The default mode network (DMN) is the most frequently investigated functional brain network in Alzheimer’s disease (AD), and the intrinsic connectivity of DMN has been reported disrupted in a variety of resting-state fMRI studies (Greicius et al., 2004; Jones et al., 2011, 2016; Petrella et al., 2011). The DMN connectivity disruption has also been identified in populations at risks for AD, such as elderly people, mild cognitive impairment (MCI) patients, and APOE-ε4 allele carriers (Sorg et al., 2007; Damoiseaux et al., 2008; Filippini et al., 2009; Petrella et al., 2011; Binnewijzend et al., 2012). While an increasing number of studies have pointed to the DMN for major functional network deficit in AD, the relationship between the functions of DMN and the early symptoms of AD is still unclear. Clinically AD first targets recent memory function, impairing the ability to remember recently acquired information, while DMN has been named after its characteristics of showing deactivation in attention-demanding tasks, and its functions have been demonstrated in vast cognitive domains (Andrews-Hanna et al., 2010; Raichle, 2015). The posterior part of DMN (i.e., posterior cingulate/precuneus and bilateral parietal cortex) are often linked to memory processing (Andrews-Hanna et al., 2010; Sestieri et al., 2011; Raichle, 2015), but the parietal regions are functionally inhomogeneous and memory functions are not unique to the DMN regions (Cavanna and Trimble, 2006; Cauda et al., 2010; Sestieri et al., 2017).

Recently, a memory-related brain network has been separated from the posterior part of the DMN, termed as the parietal memory network (PMN, Gilmore et al., 2015). Studies have demonstrated that the PMN’s activity is a reflection of the familiarity of a stimulus in both memory encoding and retrieval, and generally across task conditions (Nelson et al., 2013; Gilmore et al., 2015; McDermott et al., 2017). The PMN has been disassociated from the DMN in different types of memory retrieval tasks (McDermott et al., 2009; Chen et al., 2017), where successful recognition memory activates the PMN, while autobiographical retrieval activates the DMN. The PMN includes the middle cingulate cortex (MCC), the precuneus (PCU), and the inferior parietal lobule/angular gyrus (IPL/AG). While these regions are all adjacent to the DMN or even has been attributed to the DMN in some studies (Littow et al., 2010; Damoiseaux et al., 2012; Manoliu et al., 2014; La et al., 2015), the PMN could be reliably separated from the DMN via different methods in resting-state fMRI studies (Yeo et al., 2011; Hu et al., 2016). Further, our previous study has demonstrated that the PMN in resting-state has an obvious age-dependent inter-individual variability, while the DMN does not exhibit age-dependence (Yang et al., 2012, 2014).

The emergence of the PMN has raised our curiosity to consider its role in AD. The double dissociation between the PMN and the DMN in their roles in memory processing leads to our hypothesis that the PMN is also impaired in AD. Further, the spatial adjacency between the PMN and the posterior DMN regions implies a possibility that the deficits previously we attributed to the DMN may also be related to the PMN. The robust separation of the PMN from the DMN in resting-state fMRI provides a chance to examine this hypothesis.

We examined the above hypothesis by comparing the integrity of the PMN between AD patients and healthy controls (HCs) using independent component analysis (ICA). Two commonly used group ICA algorithms were applied and the effects of different model orders (MOs), which is an important yet undetermined parameter in ICA setting, were explored. The value of the PMN integrity in classifying AD patients was further examined. All the analysis were also performed for the DMN, which worked as a reference network.

In this study, we found that the PMN had significant lower network integrity in AD, compared with HCs. The PMN had a large effect size and exhibited at least equivalent diagnostic ability to the DMN in identifying AD patients. Further voxel-wise comparisons identified significant decreases in functional connectivity in the core regions of the PMN. These findings were robust at different MOs using two types of group ICA methods. The DMN was analyzed as a reference network, and as expected, exhibited decreased network integrity and voxel-wise functional connectivity in AD group but showed a higher variability across MOs in both methods.

ICA is the most commonly used technique to simultaneously investigate multiple brain networks, which could decompose the fMRI data into a set of inter-independent brain networks en bloc in a data-driven manner (Hyvärinen and Oja, 2000; Beckmann et al., 2005). Due to the inherent order ambiguity of the ICA algorithm, how to apply ICA to multiple subjects for group analysis has been an unsettled issue (Calhoun et al., 2001). Template matching and dual regression are two types of solutions to group ICA and have been commonly used in AD-related studies. Briefly, both methods need a set of brain network templates, but they differ in their ways of obtaining individual-level network spatial maps. Specifically, template matching aligns individual-level ICA results by measuring the similarity to the templates, while dual regression adopts a two-step multiple linear regression to reconstruct individual-level time courses and spatial maps. For the dual regression, the brain network templates are usually created from the sample under study using TCGICA, which would also generate components reflecting artifact/noise, besides brain networks. These artifact/noise components are necessary for dual regression to account for the variances from non-neural sources (Griffanti et al., 2016; Nickerson et al., 2017). Previously, template matching usually only needs the brain network templates of interest (Greicius et al., 2004). For brevity and comparability, here we also use the templates generated from the TCGICA for template matching. These two methods have advantages and disadvantages. For template matching, there are no criteria to determine whether a component is correctly matched to a template. Pearson’s correlation coefficient is a commonly employed similarity measure for template matching, which we adopted in the current study, but there are risks that mismatch could happen, especially when the similarity is low (Zuo et al., 2010). Another concern is the component splitting problem, which means a component could be split into several sub-components, especially in high MO settings (Kiviniemi et al., 2009). Splitting of brain networks is currently thought to reflect the refinement of network functions. In template matching, component splitting could result in incomparability among individual-level ICA results to some degree. In this study, we repeated analysis at different levels of MOs, attempting to mitigate this problem. It is also possible that the component splitting reflects the individuality of brain network activities, which could be reliably detected and vary from person to person. The biggest advantage of template matching is that the individual-level ICA fully preserves the variability of individuals, which are expected to be large, especially in patient populations, because the individual-level ICA maintains the data-driven property and actually we should not make too many assumptions for brain network patterns in real fMRI data. For dual regression, the accuracy of the reconstructed individual-level spatial maps and time courses would be degraded with increased individual variability (Michael et al., 2014; Nickerson et al., 2017). However, dual regression could make group comparison straightforward and showed high test-retest reliability (Zuo et al., 2010). There are theoretically improved group ICA algorithms (Lee et al., 2008; Du and Fan, 2013), though they are scarcely used in AD-related studies. In order to make the results in the current study comparable to those in previous studies, template matching and dual regression were used.

The reliability of fMRI-based studies is a great concern in recent years, therefore we tested many key factors which could potentially influence the results. MO is an undetermined parameter in the ICA algorithm, which could be the most influential factor to any unreliable findings so that a wide range of MOs were tested. Relatively high MOs were selected, based on our previous finding that only in relatively high MO, could PMN and DMN be robustly separated (Hu et al., 2016). The variances explained by PMN could be unintentionally removed in the PCA-based dimension reduction process, when the MO is set at low level (for instance, at 20–30). Nuisance regression is a common technique applied in the preprocessing of resting-state fMRI data, in which WM/ventricle signals and head motion parameters were regressed out before functional network analyses. We did not perform nuisance regression in the main analysis, because ICA could split brain networks and artifacts/noises into different components, and thus reduce the influence of non-neural signals for functional network analyses (Salimi-Khorshidi et al., 2014; Pruim et al., 2015b). ICA-based denoising has been demonstrated to outperform nuisance regression and would be less likely to affect signals of interest (Bright and Murphy, 2015; Pruim et al., 2015a). We also included head motion parameter (i.e., mean FD) as a covariate in statistical analysis to mitigate residual head motion. In verification analysis, we also demonstrated that the findings would not be significantly influenced by nuisance regression. Another notable analysis detail is the avoidance of spatial smoothing during the preprocessing. Spatial smoothing could account for the registration inaccuracy and anatomical variability and thus increase the signal-to-noise ratio. However, common spatial smoothing implementation using isotropic Gaussian kernel filter could potentially blur signals from neighboring brain regions. Due to the anatomical adjacency of PMN and DMN, we did not perform spatial smoothing at the cost of decreased statistical power in the group analysis.

In the current study, the network integrity was measured by the similarity between individual-level networks and group-level network templates. The group-level network templates were made by combining both AD patients and HCs. Making sample-specific group-level templates from all subjects is a common practice for dual regression, which would generate components representing both brain networks of interest and of no interest, as well as artifacts/noises (Griffanti et al., 2016; Nickerson et al., 2017). These components of no interest are important to accurately reconstruct the individual-level spatial maps. Although out-of-sample templates were used in some studies (Jones et al., 2016), there was no systematic examination on the influences of out-of-sample templates, and thus unexpected results could occur. In order to make results from different methods comparable, the group-level network templates used in dual regression were also used for template matching. This kind of group-level templates is not optimal to detect group difference, but we replicated the DMN’s abnormality in our sample, which has been consistently observed in AD patients (Greicius et al., 2004; Binnewijzend et al., 2012), indicating our results are valid (yet conservative). In the verification analysis, we also demonstrated that our results are still valid even if we adopted out-of-sample templates.

Our results call for considering the PMN and the DMN separately in clinical neuroimaging studies. The posterior parietal cortex is not functionally homogenous and plays an important role in episodic memory retrieval (Sestieri et al., 2017). The PMN borders on the DMN in the posterior parietal cortex anatomically, but the PMN is not a sub-component of the (posterior) DMN. These arguments are supported by two lines of previous studies (Gilmore et al., 2015): first, unlike the DMN, the core regions of the PMN are consistently and simultaneously associated with memory encoding and retrieval processes in different task conditions (Kim, 2013; Gilmore et al., 2015). The activity of the PMN is a reflection of the subjective familiarity of a stimulus (McDermott et al., 2017). Specifically, the PMN deactivates in perceiving novel stimuli while activates in familiar ones. This unique pattern differentiates the PMN from other brain networks only involved in certain tasks or in either memory encoding or retrieval. The double disassociation between the PMN and the DMN has been observed in different types of memory tasks (McDermott et al., 2009; Chen et al., 2017). In this regard, the PMN shows a more direct relationship to the early symptoms of AD than the DMN, as DMN is involved in multiple but different cognitive functions. Second, the intrinsic functional connectivity profile of the PMN is distinct from that of the DMN. The PMN and DMN could be separated in resting-state fMRI using graph theoretical methods, clustering, and ICA (Power et al., 2011; Yeo et al., 2011; Yang et al., 2012, 2014; Hu et al., 2016). Unlike the DMN, which could be split into sub-networks in high order model ICA, the PMN was immune from the effects of high MO (Kiviniemi et al., 2009; Hu et al., 2016). These evidence indicated that the PMN is not a subordinate of the DMN but an independent functional brain network. There has been a tendency to attribute any medial and lateral parietal regions to the DMN in some studies, even in situations where network analysis was not performed. Due to the spatial proximity of the PMN and the DMN, it is very likely that some findings located in posterior parietal regions have been attributed to the DMN for interpretation. It is also very possible that posterior parietal cortex contains other functional networks, besides the PMN and the DMN. Amyloid deposition, metabolism reduction, and atrophy are also co-occurring in the posterior parietal cortex at the stage of AD (Buckner et al., 2005). This line of evidence that has been used to support the dysfunction of the DMN could also support the dysfunction of the PMN, and needs further clarification using fine-grained parcellation of the posterior parietal cortex.

As AD is an irreversible process, much endeavor has been made in seeking biomarkers to diagnose AD at an earlier stage. In this study, we examined the ability of the PMN in discriminating AD group from the HC group and found that the PMN had a large effect size and at least equivalent predictability to the DMN. Further study on the prodromal stage of AD and longitudinal analysis are needed to clarify which brain network comes first, although the PMN and DMN were both disrupted in AD. We speculate that due to its close association with memory functions, the dysfunction of the PMN may emerge earlier, with the decline of memory performance before a clinical diagnosis of AD has arrived. In addition, there are different methods for network analysis besides ICA, which often make different assumptions on data and have both advantages and disadvantages. Converging evidence from different types of methods could further confirm the findings of the current study.

In conclusion, the current findings show that the integrity of the PMN is disrupted in AD, and that the PMN have a large effect size and at least equivalent diagnostic ability to the DMN in discriminating AD patients from HCs.

ZY, YHu and YHan designed the study. WD collected the data. YHu and NL analyzed the data. All authors contributed to manuscript writing.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.