Disrupted functional brain connectome in individuals at risks for AD

ARCHIVAL REPORT D o a se Jin hil an Ba fwh dis rodr Me ima the nts w bra high Re show con alyse con l key pri twor sys pat aM spec of alida Co opol fin m of me Ke ula A Am tw Nu aMCIhave impaired structural integrity (5,6) and functional connec- tiv rup ma tur con top var no mo hu dis stu ne wit tio wo aM of nan of he ph connectome in aMCI patients during a memory task. Fro Ad Rec 00 htt ity (7–10). However, whether aMCI patients also exhibit a dis- ted topological organization in their whole-brain networks re- ins largely unknown. Recent studies have suggested that human whole-brain struc- al and functional networks (i.e., the connectome [11,12]) can be structed using multimodal neuroimaging data and that their ological organization can be characterized quantitatively using ious graph theory metrics (13–15). With these metrics, many ntrivial organizational principles, including small-worldness, dularity, and highly connected hubs, have been observed in the man brain connectome. Moreover, these network properties are rupted in many neuropsychiatric disorders (13,16–18). These dies have accelerated the process of mapping the human con- Here, we employed resting-state functional magnetic reso- nance imaging (R-fMRI) to investigate the topological changes in the functional connectome in patients with aMCI. R-fMRI measures intrinsic or spontaneous neuronal activity of the brain (26,27) and has been applied to reveal aMCI-related breakdowns in functional brain synchronization (7,9,28). The current study focuses exclu- sively on the topological architecture of the intrinsic functional brain connectome in aMCI. Specifically, we sought to determine whether aMCI disrupts the topological organization of the whole- brain functional network and, if so, whether those topological ab- normalities are associatedwith individual clinical or behavioral vari- ables. Furthermore, we examined whether these abnormalities differentiated patients with aMCI from healthy elderly individuals. Methods andMaterials Participants Eighty-four right-handed participants, comprising 37 patients with aMCI (17 men and 20 women) and 47 sex-, age-, and educa- tion-matchedhealthy control subjects (HC: 20menand27women), participated in this study. The patients were recruited from the memory clinic of the neurology department of Xuanwu Hospital, CapitalMedical University, Beijing, China. The control subjectswere recruited from the local community using advertisements. At the time of the study, none of the patients had ever been treated with m the State Key Laboratory of CognitiveNeuroscience and Learning (JW, ZD, MX, YHe), Beijing Normal University; Laboratory for Functional Con- nectome andDevelopment (XZu), Key Laboratory of Behavioral Science, Magnetic Resonance Imaging Research Center, Institute of Psychology, Chinese Academy of Sciences; and Departments of Radiology (ZZ) and Neurology (XZh, JJ, YHa), Xuanwu Hospital, Capital Medical University, Beijing, China. dress correspondence to YongHe, Ph.D., BeijingNormal University, State Key Laboratory of Cognitive Neuroscience and Learning, No 19 Xinjiek- ouwai Street, Haidian District, Beijing 100875, China; E-mail: yong. he@bnu.edu.cn. eived Jan 12, 2012; revised Mar 15, 2012; accepted Mar 26, 2012. BIOL PSYCHIATRY 2012;xx:xxx06-3223/$36.00 p://dx.doi.org/10.1016/j.biopsych.2012.03.026 © 2012 Society of Biological Psychiatry isrupted Functional Brain C t Risk for Alzheimer’s Disea hui Wang, Xinian Zuo, Zhengjia Dai, Mingrui Xia, Z d Yong He ckground: Alzheimer’sdiseasedisrupts the topological architectureo ruption is present in amnesticmild cognitive impairment (aMCI), the p thods: We employed resting-state functional magnetic resonance topological organization of the functional connectome of 37 patie in networks were derived from wavelet-based correlations of both sults: In the frequency interval .031–.063 Hz, the aMCI patients nectomecomparedwith control subjects. Further graph theory an nectome in the aMCI group. Moreover, the disease targeted severa marily in the intramodule connections within the default-mode ne tems. Intriguingly, the topological aberrations correlated with the CI from healthy elderly individuals with a sensitivity of 86.5% and a our findings across different large-scale parcellation schemes and v nclusions: This study demonstrates a disruption of whole-brain t ding provides novel insights into the pathophysiological mechanis trics as a disease biomarker. yWords: Connectivity, connectomics, default-mode, MCI, mod- rity, small-world lzheimer’s disease (AD) is a progressive, neurodegenerative disease characterized by a decline in cognitive andmemory functions likely caused by aberrant neuronal circuitry (1–3). nestic mild cognitive impairment (aMCI), a transition state be- eennormal aging andAD, has ahigh risk of progressing toAD (4). merous studies have reported that the brains of patients with nnectome in Individuals ian Zhao, Xiaoling Zhao, Jianping Jia, Ying Han, ole-brainconnectivity (i.e., theconnectome);however,whether this omal stage of Alzheimer’s disease, remains largely unknown. ging and graph theory approaches to systematically investigate ith aMCI and 47 healthy control subjects. Frequency-dependent - and low-resolution parcellation units. ed an overall decreased functional connectivity of their brain s of this frequencyband revealed an increasedpath lengthof the nodes predominantly in the default-mode regions and key links k and the intermodule connections among different functional ients’ memory performance and differentiated individuals with ificity of 85.1%. Finally, we demonstrated a high reproducibility ted the test-retest reliability of our network-based approaches. ogical organization of the functional connectome in aMCI. Our aMCI and highlights the potential for using connectome-based ctome in healthy and diseased states. Specifically, in patients h AD, several research groups have reported topological altera- ns in the whole-brain connectome, including a loss of small- rldness and a redistribution of hubs (19–23). With respect to CI, only two studies have explored the topological organization the whole-brain connectome. Using structural magnetic reso- ce imaging, Yao et al. (24) found no differences in the topology cortical-thickness networks between patients with aMCI and althy control subjects. However, using magnetoencephalogra- y data, Buldu et al. (25) reported reorganization of the functional spe no ter are a s Me (CD De de set stu ap tio ob Da (Er sity siti eye 23 Da (ht Ce do vo tio spa reg con R-f (22 tw Ne wh interregional resting-state functional connectivity (RSFC). To define ne (RO atla the coe wa wa .03 ou spo Bo con Ne and tur me CW and Q˜m and sur ity) ret Ru pre Sta po by me me tain rec gro ica wh loc ity (NB firs (i.e min the der (10 line der det Va rel im MM act Ne mi ma Table 1. Demographics and Clinical Characteristics of the Participants Ge F Ag Edu MM CD CD AV R AV R AV No usi val ing hea 2 BIOL PSYCHIATRY 2012;xx:xxx J. Wang et al. ww cificmedications, such as anti-acetylcholinesterase drugs. Diag- ses of aMCI were made by experienced neurologists using Pe- sen’s criteria (4,29). The detailed inclusion and exclusion criteria described in Supplement 1. Eachparticipantwas assessedusing tandardized clinical evaluation protocol that included the Mini- ntal State Examination (MMSE) (30), the Clock Drawing Test T), theAuditoryVerbal LearningTest (AVLT) (31), and theClinical mentia Rating Scale (32). In Table 1, we present the detailed mographics and clinical characteristics of the participants. Data- s from a subset of the general population have been used to dy local brain activity in patients with aMCI (33). This study was proved by the Medical Research Ethics Committee and Institu- nal ReviewBoardof XuanwuHospital, and informedconsentwas tained from each participant. ta Acquisition All participantswere scanned using a 3.0 T Siemens Trio scanner langen, Germany) at Xuanwu Hospital, Capital Medical Univer- , within a single session (Supplement 1). During the data acqui- on, participantswere asked to lie quietly in the scannerwith their s closed. The scan lasted for 478 seconds in total and included 9 volumes for each participant. ta Preprocessing Data preprocessing was performed using the SPM8 package tp://www.fil.ion.ucl.ac.uk/spm/software/SPM8/; Wellcome Trust nter for Neuroimaging, University College London, United King- m; Supplement 1) and included the removal of the first five lumes, correction for time offsets between slices and head mo- n, spatial normalization to the Montreal Neurological Institute ce, temporal high-pass filtering (cutoff frequency� .01 Hz), and ressionofnuisance signals of sixhead-motionprofiles.Given the troversy of removing the global signal in the preprocessing of MRI data (34,35), we did not regress the global signal out ,36,37). Notably, the head-motion profiles were matched be- een the aMCI and HC groups (p� .248 in any direction). twork Construction In this study, brain networkswere constructed at themacroscale ere nodes represented brain regions and edges represented HC (n� 47) aMCI (n� 37) p Value nder (Male/ emale) 20/27 17/20 .756a e (Years) 50–79 (63.4� 7.7) 41–79 (66.8� 9.4) .184b cation (Years) 0–22 (11.4� 5.0) 0–20 (9.8� 4.2) .136b SE 20–30 (28.5� 2.0) 16–30 (24.7� 3.5) �10�7b T 1–3 (2.8 � .6) 1–3 (2.1� .8) �10�4b R 0 .5 — LT-Immediate ecall 6–14.7 (8.8� 2.0) 2.7–10.7 (5.7� 1.9) �10�9b LT-Delayed ecall 4–15 (9.8 � 2.8) 0–14 (5.1� 3.3) �10�9b LT-Recognition 3–15 (11.6� 2.7) 1–14 (8.8� 3.3) �10�4b Data are presented as the range of minimum–maximum (mean� SD). tably, there were no outliers for any characteristics of both of the groups ng the criterion of 2.5 interquartile ranges from lower/upper quartile ues of the samples. aMCI, amnesticmild cognitive impairment; AVLT, Auditory Verbal Learn- Test; CDR, Clinical Dementia Rating Scale; CDT, Clock Drawing Test; HC, lthy control subjects; MMSE, Mini-Mental State Examination. aThe p value was obtained using a two-tail Pearson chi-square test. bThe p value was obtained using a two-sample two-tail t test. w.sobp.org/journal twork nodes, we divided the brain into 1024 regions of interest Is) according to a high-resolution, randomly generated brain s (H-1024) (38). To measure interregional RSFC, we calculated Pearson correlation between any pair of ROIs in the wavelet fficients that were obtained by the maximal overlap discrete velet transform method (39). Here, we estimated RSFC in four velet scales (scale 1, .125–.250 Hz; scale 2, .063–.125 Hz; scale 3, 1–.063 Hz; and scale 4, .016–.031 Hz). To further de-noise spuri- s interregional correlations, only those correlationswhose corre- nding p values passed through a statistical threshold (p � .05, nferroni-corrected) were retained (40). Details on the network struction can be found in Supplement 1. twork Analysis For the constructed brain networks, we calculated both global regional network metrics to characterize their overall architec- e and regional nodal centrality, respectively. The global network trics included small-world attributes (clustering coefficient, and characteristic path length, LW) (41) andmodularity (Qmax) (42) their normalized versions using random networks (C˜W, L˜W, and ax). Typically, a small-world network shows C˜ W� 1 and L˜W� 1 (41) a modular network shows Q˜max� 1. For regional network mea- es, we employed nodal strength (i.e., weighted degree central- among numerous nodal metrics (43) because of its high test- est reliability (44). See Supplement 1 for the formulas and binov and Sporns (45) for a recent review on the uses and inter- tations of these network measures. tistical Analysis Between-GroupDifferences. Between-groupdifferences in to- logical attributes (bothglobal and regionalmeasures)were inferred nonparametric permutation tests (21,46). Briefly, for each network tric, we initially calculated the between-group difference of the an values. An empirical distribution of the difference was then ob- ed by randomly reallocating all of the values into two groups and omputing the mean differences between the two randomized ups (10,000permutations). The95thpercentilepointsof theempir- l distribution were used as critical values in a one-tailed test of ether the observed group differences could occur by chance. To alize the specific pairs of regions inwhich the functional connectiv- was altered in the aMCI patients, we used a network-based statistic S) approach (47). In brief, a primary cluster-defining thresholdwas tusedto identify suprathresholdconnections,withinwhich thesize ., number of edges) of any connected componentswas thendeter- ed. A corrected p value was calculated for each component using nulldistributionofmaximalconnectedcomponentsize,whichwas ived empirically using a nonparametric permutation approach ,000permutations).Notably,before thepermutationtests,multiple ar regressions were applied to remove the effects of age and gen- , the age-gender interaction, and education level (43,48–53). The ails of the statistical analyses can be found in Supplement 1. Relationships Between Network Measures and Clinical riables. Multiple linear regressions were used to assess the ationshipsbetweennetworkmetrics and clinical variables (AVLT- mediate recall, AVLT-delayed recall, AVLT-recognition, and SE score) in the aMCI group. Age, gender, the age-gender inter- ion, and education level were also controlled. twork Classification Weplotted the receiver operating characteristic curves to deter- ne whether graph-based network metrics might serve as bio- rkers for diagnosingaMCI. This analysiswasperformedusing the public MATLAB codes (http://www.mathworks.com/matlabcentral/ file ple Va pro ou bro lut no fin sis sta Ne (ht tim Int Re De ed the aMCI group had significantly lower scores on the MMSE (p � �7 �4 �9 Fig t), me fun s of a am ions, l ana ions ( J. Wang et al. BIOL PSYCHIATRY 2012;xx:xxx 3 exchange/199500 -roc-curve; Giuseppe Cardillo, Naples, Italy; Sup- ment 1). lidations: Reproducibility and Test-Retest Reliability To validate the reproducibility of our results, we adopted four cedures as follows. Preprocessing Choices. We explored the reproducibility of r results with and without regressing out white matter and cere- spinal fluid signals. Regional Parcellation Effects. We employed three low-reso- ion brain atlases (Table S1 in Supplement 1) to define network des, which allowed us to estimate the reproducibility of our dings against different parcellation schemes. Network Types Effects. Besides the weighted network analy- , we also implemented binary network analyses to assess the bility of our findings. Test-Retest Reliability. We used a public R-fMRI dataset at uroimaging Informatics Tools and Resources Clearinghouse tp://www.nitrc.org/projects/nyu_trt; New York University) to es- ate the test-retest reliability of wavelet-based network metrics. raclass correlation (54) was used (for details, see Supplement 1). sults mographic and Clinical Characteristics There were no significant differences in age, gender, or years of ucation (all p� .13) between the aMCI and HC groups. However, ure 1. (A) Between-group differences in the number of connections (lef ctional networks and distribution of wavelet correlations with respect to bin nestic mild cognitive impairment (aMCI) patients exhibited fewer connect lysis revealed that aMCI targeted more middle- and long-distance connect ), CDT (p � 10 ), AVLT-immediate recall (p � 10 ), AVLT- layed recall (p�10�9), andAVLT-recognition (p�10�4) than the group (Table 1). quency-Specific Alterations in theWavelet Correlation trix For each thresholded wavelet correlation matrix, we calculated total number of links, themean correlation, andmean anatom- l distance (defined as the Euclidean distance between stereo- ic coordinates of the centroids for two regions) for all signifi- tly (p � .05, Bonferroni-corrected) existing connections. The CI networks had a significantly lower mean wavelet correla- n (p� .048) and contained a higher proportion of short-range atomical distance � 45 mm) connections (p � .046) only in velet scale 3 (.031–.063 Hz). Additionally, trends toward fewer nections (p� .080), shorter mean anatomical distances (p� 1), and lower proportion of middle-range (p � .058) and g-range connections (p � .058) were also detected in the CI connectome (Figure 1). No significant between-group dif- ences were detected in other frequency bands (all p � .05). us, the subsequent network topological analyses focused only wavelet scale 3. bal Topological Organization of the nctional Connectome The whole-brain connectome of both the aMCI and HC groups ibited typical features of small-world topology and modular an correlation (middle), and mean anatomical distance (right) of the natomical distance (B). In the specific wavelet scale 3 (.031–.063 Hz), the ower mean correlation, and shorter mean anatomical distance. Further B). *p� .05; #.05�p�.10. HC, healthy control subjects. 10 de HC Fre Ma the ica tax can aM tio (an wa con .05 lon aM fer Th on Glo Fu exh www.sobp.org/journal structure, i.e., comparedwithmatched randomnetworks, the func- tio ide les cre 5.2 gro pa de .08 Regional Topological Organization of the Functional Fig mne diff faces AV 4 BIOL PSYCHIATRY 2012;xx:xxx J. Wang et al. ww nal brain networks had larger clustering coefficients, almost ntical shortest path lengths, and larger modularity. Neverthe- s, quantitative statistical analyses revealed significantly in- ased characteristic path lengths in the aMCI group (HC: 7.950� 36; aMCI: 14.506 � 22.250; p � .047). Additionally, the aMCI up showed trends toward increased normalized characteristic th lengths (HC: 1.673 � .412; aMCI: 1.928 � .694; p � .055) and creased modularity (HC: 3.129� 1.015; aMCI: 2.696� .632; p� 4) compared with the HC group. ure 2. Mean nodal strength in the healthy control subjects (HC) (A) and a erences (C). The nodes and connections weremapped onto the cortical sur LT, Auditory Verbal Learning Test. w.sobp.org/journal nnectome The mean nodal strength (across subjects) was distributed het- geneously across the brain. In the HC group, the most highly nected regions were located predominantly in the posterior ietal and occipital cortices, such as the bilateral precuneus UN), postcentral gyrus, superior parietal gyrus, cuneus, and cal- ine fissure and surrounding cortex (Figure 2A). This pattern was hly preserved in the aMCI patients (r� .834, p� 10�10, Figure . Further between-group comparisons revealed that 27 brain stic mild cognitive impairment (aMCI) patients (B) and between-group using the BrainNet Viewer package (http://www.nitrc.org/projects/bnv). Co ero con par (PC car hig 2B) str res sup bil mi gio bil cre con arc Qm ne mo ple ge 70 an Dis ne tur aM of ne rec the fou mo mo and po (15 cre the �.3 Re Pe Fig rics mil ing Figure 3.Modular architecture (A), amnestic mild cogni- tive impairment (aMCI)-related decrease in functional J. Wang et al. BIOL PSYCHIATRY 2012;xx:xxx 5 uctures were targeted (p� .01, uncorrected) by the disease that ided predominantly in the frontal (e.g., the bilateral dorsolateral erior frontal gyrus and middle frontal gyrus), parietal (e.g., the ateral PCUN and angular gyrus), and temporal (e.g., the bilateral ddle temporal gyrus [MTG] and left inferior temporal gyrus) re- ns (Figure 2C). In addition, several subcortical regions of the ateral caudate nucleus and right putamen also showed de- ased nodal strength in the aMCI patients (Figure 2C). We next sidered the roles of these structures in the context of modular hitecture derived from the HC group. Fivemodules were found ( ax� .536): the motor and somatosensory module, the default twork, the (ventral) attention network, the visual processing dule, and the