Diagnostic power of resting‐state fMRI for detection of network connectivity in Alzheimer's disease and mild cognitive impairment: A systematic review

Abstract Resting‐state fMRI (rs‐fMRI) detects functional connectivity (FC) abnormalities that occur in the brains of patients with Alzheimer's disease (AD) and mild cognitive impairment (MCI). FC of the default mode network (DMN) is commonly impaired in AD and MCI. We conducted a systematic review aimed at determining the diagnostic power of rs‐fMRI to identify FC abnormalities in the DMN of patients with AD or MCI compared with healthy controls (HCs) using machine learning (ML) methods. Multimodal support vector machine (SVM) algorithm was the commonest form of ML method utilized. Multiple kernel approach can be utilized to aid in the classification by incorporating various discriminating features, such as FC graphs based on “nodes” and “edges” together with structural MRI‐based regional cortical thickness and gray matter volume. Other multimodal features include neuropsychiatric testing scores, DTI features, and regional cerebral blood flow. Among AD patients, the posterior cingulate cortex (PCC)/Precuneus was noted to be a highly affected hub of the DMN that demonstrated overall reduced FC. Whereas reduced DMN FC between the PCC and anterior cingulate cortex (ACC) was observed in MCI patients. Evidence indicates that the nodes of the DMN can offer moderate to high diagnostic power to distinguish AD and MCI patients. Nevertheless, various concerns over the homogeneity of data based on patient selection, scanner effects, and the variable usage of classifiers and algorithms pose a challenge for ML‐based image interpretation of rs‐fMRI datasets to become a mainstream option for diagnosing AD and predicting the conversion of HC/MCI to AD.


| INTRODUCTION
Alzheimer's disease (AD) is a neurodegenerative disorder that is characterized by a progressive decrease in cognitive function compared to baseline performance level in one or more cognitive domains that can interfere with the ability to independently carry out activities of daily living (American Psychiatric Association, 2013). Resting-state functional magnetic resonance imaging (rs-fMRI) is a neuroimaging tool used to study the aberrations in the functional activity of different brain networks, which normally occurs in AD and its prodromal condition, mild cognitive impairment (MCI; X. Chen et al., 2017). The functional connectivity (FC) of brain networks refers to inter-regional synchrony, as detected from low-frequency fluctuations in the blood oxygen level dependent (BOLD) fMRI sequence (L. Lee, Harrison, & Mechelli, 2003). FC and other functional features of AD are studied using different molecular imaging techniques such as electroencephalography (EEG), positron emission tomography-computed tomography (PET/CT), and fMRI. Various radiotracers such as glucose analogs and amyloid detecting radiotracers have been utilized for improving the diagnostic accuracy of detecting AD (Suppiah, Didier, & Vinjamuri, 2019). Of these techniques, fMRI remains the most widely used modality because of the relative simplicity of its usage, inherent safety features, noninvasive nature, and high spatial resolution (Mier & Mier, 2015).
The default mode network (DMN) is the commonest brain network studied by rs-fMRI and is involved in memory consolidation tasks. It composed of the precuneus (Prec), posterior cingulate cortex (PCC), retro-splenial cortex, medial parietal cortex (MPC), lateral parietal cortex (LPC), and inferior parietal cortex (IPC), medial prefrontal cortex (mPFC), and the medial temporal gyrus (MTG; Mohan et al., 2016). Fundamentally, AD patients suffer from impaired DMN connectivity (Grieder, Wang, Dierks, Wahlund, & Jann, 2018). There has been consistent evidence of decreased FC in the DMN of AD patients in comparison with HCs, especially between the posterior part of the cerebral cortex (Prec and PCC) and anterior parts, for example, the anterior cingulate cortex (ACC) and mPFC (Brier et al., 2012;Gili et al., 2011;Griffanti et al., 2015). The observed decline in FC in areas within the DMN has also been reported among MCI patients (Cha et al., 2013;Ouchi & Kikuchi, 2012). This indicates that rs-fMRI detected changes in the DMN can be a noninvasive diagnostic tool for diagnosing AD. In fact, the National Institute on Aging-Alzheimer's Association (NIAAA) has listed rs-fMRI FC as a potential biomarker of neuronal injury, which is at an early stage of validation (Albert et al., 2011).
There are several methods to analyze rs-fMRI data, namely, the seed-based analysis (SBA), the independent component analysis (ICA), and the graph theory analysis (GTA). SBA or small region of interest (ROI) analysis enables temporal correlations to be made between hypotheses-based predefined seed regions. The SBA investigates the FC of a specific brain region by correlating the brain region's restingstate time series with the time series of all other regions resulting in the creation of a FC map that identifies the FC of the predefined brain region (T. Jiang, He, Zang, & Weng, 2004). The simplicity and straightforwardness of this seed-dependent analysis coupled with the clarity of the FC map, makes it popular among researchers (Buckner et al., 2009). Nevertheless, the knowledge from an FC map is restricted to the FC of a pre-defined region that requires a priori knowledge, making it hard to analyze correlations of FC in whole brain regions.
In contrast to SBA, ICA is free from any predefined seed region selection, which means one does not have to pick a seed or reference area beforehand. Hence, the entire BOLD signal is broken down to produce separate time courses and related spatial maps (De Luca, Smith, De Stefano, Federico, & Matthews, 2005). The resultant components are assumed to be non-Gaussian signals and are statistically independent of one another. ICA extracts FC information by detecting the patterns of synchronous neural activities between nodes without an a priori knowledge or pre-existing hypothesis. Thus, the signals from various nodes are temporally filtered from a sample dataset to assess the FC between two independent nodes, which is similar to the "cocktail party effect" (Li, Wang, Chen, Cichocki, & Sejnowski, 2017).
The ICA algorithm assumes a set of maximally spatially independent brain components (S), each with associated time course signals (X).
The model identifies latent sources whose elements (voxels) have the same time course and thus each component can be considered a measure of the degree to which each voxel is functional connected (correlated) to the component-time course.
Due to its ability to accommodate whole-brain FC analysis, ICA is favored over SBA. Nevertheless, the disadvantage of ICA is that there is often difficulty in differentiating useful signals from noise and variations in the separate components. Hence, this causes challenges in making between-group comparisons using ICA (Fox & Raichle, 2007).
Interestingly, both SBA and ICA can ultimately produce similar results if they are run at different experimental set-ups.
Alternatively, GTA looks at the overall brain network structure with specific spatial information. Here, the BOLD signal undergoes spatial parcellation using a topological mapping of the entire brain, and the relationships between all pairs of activated regions involving "nodes" and "edges" are determined. A "node" is a defined area in the brain, whereas "edge" signifies the direct and indirect links or FC between two defined nodes. Additionally, a "hub" is a node that has an integrative role, which reflects the diversity of a region's crossnetwork FCs. A "hub" is defined as a node that has a betweenness centrality or eigenvector centrality (ECi) that is larger than the mean plus two standard deviations (mean + 2 SD) across all nodes in a particular region (Hojjati, Ebrahimzadeh, & Babajani-Feremi, 2019).
The assessment of the relationship of nodes versus edges of the activated regions is achievable by forming a p × p square matrix. The fMRI time signal of all participants, "X" is decomposed into a set of maximally independent components, "S" such that both can be transformed to each other via the mixing matrix "A". Thus, to illustrate the concepts in fMRI, "S" or "components" is a stack of 3D images that will be "mixed" by "A" or "dimensions" that are timepoints by component. Hence, the fMRI signal of all participants, X = A × S, whereby S will be the weighted sum of all the components can be calculated to achieve the series of 3D time point images. In the most common "Dual Regression" approach, first a group ICA is run to estimate the "S" for the whole sample. Then, individual analyses are run to estimate the transformation matrix, "W" for each subject. Notably, the components in "S" then represent the common resting-state networks, that is, the DMN and visual-motor networks or physiological noise signals, for example, eyes movement, as well as heart and respiratory motion (J. E. Chen et al., 2020). Finally, one of the networks can be selected to run a multivariate regression on the individual's time courses to estimate group differences for each voxel (Salman et al., 2019).
In this way, the brain is considered as a single complex network where several global and local network topologies such as the path length, modularity of global connectedness, and clustering coefficient can be measured (Rubinov & Sporns, 2010). In GTA, the graph is directly constructed from the Fisher transformed correlation matrix by using each atlas region as the "node" and the z-value as the "edge" weight. The Fisher's r-to-z transformation is applied to the elements of the matrix to improve the normality of the correlation coefficients (Thompson & Fransson, 2016 (Cha et al., 2013;Liao et al., 2018). Typically, aMCI and AD groups had decreased FC in the left PCC and left parahippocampal gyrus as compared to HC subjects (Liu et al., 2016). Only AD patients were identified with increased FC at the right middle frontal gyrus (MFG), which was interpreted as a compensatory neural mechanism in response to the impairment of the PCC and middle temporal gyrus (MTG; Cha et al., 2013). In the PCC, MTG, and MFG regions, MMSE scores showed significant positive and negative associations with FC (Cha et al., 2013;Liao et al., 2018). Therefore, it is evident that most of the FC disruptions of the DMN occurs in the PCC and MTG (Bai et al., 2009;Zhou et al., 2008). Nevertheless, as the disease progresses, the FC disturbances spread to other brain regions (Damoiseaux, Prater, Miller, & Greicius, 2012).
Since the PCC and other DMN hubs are affected in AD and MCI, the DMN may serve as an important biomarker for the classification of patients with AD and MCI. A recent review paper by Badhwar et al. that was published in 2017, studied various patterns of rs-fMRI detected dysfunctions among patients with AD (Badhwar et al., 2017). Nevertheless, this systematic review did not report on the accuracy of the test to distinguish the disease state.
Subject classifications are made from the FC scores of the rs-fMRI datasets using machine learning (ML) methods. A commonly applied technique is the support vector machines (SVMs) methods that are applied in patient stratification studies to make inter-group classifications (Dyrba, Grothe, Kirste, & Teipel, 2015;A. Lee, Ratnarajah, Tuan, Chen, & Qiu, 2015;Park et al., 2017). Another approach is to use Gaussian process logistic regression (GP-LR) models (Challis et al., 2015). Diagnostic accuracy can be achieved by computing various classifiers that are selected from discriminating features of the multimodal imaging after performing tests using the training dataset (Teipel et al., 2016). A popular type of supervised ML is the support vector machine (SVM) method. SVM has been utilized by various researchers to boost the diagnostic results from multimodal imaging by incorporating multiple kernels in its algorithm (Jin et al., 2020;Q. Zhao, Chen, & Zhou, 2016). Apart from SVM, other more sophisticated algorithms such as convolutional neural networks have been used to discriminate between AD and HC (Qureshi, Ryu, Song, Lee, & Lee, 2019).
The main goal of this review is to examine the benefits and the issues of applying ML algorithms to assess rs-fMRI datasets for improved diagnostic accuracy of discriminating AD/MCI from HC. We also discuss the limitations of multimodal and multicenter studies, as well as recommend the future direction of research in this field. To the best of the authors' knowledge, this is the only systematic review in the existing literature that is focused on studies that perform rs-fMRI-based classification to detect AD and its prodromal stage.

| METHODS AND MATERIALS
We first provide some basic definitions with respect to the classifier methods that are utilized in evaluating the diagnostic accuracy of rs-fMRI. This is followed by the study protocol of our systematic review.
The study protocol includes the study design, search strategy used when screening the articles from the medical databases, selection criteria for identification of eligible articles, assessment of bias, and data extraction. Next, we present the data using tables in Section 3.
Afterward, the technicalities of performing ML, for example, SVM, linear regression, logistic regression, and convolutional neural networks, are described in Section 3.3. We also discuss the similarities and differences among the various articles and propose recommendations for future works.

| Study design
The systematic review method used to formulate the study design was adopted from Campbell et al. (2015). The results of this review are reported based on the Prepared Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) method (Moher, Liberati, Tetzlaff, Altman, & The, 2009 Database, and for any on-going reviews similar to this study. This review protocol has been registered with the International Prospective Register of Systematic Reviews (PROSPERO) with the registration number CRD42020181655.
Scopus, PubMed, DOAJ, and Google Scholar databases were searched for articles using a combination of the keyword using MESH terms. Our search strategy in the various database was as the following-.
SCOPUS: TITLE-ABS-KEY (("resting-state functional MRI" OR "resting-state fMRI" OR rs-fMRI) AND (Alzheimer's disease OR AD OR mild cognitive impairment OR MCI) AND (accuracy OR classification)) DOAJ TS = (("resting-state functional MRI" OR "resting-state fMRI" OR rs-fMRI) AND (Alzheimer's disease OR AD OR mild cognitive impairment OR MCI) AND (accuracy OR classification)) PubMed: (((((("resting-state functional MRI") OR "resting-state fMRI") OR rs-fMRI)) AND ((((Alzheimer's disease) OR AD) OR mild cognitive impairment OR MCI)) AND (((((accuracy) OR "classification")) We sourced for relevant published articles through December 3, 2020. The combined articles obtained from the search were screened for duplicates and the resultant articles underwent further screening as highlighted in the subsequent sections ( Figure 1).

| Inclusion criteria
The review paper included published original articles that met the following criteria: peer-reviewed articles written in the English language, the articles were sourced from journal publications until December 3, 2020, and the articles included were observational studies of human subjects, which included case-control, cohort, and crosssectional studies that utilized rs-fMRI and the DMN to quantify and correlate FC between AD or MCI with HCs. Furthermore, the articles must have used established AD or MCI diagnostic criteria, for exam-

| Exclusion criteria
We excluded articles by the following criteria: (a) Review articles, (b) case reports, (c) case series, (d) articles written in foreign languages, that is, other than the English language, (e) animal studies, and (f) articles with studies using imaging tools other than rs-fMRI, for example, structural MRI, EEG, MEG, or PET.

| Data extraction
We conducted the literature search using the databases mentioned above. Two of the co-authors (B. I., S. S.) reviewed and independently screened the articles from the search results based on the titles and abstracts for potential inclusion into this review. Only the final screened articles agreed upon by both the authors were considered for the manuscript synthesis. In accordance with the PRISMA protocol, data extracted from each primary study included: author, year, country, number of subjects (patients and controls), age of the subjects, MMSE scores, rs-fMRI imaging protocol, and analysis method, sensitivity scores, and specificity scores.

| Description of the articles included
Tables 1 and 2 summarize the main characteristics of the selected articles, which had assessed the rs-fMRI diagnostic performance for detecting DMN abnormalities among AD and MCI subjects. The majority (64%) of these articles had ≤20 subjects per group. Five articles (46%) used SBA and three studies (27%) used ICA type of analysis. While one article (9%) utilized both SBA and ICA methods, two articles (18%) used GTA as their method of analysis (Table 3).
Interestingly, large percentage of these studies (55%) were conducted in Asia, with China having 5 out of the 7 studies from the region, the other two being from Japan and Korea, respectively. -not rs-fMRI (n=10) -proceedings/ reviews (n=14) -methods design paper (n=8) -interventional study (n=2) -not evaluating AD (n=6)

Identification
No. of articles after full text assessment: -full text not in English (n=3) full text not available (n=10) -did not evaluate diagnostic accuracy of rs-fMRI in AD (n=9) No. of AD only articles included (n=13) No of records after removing duplicates (n=46) No of records remaining after removing: -not rs-fMRI (n=2) -proceedings/ reviews (n=9) -methods design paper (n=4) -not evaluating MCI (n=3) No of articles after full text assessment criteria: -full text not in English (n=2) -did not evaluate diagnostic accuracy of rs-fMRI (n=3) • 4 • 8 • 60-80 (67.9 ± 4.5) • (72.5 ± 7.9) • (29.3 ± 1.6) • (29.5 ± 0.8) • Moderate to severe AD: • (73.61 ± 4.76) • Very mild to mild AD: • (23.84 ± 3.90) • Moderate to severe AD: 3.3 | Types of machine learning methods utilized to classify Alzheimer's disease subjects Machine learning (ML) is a form of artificial intelligence application that utilizes computer algorithms. The basis of ML is dependent upon the ability of the computer program to leverage algorithms. Hence, ML can automatically learn and improve from experience gathered on an independent training dataset based on statistical models. There are various computer programming languages for ML such as Python, Java, R, and JavaScript. These programs can perform ML in any one of the three types of ML, which are supervised learning, unsupervised learning, and reinforcement learning. Supervised learning is the basic type of ML that is frequently used in classification for diagnostics and automated image interpretation. This type of learning can also be used to predict an outcome, such as the occurrence of a disease by using regression analysis. To this end, the training dataset needs to be labeled correctly and provides the algorithm with a fundamental concept of the problem, solution, and data points to be dealt with.
In rs-fMRI, the data points are the "nodes" and the FC between the "nodes" are called "edges". The probability of the number of connections or "edges" arising from the nodes gives the weightage of the

FC. A Bayesian network (BN) is a probabilistic graphical model that
represents a joint probability distribution over a set of variables.
Nodes in a BN graph represent variables of interest, and edges represent the probabilistic associations among variables. In a BN, each node has a conditional-probability distribution, which quantifies the association between that variable and the variables with which it is associ- between patients with mild AD and HCs (Balthazar et al., 2014). This indicates that even at the early stage of the AD, DMN moderately differentiates AD patients from HCs. Dai et al. (2012)  • Structural MRI, which was used to measure regional gray matter volume • rs-fMRI, using amplitude of lowfrequency fluctuations (ALFF), regional homogeneity (ReHo), • Framewise displacement (head motion) showed comparable displacements across sites, for example, cognitively impaired patients showed slightly more head motion than controls.
• The foreground-to-background energy ratio, the fractional ALFF in PCC, and the mean FC between PCC and anterior mPFC indicated no outlying center.
• tSNR was significantly different between certain centers.    (Hojjati et al., 2017). In several of these studies, rs-fMRI demonstrated a high diagnostic power in classifying abnormalities of the DMN among MCI patients compared to the HCs (Figure 2).

| DISCUSSION
Although the FC of DMN has been explored as a biomarker for distinguishing patients with AD and MCI from HCs (Brier et al., 2012;Cha et al., 2013;Griffanti et al., 2015), no compiled review about its diagnostic power has been done before this. The impaired FC of DMN may be analyzed using SBA, ICA, and GTA methods of analyses, and all these methods can be used to classify patients with AD and MCI.
To the best of our knowledge, this is the first review to determine the diagnostic power of rs-fMRI to detect impairments in the FC of the DMN, for discriminating AD and MCI subjects from HCs. The articles included in this review reported variable diagnostic powers of rs-fMRI in characterizing AD and MCI patients, by using a variety of protocols, that is, measurement of DMN FC alone (Koch et al., 2012;Miao et al., 2011), DMN FC correlated with MRI-measured cortical thickness (Balthazar et al., 2014;Park et al., 2017), DMN FC measurements along with other resting-state measures such as DMN FC with PET/CT FC (Yokoi et al., 2018) and DMN FC with regional cerebral blood flow measurements (rCBF; Zheng et al., 2019), respectively.
In differentiating AD patients from HCs, most of the primary articles used SBA analysis, all of which reported that AD patients had weaker FC between the PCC and other brain regions (Balthazar et al., 2014;Dai et al., 2012;Koch et al., 2012;Yokoi et al., 2018;Zheng et al., 2019). This imaging biomarker, that is, the PCC, is able to provide an average sensitivity of 75.2% (ranging between 65.7 and 100%), and an average specificity of 74.9% (ranging between 70 and 95%) for distinguishing patients with AD, hence, indicating a moderate diagnostic power of DMN in differentiating AD patients from HCs.

T A B L E 4 (Continued)
Author ( • 59 brain neural pathways based on a priori knowledge were analyzed • 116 nodes were identified and the FC between nodes at paired brain regions was measured by the strength of the linear relationship depicted by r • Three linear classifiers: Naïve Bayesian (NB); logistic regression; and SVM • One decision trees classifier: RF • Diagnostic performances were evaluated on a pathway-based approach and a region-based approach SVM classification model gave the best diagnostic accuracies for discriminating MCI from HC, for both the pathwaybased approach and a regionbased approach. • ROC curves were plotted The diagnostic performances of the competing methods were analyzed with HMP and without HMP. The best results were achieved with HMP in regression in the multi-spectrum analysis • Neuroanatomic volumetric indices were extracted from the segmentation and parcellation output.
• FC analyzed based on SBA. • Second network comprised four edges and five nodes, located bilaterally in precuneus as well as in the parahippocampal, fusiform, and superior temporal gyri in the right hemisphere.
• Third network comprised nine edges and eight nodes, located mostly in the left hemisphere.
Khazaee et al. (2017)  Network-based statistics were performed on the weighted raw rs-fMRI connectivity matrices to identify impaired sub-networks in the MCI-C and MCI-NC groups.
• First network had two edges and three nodes, specifically one node within the precuneus and the other two nodes within the cerebellum.
• Second network had three edges and four nodes within the vPFC, anterior insula, VFC, and occipital lobe.
• Third network had two edges and three nodes within the temporoparietal junction, occipital lobe, and lateral cerebellum. Optimal features based on sMRI data using Destrieux atlas and rs-fMRI data using the Dosenbach atlas gave the best accuracy for discriminating between MCI-C with MCI-NC.
Optimal features based on sMRI data using Destrieux atlas and rs-fMRI data using the Their method achieved a high diagnostic power, with a sensitivity of 86% and specificity of 78%, respectively, with better results achieved when using the pathway-based approach compared with the region-based approach for classifying MCI from HCs .
In essence, rs-fMRI can detect impairment of the DMN FC and can serve to identify important anatomical biomarkers for discriminating AD and MCI patients from HCs. When combined with other parameters such as cortical thickness, rCBF, or analyzed using combination of multivariate analysis, rs-fMRI has good diagnostic power for detecting AD and MCI.

| Limitations and recommendations for future works
The relatively small sample size in most of the articles leads to a reduced power of the studies. Restrictions of the studies to only include subjects with early AD had to be made due to the constraints of performing the investigation on non-cooperative patients with advanced AD. Furthermore, it is important to note that although MCI T A B L E 4 (Continued) Author ( may occur as a prodromal condition to AD, it can also occur in vascular dementia or even in cognitively healthy elderly persons without progressing to AD. Moreover, the conversion rate of MCI to AD is usually meager. Therefore, longitudinal studies, as opposed to identifying neural FC changes using a single time-point rs-fMRI study, can best assess whether an MCI patient will develop full-blown AD. Additionally, there is a need for further improvement and standardization of rs-fMRI patient selection criteria, acquisition, imageprocessing, and data analysis. The establishment of local populationbased database of fMRI studies involving AD subjects can also help in improving the suitability of comparison. Multicenter rs-fMRI using SBA FC has limited accuracy in the discrimination of AD and MCI cases from HC and requires careful data quality checks beyond the evaluation of global quality metrics, including visual inspection of all the data (Teipel et al., 2017). Furthermore, the combination and integration of multimodal imaging and clinical markers, introduce innumerable classifiers for the improved diagnostic accuracy of detecting and predicting AD. Although it appears very enticing to incorporate numerous multimodal features, nevertheless, this poses a challenge for ensuring homogeneity of datasets and hinders consistency of results. Other rs-fMRI features of engineering models that go beyond the classical Pearson correlation FC and ICA, that is, regional homogeneity (ReHo), fractional amplitude of low-frequency fluctuation ((f)ALFF), and dynamic FC need to be explored further to optimize the wealth of information available on rs-fMRI datasets.
Additionally, novel computational models using convolutional neural networks that use 3D-deep learning frameworks are the way forward.
There is potential for developing the utility of this technique by incorporating biomarker-based serial labeled data and domain transfer learning methods.

| CONCLUSION
The assessment of the DMN FC based on rs-fMRI analytic methods, has an excellent potential as a diagnostic tool for AD, particularly when using multivariate analysis to combine SBA and ICA methods of analyses. Nevertheless, the rs-fMRI protocols and analytical methods need to be more standardized to achieve uniformity in reporting improved diagnostic power.

ACKNOWLEDGMENTS
The research grant was awarded to Associate Professor Dr Subapriya Suppiah by the Malaysian Ministry of Education, that is, the

CONFLICT OF INTEREST
The authors declare and report no conflict of interest.

Subapriya Suppiah conceptualized the study design. Buhari Ibrahim and
Nisha Syed Nasser carried out the literature search, data extraction, and quality assessment. Buhari Ibrahim wrote the manuscript first draft.

Subapriya Suppiah, Normala Ibrahim, Mazlyfarina Mohamad, Hasyma
Abu Hassan, and M Iqbal Saripan edited the manuscript, verified the data, and provided critical feedback to help shape the research.

DATA AVAILABILITY STATEMENT
Data sharing is not applicable to this article as no new data were created or analyzed in this study.