Alzheimer’s disease (AD) is a typical neurodegenerative disease among older people, and its incidence is increasing every year. In the early stages of AD, individuals experience a decline in memory function. This decline is accompanied by progressive damage to various higher cognitive functions, ultimately leading to a gradual loss of the ability to live independently and, eventually, death. The prevailing conventional view is that AD is mainly associated with excessive deposition of extracellular beta-amyloid protein (Aβ) in neurons and hyperphosphorylation of Tau protein in neuronal cells. The excessive accumulation of Aβ protein in the brain forms extracellular plaques or senile plaques, which have toxic effects on neurons. They can have a complex impact on a patient’s brain function (Masters et al., 2015). Therefore, the identification of brain abnormalities specific to the early stages of AD is crucial to the study of brain injury in AD.

Functional magnetic resonance imaging (fMRI) provides spatial and temporal information about the spontaneous activity of the brain, and its high spatial resolution allows us to explore the interactions between brain regions. Resting-state functional MRI (rs-fMRI) is the measurement of low-frequency fluctuations (LFFs) in the brain’s blood oxygen level-dependent (BOLD) signal. This helps reflect the brain’s spontaneous neural activity in the absence of a specific task (Raichle et al., 2001; Grodd and Beckmann, 2014). It has become one of the main imaging tools for exploring the neural mechanisms of the brain. Resting-state functional MRI-based brain function analysis is frequently employed in clinical research for various neurological disorders, including Alzheimer’s disease (Barkhof et al., 2014), Parkinson’s disease, schizophrenia, depression, and more (Griffanti et al., 2016; Khazaee et al., 2016, 2017; Wang et al., 2016). Especially for some patients with cognitive impairment, it is difficult to perform cognitive tasks during MRI scanning. Therefore, resting-state fMRI provides an effective way to study the brain function of AD patients (Dosenbach et al., 2010).

Current methods of analyzing resting-state fMRI data, such as functional connectivity (FC), are often based on the assumption that brain activity is at a steady state (Arbabshirani et al., 2013). In recent years, it has been found that resting brain activation is not stable; instead, it transitions among multiple patterns. Moreover, the corresponding activation patterns and transition behaviors of patients with brain dysfunction were significantly different from those of healthy subjects (Griffanti et al., 2016; Khazaee et al., 2016, 2017; Wang et al., 2016). Therefore, considering only static functional connectivity is not sufficient to explain the time-varying dynamic information interactions of the brain; rather, the dynamics of functional connectivity should be investigated to reveal the complex and variable mechanisms of brain networks. In recent years, a number of dynamic functional connectivity (dFC) methods have been developed that focus on the time-varying properties of functional connectivity in the brain. Sliding window correlation (SWC) is the most widely used method (Chang and Glover, 2010; Allen et al., 2014). It reflects the dynamics of brain networks by calculating functional connectivity over different periods of time. However, the sliding window method suffers from many limitations (Hindriks et al., 2016). In particular, the choice of window length is crucial. Long windows are insensitive to rapid changes in connectivity, while short windows are more susceptible to high-frequency noise fluctuations (Leonardi and Van De Ville, 2015). Therefore, many alternatives to the sliding window method have emerged, such as time-frequency domain-based analysis, point process analysis (PPA) methods (Liu and Duyn, 2013; Chen et al., 2015), sliding-window methods based on temporal functional modes (TFMs; Smith et al., 2012), PPA-inspired coactivation patterns (CAPs; Liu and Duyn, 2013; Liu et al., 2018), etc.

In contrast to the sliding window-based approach, which considers spontaneous brain activity to be slowly changing, Tagliazucchi et al. proposed that information related to resting brain activity can be tied to discrete events and that point process analysis (PPA) which includes only relevant time points can contain the same information as conventional full-time course analysis (Tagliazucchi et al., 2011, 2012, 2016). Based on this idea that meaningful information can actually be obtained from the observation of individual frames, a new alternative to functional connectivity analysis is proposed.

Inspired by the PPA method, Liu and Duyn (2013) and Liu et al. (2018) extracted time frames with activation levels above a specific threshold in the region of interest (ROI) and obtained CAPs by a clustering algorithm. These CAPs are highly similar to those obtained by correlation analysis. Subsequent improvements to the CAP technique have been developed, such as the extension to a data-driven whole brain analysis method (Liu et al., 2013). This allows for the generation of CAPs with different activation patterns and avoids the effects of selecting a specific activation threshold of the ROI.

In addition, the brain has complex dynamic properties and is capable of generating oscillatory waves at many different frequencies. Synchronous oscillatory activities of neural networks are regarded as the key to realizing the function. Frequency-specific studies initiated by Buzsaki et al. found that neural networks in the mammalian forebrain display several oscillatory bands, and the average frequency of oscillations forms a linear relationship on the natural logarithmic scale. The adjacent frequencies were separated with a constant ratio. They suggested that the separate bands are generated by different oscillators, each possessing specific properties and physiological functions (Penttonen and Buzsáki, 2003; Buzsáki and Draguhn, 2004). This extended Buzsaki’s framework to low-frequency oscillations of the BOLD signal, subdividing the spontaneous BOLD waves into four different frequency ranges. Slow-2 (0.198–0.25 Hz) and slow-3 (0.073–0.198 Hz) were mainly confined to the white matter. Slow-4 (0.027–0.073 Hz) and slow-5 (0.01–0.027 Hz) oscillations were clearly detected in the gray matter. The respiratory and heart signals were in the range of slow 2–3, while the oscillations underlying the resting-state functional connectivity were mainly in the range of slow 4–5 (Zuo et al., 2010). Several studies have shown that the low-frequency oscillation pattern and functional connectivity in MCI and AD is frequency dependent (Han et al., 2011; Wang et al., 2011; Liu et al., 2014; Li et al., 2017). The pattern of disruption of spontaneous neural activity and the ALFF/fALFF of brain regions may also differ between subbands (slow-4 and slow-5; Han et al., 2011; Liu et al., 2014). It has also been proposed that classifiers trained by fusing multiband ALFF and fALFF features have good performance in distinguishing NC, MCI and AD (Yang et al., 2018). In addition, it has been found that band segmentation significantly improves the accuracy of MCI classifiers using graph theory-based functionally connected networks (Wee et al., 2012). Another study found that the coupling between global functional connectivity (FC) and low frequency oscillatory amplitude (ALFF) in the brains of dementia patients showed abnormalities/losses in certain brain regions compared to healthy subjects and that changes in this coupling were sensitive to band range (Mascali et al., 2015). Therefore, it seems necessary to further investigate the specific changes in brain regions in different frequency bands.

In this study, we used a multiband CAP approach to model the state space and study brain dynamics in AD and NC. Unlike previous studies, we focus on the frequency dependence of the CAP. fMRI data in the original frequency band of 0.01–0.1 Hz (LFO band) were subdivided into the slow-5 band (0.01–0.027 Hz) and slow-4 band (0.027–0.073 Hz) to calculate the CAP. In addition, we used classical dynamic features of state occupancy (Occurrence), mean duration (Duration) and frequency of occurrence (EntryRate) of CAPs to measure changes in the dynamics of resting-state brain spontaneous activity. We further analyzed the correlation between the dynamic characteristics and the subjects’ clinical indices to explore potential relationships between altered brain dynamics and disease pathology.

A total of 21 CAPs in three different frequency bands are shown in Figure 2. The seven CAP patterns in different frequency bands are shown vertically from left to right, where the order is determined by the proportion of occurrence of CAPs among NC subjects in descending order, i.e., CAP 1 has the highest proportion, and CAP 7 has the lowest proportion. We calculated the Z score for each CAP, and the value of voxels above positive or negative 1.5 times the standard deviation are shown in Figure 2.

Our study focused on resting-state fMRI data of patients with AD using whole-brain CAP analysis in multiple frequency bands. The findings revealed similar CAP patterns at different frequency bands, but the occurrence of patterns was different. Furthermore, there were significant alterations in CAPs associated with the default mode network (DMN) and the ventral/dorsal visual network (dorsal/ventral VN) when comparing the AD and NC groups. The findings also revealed that delineating subbands was more helpful in distinguishing AD from NC in terms of CAP.

The spatial mapping of all CAPs at different frequency bands and the results of the cross-band correlation matrix indicated that many CAPs between different frequency bands were highly similar. However, the occurrence of similar CAPs between different frequency bands was not consistent. For example, CAP 1 in the LFO band, CAP 2 in the slow-5 band, CAP 7 in the LFO band and CAP 4 in the slow-5 band exhibit highly similar spatial patterns, but their occurrence patterns differ. Conversely, CAPs like CAP 3 in the LFO band and CAP 6 in the slow-5 band share similar patterns and the same occurrence order. Similar results are seen between the LFO-slow-4 and slow-5-slow-4 bands. The resting brain constantly switches between multiple steady states (Chang and Glover, 2010; Smith et al., 2012; Liu et al., 2013), and CAP divide spontaneous activity in the resting brain into multiple substates, i.e., patterns of repeated coactivation or deactivation of different brain regions, based on differences in the transient spatial activation patterns of the brain. Previous studies have shown that there are both similarities and differences in the activation of brain regions and network connections of different frequency bands in AD patients. Our results were consistent with these results. Each CAP represents a state of spontaneous brain activity in the resting state of the brain, and a similar pattern of brain activity can be repeated in different frequency bands, but the occurrence was altered. At present, there are relatively few studies on the use of the frequency dependent whole-brain CAP method in the analysis of resting-state fMRI data for neurological diseases. Hang et al. used the Frequency dependent CAP method to study schizophrenia and found that the spatial patterns of CAP remained consistent across different frequency bands, which is familiar with our findings (Yang et al., 2022).

The characteristics of CAPs associated with the DMN and ventral/dorsal visual network were altered significantly between the AD and NC groups. Some CAP patterns with activation or inhibition of DMN network-related brain areas were significantly different between AD and NC subjects. The CAP3 spatial map in the LFO band showed activation and inhibition in the primary visual cortex, the default mode network and prefrontal-related brain regions. The activated brain areas include the talar sulcus, lingual gyrus, occipital gyrus, cuneus and precuneus, which are mainly associated with some higher brain functions, such as episodic memory, visuospatial processing, self-reflection and consciousness (Cavanna and Trimble, 2006; Xia et al., 2013). In CAP3, inhibitory brain areas encompass portions of the prefrontal lobe, including the superior frontal gyrus, medial frontal gyrus, and middle frontal gyrus, primarily concentrated in Brodmann areas 9 and 10. These regions are associated with higher-order cognitive functions like working memory and self-awareness (Boisgueheneuc et al., 2006). In addition, there was a significant positive correlation between the state percentage of CAP3 and the cognitive level of the subjects. The patterns that were altered in the slow-5 band were CAP2, CAP3 and CAP5. Although, CAP3 also showed inhibition in the temporal gyrus and angular gyrus of the temporal cortex (area 39 of Brodmann’s subdivision) and in the posterior cingulate cortex (area 23 of Brodmann’s subdivision). The posterior cingulate cortex is the central node of the default mode network, which is connected to intrinsic control networks and more active during periods of inattention (e.g., recall of episodic memories, self-programming, daydreaming). Conversely, the network is inhibited when attention is externally focused (e.g., working memory, meditation; Brewer et al., 2013). It has been found that the PCC is not only related to emotional processing but also plays a key role in cognitive functioning and that abnormalities in the PCC are often associated with cognitive impairments, including memory function and concentration problems (e.g., Alzheimer’s disease, traumatic brain injury, ADHD; Leech and Sharp, 2014). At a physiological level, the posterior cingulate cortex shows abnormalities in metabolic levels and synaptic connections early in the brain of AD patients (Scheff et al., 2015). The slow-5 band CAP5 shows activation in the posterior cingulate cortex, angular gyrus and occipital gyrus and inhibition in the precentral gyrus and postcentral gyrus (Brodmann areas 3 and 4), which are associated with the primary motor and sensory cortex of the brain. However, the inhibitory brain regions in this band also contained several of the typical DMN-related brain regions described above. Furthermore, the correlation analysis revealed that all of the above CAPs showed significant correlations with subjects’ MMSE scores. Our findings are consistent with previous studies that have found differences in some dynamic indicators of the default mode network (DMN) and ventral/dorsal visual network (dorsal/ventral VN)-related CAP between AD and NC (Ma et al., 2020). The PCC, precuneus, angular gyrus and medial prefrontal lobes mentioned above belong to the typical DMN network, whereas the lingual gyrus, talar sulcus and cuneus belong to the typical dorsal and ventral visual networks. Damage to the DMN network and deterioration of higher visuospatial abilities have been recognized as early and prominent clinical signs of AD. Although the neurophysiological basis remains controversial, most previous visual task-based studies have concluded that this is due to dysfunction of the visual perceptual stream associated with cognition (Franceschi et al., 2007; Paxton et al., 2007). In addition, another fMRI study showed altered interactions between the PCC or precuneus and the visual perceptual network in MCI and AD compared to NC in a visual task (Krajcovicova et al., 2017). Our result is consistent with this finding and suggests that the interaction between the PCC or precuneus and the visual perceptual network is altered in AD and MCI patients. This has shown that the higher-order visual processing dysfunction seen in AD and MCI patients may arise from inefficient communication mechanisms between these regions.

Based on CAP analysis, delineating subbands was more helpful in distinguishing AD from NC. Our dynamic characterization of the CAP in the two groups of subjects in different frequency bands revealed that the only CAP that differed between AD and NC in the LFO band was CAP 3, while the number of CAP patterns with component differences increased to three in the slow-5 band and to five in the slow-4 band. These results suggest that the subband division may help to reveal differences in brain activity. Several previous studies (Han et al., 2011; Wee et al., 2012; Mascali et al., 2015; Yang et al., 2018) found that decomposing the BOLD signal into smaller bands and analyzing subbands would provide a more sensitive representation of the spatiotemporal information of brain activity, allowing for better identification of abnormalities in brain function in AD patients. Furthermore, our correlation analysis revealed that dynamic characterization of CAPs that showed significantly different between AD and NC groups almost realized significant correlations (positive or negative) with MMSE score. These studies suggest that CAP analysis has the potential to be a clinical indicator of AD.

Various limitations need to be taken into consideration in future research. First, the current study investigated the subbands CAPs, the subbands definition might constrain the CAP results. More research is needed to investigate the subbands properties of CAPs by using wavelet-based or other frequency bands in the future. Second, further study should focus on using machine learning based on subbands CAP to distinguish the NC, AD, which result may be useful in the clinical diagnosis of AD. Finally, the relationship between subbands CAPs and physiological meaning is still unclear. More studies are needed to fully understand the physiological meaning of the CAPs. Hence, further studies that combine CAPs with genotype or bio-analyses are necessary to identify main reason behind the altered CAPs We also plan to apply frequency dependent whole-brain CAP method to other diseases, such as autism and major depressive disorder.

In this study, we used whole-brain CAP analysis in multiple frequency bands. The findings revealed that similar CAP spatial patterns are shown in different frequency bands, but the dynamic characterization of similar patterns in subbands was different. In addition, CAPs associated with the default mode network (DMN) and the ventral/dorsal visual network (dorsal/ventral VN) were altered significantly between the AD and NC groups. The findings also revealed that delineating subbands was more helpful in distinguishing AD from NC in terms of CAP.

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding authors.

S-PZ, YL, NY, and Z-GH designed the project. BM, S-PZ, CL, JJ, and TZ analyzed the data. S-PZ, YL, BM, C-WS, SA, NY, and Z-GH interpreted the results and wrote the manuscript. All authors participated in the revision of the manuscript, contributed to the article, and approved the submitted version. All authors provide approval for publication of the content and agree to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) data base (https://adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at https://adni.loni.usc.edu/wpcontent/uploads/howtoapply/ADNI AcknowledgementList.pdf.