Inference of brain pathway activities for Alzheimer’s disease classification
Subjects
This study was approved by the Institutional Review Board of Samsung Medical Center.
Written informed consent was obtained from aMCI, AD patients, and cognitively normal
(CN) subjects. One hundred twenty right-handed subjects were recruited through the
Samsung Medical Center: 22 cognitively normal (CN) subjects, 37 aMCI patients, 61
AD patients (21 patients with very mild AD, 27 patients with mild AD, 13 patients
with moderate AD). The AD stages were categorized by the National Institute of Neurological
and Communicative Disorders and Stroke/ Alzheimer’s Disease and Related Disorders
Association (NINCDS-ADRDA) criteria 16]. For the diagnosis of aMCI patients, Mayo Clinic criteria were used 17].
fMRI imaging
Magnetic resonance imaging (MRI) examination was conducted on a 3.0-T MR scanner 3.0
T scanner (Model; Philips Intera Achieva, Phillips Healthcare, Netherlands). Scans
involved the acquisition of 35 axial slices using a gradient echo planar imaging pulse
sequence: repetition time (TR) = 3000 ms; echo time (TE) = 35 ms; acquisition time
(TA) = 5 minutes; flip angle (FA) = 90°; field of view (FOV) (RL, AP, FH) = 220 ×
220 × 140 mm; voxel size (RL, AP) = 2.875 mm × 2.875 mm with a slice thickness of
4 mm. During the scan, participants were instructed to lie still with their eyes open.
Additionally T1-weighted anatomical images were obtained for each subject: TR = 1114
ms; TE = 10 ms; FA = 8°; FOV (RL, AP, FH) = 220 × 220 × 132 mm; REC voxel size = 0.43
mm × 0.43 mm × 0.43 mm.
Preprocessing of MR image data was performed by the FMRIB Software Library, FSL 4.1
18]. The first 6 volumes from the functional MRI runs were discarded to avoid T1 equilibrium
effects. Then, the following pre-statistic processing steps were done: deleting non-brain
tissues from images using a Brain Extraction Tool (BET), motion correction using MCFLIRT
19]. Grand mean intensity normalization of the whole 4D data set by a single multiplicative
factor, spatial smoothing using a Gaussian kernel of FWHM 5 mm, high pass temporal
filtering (Gaussian-weighted least-squares straight line fitting, with sigma = 50
to ensure at least half power was preserved for frequencies down to 0.01 Hz). The
corrected MR images were registered into the Montreal Neurological Institute space
(MNI-152 stereotactic template) using FLIRT, FMRIB’s linear image registration tool.
Integration of brain pathways
Brain pathways are comprised of anatomically separated regions, but functionally connected
regions. Brain pathways were characterized by biological functions and behavioral
domains such as perception, motor, cognition, emotion, and sensation. These brain
pathways have been discovered and revised by in vivo and in vitro experiments. For
example, the Papez pathway was regarded as the emotional pathway, but it was revised
as the limbic system pathway through other experimental validations 20].
In our study, the 59 brain pathways were selected based on the behavioral domains,
the associated functions, and lateralization (Table 1). Brain pathways were divided into the left (L) and the right (R) brain hemisphere
except for 7 pathways. The 7 pathways are dominantly lateralized in the left or the
right brain hemisphere. In the default mode network (DMN), the positively correlated
networks of both ventromedial prefrontal cortex (vmPFC) and posterior cingulate cortex
(PCC) regions were used as the brain pathways. Among 59 brain pathways, 38 pathways
are well-known brain pathways covering systemic neuroscience 21,22]. Other 21 pathways were manually curated from literature to supplement 38 well-known
pathways, considering specialized functional and structural systems of the brain in
cognitively normal subjects 23-39]. For example, emotional domains of 38 pathways describe general features of the emotion,
however manually curated Krolak-Salmon (2004) pathway explain more detailed phenomena
of fear spreading in emotional domain. These 21 pathways were aggregated with in vitro
imaging studies, such as diffusion tensor imaging (DTI), electroencephalography (EEG),
structural MRI, and functional MRI. The manually curated 21 brain pathways were named
according to last name of first authors and publication years. The regional connectivity
and lateralization of 59 brain pathways are described in Additional file 1.
Table 1. The 59 Brain pathways with behavioral domains, associated functions, and lateralization.
Additional file 1. The regional connectivity and lateralization of 59 brain pathways.
Format: PDF
Size: 91KB Download file
This file can be viewed with: Adobe Acrobat Reader
Functional connectivity
The whole brain was divided into 116 brain regions based on a gray matter mask with
atlas labels by using the automated anatomical labeling (AAL) atlas 40]. A hierarchical segmentation of the whole brain covers the cortical regions (frontal
cortex, temporal cortex, parietal cortex, and occipital cortex), subcortical regions
(limbic regions, insula, basal ganglia, thalamus), and cerebellar hemisphere (Additional
file 2). Using gray matter masks, the averaged MR signals of 116 brain regions of both CN
subjects and patients were extracted by FSL tool.
Additional file 2. Automatically parcellated 116 brain regions.
Format: PDF
Size: 25KB Download file
This file can be viewed with: Adobe Acrobat Reader
The functional connectivities between paired brain regions were measured by the Pearson’s
correlation coefficient (r). The types of functional connectivity were categorized according to strengths of
linear relationships. Positive correlations between paired regions indicate that MR
signals of one brain region were increased, and the other brain region has a tendency
to also increase, while negative correlations has a tendency to decrease. Also, no
correlations between paired regions show that the other brain region does not tend
to either decrease or increase. As a result, the 6670 (116 × 115 / 2) r values between paired brain regions were produced of both CN subjects and patients,
and these r values were arranged into the functional connectivity (R) matrix (Figure 1). Strengths of functional connectivities between paired regions were represented
by color bars: red color indicates positive correlations; blue color indicates negative
correlations; green color indicates no correlations.
Figure 1. Functional connectivity (R) matrix between paired ROIs. (A) CN subjects (B) aMCI patients (C) AD patients. Each matrix has the 6670 r values calculated from combination of 116 brain regions. The color bar indicates the
Pearson’s correlation coefficient (r) values.
Inferring pathway activity for the brain pathway-based approach
In this work, we use three different data sets described in Method, and each data
set consists of two different groups: the first data set (between 22 CN subjects and
37 aMCI patients); the second data set (between 22 CN subjects and 61 AD patients);
the third data set (between 37 aMCI and 61 AD patients). Pearson’s correlation coefficient
(r) values of all sample i over functional connectivity j were arranged into the connectivity (F) matrix to aggregate the functional connectivities
between paired brain regions corresponding the brain pathway P (Figure 2A). For example, the orbitofrontal pathway has sequential connections from orbitofrontal
cortex to caudate, globus pallidus, thalamus, and orbitofrontal cortex 23]. The r values of each data set (r1 between orbitofrontal cortex and caudate, r2 between caudate and globus pallidus, r3 between globus pallidus and thalamus, r4 between thalamus and orbitofrontal cortex) were arranged in the connectivity (F) matrix.
Fisher’s z transformation was applied to r values to obtain the normal distributed values Rij of all samples i over functional connectivity j. As a result, the connectivity (F) matrix was acquired between two groups in each
data set.
Figure 2. Schematic diagram of the brain pathway activity inference. For the given brain pathway P, the activity Ap matrix was acquired through the exhaustive search between two different groups.
The exhaustive search was performed to identify the discriminative connectivity set
in connectivity (F) matrix (Figure 2B). To detect dominant signals between two groups, all the possible connectivities
in connectivity (F) matrix were considered through the exhaustive search, and optimally
discriminative connectivity set between two groups was selected according to their
statistical significance. All the possible combinations of the functional connectivities
in the connectivity set Bn were considered from k = 1 to k = n. All Rij values in the connectivity set Bk was transformed into averaged R values which were designated the activity score apk.
Activity(apk)score= ?s=1kRisk
All the possible combinations in the connectivity set Bn, and activity matrices Apk
k=1?{b1},{b2},……,{bn}?A1p1,A2p1,……,Anp1
k=2?{b1,b2},{b1,b3},…,{bn – 1,bn}?A1p2,A2p2,…,An!k!(n-k)!p2
……
k=n?{b1,b2,……,bn}?A1pn
We defined the t-score T(Bpk) of the activity matrices Apk derived from connectivity set Bk from k = 1 to k = n. The normality of the activity matrices Apk was confirmed by Kolmogorov-Smirnov tests (p-value 0.05, normal distribution), and
the equality of variance was assessed by Levene’s tests (p-value 0.05, unequal variance)
for the 2-tail t-test. In the brain pathway P, the discriminative connectivity set
was defined when the T(Bpk) score reaches much higher than other t-score values among the activity matrices
Apk. The higher T(Bpk) score between two groups indicates statistical differences, and their connectivity
set was regarded as quantitative activity indices of brain pathways. The activity
matrix Ap acquired from the discriminative connectivity set between two groups was assigned
as the brain pathway P activity. If k = 2 with connectivity set {b1, bn} shows the highest T(Bp2) score among all the possible combinations of connectivity set Bn, and the activity matrix Ap of connectivity set {b1, bn} was designated as the brain pathway P activity (Figure 2C).
The discriminative connectivity set between two groups indicates brain malfunctions
corresponding specific brain pathway. For example, if functions of memory in the brain
are disrupted during AD progression, functional connectivities of memory pathways
in AD patients show abnormal patterns versus cognitively normal subjects. These unusual
patterns of functional connectivities are selected as the discriminative connectivity
set between cognitively normal subjects and AD patients. As a result, we obtained
the 59 activity matrices between two groups across 59 brain pathways (Figure 2D). In three different data sets, the 59 activity matrices were generated between two
groups. To clearly differentiate 59 brain activities between two groups, they were
rearranged according to statistical significance (Figure 3).
Figure 3. Inferred activity matrices using 59 brain pathways between two different groups. (A) between CN subjects and aMCI patients (B) between CN subjects and AD patients
(C) between aMCI and AD patients.
Regional functional correlation strength for the region-based approach
The activities of 59 brain pathways were used as input features for the brain pathway-based
classification. In order to compare the brain pathway-based approach with the region-based
approach in an unbiased manner, we used the same number of input features for the
region-based approach. For selecting 59 input features of the region-based approach,
we estimated regional functional correlation strength (RFCS) using previously described
method in imaging study 41]. In region a, the correlation strength was defined as:
CorrelationStrengthregion(a)=1N-1 ?i?j|rab|
where rab is the Pearson’s correlation coefficient (r) values between brain region a and b, and N is the number of brain regions.
From the functional connectivity (R) matrix, we acquired the RFCS scores of 116 regions
between two different groups, and then 2-tail t-test was performed between two different
groups to select the RFCS scores of 59 brain regions as input features for the seed-based
approach. The RFCS scores of 116 brain regions were rearranged by their t-test score
in ascending order, and then the RFSC scores of top 59 brain regions were selected
as input features for the seed-based approach. As a result, we acquired the 59 RFCS
matrices between two groups across 59 brain regions in three different data sets (Figure
4).
Figure 4. The 59 RFCS matrices between two different groups among 116 brain regions. (A) between CN subjects and aMCI patients (B) between CN subjects and AD patients
(C) between aMCI and AD patients.
Evaluation methods for classification
We used the four supervised machine learning algorithms for evaluating the classification
performance. In three different data sets, we trained on both 59 activity matrices
(pathway-based approach) and 59 RFSC matrices (region-based approach) by using three
linear classifiers: Naïve Bayes (NB); logistic regression; support vector machine
(SVM) and one decision trees classifier: random forest (RF).
All samples in both 59 activity matrices and 59 RFSC matrices were randomly partitioned
into ten equivalent subsamples. Among the ten subsamples, nine subsamples were used
as training data set for building classifiers, and one subsample was remained for
testing the classification models. The process of cross-validation was repeated 10
times, and each of ten subsamples employed just once as the test set (10-fold cross-validation).
Each fold calculated classification accuracy, and results of ten folds were averaged
to create a single evaluation. Performance of four classifiers was estimated by the
area under the curve (AUC) in the receiver operating characteristics (ROC), accuracy,
sensitivity and specificity.
Accuracy=(TP+TN)(TP+FP+TN+FN)
Sensitivity=TP(TP+FN)
Specificity=TN(FP+TN)
TP: True Positive, FP: False Positive, TN: True Negative, FN: False Negative
Feature selection
The quantification of the brain pathways importance is necessary for interpretation
of pathological symptoms during AD progression. For identifying discriminatory pathways
between two groups, the feature selection was performed by using the random forest
(RF). RF is the efficient algorithm for solving classification problems, because classification
performance of RF model is enhanced by growing an ensemble of trees and letting them
vote for the most preferable group 42].
There are two scoring methods for measuring variable importance with RF: mean decrease
accuracy (MDA), mean decrease Gini (MDG). Between two scoring methods, we adopted
the MDA for evaluating variable importance within two groups; variables having higher
MDA values contribute importantly toward the classification, and variables having
lower MDA values could not affect the classification. After calculating MDA scores
between two groups, we ranked the variables (59 brain pathways) according to MDA scores,
and selected the top-K variables as significant features. The K values were determined
by the distribution of MDA scores, and we defined the decreasing points of MDA scores
when slopes of distributions is dramatically changed, and selected the variables before
the first changing point in the distribution as important features.