is compressedsuch that it contains new and shorter labels This processiterates until the node label sets are identical or the numberof iterations reaches its predefined maximum value For adetailed description of the Weisfeiler-Lehman isomorphismtest and pseudocode see Supplemental Text S4

As this study involves two different types of kernel(vector-based kernels and graph kernels) a normalizationstep must be performed individually before combining themThis normalization step can be accomplished using thefollowing formula

119896lowast (119909 119910) = 119896 (119909 119910)radic119896 (119909 119909) 119896 (119910 119910)

(10)

Note that unlike the previous multikernel learning methodin which the weighting parameters 119886119894 are jointly optimizedtogether with other classifier parameters in this study theoptimal weighting parameters 119886119894 are determined via a gridsearch of the training data Once the optimal weightingparameters 119886119894 are obtained the multikernel learning-basedclassifier can be naturally embedded into the conventionalsingle-kernel classifier framework In this paper we selectedthe SVM as the classifier framework

As described above we used multikernel learning meth-ods to perform classification As different types of kernelsrepresent different properties of the network we combinedmultiple features through multikernel learning Specificallythe vector-based kernel describes the correlation betweenpairwise brain regions according to degree betweenness cen-trality and eccentricity and the graph-based kernel describesthe topological information contained in the whole network

3 Results

We performed two types of feature extraction on the con-structed network The first involved the quantifiable localnetwork features namely degree betweenness centrality andeccentricity The second involved the extraction of discrimi-native subgraph patterns from the HC and MDD groups


31 Abnormal Functional Connectivities The high-orderfunctional connectivity network was 4005 lowast 4005 so therewere 4004 edges in the high-order minimum spanningtree network After constructing the network we analyzedthree kinds of traditional quantifiable network propertiesWe selected the high-order functional connectivities of atleast two network properties with 119901 lt 005 (false discov-ery rate corrected) We obtained 40 abnormal functionalconnectivities in total encompassing a total of 42 abnormalregions (Table 3) All 40 significant abnormal functional con-nectivities and frequency-corresponding nodes are shownin Supplemental Figure S3 These significant regions wereconcentrated in the limbic-cortical networks (left anteriorcingulate and paracingulate gyri bilateral median cingulateand paracingulate gyri right posterior cingulate gyrus bilat-eral caudate nucleus bilateral lenticular nucleus bilateralputamen bilateral thalamus bilateral hippocampus bilateralparahippocampal gyrus and bilateral amygdala) frontal lobe(bilateral precentral gyrus bilateral dorsolateral superiorfrontal gyrus bilateral superior frontal gyrus orbital partright middle frontal gyrus bilateral middle frontal gyrusorbital part bilateral inferior frontal gyrus triangular partand bilateral inferior frontal gyrus opercular part) temporallobe (right temporal pole middle temporal gyrus left Heschlgyrus) and cuneus (bilateral cuneus bilateral lingual gyrusbilateral precuneus right postcentral gyrus left calcarinefissure and surrounding cortex)

We selected the top 10 brain regions with the mostsignificant differences in terms of frequency (Table 4) Thesemainly included the bilateral temporal pole middle temporalgyrus left superior frontal gyrus orbital part left thalamusright lenticular nucleus putamen left lingual gyrus rightcuneus left posterior cingulate gyrus and right dorsolateralsuperior frontal gyrus

32 Frequent Subgraph Patterns We analyzed the frequentdiscriminative subnetworks We mined two sets of frequentsubnetworks from the functional connectivity networks ofthe MDD and HC groups with respective frequencies fq =0286 and fq = 0211 Specifically we mined 4057 subgraphsfrom the HC group and 4078 from the MDD group Forstatistical information regarding the number of edges in thesubgraphs see Supplemental Table S3 We calculated thediscriminative scores of the frequent subgraphs and found16 subgraphs for the HC group and 37 for the MDD groupTo ensure that the features were balanced we selected 16discriminative subnetworks to assess the subgraph patternsfrom the two groups

To analyze the connected patterns connections in the 16subgraph connected patterns for each of the HC and MDDgroups were merged with the subgraph in SupplementalFigure S4 By analyzing the subgraphs of the HC and MDDgroups we found some nodes that were significantly differentbetween the two groups These significantly different nodeswere mainly concentrated in the bilateral lenticular nucleusthe putamen bilateral lingual gyrus bilateral amygdala bilat-eral thalamus bilateral median cingulate and paracingulategyri right posterior cingulate gyrus bilateral cuneus leftanterior cingulate and paracingulate gyri right superior

frontal gyrus orbital part right middle frontal gyrus righttemporal pole middle temporal gyrus left precentral gyrusright lenticular nucleus pallidum and so forth

We also analyzed these significantly different brainregions and ranked them according to the frequency inwhichthey appeared in the HC and MDD groups The significantlydifferent brain regions and the frequency-correspondingnodes are given in Supplemental Figure S5 We selected thetop 10 regions as the most discriminative (Table 5)

33 Classification Results We evaluated the classificationperformance of the proposed method by measuring classi-fication accuracy sensitivity specificity and area under thecurve (Table 6) Table 6 also compares the classificationperformance of the partial correlation functional connec-tivity network Pearson functional connectivity networkhigh-order functional connectivity network and frequentsubgraph mining methods The results indicate that our pro-posed method achieves good results in terms of classificationaccuracy sensitivity specificity and area under the curve

Specifically to compare the method proposed in thispaper with those used previously we constructed partial andPearson correlation networks and a high-order functionalconnectivity network without minimum spanning tree anal-ysis (see Supplemental Text S5 for the detailed information ofother contrast networks) In addition we used a high-orderminimum spanning tree network for assessing quantifiablelocal network features and subgraph patterns as features Ourexperimental results showed that the proposed method ofclassification performed significantly better than the partialcorrelation network Pearson correlation network and high-order functional connectivity network and also better thanthe method in which only the quantifiable local networkfeatures and subgraph pattern features were assessed Thisillustrates the potential for the integration of the two differenttypes of features to significantly improve classification per-formance We used the Relief method [49] to calculate theaverage weights of subgraph pattern features the minimumspanning tree of quantifiable local network features andboth types of features combined (Figure 3) The averageweight of the subgraph pattern features was 55031 that ofthe minimum spanning tree of the quantifiable local networkfeatures was 91542 and that for both feature types togetherwas 94516 Figure 4 shows the receiver operating character-istic curve of the proposed method the partial correlationnetwork the Pearson correlation network the higher-orderfunctional connectivity network and the approach with onlysubgraph patterns as features and quantifiable local networkfeatures These results indicate that our proposed methodclearly improved classification performance

4 Discussion

41 Abnormal Brain Regions We extracted quantifiable localnetwork features and frequent subgraph features to explorebrain regions with significantly abnormal connectivitiesbetween the HC and MDD groups By calculating thequantifiable local network features we were able to obtain40 significant abnormal connectivities involving 42 brain


Table 3 The 40 functional connectivities and associated statistical significance

Number FC connectivities 119901 valuesBetweenness Degree Eccentricity

(1) PreCGL SFGdorR 00135 00110 00506(2) PreCGL PUTR 00199 06648 00303(3) PreCGR ORBsupL 08049 00379 00379(4) SFGdorR CUNL 00491 00412 06491(5) SFGdorR TPOmidR 00379 02216 00122(6) ORBsupL PUTL 07558 00396 00026(7) ORBsupL PCLR 00252 05730 00252(8) ORBsupL CAUR 08504 00135 00135(9) ORBsupR DCGR 00252 02943 00252(10) MFGL STGL 00077 00470 04780(11) MFGR PCGL 00276 04262 00362(12) MFGR PCGL 00135 00135 01226(13) PCUNL CAUR 00149 08049 00149(14) IFGtriangR ORBmidR 00009 00129 03289(15) ORBinfR LINGL 04039 00063 00173(16) ORBinfR ACGR 00105 00362 00105(17) AMYGR LINGR 00190 07431 00029(18) AMYGL TPOmidR 00029 00264 06648(19) SFGmedL ACGL 01264 00470 00264(20) THAR ACGL 00470 00036 09172(21) DCGL THAL 01752 00210 00077(22) ACGR LINGR 00209 00167 00276(23) DCGR PUTR 00029 00252 01029(24) PHGL TPOmidR 00180 04377 00252(25) AMYGL HESL 00470 00095 05470(26) AMYGR MTGR 00368 01368 00095(27) CUNL ORBmidL 00379 00243 00379(28) CUNR PUTR 00157 00241 03930(29) CUNL HESL 00209 00053 07683(30) PCGR LINGR 00033 01693 00173(31) ORBsupR TPOmidR 00241 00122 00241(32) PCGR THAL 08282 00396 00122(33) PCGR TPOmidL 00396 00258 07558(34) PreCGR THAL 00105 08049 00379(35) HIPL PALL 00662 00382 00289(36) HIPL HESL 02379 00431 00379(37) PCLR PALL 00317 08282 00317(38) PUTL TPOmidR 00004 00338 00027(39) PALL THAR 01824 00048 00396(40) ROLR ITGR 00095 00432 05220Values in bold indicate significance (119901 lt 005) For all brain region abbreviations see Supplemental Table S2

regions Then we selected the top 10 most frequently impli-cated brain regions as those were the most significantlydifferent between the two groups Consistent with previousstudies these regions included the bilateral temporal polethe middle temporal gyrus the left superior frontal gyrusorbital part the right lateral dorsolateral frontal gyrus the leftthalamus the right putamen the left lingual gyrus the rightcuneus and the left posterior cingulate gyrus

The present results are consistent with our previousfindings Ma et al [54] adopted voxel-based morphometryto investigate brain regions with gray-matter abnormal-ity in patients with treatment-resistant depression and inthose with treatment-responsive depressionThey found thatpatients in both groups showed clear gray-matter abnormali-ties in the right temporal pole of the temporal gyrus specif-ically the middle temporal gyrus Qiu et al [55] examined


Table 4 Top 10 regions of interest in the minimum spanning tree

Number ROIs name Citations(1) Temporal Pole Mid R Li et al 2005 [28](2) Frontal Sup Orb L Veer et al 2009 [29](3) Thalamus L Greicius et al 2007 [30](4) Pallidum L Anand et al 2005 [31](5) Lingual L Veer et al 2009 [29](6) Cuneus R Tao et al 2013 [32](7) Cingulum Post L Wang et al 2016 [33](8) Frontal Sup R Wang et al 2016 [33](9) Putamen R Anand et al 2005 [31](10) Temporal Pole Mid L Yue et al 2013 [34]For all brain region abbreviations see Supplemental Table S2

Table 5 Top 10 regions of interest in the subgraph patterns

Number ROIs name Citations(1) Putamen R Anand et al 2005 [31](2) Lingual R Veer et al 2009 [29](3) Amygdala L Yue et al 2013 [34](4) Lingual L Veer et al 2009 [29](5) Cingulum Mid L Sheline et al 2010 [35](6) Thalamus L Greicius et al 2007 [30](7) Thalamus R Greicius et al 2007 [30](8) Cingulum Post R Wang et al 2016 [33](9) Putamen L Anand et al 2005 [31](10) Amygdala R Yue et al 2013 [34]For all brain region abbreviations see Supplemental Table S2

cortical thickness and surface area in first-episode treatment-naıve mid-life MDD and observed a significant increase ingray-matter volume in the left superior frontal gyrus leftthalamus and right cuneus Sacchet et al [56] obtainedwhole-brain T1-weighted images in their own HC and MDDgroups and evaluated gray-matter volumes in the basalganglia (specifically the caudate nucleus lentiform pallidumand putamen) They reported that gray-matter volumes inthe bilateral lenticular nucleus and putamen were signifi-cantly different between patients with depression and healthycontrols Jung et al [57] used voxel-based morphometry todetect structural changes in healthy subjects and patientswithdepression who underwent 8 weeks of antidepressant treat-ment The results showed significantly different gray-mattervolume in the left lingual gyrus between the two participantgroups Fang et al [58] measured spontaneous whole-brainhomodynamic responses using amplitude of low-frequencyfluctuation (ALFF) and fractional ALFF (fALFF) and foundthat the ALFF and fALFF decreased in the left posteriorcingulate gyrus the right cuneus and the superior frontalgyrus after depression treatment Cotter et al also foundabnormalities in the dorsal prefrontal cortex of patients withMDD [59]

In the present study frequent subgraph mining revealeda total of 32 discriminative patterns (16 in the HC groupand 16 in the MDD group) We found 19 common brain

Proposed Local network featuresSubgraph features0

300

600

900

1200

1500

1800

2100

Aver

age v

alue

of w

eigh

t

lowast

lowastlowastlowast

lowastp lt 005

lowastlowastlowastp lt 0001

Figure 3 Average weight of different types of feature Statistical anal-ysis of the average weight for subgraph pattern features minimumspanning tree of quantifiable local network features and featuresused in this study The average weight of subgraph pattern featureswas 55031 theminimumspanning tree of quantifiable local networkfeatures was 91542 and that for both feature types together was94516 The combination of the two different types of features hadthe greatest weight

regions in the 32 connected patterns in the two groupsAccording to the frequency of each region in the connectedpatterns we selected the 10most discriminative brain regionsThese included the bilateral lenticular nucleus putamenbilateral lingual gyrus bilateral amygdala left median cin-gulate paracingulate gyri right posterior cingulate gyrusand bilateral thalamus Anand et al [31] studied the dif-ferences between limbic-cortical activities and connectionsin patients with MDD versus HCs They found significantdifferences between patients and controls in the bilateralanterior cingulate cortex bilateral amygdala and bilateralthalamus The top 10 most frequent regions in our studyincluded the amygdala which is part of the limbic systemand is involved in the formation of emotional behaviorspontaneous activity and endocrine integration processesPrevious studies [60ndash62] have indicated that the amygdalaplays a significant role in the pathogenesis of depression Veeret al [29] used independent component analysis to assessrs-fMRI data from 19 medication-free patients with a recentdiagnosis of MDD (within 6 months prior to inclusion) andno comorbidity and 19 age- and gender-matched controlsThey found decreased activation in the bilateral amygdalawhich is associated with emotional behavior the frontal lobewhich is associated with attention and working memoryand the lingual gyrus which is related to visual processingThe other discriminative brain regions that we identified viafrequent subgraph mining are also consistent with previousresults such as the bilateral lenticular putamen [56] the left


Table 6 Comparison of classification results from different methods

Method Research Disease Accuracy Sensitivity Specificity AUC

Partial FCGuo et al 2013 [36] MDD 8601 - - -Qiao et al 2016 [37] MCI 8901 8667 9130 -

This study MDD 6306 5056 8737 7102

Pearson FCWong et al 2012 [38] MDD 6300 4000 8300 -Liu et al 2015 [39] SAD 8250 8500 8000 -

This study MDD 6667 4643 8158 7446

High-order FC Chen et al 2016 [14] MCI 8814 8621 9000 9299This study MDD 9251 8851 9319 9283

Frequent subgraph Du et al 2016 [26] ADHD 9491 9322 9694 9690Fei et al 2014 [25] MCI 9730 - - 9583

Frequent and local cluster coefficient Wang et al 2014 [27] MCI 9727 - - 9200

High-order MST FCSubgraph features MDD 7332 8036 6758 7567

Minimum spanning tree features MDD 9404 9826 9250 9784Proposed MDD 9754 10000 9667 9906

FC functional connectivity MST minimum spanning tree SAD social anxiety disorder MCI mild cognitive impairment ADHD attention deficithyperactivity disorder MDD major depressive disorder AUC area under receiver operating characteristic (ROC) curve

Frequent subgraphs

Local network featuresProposed

High orderPartial Pearson

Random

001

02

03

04

05

06

07

08

09

1

True

pos

itive

rate

0 01 02 03False positive rate

04 05 06 07 08 09

Figure 4 ROC curves of the different methods Receiver operatingcharacteristic (ROC) curves of the proposed method the partialcorrelation network the Pearson correlation network the higher-order functional connectivity network and the method usingonly subgraph patterns as features and quantifiable local networkfeatures The proposed method has the greatest ROC

median cingulate and paracingulate gyri [36] and the rightposterior cingulate gyrus [57]

Three of the brain regions with the most significant dif-ferences were obtained by two analysis methods quantifiablelocal network features and discriminative subgraph patterns

These included the right lenticular nucleus (putamen) leftlingual gyrus and left thalamus The right lenticular nucleus(putamen) and the left thalamus are key regions of the limbic-cortical circuit and the default network Meng et al [63]suggested that reward-related dysfunction of the right puta-men within a striatum-centered limbic-cortical circuit mayinhibit learning related to appreciating and enjoying positivelife experiences which is critical for depression recoveryAs the thalamus makes a critical connection between theamygdala and the prefrontal cortex it is well positioned forinvolvement in MDD pathophysiology Similarly the rightlingual gyrus is a key region of the visual network Jung etal reported that the volume of the lingual gyrus is associatedwith neuropsychological features of depression [57] Thelingual gyrus plays an important role in visual processingTherefore the results of this studymay be helpful in the searchfor biomarkers of MDD

42 Classification Result Analysis To study dynamic changesin functional connectivities between brain regions Chen etal used sliding time windows to construct a high-order func-tional connectivity network that can be used for classification[14] Their method has a high accuracy for diagnosing mildcognitive impairment (MCI) To show that features obtainedfrom subgraph patterns can better reflect the topologicalinformation among brain regions Du et al adopted frequentsubgraph mining technology to mine frequent subnetworksin fMRI data from people with ADHDThey used a frequent-scoring feature selection method to choose discriminativesubnetworks and kernel principal component analysis toextract features before using LIBSVM (library for supportvector machines) for classification [26] Wang et al also usedfrequent subgraphmining techniques to mine discriminativesubnetworks that were based on fMRI data from people withMCI [27] They combined traditional quantifiable proper-ties with local clustering coefficients as features and thenused a multikernel SVM for classification And Fei et al


Table 7 Classification results of different frequencies

Frequency Accuracy Sensitivity Specificity AUCHC MDD029 021 9754 10000 9667 9906014 011 8590 9217 7000 8798007 006 8029 9150 6247 8449HC healthy control group MDD major depressive disorder group AUC area under receiver operating characteristic (ROC) curve

used frequent subgraph mining techniques combined witha discriminative subnetwork mining algorithm [25] Theythen used a graph kernel-based SVM for classification Theirresults showed that frequent subgraph patterns are highlyaccurate as features of classification

Table 6 compares the accuracy specificity sensitivity andarea under the curve of the methods used in the presentstudy Different methods can produce different results usingthe same data set and similar methods can produce differentresults with different data sets Therefore we constructed apartial correlation network Pearson correlation network andhigh-order functional connectivity network with the samedata setOurmethod for constructing a high-orderminimumspanning tree network is superior to the other networks(Table 6) Compared with the traditional method the high-order minimum spanning tree network can reveal strongerand more complex interactions between brain regions andthus may significantly improve the diagnostic accuracy ratein patients withMDD Likewise construction of a high-orderminimum spanning tree functional connectivity networkmay result in better extraction of information regarding theinteractions between brain regions in original rs-fMRI timeseries

We independently classified quantifiable local networkfeatures and subgraph patterns as features on the same dataset Regardless of classification accuracy sensitivity speci-ficity or area under the curve the classification result pro-duced by themethod proposed in this study performed betterthan an analysis with only quantifiable local network featuresor with only the subgraph pattern as features (Table 6)We used the Relief method to calculate the average weightsfor subgraph pattern features quantifiable local networkfeatures and both types of feature combined As a resultour proposed method obtained the highest average weightsThe Relief algorithm is a feature-weighting algorithm inwhich weights are continuously adjusted to show correlationbetween features until the feature with the largest weight canbe identified Because this algorithm has high efficiency andcan be used to make an accurate selection of discriminativefeatures it has been widely used in many fields includingbiomedicine [64] Given the results from the classificationand feature analysis using the Reliefmethod the combinationof quantifiable local network features and subgraph patternsas features appears to effectively reflect the informationcontained in a single brain region while simultaneouslyreflecting the topological information contained in multiplebrain regions Combining these two different types of featureis likely to substantially improve diagnostic accuracy forpatients with MDD

43 Influence of Frequency on Graph Features In this exper-iment we mined frequent subnetworks from the functionalconnectivity network whichwas constructed using data fromthe HC and MDD groups This construction involved theselection of frequency which can control the number ofselected graph features However the high-order functionalconnectivity network indicates that there exists a temporalcorrelation between different low-order dynamic functionalconnectivity networksThus the size of the network can reach4005 lowast 4005 Even if the intercept associated with sparsity isvery small (01 or 005) the size of the network can reach tensof thousands of edges and the data for each subject will alsohave tens of thousands of edges Hence the number of sub-graphs will be larger whenmining frequent subgraphs whichis not conducive to the selection or analysis of subgraphfeatures Therefore we constructed the minimum spanningtree network after constructing the high-order functionalconnectivity network This method guarantees the integrityof the topological information associated with the high-orderfunctional connectivity network and reduces the complexscale of the network However each subject in the minimumspanning tree network has only 4004 edges which accountfor only 002 of high-order functional connectivity Thusfrequent subgraphmining frequencymust not be too big if itis subgraph pattern mining will not be possible In contrastif the frequency selection is too small the subgraph patternsmay be too large increasing the chance of abandoning a largenumber of discriminative subnetwork patterns in the processof frequent subgraphmining In the present study we selectedthe frequencies of the HC and MDD groups as 029 and 021respectively We left the minimum spanning tree quantifiablelocal network features unchanged and only changed thefeatures of the subgraph patterns The different frequenciesof the HC and MDD groups were used for classificationThe classification results were optimal when the frequencieswere 029 and 021 for the HC and MDD groups respectively(Table 7)

44 Influence of OptimalWeighted Parameters 119886119894 on Classifica-tion Multikernel SVMs are widely adopted in neuroimagingclassification [27] Optimizing the weighting parameter 119886119894 isvery important in such classification and optimal parameterselection will affect the classification results We tested opti-mal parameters from 0 to 1 with a step size of 01 Figure 5shows the classification accuracy of different parametersClassification accuracy was 94ndash98when different optimalparameters were used and was highest (9754) when theoptimal parameter was 04


9754

0804 06 1000 02

Optimal weighting parameter

85

90

95

100

Accu

ranc

y (

)

Figure 5 Classification accuracy of different optimal parametersOptimal parameters were selected from a range of 0-1 step size01 Accuracy of classification with different optimal parameters was94ndash97 Classification accuracy was greatest (9754) when theoptimal parameter was 04

5 Conclusion

The high-order functional connectivity network is relativelylarge making it computationally expensive to use certainelements of complex network or graph theory to calculatetopological properties In the construction of a networkprevious classification methods are based on local networkfeatures so some useful network topology information maybe lost To address this we proposed and tested the high-order minimum spanning tree to reduce computationalconsumption We combined quantifiable local network fea-tures with discriminative subgraph patterns as features andthen used a multikernel SVM for classification The resultsshowed that the high-order minimum spanning tree func-tional connectivity network could reflect dynamic changesin functional connectivities between brain regions Addi-tionally the high-order network takes into account time-varying characteristics such that the functional connectivitycan reflect stronger and more complex interactions amongmore brain regions The consistency of the results from thetwo different types of feature that is quantifiable local net-work features and frequent discriminative subgraph patternsindicates that the detected significant differences betweenbrain regions were consistent More importantly comparedwith traditional methods the proposed method appears tooffer better classification performance and thus may greatlyimprove the accuracy of MDD diagnosis In future work weplan to explore the impact of these functional connectivitiesand the relationship between the various ROIs with the aimof further improving classification performance and betterexplaining pathology

Ethical Approval

This study was approved by the medical ethics committeeof Shanxi Province and the approval certification number is2012013

Consent

All subjects have given written informed consent in accor-dance with the Declaration of Helsinki

Disclosure

The sponsors had no role in the design or execution of thestudy the collection management analysis or interpretationof the data or the preparation review or approval of themanuscript All authors have read through the manuscriptand approved it for publication Hao Guo had full access toall data in the study and takes responsibility for its integrityand the accuracy of data analysis

Conflicts of Interest

The authors declare that there are no conflicts of interest

Acknowledgments

This study was supported by research grants from theNational Natural Science Foundation of China (6137310161472270 61402318 and 61672374) theNatural Science Foun-dation of Shanxi Province (201601D021073) and the Scientificand Technological Innovation Programs ofHigher EducationInstitutions in Shanxi (2016139)

Supplementary Materials

Supplementary 1 Supplemental Text S1 Image acquisitionSupplementary 2 Supplemental Text S2 Frequent subgraphmining algorithmSupplementary 3 Supplemental Text S3 Kruskalrsquos algorithmSupplementary 4 Supplemental Text S4 Weisfeiler-LehmanalgorithmSupplementary 5 Supplemental Text S5 The methods andresults of other contrast networksSupplementary 6 Supplemental Table S1 Results of multiplelinear regression analysis between network properties andconfounding variablesSupplementary 7 Supplemental Table S2 All regions ofinterest (abbreviations and full names)Supplementary 8 Supplemental Table S3 Number of fre-quent subgraph edgesSupplementary 9 Supplemental Figure S1 Sliding windowSupplementary 10 Supplemental Figure S2 Discriminationof different subgraphsSupplementary 11 Supplemental Figure S3 Minimum span-ning tree functional connectivities and degree of the corre-sponding nodeSupplementary 12 Supplemental Figure S4 Subgraphs andconnected patterns in HC and MDD groupsSupplementary 13 Supplemental Figure S5 Discriminativebrain regions and corresponding degree


References

[1] Y Liu M Liang Y Zhou et al ldquoDisrupted small-worldnetworks in schizophreniardquo Brain vol 131 no 4 pp 945ndash9612008

[2] M-E Lynall D S Bassett R Kerwin et al ldquoFunctionalconnectivity and brain networks in schizophreniardquoThe Journalof Neuroscience vol 30 no 28 pp 9477ndash9487 2010

[3] S Micheloyannis E Pachou C J Stam et al ldquoSmall-world net-works and disturbed functional connectivity in schizophreniardquoSchizophrenia Research vol 87 no 1ndash3 pp 60ndash66 2006

[4] M Rubinov S A Knock C J Stam et al ldquoSmall-worldproperties of nonlinear brain activity in schizophreniardquoHumanBrain Mapping vol 30 no 2 pp 403ndash416 2009

[5] K Supekar VMenon D RubinMMusen andM D GreiciusldquoNetwork analysis of intrinsic functional brain connectivity inAlzheimerrsquos diseaserdquo PLoS Computational Biology vol 4 no 6Article ID e1000100 2008

[6] Y He Z Chen and A Evans ldquoStructural insights into aber-rant topological patterns of large-scale cortical networks inAlzheimerrsquos diseaserdquoThe Journal of Neuroscience vol 28 no 18pp 4756ndash4766 2008

[7] C J Stam ldquoUse of magnetoencephalography (MEG) to studyfunctional brain networks in neurodegenerative disordersrdquoJournal of the Neurological Sciences vol 289 no 1-2 pp 128ndash134 2010

[8] D E Van et al ldquoLong-term effects of temporal lobe epilepsy onlocal neural networks a graph theoretical analysis of corticog-raphy recordingsrdquo Plos One vol 4 no 11 2009

[9] S Pieper et al ldquoNetwork-level analysis of cortical thickness ofthe epileptic brainrdquo Neuroimage vol 52 no 4 pp 1302ndash13132010

[10] M-T Horstmann S Bialonski N Noennig et al ldquoState depen-dent properties of epileptic brain networks Comparative graph-theoretical analyses of simultaneously recordedEEGandMEGrdquoClinical Neurophysiology vol 121 no 2 pp 172ndash185 2010

[11] L Wang C Zhu Y He et al ldquoAltered small-world brain func-tional networks in children with attention-deficithyperactivitydisorderrdquo Human Brain Mapping vol 30 no 2 pp 638ndash6492009

[12] F D V Fallani L Astolfi F Cincotti et al ldquoEvaluation of thebrain network organization from EEG signals a preliminaryevidence in stroke patientrdquo Anatomical Record Advances inIntegrative Anatomy Evolutionary Biology vol 292 no 12 pp2023ndash2031 2009

[13] J Wang L Wang and Y Zang ldquoParcellation-dependent small-world brain functional networks a resting-state fMRi studyrdquoHuman Brain Mapping vol 30 no 5 pp 1511ndash1523 2009

[14] X Chen H Zhang Y Gao C-Y Wee G Li and D ShenldquoHigh-order resting-state functional connectivity network forMCI classificationrdquo Human Brain Mapping vol 37 no 9 pp3282ndash3296 2016

[15] E Damaraju E A Allen A Belger et al ldquoDynamic functionalconnectivity analysis reveals transient states of dysconnectivityin schizophreniardquo NeuroImage Clinical vol 5 pp 298ndash3082014

[16] E A Allen E Damaraju SM Plis E B Erhardt T Eichele andVD Calhoun ldquoTrackingwhole-brain connectivity dynamics inthe resting staterdquo Cerebral Cortex vol 24 no 3 pp 663ndash6762014

[17] H Zhang X Chen F Shi et al ldquoTopographical Information-Based High-Order Functional Connectivity and Its Application

in Abnormality Detection for Mild Cognitive ImpairmentrdquoJournal of Alzheimerrsquos Disease vol 54 no 3 pp 1095ndash1112 2016

[18] R M Hutchison T Womelsdorf E A Allen et al ldquoDynamicfunctional connectivity promise issues and interpretationsrdquoNeuroImage vol 80 pp 360ndash378 2013

[19] C Vikas ldquoMinimum Spanning Tree Algorithmrdquo InternationalJournal of Computer Applications vol 1 no 8 pp 39ndash45 2010

[20] U Lee S Kim and K-Y Jung ldquoClassification of epilepsy typesthrough global network analysis of scalp electroencephalo-gramsrdquo Physical Review E Statistical Nonlinear and SoftMatterPhysics vol 73 no 4 Article ID 041920 2006

[21] V D Edwin et al ldquoEpilepsy surgery outcome and functionalnetwork alterations in longitudinalMEG aminimum spanningtree analysisrdquo Neuroimage vol 86 no 1 pp 354ndash363 2014

[22] P Tewarie A HillebrandMM Schoonheim et al ldquoFunctionalbrain network analysis using minimum spanning trees inMultiple Sclerosis an MEG source-space studyrdquo NeuroImagevol 88 pp 308ndash318 2014

[23] C McDiarmid T Johnson and H S Stone ldquoOn finding aminimum spanning tree in a network with random weightsrdquoRandom Structures amp Algorithms vol 10 no 1-2 pp 187ndash2041997

[24] P Tewarie E Van Dellen A Hillebrand and C J Stam ldquoTheminimumspanning tree an unbiasedmethod for brain networkanalysisrdquo NeuroImage vol 104 pp 177ndash188 2015

[25] F Fei B Jie and D Zhang ldquoFrequent and discriminative sub-network mining for mild cognitive impairment classificationrdquoBrain Connectivity vol 4 no 5 pp 347ndash360 2014

[26] J Du L Wang B Jie and D Zhang ldquoNetwork-based classifica-tion of ADHD patients using discriminative subnetwork selec-tion and graph kernel PCArdquoComputerizedMedical Imaging andGraphics vol 52 pp 82ndash88 2016

[27] L Wang F Fei B Jie and D Zhang ldquoCombining multiplenetwork features for mild cognitive impairment classificationrdquoin Proceedings of the 14th IEEE International Conference on DataMining Workshops ICDMW pp 996ndash1003 2014

[28] Y Li S Dou EWang et al ldquoEvaluation of brain default networkfMRI of insomnia with depression patients at resting staterdquo LifeScience Journal vol 11 no 8 pp 794ndash801 2014

[29] I M Veer C F Beckmann M-J van Tol et al ldquoWhole brainresting-state analysis reveals decreased functional connectivityin major depressionrdquo Frontiers in Systems Neuroscience vol 4article 41 2009

[30] M D Greicius B H Flores and VMenon ldquoResting-state func-tional connectivity in major depressionn abnormally increasedcontributions from subgenual cingulate cortex and thalamusrdquoBiological Psychiatry vol 62 no 5 pp 429ndash437 2007

[31] A Anand Y Li Y Wang et al ldquoActivity and connectivityof brain mood regulating circuit in depression a functionalmagnetic resonance studyrdquo Biological Psychiatry vol 57 no 10pp 1079ndash1088 2005

[32] H Tao S Guo T Ge et al ldquoDepression uncouples brain hatecircuitrdquoMolecular Psychiatry vol 18 no 1 pp 101ndash111 2013

[33] X Wang Y Ren Y Yang W Zhang and N N Xiong ldquoAweighted discriminative dictionary learningmethod for depres-sion disorder classification using fMRI datardquo in Proceedingsof the IEEE International Conferences on Big Data and CloudComputing pp 618ndash623 IEEE October 2016

[34] Y Yue Y Yuan Z Hou W Jiang F Bai and Z Zhang ldquo2129ndash Abnormal functional connectivity of amygdala in late onsetdepression was associated with cognitive deficits but not withdepressive severityrdquo European Psychiatry vol 28 p 1 2013


[35] Y I Sheline J L Price Z Yan andMAMintun ldquoResting-statefunctional MRI in depression unmasks increased connectivitybetween networks via the dorsal nexusrdquo Proceedings of theNational Acadamy of Sciences of the United States of Americavol 107 no 24 pp 11020ndash11025 2010

[36] H Guo X Cao Z Liu and J Chen ldquoAbnormal functional brainnetwork metrics for machine learning classifier in depressionpatients identificationrdquo Research Journal of Applied SciencesEngineering amp Technology vol 5 no 10 pp 3015ndash3020 2013

[37] L Qiao H Zhang M Kim S Teng L Zhang and DShen ldquoEstimating functional brain networks by incorporatinga modularity priorrdquo NeuroImage vol 141 pp 399ndash407 2016

[38] M-L Wong C Dong V Andreev M Arcos-Burgos and JLicinio ldquoPrediction of susceptibility to major depression by amodel of interactions of multiple functional genetic variantsand environmental factorsrdquoMolecular Psychiatry vol 17 no 6pp 624ndash633 2012

[39] F Liu W Guo and J-P Fouche ldquoMultivariate classification ofsocial anxiety disorder using whole brain functional connectiv-ityrdquo Brain Structure amp Function vol 220 no 1 pp 101ndash115 2015

[40] N Tzouriomazoyer et al ldquoAutomated anatomical labeling ofactivations in SPMusing amacroscopic anatomical parcellationof the MNI MRI single-subject brainrdquo Neuroimage vol 15 no1 p 273 2002

[41] J Kruskal ldquoOn the shortest spanning subtree of a graph andthe traveling salesman problemrdquo Proceedings of the AmericanMathematical Society vol 7 no 1 pp 48ndash50 1956

[42] A M Elfeki and J Bahrawi ldquokolmogorov-smirnov testrdquo Inter-national Encyclopedia of Statistical Science vol 10 no 1 pp 718ndash720 2014

[43] A Roberto et al BenjaminindashHochberg False Discovery Rate(FDR) as A Function of The P-Value 2014

[44] M Polajnar and J Demsar ldquoSmall network completion usingfrequent subnetworksrdquo Intelligent Data Analysis vol 19 no 1pp 89ndash108 2014

[45] W Lin X Xiao and G Ghinita ldquoLarge-scale frequent subgraphmining inMapReducerdquo in Proceedings of the 30th IEEE Interna-tional Conference on Data Engineering (ICDE rsquo14) pp 844ndash855April 2014

[46] M Kuramochi and G Karypis ldquoFrequent subgraph discoveryrdquoin Proceedings of the 1st IEEE International Conference onData Mining (ICDM rsquo01) pp 313ndash320 Piscataway NJ USADecember 2001

[47] A Inokuchi T Washio and H Motoda ldquoAn apriori-basedalgorithm for mining frequent substructures from graph datardquoLecture Notes in Computer Science (including subseries LectureNotes in Artificial Intelligence and Lecture Notes in Bioinformat-ics) Preface vol 1910 pp 13ndash23 2000

[48] M Kuramochi and G Karypis ldquoAn efficient algorithm for dis-covering frequent subgraphsrdquo IEEE Transactions on Knowledgeand Data Engineering vol 16 no 9 pp 1038ndash1051 2004

[49] S F Rosario and K Thangadurai RELIEF Feature SelectionApproach 2015

[50] X Yan and J Han ldquogSpan graph-based substructure patternminingrdquo inProceedings of the 2nd IEEE International Conferenceon Data Mining (ICDM rsquo02) pp 721ndash724 Maebashi City JapanDecember 2002

[51] X Kong et al Discriminative Feature Selection for UncertainGraph Classification 2013 Discriminative Feature Selection forUncertain Graph Classification

[52] G R G Lanckriet et al ldquoLearning the kernel matrix withsemi-definite programmingrdquo in Proceedings of the NineteenthInternational Conference 2002

[53] N Shervashidze P Schweitzer E J van Leeuwen K Mehlhornand K M Borgwardt ldquoWeisfeiler-Lehman graph kernelsrdquoJournal of Machine Learning Research (JMLR) vol 12 no 3 pp2539ndash2561 2011

[54] C Ma J Ding J Li et al ldquoResting-state functional connectivitybias of middle temporal Gyrus and Caudate with altered Graymatter volume in major depressionrdquo PLoS ONE vol 7 no 9Article ID e45263 2012

[55] L Qiu S Lui W Kuang et al ldquoRegional increases of corticalthickness in untreated first-episode major depressive disorderrdquoTranslational Psychiatry vol 4 article e378 2014

[56] M D Sacchet M C Camacho E E Livermore E A Thomasand I H Gotlib ldquoAccelerated aging of the putamen in patientswith major depressive disorderrdquo Journal of Psychiatry amp Neuro-science vol 42 no 3 pp 164ndash171 2017

[57] J Jung J Kang E Won et al ldquoImpact of lingual gyrus vol-ume on antidepressant response and neurocognitive functionsin Major Depressive Disorder A voxel-based morphometrystudyrdquo Journal of Affective Disorders vol 169 pp 179ndash187 2014

[58] F Junfang W Qian and W Bin ldquoAmplitude of low-frequencyoscillations in major depression as revealed by resting statefunctional magnetic resonance imagingrdquo Journal of ClinicalRadiology 2015

[59] D Cotter D Mackay G Chana C Beasley S Landau and IP Everall ldquoReduced neuronal size and glial cell density in area9 of the dorsolateral prefrontal cortex in subjects with majordepressive disorderrdquoCerebral Cortex vol 12 no 4 pp 386ndash3942002

[60] Y-T Chen M-W Huang I-C Hung H-Y Lane and C-JHou ldquoRight and left amygdalae activation in patientswithmajordepression receiving antidepressant treatment as revealed byfMRIrdquo Behavioral and Brain Functions vol 10 article 101 2014

[61] H J Rachel Differences between Healthy Controls and remittedMajor Depression for left amygdala seed 2014

[62] M Ye T Yang P Qing X Lei J Qiu and G Liu ldquoChangesof functional brain networks in major depressive disorder Agraph theoretical analysis of resting-state fMRIrdquo PLoSONE vol10 no 9 Article ID e0133775 2015

[63] C Meng F Brandl M Tahmasian et al ldquoAberrant topologyof striatumrsquos connectivity is associated with the number ofepisodes in depressionrdquo Brain vol 137 no 2 pp 598ndash609 2014

[64] X Liu J Liu Z Feng X Xu and J Tang ldquoMass classificationin mammogram with semi-supervised relief based featureselectionrdquo in Proceedings of the 5th International Conference onGraphic and Image Processing ICGIP 2013 China October 2013

Submit your manuscripts athttpswwwhindawicom

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


MEDIATORSINFLAMMATION

of


Behavioural Neurology

EndocrinologyInternational Journal of



Disease Markers


BioMed Research International

OncologyJournal of



Oxidative Medicine and Cellular Longevity


PPAR Research

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

ObesityJournal of



Computational and Mathematical Methods in Medicine

OphthalmologyJournal of


Diabetes ResearchJournal of



Research and TreatmentAIDS


Gastroenterology Research and Practice


Parkinsonrsquos Disease

Evidence-Based Complementary and Alternative Medicine

Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom


Themost commonmethod for constructing a high-orderfunctional network is the dynamic sliding window methodin which the whole rs-fMRI time series is divided into severaltime windows [18] A low-order functional connectivity net-work is built in each time window and then all the low-ordernetworks are stacked A clustering algorithm is performedto divide all relevant time series into several clusters Theaverage time series of each cluster is then taken as a newnodeand the Pearson correlation coefficient is calculated betweeneach node pair as the weight of connectivity [14]

In this method clustering is employed to decrease theassociated computational costs and classification accuracyis greatly influenced by the randomness of the selectionof initial clustering centers and the number of clustersHowever because the time series of all connectivities withineach cluster are averaged the network loses neurologicalinterpretability

In the present study we used the minimum spanningtree method [19] to reduce the computational cost while pre-serving the core framework of networks A classic approachin graph theory this unbiased method greatly simplifies thenetwork structure while preserving its core framework thusavoiding the influences of network sparseness and otherparameters on network structure It also guarantees thenetworkrsquos neurological interpretability and has been widelyused in previous studies [20ndash22]

Furthermore the traditional feature extraction methodin minimum spanning tree networks uses a quantifiablenetworkwith local features for classification of brain diseasessuch as degree clustering coefficient minimum path lengthand eccentricity [23 24] However a clear shortcomingof this method was the chance that some of the usefultopology information in the network (including connectionpatterns in the sample itself and the common connectionpatterns between the samples) would be lost resulting inreduced classifier performance Frequent subgraph miningtechnology was proposed to mine discriminative subgraphpattern features for machine-learning classification of braindiseases [25 26] Subgraph pattern features could account forthe connection pattern information between multiple brainregions but it was not sensitive to changes in single brainregions [27] Therefore both methods can lead to loss ofsample information

Here we propose a novel feature extraction method thatcombines quantifiable local network features with subgraphpattern features Specifically we computed degree eccentric-ity and betweenness centrality of each brain region as localnetwork features and extracted the subgraph features usinga frequent subgraph mining method for a group of healthycontrols (HC) and a group of people with major depressivedisorder (MDD)Then a kernel function for each type of fea-ture was constructed namely a vector kernel (local networkfeatures) and a graph kernel (subgraph features) Finallythe two kernel matrices were combined and a multikernelsupport vector machine was constructed as a classifier Theproposed method achieves better classification performancethan traditionalmethods that use only a single type of feature

Table 1 Subject demographics and clinical characteristics

HC MDD 119901 values

Age 17ndash51(266 plusmn 94)

17ndash49(284 plusmn 968) 044a

Sex (malefemale) 1315 1523 057b

Handedness (RL) 280 380

HAMD NA 15ndash42(228 plusmn 133)

Data are minimumndashmaximum (mean plusmn standard deviation) HAMD 24-item Hamilton scale atwo-sample two-tailed 119905-test btwo-tailed Pearsonrsquoschi-squared test

2 Materials and Methods

21 Proposed Framework Figure 1 shows the flowchart ofthe proposed method which includes four main steps (1)data acquisition andpreprocessing (2) network constructionin which a high-order functional connectivity network isconstructed first followed by construction of a minimumspanning tree network (3) feature extraction and selectionin which two types of feature are extracted and selected(the first is used to calculate quantifiable local networkfeatures (degree betweenness centrality and eccentricity)and uses the KolmogorovndashSmirnov test for feature selectionand the second is used to mine frequent subgraphs from theHC and MDD groups and selects the most discriminativesubnetworks as the subgraph patterns) (4) constructionof a classification model in which the kernel matrix ofthe two types of feature is calculated The multiple-kernelsupport vectormachine (SVM) is adopted to combine the twoheterogeneous kernels enabling the distinction of individualswith MDD from healthy controls

22 Data Acquisition and Preprocessing The study was car-ried out in accordance with the recommendations of themedical ethics committee of Shanxi Province (referencenumber 2012013) All subjects provided written informedconsent in accordance with the Declaration of HelsinkiTwenty-eight healthy subjects and thirty-eight people withMDD underwent rs-fMRI in a 3T scanner (Siemens Trio3-Tesla scanner Siemens Erlangen Germany) Participantdemographic information is shown in Table 1

Data collection was completed at the First Hospital ofShanxi Medical University Radiologists familiar with fMRIperformed all scans During each scan the participant wasasked to relax with their eyes closed and not think aboutanything in particular but to stay awake and avoid fallingasleep Each scan consisted of 248 contiguous echo-planarimaging (EPI) functional volumes (33 axial slices repetitiontime (TR) = 2000ms echo time (TE) = 30ms thicknessskip= 40mm field of view (FOV) = 192 times 192mm matrix =64 times 64mm and flip angle = 90∘) The first 10 volumes inthe time series were discarded to account for magnetizationstabilization See Supplemental Text S1 for detailed scanningparameters

Data preprocessing was performed in SPM8 (httpwwwfilionuclacukspm)with slice-timing and head-movementcorrections Two samples containing a translation of more


gSpan

gSpan


Selectedfeatures

NC

MDD





subgraphs

Subgraph features


Graph kernel matrix


Multikernel SVM

Degree

Eccentricity




















119909(119897) (119896) = [119909(119897)1 (119896) 119909(119897)2 (119896) 119909(119897)119877 (119896)] isin 119877119873times119877 (2)


119862(119897)119871 (119896) = (119909(119897)119894 (119896) 119862(119897)119894119895 (119896)) (119896 = 1 2 119896) (3)


119910(119897)119894119895 = [119862(119897)119894119895 (1) 119862(119897)119894119895 (2) 119862(119897)119894119895 (119896)] isin 119877119870 (4)



119867(119897)119894119895119901119902 = corr (119910(119897)119894119895 119910(119897)119901119902) (5)


119866(119897)119867 = (119910(119897)119894119895 119867(119897)119894119895119901119902) (6)




119895isin119873

119886119894119895





120588(119894)ℎ119895















119866 (7)



119878 (119892119904) = 10038161003816100381610038161003816119891119902 (119892119904 | 119866119875) minus 119891119902 (119892119904 | 119866119899)10038161003816100381610038161003816 (8)



119896 (119909 119910) =119872

sum119894=1

119886119894119896119894 (119909 119910) st119872

sum119894=1

119886119894 = 1 (9)




119896lowast (119909 119910) = 119896 (119909 119910)radic119896 (119909 119909) 119896 (119910 119910)

(10)



3 Results











4 Discussion
















300

600

900

1200

1500

1800

2100

Aver

age v

alue

of w

eigh

t

lowast

lowastlowastlowast

lowastp lt 005








This study MDD 6306 5056 8737 7102


This study MDD 6667 4643 8158 7446







Frequent subgraphs



Random

001

02

03

04

05

06

07

08

09

1

True

pos

itive

rate


04 05 06 07 08 09















9754

0804 06 1000 02


85

90

95

100

Accu

ranc

y (

)


5 Conclusion


Ethical Approval


Consent


Disclosure




Acknowledgments





References








































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















gSpan

gSpan


Selectedfeatures

NC

MDD





subgraphs

Subgraph features


Graph kernel matrix


Multikernel SVM

Degree

Eccentricity




















119909(119897) (119896) = [119909(119897)1 (119896) 119909(119897)2 (119896) 119909(119897)119877 (119896)] isin 119877119873times119877 (2)


119862(119897)119871 (119896) = (119909(119897)119894 (119896) 119862(119897)119894119895 (119896)) (119896 = 1 2 119896) (3)


119910(119897)119894119895 = [119862(119897)119894119895 (1) 119862(119897)119894119895 (2) 119862(119897)119894119895 (119896)] isin 119877119870 (4)



119867(119897)119894119895119901119902 = corr (119910(119897)119894119895 119910(119897)119901119902) (5)


119866(119897)119867 = (119910(119897)119894119895 119867(119897)119894119895119901119902) (6)




119895isin119873

119886119894119895





120588(119894)ℎ119895















119866 (7)



119878 (119892119904) = 10038161003816100381610038161003816119891119902 (119892119904 | 119866119875) minus 119891119902 (119892119904 | 119866119899)10038161003816100381610038161003816 (8)



119896 (119909 119910) =119872

sum119894=1

119886119894119896119894 (119909 119910) st119872

sum119894=1

119886119894 = 1 (9)




119896lowast (119909 119910) = 119896 (119909 119910)radic119896 (119909 119909) 119896 (119910 119910)

(10)



3 Results











4 Discussion
















300

600

900

1200

1500

1800

2100

Aver

age v

alue

of w

eigh

t

lowast

lowastlowastlowast

lowastp lt 005








This study MDD 6306 5056 8737 7102


This study MDD 6667 4643 8158 7446







Frequent subgraphs



Random

001

02

03

04

05

06

07

08

09

1

True

pos

itive

rate


04 05 06 07 08 09















9754

0804 06 1000 02


85

90

95

100

Accu

ranc

y (

)


5 Conclusion


Ethical Approval


Consent


Disclosure




Acknowledgments





References








































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of




























119909(119897) (119896) = [119909(119897)1 (119896) 119909(119897)2 (119896) 119909(119897)119877 (119896)] isin 119877119873times119877 (2)


119862(119897)119871 (119896) = (119909(119897)119894 (119896) 119862(119897)119894119895 (119896)) (119896 = 1 2 119896) (3)


119910(119897)119894119895 = [119862(119897)119894119895 (1) 119862(119897)119894119895 (2) 119862(119897)119894119895 (119896)] isin 119877119870 (4)



119867(119897)119894119895119901119902 = corr (119910(119897)119894119895 119910(119897)119901119902) (5)


119866(119897)119867 = (119910(119897)119894119895 119867(119897)119894119895119901119902) (6)




119895isin119873

119886119894119895





120588(119894)ℎ119895















119866 (7)



119878 (119892119904) = 10038161003816100381610038161003816119891119902 (119892119904 | 119866119875) minus 119891119902 (119892119904 | 119866119899)10038161003816100381610038161003816 (8)



119896 (119909 119910) =119872

sum119894=1

119886119894119896119894 (119909 119910) st119872

sum119894=1

119886119894 = 1 (9)




119896lowast (119909 119910) = 119896 (119909 119910)radic119896 (119909 119909) 119896 (119910 119910)

(10)



3 Results











4 Discussion
















300

600

900

1200

1500

1800

2100

Aver

age v

alue

of w

eigh

t

lowast

lowastlowastlowast

lowastp lt 005








This study MDD 6306 5056 8737 7102


This study MDD 6667 4643 8158 7446







Frequent subgraphs



Random

001

02

03

04

05

06

07

08

09

1

True

pos

itive

rate


04 05 06 07 08 09















9754

0804 06 1000 02


85

90

95

100

Accu

ranc

y (

)


5 Conclusion


Ethical Approval


Consent


Disclosure




Acknowledgments





References








































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of



















119895isin119873

119886119894119895





120588(119894)ℎ119895















119866 (7)



119878 (119892119904) = 10038161003816100381610038161003816119891119902 (119892119904 | 119866119875) minus 119891119902 (119892119904 | 119866119899)10038161003816100381610038161003816 (8)



119896 (119909 119910) =119872

sum119894=1

119886119894119896119894 (119909 119910) st119872

sum119894=1

119886119894 = 1 (9)




119896lowast (119909 119910) = 119896 (119909 119910)radic119896 (119909 119909) 119896 (119910 119910)

(10)



3 Results











4 Discussion
















300

600

900

1200

1500

1800

2100

Aver

age v

alue

of w

eigh

t

lowast

lowastlowastlowast

lowastp lt 005








This study MDD 6306 5056 8737 7102


This study MDD 6667 4643 8158 7446







Frequent subgraphs



Random

001

02

03

04

05

06

07

08

09

1

True

pos

itive

rate


04 05 06 07 08 09















9754

0804 06 1000 02


85

90

95

100

Accu

ranc

y (

)


5 Conclusion


Ethical Approval


Consent


Disclosure




Acknowledgments





References








































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of



















119866 (7)



119878 (119892119904) = 10038161003816100381610038161003816119891119902 (119892119904 | 119866119875) minus 119891119902 (119892119904 | 119866119899)10038161003816100381610038161003816 (8)



119896 (119909 119910) =119872

sum119894=1

119886119894119896119894 (119909 119910) st119872

sum119894=1

119886119894 = 1 (9)




119896lowast (119909 119910) = 119896 (119909 119910)radic119896 (119909 119909) 119896 (119910 119910)

(10)



3 Results











4 Discussion
















300

600

900

1200

1500

1800

2100

Aver

age v

alue

of w

eigh

t

lowast

lowastlowastlowast

lowastp lt 005








This study MDD 6306 5056 8737 7102


This study MDD 6667 4643 8158 7446







Frequent subgraphs



Random

001

02

03

04

05

06

07

08

09

1

True

pos

itive

rate


04 05 06 07 08 09















9754

0804 06 1000 02


85

90

95

100

Accu

ranc

y (

)


5 Conclusion


Ethical Approval


Consent


Disclosure




Acknowledgments





References








































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

























4 Discussion
















300

600

900

1200

1500

1800

2100

Aver

age v

alue

of w

eigh

t

lowast

lowastlowastlowast

lowastp lt 005








This study MDD 6306 5056 8737 7102


This study MDD 6667 4643 8158 7446







Frequent subgraphs



Random

001

02

03

04

05

06

07

08

09

1

True

pos

itive

rate


04 05 06 07 08 09















9754

0804 06 1000 02


85

90

95

100

Accu

ranc

y (

)


5 Conclusion


Ethical Approval


Consent


Disclosure




Acknowledgments





References








































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of






























300

600

900

1200

1500

1800

2100

Aver

age v

alue

of w

eigh

t

lowast

lowastlowastlowast

lowastp lt 005








This study MDD 6306 5056 8737 7102


This study MDD 6667 4643 8158 7446







Frequent subgraphs



Random

001

02

03

04

05

06

07

08

09

1

True

pos

itive

rate


04 05 06 07 08 09















9754

0804 06 1000 02


85

90

95

100

Accu

ranc

y (

)


5 Conclusion


Ethical Approval


Consent


Disclosure




Acknowledgments





References








































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of




















This study MDD 6306 5056 8737 7102


This study MDD 6667 4643 8158 7446







Frequent subgraphs



Random

001

02

03

04

05

06

07

08

09

1

True

pos

itive

rate


04 05 06 07 08 09















9754

0804 06 1000 02


85

90

95

100

Accu

ranc

y (

)


5 Conclusion


Ethical Approval


Consent


Disclosure




Acknowledgments





References








































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

























9754

0804 06 1000 02


85

90

95

100

Accu

ranc

y (

)


5 Conclusion


Ethical Approval


Consent


Disclosure




Acknowledgments





References








































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















References








































































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of




















































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
















machine-learning classifier for patients with major depressive...

Documents