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

1Alzheimer'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).

2The 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).

3There 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 resting‐state 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.

4In 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.

5Due 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.

6Alternatively, 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 cross‐network 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).

7The 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).

8In 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). Whereas in SBA, the extracted time courses of the selected predefined regions are correlated with the time course of all other voxels.

9Several studies have analyzed the relationship between the FC of the DMN, Mini‐Mental State Examination (MMSE) test scores, and the development of disease among amnestic MCI (aMCI) and AD patients and compared them with HC subjects (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.

10Subject 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).

11The 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.

12We 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.

13The 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).

14A preliminary search was conducted to check for existing reviews in the Cochrane central register, Centre for Reviews and Dissemination such as Database of Abstract of Reviews of Effectiveness, National Health Service Economic Evaluation Database and the Health Technology Assessment Database, Turning Research Into Practice (TRIP) 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.

15Scopus, 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‐.

16SCOPUS: 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))

17DOAJ 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))

18PubMed: ((((((“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”))

19We 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).

20The 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 cross‐sectional 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 example, Diagnostic and Statistics Manual of Mental Disorders‐IV or V (DSM‐4 or DSM‐5; American Psychiatric Association, 2013) or the revised National Institute of Neurological and Communicative Disorders and Stroke and Alzheimer's Disease and Related Disorders Association (NINCDS‐ADRDA; American Psychiatric Association, 2013).

21We 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.

22We 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.

23The quality of the methodology of the primary studies used in this review was assessed using the Quality Assessment of Diagnostic Accuracy Study 2 (QUADAS‐2) tool (Beynon et al., 2012). The tool is designed with 14 questions to determine the risk of bias and applicability in terms of patient selection, index test, reference standard, and flow and timing domains. Based on the 13 questions, which served as a reference guide, each domain was rated low, high, or unclear (Table S1). A study was declared to have a low risk of bias in a domain if many of the questions were positively scored for that domain.

24The summary of the literature search results is in Figure 1. Out of the 226 primary articles obtained for AD studies, 142 duplicates were discarded. Then, 40 nonrelevant articles were removed after screening based on titles and abstracts. Furthermore, 22 articles were removed during the screening of full texts because of noncompliance with our study protocol. Out of the 176 primary articles related to MCI, we removed 130 duplicates and 18 nonrelevant articles during screening. Finally, five articles were removed because they failed to conform to the study inclusion criteria. As a result, 36 articles met our inclusion criteria (13 AD, 14 MCI, and 9 combined AD/MCI articles) and were included in the quantitative synthesis. Thus, the total number of the final included articles was 36 (Balthazar, de Campos, Franco, Damasceno, & Cendes, 2014; Bi, Shu, Sun, & Xu, 2018; Challis et al., 2015; X. Chen et al., 2017; Dai et al., 2012; de Marco et al., 2017; de Vos et al., 2018; Dyrba et al., 2015; Hojjati et al., 2019; Hojjati, Ebrahimzadeh, Khazaee, & Babajani‐Feremi, 2017, 2018; X. Jiang, Zhang, & Zhu, 2014; Jin et al., 2020; Khazaee, Ebrahimzadeh, & Babajani‐Feremi, 2017; Kilian, Bröckel, Overmeyer, Dieterich, & Endrass, 2020; Koch et al., 2012; J. Lee et al., 2015; Y. Li, Wee, Jie, Peng, & Shen, 2014; Lisowska & Rekik, 2019; J. Liu, Pan, Wu, & Wang, 2020; Miao, Wu, Li, Chen, & Yao, 2011; Park et al., 2017; Qian, Zheng, Shang, Zhang, & Zhang, 2018; Qureshi et al., 2019; Schouten et al., 2016; Son, Kim, & Park, 2017; Suk, Wee, Lee, & Shen, 2015; Teipel et al., 2016; Teipel et al., 2017; Wee et al., 2012; Yokoi et al., 2018; Yu et al., 2017; L. Zhang et al., 2020; Y. Zhang, Zhang, Chen, Lee, & Shen, 2017; J. Zhao, Ding, Du, Wang, & Men, 2019; Zheng et al., 2019; Zhu et al., 2014) Overall, the articles had low risk of bias.

25Tables 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.

26Machine 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.

27In 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 associated. There are different types of BNs, based on constraints on allowed model structures. Several articles included in our review demonstrated the diagnostic power of rs‐fMRI in detecting DMN abnormalities for distinguishing AD and MCI patients from HCs, based on the brain regions that demonstrated weaker FC (Tables 3 and 4; Balthazar et al., 2014; Dai et al., 2012; de Vos et al., 2018; Koch et al., 2012; Y. Li et al., 2014; Miao et al., 2011; Park et al., 2017; Yokoi et al., 2018; Zheng et al., 2019). Examples of salient regions of the DMN that demonstrated impaired FC among AD subjects in selected studies are shown in Figure 2.

28Balthazar et al. (2014) evaluated the effect of FC of the DMN, namely impaired FC of the PCC together with cortical atrophy in discriminating 22 patients with mild AD, with 26 age and gender‐matched HC subjects. The authors found a moderate diagnostic power in the FC of the DMN (77.3% sensitivity and 70% specificity) 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) performed discriminative analysis using FC, ReHo, ALFF, and structural MRI gray matter density (GMD) to differentiate patients with mild AD from HCs. Some DMN nodes (including PCG, mPFC, hippocampus [HIPP], and parahippocampus) were found to be the most prominent distinguishing feature between patients with mild AD and HC (Dai et al., 2012). Additionally, in this study, the diagnostic power was 81.25% sensitivity and 68.18% specificity suggesting a high discriminating power of DMN FC among the patients with AD. de Vos et al. (2018) assessed the DMN impairments in AD patients, in a sample of 70 AD and 173 HCs, using GTA to analyze the FC. They reported a low diagnostic power of DMN between AD and controls with 66 and 67% sensitivity and specificity, respectively (de Vos et al., 2018).

29Koch et al. (2012) reported 100% sensitivity and 95% specificity when they utilized ML by combining DMN features from ICA and SBA, to determine the diagnostic power of rs‐fMRI in distinguishing 15 AD patients from 21 HCs. This is an indication that a combination of two methods of analysis yielded a better diagnostic power of DMN abnormalities as a biomarker of AD (Koch et al., 2012). Li et al. (2014) reported that the PCC can serve as a sensitive (73.3%) and specific (93.7%) biomarker for discriminating AD from HC, as shown in ICA analysis of 15 AD patients and 16 HCs (Y. Li et al., 2014). A particular article in our review used Granger causality analysis to study the FC of the DMN among 15 AD and 16 HC subjects, in which they identified impaired connectivity in various DMN hubs among AD subjects (80% sensitivity and 81.25% specificity; Miao et al., 2011). Their results also demonstrated the high diagnostic power of DMN in distinguishing AD from HCs. Furthermore, in their attempt to evaluate the diagnostic power of MRI to detect the alterations in the cortical thickness and DMN FC to classify 41 AD patients and 22 HCs, Park et al. (2017) concluded that both parameters were significant biomarkers of AD, with the latter having 81.3% sensitivity and 74.7% specificity to classify AD from HCs (Park et al., 2017).

30In evaluating the difference between patients with early AD and HCs, Yokoi et al. (2018) compared the spatial retention of an amyloid marker, 18F‐THK5351, and the DMN FC. It was noted that the Prec/PCC is a specific hub for retention of 18F‐THK5351 and that the disruption of both the PCC and the DLPFC account for fMRI diagnostic values of 82.6% sensitivity and 79.1% specificity (Yokoi et al., 2018). Zheng et al. (2019) reported the presence of disruption of the FC of the DMN and cerebral blood flow in the brain regions of 40 AD patients compared with 30 HC: namely, in the PCC, DLPFC, inferior parietal lobule (IPL), MTG, MOG, and Prec regions (Zheng et al., 2019). It was observed that the disruption of the FC of the DMN discriminated AD from HCs, with a sensitivity of 65.7% and a specificity of 73.1%. This particular result showed lower discriminating power of fMRI in the classification of AD patients compared to the previously reported results. Furthermore, this drop in FC had a significant positive relationship with the decrease in the patients' MMSE test scores (Zheng et al., 2019).

31The articles that met our inclusion criteria for studying MCI subjects reported the diagnostic power of rs‐fMRI in detecting impaired DMN FC. Koch et al. (2012) indicated that fMRI identified impaired FC of the DMN and correctly classified 17 MCI patients from 21 HCs (Koch et al., 2012). Using the combined methods of ICA and SBA, this yield a sensitivity of 86.5% and a specificity of 85.1%. This result shows that rs‐fMRI has a high diagnostic power in distinguishing MCI from HCs especially when using multiple comparators. Krajcovicova et al. (2017), by utilizing the SBA method, reported a moderate to high diagnostic power of rs‐fMRI to correctly classify MCI patients (sensitivity and specificity of 76.81% and 88.55%, respectively; Krajcovicova et al., 2017). Hojjati et al. (2017) demonstrated that disruptions in several nodes of the DMN can act as a biomarker for classifying MCI patients, giving a sensitivity and specificity of 83.24% and 90.1%, respectively (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).

32Although 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.

33To 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.

34In 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. Even though SBA yielded good results, nevertheless, the applicability of solely evaluating the PCC is limited (Koch et al., 2012) because the DMN has numerous hubs that are frequently disrupted in AD (Mohan et al., 2016). Other regions of the DMN that were reported to have weaker FC in AD patients included the ACC, LPC, superior frontal gyrus (SFG), medial temporal cortex (MTC), HIPP, DLPFC, IPL, MTG, MOG, and the Prec. The decreased FC among these regions is consistent with those reported in previous studies (Grieder et al., 2018; Griffanti et al., 2015; Rombouts et al., 2009). The deficient FC within regions of the DMN was correlated with MMSE scores, a global cognition and episodic memory measurement test (Balthazar et al., 2014; Yokoi et al., 2018). Moreover, this pattern of impaired DMN FC is in line with the course of early AD pathology, beginning from the MTG and involving the entorhinal cortex, HIPP, parahippocampus, and fusiform gyri (Du et al., 2004; X. Li, Coyle, Maguire, Watson, & McGinnity, 2011). Advantageously, large‐scale network (LSN) of rs‐fMRI brain networks (DMN and dorsal attention network) can be studied using the ICA method. Articles that met our inclusion criteria, which utilized ICA (Miao et al., 2011; Park et al., 2017), reported that rs‐fMRI had moderate to high diagnostic power to distinguished AD patients from HCs, having an average sensitivity of 78.2% (ranging between 73.3 and 81.3%) and an average specificity of 83.22% (ranging between 74.7 and 93.7%). The average sensitivity and specificity of these results were compared with that of the SBA method. This is likely due to the same data being used for SBA was also used to construct independent component networks of the brain. By combining both the ICA and SBA rs‐fMRI methods and multivariate analysis to evaluate the ACC and PCC, Koch et al. (2012) achieved a high sensitivity and specificity, that is, 100 and 95%, respectively, in discriminating AD from HC subjects (Koch et al., 2012). Apart from the SBA and ICA methods, GTA was an additional method used to analyze fMRI findings in AD, which resulted in a moderate sensitivity of 67% and specificity of 66%, respectively (de Vos et al., 2018).

35In comparison, Koch et al. (2012) failed to produce any statistical difference between the FC of the DMN regions, that is, in the ACC, PCC, LPC, SFG, MTC, and HIPP, of the AD and HCs. They also reported a lower diagnostic power of rs‐fMRI, that is, sensitivity of 64.7% and specificity of 95.2%, respectively (Koch et al., 2012). Lee et al. (2015) using SVM determined the diagnostic power of rs‐fMRI to detect DMN FC abnormalities in classifyingaMCI from HCs. 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 (Lee et al., 2015).

36In 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.

37The 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 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, image‐processing, and data analysis. The establishment of local population‐based 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.

38The 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.

39The authors declare and report no conflict of interest.

40Subapriya 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.