Classification of Alzheimer Disease, Mild Cognitive Impairment, and Normal Cognitive Status with Large-Scale Network Analysis Based on Resting-State Functional MR Imaging

We have shown that large-scale network connectivity changes can be used to classify subjects with Alzheimer disease (AD), those with amnestic mild cognitive impairment, and those with normal cognitive function; this method has the potential to assist clinicians in disease assessment, to serve as a biomarker in the prediction of the risk of AD progression and to be used to monitor the efficacy of disease-modifying therapies.

Abstract

Purpose:

To use large-scale network (LSN) analysis to classify subjects with Alzheimer disease (AD), those with amnestic mild cognitive impairment (aMCI), and cognitively normal (CN) subjects.

Materials and Methods:

The study was conducted with institutional review board approval and was in compliance with HIPAA regulations. Written informed consent was obtained from each participant. Resting-state functional magnetic resonance (MR) imaging was used to acquire the voxelwise time series in 55 subjects with clinically diagnosed AD (n = 20), aMCI (n =15), and normal cognitive function (n = 20). The brains were divided into 116 regions of interest (ROIs). The Pearson product moment correlation coefficients of pairwise ROIs were used to classify these subjects. Error estimation of the classifications was performed with the leave-one-out cross-validation method. Linear regression analysis was performed to analyze the relationship between changes in network connectivity strengths and behavioral scores.

Results:

The area under the receiver operating characteristic curve (AUC) yielded 87% classification power, 85% sensitivity, and 80% specificity between the AD group and the non-AD group (subjects with aMCI and CN subjects) in the first-step classification. For differentiation between subjects with aMCI and CN subjects, AUC was 95%; sensitivity, 93%; and specificity, 90%. The decreased network indexes were significantly correlated with the Mini-Mental State Examination score in all tested subjects. Similarly, changes in network indexes significantly correlated with Rey Auditory Verbal Leaning Test delayed recall scores in subjects with aMCI and CN subjects.

Conclusion:

LSN analysis revealed that interconnectivity patterns of brain regions can be used to classify subjects with AD, those with aMCI, and CN subjects. In addition, the altered connectivity networks were significantly correlated with the results of cognitive tests.

© RSNA, 2011

Supplemental material: http://radiology.rsna.org/lookup/suppl/doi:10.1148/radiol.10100734/-/DC1

Introduction

Alzheimer disease (AD) is the seventh leading cause of death in the United States (1). Substantial progress has been made in unraveling the causes and pathophysiology of AD, developing animal models of AD, and designing treatments for AD. There is great interest in developing objective biologically based markers that can be used to predict AD risk, diagnose the disease, track its progression, and monitor the efficacy of treatment.

In the quest for early AD biomarkers, resting-state functional magnetic resonance (MR) imaging has been used. The term resting state refers to study subjects who do not perform any tasks or respond to any stimulus during imaging. Resting-state functional MR imaging yields new insights into how structurally segregated and functionally specialized brain networks are interconnected (2,3). A new network dysfunction perspective on neurodegenerative diseases has been proposed (4–9). Several previous studies have focused on specific hypothesis-driven tests, such as hippocampus networks (10,11), the default mode network (12,13), and the small-world network (14). These studies yielded a new large-scale view of the neural network of intrinsic activity in the brain. So far, however, these approaches have yielded limited success in the classification of subjects with AD, amnestic mild cognitive impairment (aMCI), or normal cognitive function in a meaningful way at the single-subject level (6,9,15). A more sophisticated analysis of the large-scale network (LSN) patterns of intrinsic activity in the brain is needed.

Since resting-state functional MR imaging can provide new insights into how structurally segregated and functionally specialized brain networks are interconnected, it is hypothesized that the disease-specific neural circuitry underlying the interactive pathophysiology during the neural degenerative processes can be identified. The purpose of this study was to use LSN analysis to classify subjects with AD, those with aMCI, and cognitively normal (CN) subjects (16,17).

Materials and Methods

Human Subjects

The study was conducted with institutional review board approval and was in compliance with Health Insurance Portability and Accountability Act regulations. Written informed consent was obtained from each participant. This ongoing prospective research project was started in 2004, with the goal being to identify functional neuroimaging markers for AD. Fifty-seven subjects (21 with mild AD, 16 with aMCI, 20 CN subjects) were recruited through the Memory Disorders Clinic at the Medical College of Wisconsin. The detailed inclusion and exclusion criteria for the three groups of subjects (those with AD, those with aMCI, and CN subjects) have been described previously (10,18). One subject with AD was excluded from analysis because of incomplete brain coverage, and one subject with aMCI was excluded because of excessive motion artifacts. As a result, 20 subjects with AD, 15 subjects with aMCI, and 20 CN subjects were included in the final analysis. The characteristics of these subjects are listed in Table 1. The Mini-Mental State Examination score and the Rey Auditory Verbal Leaning Test delayed recall score revealed significant differences between the compared groups, as shown in Table 1.

Table 1.

Characteristics of Subjects with AD, Subjects with aMCI, and CN Subjects

Note.—Data are means ± standard deviations. Data in parentheses are numbers of patients. Overall mean age was 77 years (range, 65–92 years). Mean age of all men was 76 years (range, 65–92 years), while mean age of all women was 77 years (range, 68–85 years). NA = not applicable.

^*P = .09 (one-way analysis of variance).

^†P < .001 (one-way analysis of variance).

^‡P < .001 (two-sample t test).

Data Acquisition

Imaging was performed with a whole-body 3-T MR imager (Signa; GE Medical Systems, Milwaukee, Wis) with a standard transmit-receive head coil. During the resting-state acquisitions, no specific cognitive tasks were performed, and study participants were instructed to close their eyes and relax inside the imager. Sagittal resting-state functional MR imaging data sets of the whole brain were obtained in 6 minutes with a single-shot gradient echo-planar imaging pulse sequence. The functional MR imaging parameters were as follows: repetition time msec/echo time msec, 2000/25; 90° flip angle; 36 sections obtained without a gap; section thickness, 4 mm; matrix, 64 × 64; and field of view, 24 × 24 cm. High-spatial-resolution three-dimensional spoiled gradient-recalled acquisition in the steady state axial images were acquired for anatomic reference. The parameters were as follows: repetition time msec/echo time msec/inversion time msec, 10/4/450; flip angle, 12°; 144 sections obtained; section thickness, 1 mm; and matrix, 256 × 192.

To ensure that cardiac and respiratory frequencies did not account for any significant artifacts in the low-frequency spectrum, we used a pulse oximeter and respiratory belt to measure these physiologic noise sources, and we further processed data to minimize the potential aliasing effects (19,20).

Each subject’s functional MR image was automatically parcellated into 116 regions of interest (ROIs) (Fig E1, Appendix E1[online]) (21). A series of preprocessing steps that are common to most functional MR analyses were conducted to obtain the averaged low-frequency blood oxygen level–dependent (BOLD) signal time course for each region (Appendix E1 [online]). Preprocessing involves allowing for T1 equilibration effects; section acquisition–dependent time shifts correction; despiking; motion correction; detrending; removal of cardiac, respiratory, white matter, cerebrospinal fluid, and global signal effect (22); and low-frequency band-pass filtering.

Functional Connectivity Matrix and W Matrix

The functional connectivity between any two ROIs (paired ROIs) in the brain was assessed with the Pearson product moment correlation coefficient (r) (23). The partial r values were calculated by removing the effect-of-age variable, since age is a risk factor for AD (Appendix E2 [online]). There were 6670 (116 × 115/2) pairwise partial r values among 116 ROIs for each subject. These r values were arranged in an r matrix, as shown in Figure E2 (online). In the between-group analysis (AD group vs non-AD group), the nonparametric two-sample Wilcoxon rank-sum test (24) was used to compare the corresponding r value distributions for each pair of ROIs. We used the nonparametric two-sample Wilcoxon rank-sum test instead of parametric tests (such as the t test) to avoid making any assumption about the distribution of r values. The Wilcoxon rank-sum test yielded a matrix with 6670 z values, one for each pair of ROIs, as shown in Figure 1a. This is the W matrix. The absolute z value indicates statistical significance, while the sign of the z value indicates which group has the higher mean r value.

Figure 1a:

Images obtained with LSN classification. (a) W matrix obtained when AD and non-AD groups were compared. In the W matrix, the upper right 6670 z values are symmetrical to the lower left 6670 values along the diagonal line, similar to the r matrices in Figure E2 (online). Color bar shows z values. (b) Distribution of z values in the W matrix between AD and non-AD groups. In AD and non-AD classification, a threshold (z < −1.96, uncorrected for multiple comparison) was empirically used. (c) Distribution of z values in the W matrix between aMCI and CN groups. In aMCI and CN classification, four decreased connections and two increased connections were empirically used to enable better classification. (d) Result from LOO procedure during aMCI versus CN classification. In the LOO procedure, one subject with aMCI was left out; therefore, 14 subjects with aMCI and 20 CN subjects were used for training. Thus, classification criteria were determined by using all subjects except the one to be evaluated. As a result, Fisher linear discriminant function (diagonal line) was used to identify the subject with aMCI who was left out in the beginning. Ellipses represent 50% and 90% probability containment for the aMCI and CN groups, respectively. B.G. 1 Tha = basal ganglia and thalamus, DCI = decreased connectivity index, ICI = increased connectivity index.

Classification Method

To evaluate the accuracy of classification with the LSN method, we used the leave-one-out (LOO) method to make up for the lack of independence between the training and testing sets because of the limited sample size in the present study. The LOO error estimate (25) uses one subject at a time for evaluation; the remaining subjects are used for training. That is, the classification criteria are determined by using all subjects except the one to be evaluated. This entire process is repeated for each subject, thereby yielding an unbiased estimate of the classification error rate.

In the trichotomous classification scenario, we used a two-step approach. In the first step, we separated subjects with AD from those who did not have AD (subjects with aMCI and CN subjects). In the second step, we separated subjects with aMCI from CN subjects with the LOO method.

In the first step, we performed 55 LOO processes. For example, in the AD group, subject 1 was left out in the first LOO process. A W matrix was produced for the AD group, which consisted of the remaining 19 subjects with AD and the 35 subjects without AD. The 6670 pairwise z values in the W matrix were sorted, as shown in Figure 1b and 1c. Thereafter, a number of connections with the largest negative z values were defined as the decreased connection set. The averaged r value in the decreased connection set for each subject was obtained as the decreased connectivity index. Similarly, a number of connections with the largest positive z values were defined as the increased connection set, and the averaged r value in the increased connection set was defined as the increased connectivity index. However, the numbers of decreased and increased connections that could be used to classify the AD and non-AD groups were not known a priori and could have affected the classification. To optimize classification accuracy, different combinations of decreased and increased connections were evaluated to assess their classification performances (Table E1 [online]). Fisher linear discriminant analysis was performed with the decreased and increased connectivity indexes to classify a subject as having AD or not having AD with use of the MatLab (Mathworks, Natick, Mass) function “classify” (25). The process was repeated 55 times with different LOO subjects.

Figure 1b:

Images obtained with LSN classification. (a) W matrix obtained when AD and non-AD groups were compared. In the W matrix, the upper right 6670 z values are symmetrical to the lower left 6670 values along the diagonal line, similar to the r matrices in Figure E2 (online). Color bar shows z values. (b) Distribution of z values in the W matrix between AD and non-AD groups. In AD and non-AD classification, a threshold (z < −1.96, uncorrected for multiple comparison) was empirically used. (c) Distribution of z values in the W matrix between aMCI and CN groups. In aMCI and CN classification, four decreased connections and two increased connections were empirically used to enable better classification. (d) Result from LOO procedure during aMCI versus CN classification. In the LOO procedure, one subject with aMCI was left out; therefore, 14 subjects with aMCI and 20 CN subjects were used for training. Thus, classification criteria were determined by using all subjects except the one to be evaluated. As a result, Fisher linear discriminant function (diagonal line) was used to identify the subject with aMCI who was left out in the beginning. Ellipses represent 50% and 90% probability containment for the aMCI and CN groups, respectively. B.G. 1 Tha = basal ganglia and thalamus, DCI = decreased connectivity index, ICI = increased connectivity index.

Figure 1c:

Images obtained with LSN classification. (a) W matrix obtained when AD and non-AD groups were compared. In the W matrix, the upper right 6670 z values are symmetrical to the lower left 6670 values along the diagonal line, similar to the r matrices in Figure E2 (online). Color bar shows z values. (b) Distribution of z values in the W matrix between AD and non-AD groups. In AD and non-AD classification, a threshold (z < −1.96, uncorrected for multiple comparison) was empirically used. (c) Distribution of z values in the W matrix between aMCI and CN groups. In aMCI and CN classification, four decreased connections and two increased connections were empirically used to enable better classification. (d) Result from LOO procedure during aMCI versus CN classification. In the LOO procedure, one subject with aMCI was left out; therefore, 14 subjects with aMCI and 20 CN subjects were used for training. Thus, classification criteria were determined by using all subjects except the one to be evaluated. As a result, Fisher linear discriminant function (diagonal line) was used to identify the subject with aMCI who was left out in the beginning. Ellipses represent 50% and 90% probability containment for the aMCI and CN groups, respectively. B.G. 1 Tha = basal ganglia and thalamus, DCI = decreased connectivity index, ICI = increased connectivity index.

One metric commonly used to evaluate the accuracy of a classifier is the receiver operating characteristic (ROC) curve and the area under the ROC curve. To demonstrate the effect of weighting with the classify function, we assigned 1001 weights that ranged from e¹⁵:1 to e⁻¹⁵:1 (e^{15−30·n/1000}:1, n = 0,1,...,1000. e = Euler constant). LOO analysis was repeated with the 1001 weights to obtain a reasonably smooth curve that resembled an ROC curve.

There were 484 combinations, with the highest area under the ROC curve being 0.93. In consideration of the fact that AD could cause a large amount of LSN connection dysfunctions instead of a limited amount of dysfunctions in a few connections, individual connections (threshold, z < −1.96; uncorrected for multiple comparisons) were used in this study to separate subjects with AD from those without AD, as shown in Figure 1b. (For the theoretical continuous normal distribution, a z value of 1.96 is associated with a two-sided P value of .05.) In consideration of the fact that those increased connection sets (connections with positive z values, where connections in the AD group were stronger than those in the non-AD group) may be related to compensatory effects and may decrease with disease progression, the increased connection set was not used. The Fisher linear discriminant analysis with LOO cross-validation was performed with the decreased connectivity index to classify the 55 subjects as having AD or not having AD. The ROC curve was obtained in the same fashion, as described previously.

Similarly, in the second step, 35 LOO processes were performed for the CN and aMCI groups. In this step, 361 combinations of decreased and increased connections were evaluated to assess their classification performance (Table E2 [online]). To generate the ROC curve between aMCI and CN classifications, we used a method that was similar to that described previously. The highest area under the ROC curve occurred at one decreased connection combined with two increased connections (area under the ROC curve, 0.96). Because only one decreased connection may result in undertraining of the classifier, a set of four decreased connections with a set of two increased connections was used (area under the ROC curve, 0.95), as shown in Figure 1c. For those subjects with AD who were classified as non-AD subjects in the first-step classification, the same classifier (the set with four decreased connections and two increased connections) was used to classify them as subjects with aMCI or CN subjects.

In the present study, the two-step approach was performed in the trigroup classification scenario. The decision to use the two-step approach was justified because the involved LSN functional connectivity changes between the AD and non-AD groups may be different than those between the aMCI and CN groups. Indeed, as will be shown later in this article, the decreased connections that enabled us to separate subjects with AD from those without AD did not contain the connections that enabled us to separate subjects with aMCI from CN subjects.

Results

Each subject has a functional connectivity matrix. Figure E2 (online) shows a functional connectivity matrix in a CN subject, a subject with aMCI, and a subject with AD. To compare group differences in functional connectivity in all of the pairwise ROIs, a W matrix was obtained. Figure 1a shows the W matrix when the AD and non-AD groups were compared. Figure E3 (online) shows cross correlation between ROI 84 (the right superior temporal pole) and ROI 88 (the right middle temporal pole). The preprocessed time courses extracted from ROI 84 and ROI 88 were strongly correlated (r = 0.90) in a CN subject (Fig E3a [online]) but were not strongly correlated in a subject with AD (r = −0.02) (Fig E3b [online]). The z value for this pair (ROIs 84 and 88) between the AD group and the non-AD group was −3.51, demonstrating weaker functional connectivity between the two ROIs in the AD group.

Figure 1b and 1c shows the distribution of z values in the W matrix between the AD and non-AD groups (Fig 1b) and between the aMCI and CN groups (Fig 1c), as described in the Materials and Methods section. The threshold for classification of AD versus non-AD was a z value less than −1.96 (individual connections are shown in Fig E4 [online]). The classifier for aMCI versus CN was a combination of four decreased connections and two increased connections (individual connections are shown in Fig E5 [online]). Figure 1d shows a result of the LOO cross-validation method between subjects with aMCI and CN subjects. One subject with aMCI was excluded (labeled with a filled circle), and the other 14 subjects with aMCI and 20 CN subjects were used to establish classification criteria with Fisher linear discriminant analysis. Clearly, the subject with clinically defined aMCI was correctly classified with the altered functional connectivity in the LSN. This LOO procedure was performed once for each subject.

Figure 1d:

Images obtained with LSN classification. (a) W matrix obtained when AD and non-AD groups were compared. In the W matrix, the upper right 6670 z values are symmetrical to the lower left 6670 values along the diagonal line, similar to the r matrices in Figure E2 (online). Color bar shows z values. (b) Distribution of z values in the W matrix between AD and non-AD groups. In AD and non-AD classification, a threshold (z < −1.96, uncorrected for multiple comparison) was empirically used. (c) Distribution of z values in the W matrix between aMCI and CN groups. In aMCI and CN classification, four decreased connections and two increased connections were empirically used to enable better classification. (d) Result from LOO procedure during aMCI versus CN classification. In the LOO procedure, one subject with aMCI was left out; therefore, 14 subjects with aMCI and 20 CN subjects were used for training. Thus, classification criteria were determined by using all subjects except the one to be evaluated. As a result, Fisher linear discriminant function (diagonal line) was used to identify the subject with aMCI who was left out in the beginning. Ellipses represent 50% and 90% probability containment for the aMCI and CN groups, respectively. B.G. 1 Tha = basal ganglia and thalamus, DCI = decreased connectivity index, ICI = increased connectivity index.

Table 2 summarizes the overall classification results. In the first step, we separated subjects with AD from those who did not have AD. Among the 20 subjects with clinically defined AD and the 35 subjects who did not have clinically defined AD, the LSN classifier yielded 85% sensitivity and 80% specificity, as well as a 20% false-positive rate and a 15% false-negative rate, with an accuracy of 82%. In the second step, we separated subjects with aMCI from CN subjects in the subset of 35 subjects who did not have clinically defined AD. Among these 35 subjects, the large LSN classifier yielded 93% sensitivity and 90% specificity, as well as a 10% false-positive rate and a 7% false-negative rate, with an accuracy of 91%. Of the three subjects with clinically defined AD who were classified as non-AD subjects, two were classified as having aMCI, and one was classified as a CN subject. Among the 14 subjects with aMCI, three were classified as having AD in the first step; among 18 CN subjects, three were classified as having AD. Two CN subjects were classified as having aMCI; one of these subjects was classified as having AD in the first step.

Table 2.

Two-Step Classification Results with LSN Analysis in Subjects with AD, Subjects with aMCI, and CN Subjects

Note.—Data are numbers of patients. Data in parentheses are percentages.

When we combined the results from steps 1 and 2, we obtained the following findings: Among the 20 subjects with clinically defined AD, AD was confirmed in 17 subjects; two subjects were classified as aMCI subjects, and one was ruled out as a CN subject. Among the 15 subjects with clinically defined aMCI, aMCI was confirmed in 11 subjects, three were classified as having AD, and one was a CN subject. Among the 20 CN subjects, 15 were confirmed to have normal cognitive function, four were classified as having AD, and one was classified as having aMCI. In noting that subjects with aMCI are at great risk to develop AD and are defined as the diseased category with the AD group, Table 2 shows the overall classification results: Among the 35 subjects with clinically defined disease (20 subjects with AD, 15 subjects with aMCI), findings were confirmed in 33 (94%) subjects and false negative in two (6%). Among the 20 CN subjects, findings were confirmed in 15 (75%) and false positive in five (25%).

Figures 2a and 3a show the ROC curves, which demonstrate the distributions of sensitivity and specificity of the LSN method. We investigated whether the measured decreased connectivity index and increased connectivity index correlate with behavioral scores. Figure 2b shows that the decreased connectivity index significantly correlated with the Mini-Mental State Examination scores (P < .001). The more decreased the connectivity, the more severe the cognitive impairment and the lower the Mini-Mental State Examination score. Similarly, the measured increased connectivity index and decreased connectivity index were significantly correlated with memory scores measured with the Rey Auditory Verbal Leaning Test delayed recall score, as shown in Figure 3b (P < .004).

Figure 2a:

ROC curve and behavioral importance of altered network connectivity strengths. (a) ROC curve shows that use of the LSN method resulted in accurate classification of subjects with AD and those without AD (area under the ROC curve, 0.87). (b) Relationship between Mini-Mental State Examination (MMSE) scores and decreased connectivity index (DCI). MMSE = 30/[1 + exp(−8.1·DCI − 2.7)]. F = 61.26; df = 1, 53; P < .001.

Figure 3a:

ROC curve and behavioral importance of altered network connectivity strengths. (a) ROC curve shows that use of the LSN method resulted in accurate classification of subjects with aMCI and CN subjects (area under the ROC curve, 0.95). (b) Relationship between Rey Auditory Verbal Leaning Test (RAVLT) score and altered connectivity indexes. DCI = decreased connectivity index, ICI = increased connectivity index. (RAVLT = 15/[1 + exp(−1.2·DCI + 3.2·ICI−1.4)]; F = 6.70; df = 2, 32; P < .004.

Figure 2b:

ROC curve and behavioral importance of altered network connectivity strengths. (a) ROC curve shows that use of the LSN method resulted in accurate classification of subjects with AD and those without AD (area under the ROC curve, 0.87). (b) Relationship between Mini-Mental State Examination (MMSE) scores and decreased connectivity index (DCI). MMSE = 30/[1 + exp(−8.1·DCI − 2.7)]. F = 61.26; df = 1, 53; P < .001.

Figure 3b:

ROC curve and behavioral importance of altered network connectivity strengths. (a) ROC curve shows that use of the LSN method resulted in accurate classification of subjects with aMCI and CN subjects (area under the ROC curve, 0.95). (b) Relationship between Rey Auditory Verbal Leaning Test (RAVLT) score and altered connectivity indexes. DCI = decreased connectivity index, ICI = increased connectivity index. (RAVLT = 15/[1 + exp(−1.2·DCI + 3.2·ICI−1.4)]; F = 6.70; df = 2, 32; P < .004.

Discussion

The current LSN analysis described herein has demonstrated that interconnectivity patterns of brain regions may be helpful in the classification of subjects with AD, those with aMCI, and CN subjects. In addition, the altered connectivity networks were significantly correlated with the results of cognitive tests. This LSN classification method has at least three advantages: First, it does not depend on the default mode or the hippocampus hypothesis of AD progression; therefore, there is no limit to the brain regions involved (10,12). Thus, the global LSN classification could yield a high level of accuracy. Second, in previous studies, researchers used dichotomous categorization based on brain atrophy to separate subjects with AD from CN subjects, CN subjects from subjects with MCI, and subjects with MCI from subjects with AD (11,26–29). However, in trigroup classification, such dichotomous categorization could be problematic because the classifiers can be different between any two groups of subjects. To our knowledge, there have been no reports on the classification accuracy among subjects with AD, subjects with aMCI, and CN subjects. In the present study, we presented the results of the trigroup classification for subjects with AD, subjects with aMCI, and CN subjects on the basis of the two-step approach. Third, in comparison with the Pittsburgh compound B positron emission tomography method, in which the clinical symptoms were not coupled to amyloid deposition (30), the present LSN classification method resulted in quantitative relationships between network connectivity strengths and behavioral changes. Thus, the changes in connectivity strengths can serve as a biomarker at the preclinical stage and can be used to predict and monitor disease progression.

Classification with LSN analysis based on the intrinsic spontaneous BOLD signal acquired in the resting state is a relatively new approach, and several technologic issues need to be addressed. First, while the resting-state functional MR paradigm has major advantages in the study of patients with dementia (no confounding due to task performance), it is not clear how arousal, anxiety, or unfocused thoughts could affect functional connectivity and classification. To address this question, it is necessary to understand the neural basis of the intrinsic spontaneous BOLD signal. A recent review conducted by Raichle (31) suggested that the functional MR BOLD signal is best correlated with local field potentials (32–34), particularly the γ band (35–37). The intrinsic spontaneous BOLD signals are best correlated with local field potential activity of the slow cortical potentials in the range of 0.1–4.0 Hz, including the d band (1–4 Hz) (38,39). It has been reported that the intrinsic spontaneous BOLD signal and the task-induced functional MR BOLD signal are linearly superimposed (40) and that functional connectivity can be affected (41). However, because of the randomness of mental activities, the possible functional MR BOLD signals associated with arousal, anxiety, or unfocused thoughts during the resting state would not be synchronized with the intrinsic spontaneous BOLD signal. Thus, the functional connectivity determined with the pairwise Pearson cross-correlation coefficients would not be significantly affected. The present classification results with high classification accuracy support this view.

Second, the present resting-state functional MR images acquired with 4-mm section thickness and 25 msec of echo time with the gradient-recalled-echo pulse sequence could suffer from the susceptibility artifacts of imaging distortion and signal dropout in the orbitofrontal cortex and middle inferior temporal cortex regions. It should be emphasized that the existence of susceptibility induced artifacts can severely affect voxel pairwise connectivity because of lower signal-to-noise ratio. In fact, functional connectivity in the present study was determined with pairwise regional signals instead of pairwise voxels. The averaged time courses of a region have a high signal-to-noise ratio. For example, the signal-to-noise ratio of the averaged regional time course in the orbitofrontal cortex was 210 in this study. That was sufficient to obtain reliable connectivity measurements (42). In addition, because the cross-correlation coefficient is magnitude independent, the susceptibility induced signal dropout may not play a pivotal role.

Third, it is conceivable that the intrinsic spontaneous BOLD signal may depend on individual hemodynamics, which can be affected by various vascular diseases, stroke, drug abuse, and various psychiatric disorders, such as depression. In this study, these cases were excluded on the basis of study exclusion and inclusion criteria. In the future, it is suggested that a control task, such as a simple visual task or a breath-holding task, be used to characterize individual hemodynamics in the brain.

The current study had the following limitations. A relatively small cohort was used to evaluate the classification accuracy. Although a relatively unbiased LOO method was applied to error estimation, it was imperfect. The classification method (the selection of decreased and increased connections) can be further validated with a test cohort and a different validation cohort. Collaboration with other research groups is being pursued to build a large cohort with which to cross-sectionally validate the LSN method. Another limitation was that the clinically defined groups of subjects were considered a reference standard, without postmortem confirmation to produce ROC curves. As a result, a subject could wrongly be classified as having AD and then used in the analysis of subjects with aMCI versus CN subjects, while a subject with AD who had been misclassified as a non-AD subject would be used. As a consequence, both sensitivity and specificity in the analysis of subject with aMCI versus CN subjects would be overestimated. Thus, a longitudinal study is needed to enable researchers to confirm or validate the classification and disease predictions. Furthermore, it is desirable to use postmortem studies as the reference standard when validating the LSN method.

To summarize, we have shown that LSN connectivity changes can be used to classify subjects with AD, those with aMCI, and those with normal cognitive function; this method has the potential to assist clinicians in disease assessment, to serve as a biomarker in the prediction of AD progression risk, and to be used to monitor the efficacy of disease-modifying therapies.

Advances in Knowledge.

• Resting-state functional connectivity MR imaging can be used to detect reorganization of large-scale network (LSN) brain functional connectivity in subjects with dementia.

• LSN functional connectivity might be useful in differentiation of subjects with Alzheimer disease (AD) from subjects with amnestic mild cognitive impairment (aMCI) and cognitively normal (CN) subjects.

• LSN connectivity strengths were significantly correlated with behavioral scores; they have the potential to serve as biomarkers at the preclinical stage, and they can be used to predict disease progression and monitor therapeutic effects.

Implications for Patient Care.

• LSN functional connectivity has the potential to assist primary care clinicians in the classification of subjects with AD, those with aMCI, and CN subjects.

• Changes in LSN connectivity can be useful in the development of surrogate markers for drug discovery and diagnostic tests.

The authors thank C. M. O’Connor, MA, for editorial assistance.

Disclosures of Potential Conflicts of Interest: G.C. Financial activities related to the present article: institution received funds from Pfizer. Financial activities not related to the present article: none to disclose. Other relationships: none to disclose. B.D.W. Financial activities related to the present article: institution received funds from Pfizer. Financial activities not related to the present article: none to disclose. Other relationships: none to disclose. C.X. Financial activities related to the present article: institution received funds from Pfizer. Financial activities not related to the present article: received support from China Scholarship Council. Other relationships: none to disclose. W.L. Financial activities related to the present article: institution received funds from Pfizer. Financial activities not related to the present article: none to disclose. Other relationships: none to disclose. Z.W. Financial activities related to the present article: institution received funds from Pfizer. Financial activities not related to the present article: none to disclose. Other relationships: none to disclose. J.L.J. Financial activities related to the present article: institution received funds from Pfizer. Financial activities not related to the present article: none to disclose. Other relationships: none to disclose. M.F. Financial activities related to the present article: institution received funds from Pfizer. Financial activities not related to the present article: none to disclose. Other relationships: none to disclose. P.A. Financial activities related to the present article: institution received funds from Pfizer. Financial activities not related to the present article: served on the speakers bureau of Pfizer and Novartis; received a research grant from Bristol-Myers Squibb and Octapharma. Other relationships: none to disclose. S.J..L. Financial activities related to the present article: institution received funds from Pfizer. Financial activities not related to the present article: none to disclose. Other relationships: none to disclose.