Topographic metabolism‐function relationships in Alzheimer's disease: A simultaneous PET/MRI study

Abstract

Disruptions of neural metabolism and function occur in parallel during Alzheimer's disease (AD). While many studies have shown diverse metabolic‐functional relationships in specific brain regions, much less is known about how large‐scale network‐level functional activity is associated with the topology of metabolism in AD. In this study, we took the advantages of simultaneous PET/MRI and multivariate analyses to investigate the associations between AD‐related stereotypical spatial patterns (topographies) of glucose metabolism, measured by fluorodeoxyglucose PET, and functional connectivity, measured by resting‐state functional MRI. A total of 101 participants, including 37 patients with AD, 25 patients with mild cognitive impairment (MCI), and 39 cognitively normal controls, underwent PET/MRI scans and cognitive assessments. Three pairs of distinct but optimally correlated metabolic and functional topographies were identified, encompassing large‐scale networks including the default‐mode, executive and control, salience, attention, and subcortical networks. Importantly, the metabolic‐functional associations were not only limited to one‐to‐one‐corresponding regions, but also occur in remote and non‐overlapping regions. Furthermore, both glucose metabolism and functional connectivity, as well as their linkages, exhibited various degrees of disruptions in patients with MCI and AD, and were correlated with cognitive decline. In conclusion, our results support distributed and heterogeneous topographic associations between metabolism and function, which are jeopardized by AD. Findings of this study may deepen our understanding of the pathological mechanism of AD through the perspectives of both local energy efficiency and long‐term interactions between synaptic disruption and functional disconnection contributing to the clinical symptomatology in AD.

The topographic association between glucose metabolism and functional connectivity was investigated in Alzheimer's disease using simultaneous PET/MRI and multivariate analyses. We identified three distinct yet optimally correlated patterns, spanning both one‐to‐one corresponding regions and distant non‐overlapping regions. The metabolism‐function association was disrupted in AD and correlated with cognitive decline.

Practitioner Points.

1. This study systematically investigates the topographic association between glucose metabolism and functional connectivity in Alzheimer's disease.

2. In three pairs of distinct but optimally correlated topographic patterns, glucose metabolism and functional connectivity were associated with each other not only in one‐to‐one corresponding regions, but also in distant and non‐overlapping regions.

3. The metabolism‐function association was disrupted in AD and correlated with cognitive decline.

1. INTRODUCTION

Alzheimer's disease (AD) is characterized by gradually impaired episodic memory and other cognitive functions. Apart from amyloid and tau aggregations, which are considered the hallmarks of AD, it is widely recognized that disruptions of energy metabolism and network function occur and contribute to progressive neurodegeneration and cognitive decline (Perovnik et al., 2023; Sala et al., 2020). For example, glucose hypometabolism, an indicator of neural and synaptic dysfunction, shows a particular topographic pattern in AD and has a predictive value of future cognitive decline (Kato et al., 2016). Moreover, disruptions of large‐scale brain networks, with a main focus on the reduced intra‐ and inter‐network functional connectivity of the default mode network (DMN), present in patients with mild cognitive impairment (MCI) and AD and are associated with the earliest accumulation of β‐amyloid (Brier et al., 2012; Palmqvist et al., 2017). Nevertheless, the precise nature of how metabolism is related to function in AD and along with cognitive decline remains not fully understood. A better understanding of this relationship would be valuable for unraveling the pathological mechanism of AD.

In the brain, a substantial amount of energy derived from glucose metabolism is used to support intrinsic functional activity and information processing in an efficient and economical fashion (Achard & Bullmore, 2007; Logothetis et al., 2001; Niessing et al., 2005), which implicates a fundamental link between glucose metabolism and functional connectivity (Fouquet et al., 2009; Sperling et al., 2010; Teipel et al., 2015). With the development of multimodal imaging techniques, especially hybrid PET/MRI, a great effort has been made to explore the spatial correlation between glucose metabolism and functional connectivity and generated complex results. Among them, direct correlations have been demonstrated in specific brain regions such as the DMN and the salience network in both healthy subjects and patients with MCI and AD (Aiello et al., 2015; Ding et al., 2021; Drzezga et al., 2011; Marchitelli et al., 2018; Passow et al., 2015; Riedl et al., 2014). On the other hand, Manza et al. did not find metabolic‐functional associations in the association or sensorimotor cortices (Manza et al., 2020); and Shokri‐Kojori et al. demonstrated a mismatch between energy consumption and neuroglial activity in particular brain regions (Shokri‐Kojori et al., 2019). Furthermore, several studies, including our previous research, have shown that the metabolic‐functional relationship could be altered during the progression of AD and might exert a detrimental effect on cognitive performance (Ding et al., 2021; Maleki Balajoo et al., 2022; Marchitelli et al., 2018; Martin Scherr et al., 2021; Zhang et al., 2022).

Most of the above studies utilized regional‐based univariate analytical approaches to investigate the correlations between metabolism and function in colocalized regions. Although valuable evidence has been provided, a study from a topographic perspective is still needed to systematically understand the associations that may be hidden in high‐dimensional metabolic and functional information and settle the discrepancy. The complex cognitive functions rely on the brain's hierarchical organization, which consists of segregated networks interacting between distributed brain regions (He et al., 2009; Wig, 2017). Under this framework, changes in local energy consumption may be the cause and consequence of functional alterations in remote brain regions (Strom et al., 2022). We thus hypothesized that the association between glucose metabolism and functional connectivity should not be limited to one‐to‐one correspondence, but also occur through a distributed and heterogeneous pattern. In addition, the metabolism‐function relationship should be disrupted by AD, along with cognitive decline.

In this study, we applied a multivariate analytical approach (i.e., sparse canonical correlation analysis) to evaluate the topographic relationship between glucose metabolism and functional connectivity and its linkage with cognitive decline in a cohort of cognitively healthy subjects and patients with AD and MCI. The multivariate methods are able to capture interrelated patterns, making them ideally suited for detecting complex effects hidden in high‐dimensional datasets, providing better interpretation and robustness to noise compared to univariate methods (Calhoun & Sui, 2016). In addition, data of ¹⁸F‐fluorodeoxyglucose (¹⁸F‐FDG) PET and functional MRI (fMRI) were simultaneously acquired on a hybrid PET/MRI scanner, which could minimize the confounds caused by image registration and intra‐subject variations of physiologic and cognitive conditions (Harrison et al., 2008; Waites et al., 2005). Findings of this study could deepen our understanding of the association between energy metabolism and network function along with cognitive decline in AD.

2. MATERIALS AND METHODS

2.1. Participants

A total of 101 right‐handed subjects were included in this study, including 25 patients with MCI [mean age: 69.32 (range: 58–76)], 37 patients with A [mean age: 67.14 (range: 46–83)], and 39 cognitively normal (CN) subjects [mean age: 64.59 (range: 50–83)]. The patients with MCI and AD were recruited from the Memory Clinic of Ruijin Hospital, Shanghai Jiao Tong University School of Medicine. The CN subjects were recruited from local community through advertisements. All participants underwent the global clinical dementia rating (CDR = 0.5 for MCI diagnosis, CDR >0.5 for AD diagnosis) (Morris, 1991), the Mini‐Mental State Examination (MMSE, Chinese Version) (Folstein et al., 1975), and activity of daily living questionnaire. All AD patients met the National Institute of Neurological and Communicative Diseases and Stroke/Alzheimer's Disease and Related Disorders Association criteria (G. McKhann et al., 1984; G. M. McKhann et al., 2011). The participants in the CN group had CDR scores equal to 0. We excluded participants with MRI or PET contraindications (e.g., pregnancy or renal failure), alcohol or drug addiction, major systemic disease, brain parenchyma diseases (e.g., traumatic brain injury), psychiatric disorders, or other neurological disorders that potentially affected cognitive performance. All participants or their designees provided written informed consent to this study. Ethical approval was obtained from the Institutional Review Board of Ruijin Hospital and is in line with the ethical standards of the Helsinki declaration and its later revisions.

2.2. Data acquisition

All imaging data were collected on a 3T integrated Siemens Biograph mMR scanner (Siemens Healthcare, Erlangen, Germany) equipped with a 12‐channel phase‐array head coil. Participants were required to lay supine with eyes closed and remain awake without systematic thinking during scanning. The PET/MRI imaging protocols was the same as our previous studies (Zhang et al., 2022; M. Zhang, Sun, et al., 2021). After fasting for at least 6 h, participants received a bolus injection of the ¹⁸F‐FDG using a mean dose of 217.2 MBq (range: 140.6–314.5 MBq). PET/MRI images were obtained at 40–60 min post‐injection. The PET emission data were acquired in sinogram mode for 15 min. To enhance the performance of the standardized uptake value estimation (Koesters et al., 2016), attenuation correction was conducted based on the Dixon method with an additional model‐based bone compartment. Corrections of random coincidences, dead‐time, scatter, and photon attenuation correction were also performed. The ¹⁸F‐FDG PET images (127 slices, matrix size = 344 × 344, voxel size = 2.1 × 2.1 × 2.0 mm³) were then reconstructed using an ordered subset expectation–maximization algorithm (4 iterations, 21 subsets, and full width at half maximum (FWHM) of a Gaussian filter of 2.0 mm). High‐resolution T1‐weighted MR images were acquired using the 3D MPRAGE sequence (TR/TE = 1900/2.44 ms, FOV = 256 × 256 mm², voxel size = 0.5 × 0.5 × 1.0 mm³, number of slices = 192). Resting‐state fMRI data were obtained using the gradient‐echo echo‐planar imaging (EPI) sequence (TR/TE = 2000/22 ms, voxel size = 3.0 × 3.0 × 3.0 mm³, FOV = 192 × 192 mm², number of slices = 36, number of volumes = 200).

2.3. Resting‐state fMRI data processing

The fMRI images were preprocessed using the Analysis of Functional Neuroimaging (AFNI) software (Cox, 1996) and SPM12 (http://www.fil.ion.ucl.ac.uk/spm/). The first 10 frames of each subject were removed to alleviate the impact of artefacts introduced by magnetic field stabilization and subjects' inadaptation to the scanning environment. The remaining 190 volumes were then despiked using AFNI's 3dDespike program. Slice‐timing correction and realignment to the mean EPI image were performed using SPM12 to correct the interleaved acquisitions and head motion. The T1‐weighted anatomical images were co‐registered to the mean EPI image. Unified segmentation and normalization to the Mayo Clinic Adult Lifespan Template (MCALT) (https://www.nitrc.org/projects/mcalt/) were performed to obtain white matter (WM) and cerebrospinal fluid (CSF) segmentations in the subject space. The MCALT template was constructed from a collection of 202 subjects including 80 clinically unimpaired subjects aged 51–89 and 83 subjects with probable AD aged 51–92 (Schwarz et al., 2017), which makes it suitable for the analysis of data concerning older adults. To remove nuisance signals, WM and CSF masks were created by binarizing the probabilistic segmentations at the 0.9 probability threshold and eroding the remaining masks by two voxels in each direction using the function ‘erode’ in MATLAB (Jones et al., 2016). Voxel‐wise time‐series within the WM and CSF masks were then extracted to be used in a principal component analysis. Nuisance regressor matrices were then constructed by combining six motion parameters and the first six principal components of WM and CSF signals, which were used for nuisance regression. Next, we performed detrending, band‐pass filtering (0.009–0.08 Hz), nuisance regression, time series variance normalization, masking, and smoothing with a 6 mm FWHM Gaussian kernel. To eliminate the potential effect of head motion, we removed high‐motion frames with framewise displacement over 0.5 mm (Power et al., 2012), as well as its one preceding and two subsequent frames. The entire fMRI data would be excluded if more than 40% of frames were censored.

The functional connectivity matrices were constructed for each participant based on the Power et al. (2011) parcellation. Specifically, we extracted the mean fMRI time series of 264 spherical nodes and calculated the Fisher‐z transformed Pearson's correlation coefficients between the time series. The communities for these nodes were defined a priori based on the Infomap algorithm (Rosvall & Bergstrom, 2008).

2.4. ¹⁸F‐FDG PET data processing

The ¹⁸F‐FDG PET data was processed using SPM12. All PET images were co‐registered to each subject's T1‐weighted images and normalized to the Mayo Clinic Adult Lifespan Template. To correct for partial volume effect, the Müller‐Gärtner method was applied to the co‐registered PET images using the PETPVE12 toolbox (Gonzalez‐Escamilla et al., 2017; Muller‐Gartner et al., 1992). To calculate the standard uptake value ratio (SUVR), the PET images were normalized to the mean ¹⁸F‐FDG uptake within a cerebellum mask, which was generated from the automated anatomical labeling (AAL) atlas (Tzourio‐Mazoyer et al., 2002). The mean ¹⁸F‐FDG SUVR within each ROI of the Power et al. (2011) parcellation was then extracted for each subject.

2.5. Sparse canonical correlation analysis

The workflow of feature extraction and sparse canonical correlation analysis (sCCA) are illustrated in Figure 1. For each subject, there are 34,716 connectivity features: upper triangular elements = 264 × (264–1)/2. Before the sCCA analysis, we performed dimensionality reduction for the functional connectivity matrix using the median absolute deviation (MAD), which is a robust statistic and is defined as $\text{median}\left(X_i-\text{median}\left(X\right)\right)$ (Xia et al., 2018), where $X_i$ represent the connectivity feature vector for subject $i$. The top 10% connectivity features were selected and used for the subsequent sCCA analysis.

FIGURE 1.

Graphical illustration of the feature extraction and sCCA analysis. (a) Schematic diagram of the resting‐state fMRI, the ¹⁸F‐FDG PET data and the parcellation defined by Power et al. (2011). (b) Workflow of resting‐state fMRI analysis. The BOLD time series of each ROI defined by Power et al. were extracted and correlation analysis were performed on the mean time series between different ROIs to obtain the FC matrices. The FC matrices were then re‐organized by stacking the upper triangle values from each subject's FC matrix to form a subject × FC edge matrix and z‐scored for the following sCCA analysis. (c) Workflow of ¹⁸F‐FDG PET analysis. After preprocessing, mean ¹⁸F‐FDG uptake of each ROI was extracted for each subject. The mean ¹⁸F‐FDG SUVR values were stacked into a subject × FDG SUVR matrix and z‐scored columnwise for the sCCA analysis. (d) The sCCA analysis finds linear combinations of mean FDG uptake and functional connectivity that maximize the metabolic‐functional relationships, resulting a set of canonical variates and corresponding loadings for all FC and FDG SUVR features. BOLD, blood oxygen‐level dependent; CVs, canonical variates; FC, functional connectivity; FDG, fluorodeoxyglucose; sCCA, sparse canonical correlation analysis; SUVR, standardized uptake value ratio.

We used the penalized multivariate analysis (PMA) in R package for sCCA analysis (Witten et al., 2009), which is a regularized multivariate method that finds the maximum correlation between the linear combination of two modalities. Given the feature matrices of the two modalities, ${\mathit{X}}_1$ and ${\mathit{X}}_2$, sCCA seeks the canonical loading vectors ${\mathit{u}}_1$ and ${\mathit{u}}_2$ to maximize the correlation between canonical variates ${\mathit{X}}_1{\mathit{u}}_1$ and ${\mathit{X}}_2{\mathit{u}}_2$, the optimization function is defined as:

$$\begin{array}{c}\underset{{\mathit{u}}_1\mathbf{,}{\mathit{u}}_2}{\max }{\mathit{u}}_1^T{\mathit{X}}_1^T{\mathit{X}}_2{\mathit{u}}_2\\ \text{subject to}{\left|\left|{\mathit{u}}_1\right|\right|}_2^2\le 1,{\left|\left|{\mathit{u}}_2\right|\right|}_2^2\le 1,{\left|\left|{\mathit{u}}_1\right|\right|}_1\le c_1,{\left|\left|{\mathit{u}}_2\right|\right|}_1\le c_2\end{array}$$ (1)

where and denote the $L^1$ and $L^2-\mathrm{norm}$, respectively. The optimal regularization parameters $c_1$ and $c_2$ were determined based on grid search. More specifically, we conducted sCCA analyses with sparsity parameters range from 0 and 1 (interval: 0.1), where 0 leads to the highest level of sparsity (i.e., the weights of most features may be set to zero) and 1 result in the lowest level of sparsity (i.e., most features may be retained). The regularization parameters that derive the maximum correlation between the first canonical covariates were selected.

After the sCCA analysis, we performed the permutation testing to evaluate the statistical significance of the canonical variates (Xia et al., 2018). More specifically, we fixed the connectivity features and randomly permuted the metabolic features for 1000 times, and then conducted the sCCA analysis after each permutation. The derived correlations were compared with the correlation obtained from the original connectivity and metabolic data. The p‐value was calculated as the number of null correlations over the value of the sCCA correlation of the original dataset divided by the number of permutations (i.e., 1000 times). The FC and FDG scores of each subject were calculated as the linear combination of the FC and FDG features (canonical loading as weighting for each feature), respectively.

We used a bootstrap re‐sampling strategy to further determine features that consistently contribute to the canonical variates (Misic et al., 2016; Xia et al., 2018). Specifically, one‐thirds of the training samples were replaced by numbers randomly sampled from the other two‐thirds of the samples for 1000 times, and the sCCA were then performed after each sampling procedure. Since the re‐sampling strategy may cause changes in the order of canonical variates or the sign for the weights of features, the canonical variates that derived from the re‐sampled data were matched with the canonical variates generated from original data using Procrustes rotations (McIntosh & Lobaugh, 2004). The features were considered significant if their 95% and 99% confidence intervals did not contain zero for FDG and FC, respectively (Xia et al., 2018).

2.6. Statistical analyses

Statistical analyses were performed using SPSS 19 (https://www.ibm.com/products/spss-statistics) and MATLAB 2014a (MathWorks, USA). All continuous variables were tested for normality using Kolmogorov–Smirnov tests. To characterize group differences of FDG SUVR and FC for each network, the within‐network FDG SUVR and FC were averaged for each subject. One‐way analysis of covariance (ANCOVA) with age and sex as covariates were then performed to compare network‐level mean FDG SUVR and FC among CN, MCI and AD groups, followed by post hoc two‐sample t‐tests. Multiple comparisons were corrected using false discovery rate (FDR) correction at p < .05. Chi‐square test was used to compare the categorical measure (i.e., sex) of the three groups. To investigate how the linked patterns between glucose metabolism and functional connectivity are associated with cognitive decline in AD, Spearman's partial correlation analyses between the canonical variates and MMSE scores were performed for all subjects (age, sex and education are as covariates).

3. RESULTS

3.1. Demographics

Thirty‐nine CN participants, 25 MCI and 37 AD patients were studied. The detailed demographics of each group are presented in Table 1. There were no significant differences among AD, MCI and CN groups in age, sex, or education. The MMSE scores of AD patients were significantly lower than MCI and CN subjects (p < .001). The MMSE scores of the MCI patients were also significantly lower than the CN subjects (p = .003). The fMRI data of two MCI and one AD patients were excluded due to excessive head motion.

TABLE 1.

Demographic and clinical data.

	CN(n = 39)	MCI(n = 25)	AD(n = 37)	p‐value
Age, years	64.59(8.38)	69.32(4.43)	67.14(8.84)	.060 a
Sex (male/female)	13/26	6/19	14/23	.519 b
Education, years	12.77(2.91)	10.76(3.83)	11.78(3.02)	.052 a
MMSE	29.49(0.72)	26.84(1.97)	20.03(5.21)	<.001 a

Note: Values are listed as mean (SD).

Abbreviations: AD, Alzheimer's disease; CN, cognitively normal; MCI, mild cognitive impairment; MMSE, Mini‐Mental State Exam.

^a One‐way ANOVA.

^b Chi‐square test.

3.2. Functional connectivity and glucose uptake in AD, MCI and CN groups

For each subject, mean FDG SUVR was extracted for each ROI defined in the Power et al. parcellation. The average FDG SUVR maps for AD, MCI, and CN groups are shown in Figure 2a. In addition, subject‐specific functional connectivity matrices were obtained using the Power et al. parcellation. Subsequently, they were averaged to obtain a group‐mean connectivity matrix for each of the AD, MCI, and CN group, as shown in Figure 2c. Trends of widespread reductions in FDG SUVR and functional connectivity can be observed in the MCI and AD groups. Group comparisons of the network‐level mean FDG SUVR and FC highlighted significantly decreased mean FDG SUVR within the AD group across most networks (Figure 2b), except the SMT and VIS, and decreased mean FC in the SBC, DMN and COP (Figure 2d). Additionally, the MCI group demonstrated notably lower DMN FDG SUVR compared to the CN group.

FIGURE 2.

Group comparison of the FDG SUVR and FC in CN, MCI, and AD groups. (a) Group‐average FDG SUVR of CN, MCI, and AD groups. (b) Group comparisons of the network‐level mean FDG SUVR among CN, MCI, and AD groups. (c) Group‐average FC matrices of CN, MCI, and AD groups. (d) Group comparisons of the network‐level mean FC among CN, MCI, and AD groups. AD, Alzheimer's disease; AUD, auditory network; CN, cognitively normal; COP, cingulo‐opercular network; DAT, dorsal attention network; DMN, default mode network; FC, functional connectivity; FPT, fronto‐parietal network; MCI, mild cognitive impairment; SAL, salience network; SBC, subcortical network; SMT, somatosensory/motor network; SUVR, standard uptake value ratio; VAT, ventral attention network; VIS, visual network. *p < 0.05, **p < 0.01, ***p < 0.001.

3.3. Topographic associations between glucose metabolism and functional connectivity

We used sCCA to identify multivariate relationships between functional connectivity and glucose metabolism in CN, MCI and AD subjects. To reduce the high dimensionality of functional connectivity features, we selected the first 10% of connectivity features with the highest median absolute deviation, resulting 3410 functional connectivity features. The input data thus included 3410 functional connectivity features and 264 glucose metabolic features. To achieve a sparse and interpretable model, grid search was performed to identify optimal regularization parameters, resulting a LASSO regularization of 0.7 for fMRI data and 0.4 for FDG‐PET data. The sCCA analysis revealed 98 pairs of covarying patterns of functional connectivity and glucose uptake, each pair of canonical variates captured specific pattern that linked a weighted set of functional connectivity features to a weighted set of glucose metabolic features. The first three pairs of canonical variates were selected for further analysis since they are statistically significant by permutation test.

We then explored the topographic patterns of the selected canonical variates. The first canonical dimension prominently captures glucose metabolism of subcortical network (SBC), including regions of thalamus and putamen (Figure 3a). For ease of visualization and interpretation, the canonical loadings of connectivity features were averaged for each network (Figure 3c). The corresponding connectivity pattern suggested that this metabolic pattern mostly covaries with the functional connectivity between cingulo‐opercular network (COP) and DMN, as well as between SBC and ventral attention network (VAT) (Figure 3b,c). The second canonical variate predominantly captures the emergence of glucose metabolism of COP, DMN, sensory‐motor network (SMN), dorsal attention network (DAT) and salience network (SAL), encompassing precentral, postcentral, anterior and middle cingulum, medial orbitofrontal cortex and rolandic regions (Figure 3a). These metabolic patterns are correlated with COP‐DMN, COP‐auditory network, and COP‐visual network functional connectivity, as well as the SBC‐VAT and SAL‐DAT functional connectivity (Figure 3b,c). The third pair of canonical variates was characterized by a marked change in glucose uptake in DMN and fronto‐parietal network (FPT) (Figure 3a), including regions of precuneus, angular, middle temporal lobe, middle and superior frontal gyrus, which prominently link to the functional connectivity between DMN and COP, and between SBC and VAT (Figure 3b,c).

FIGURE 3.

Spatial configurations (topographies) of optimally covarying FDG SUVR and FC. sCCA analysis revealed three pairs of statistically significant canonical variates and the associated loadings of FDG SUVR (a) and FC (b). FC with positive weight is shown in warm color and FC with negative weight is shown in cold color. (c) Averaged contribution of within‐ and between‐network functional connectivity for the three canonical dimensions. The color bar shows the magnitude of contributions of each connectivity and metabolic feature. AUD, auditory network; COP, cingulo‐opercular network; DAT, dorsal attention network; DMN, default mode network; FC, functional connectivity; FPT, fronto‐parietal network; SAL, salience network; SBC, subcortical network; SMT, somatosensory/motor network; SUVR, standard uptake value ratio; VAT, ventral attention network; VIS, visual network.

3.4. Changes of FDG and FC scores in patients and correlations with the MMSE scores

Group comparisons were performed for the selected FDG and FC scores among CN, MCI, and AD groups, as shown in Figure 4a. For the first two canonical dimensions, the AD group showed significantly lower FC and FDG scores than the CN group, as well as lower FDG scores than the MCI group. For the third canonical dimension, both the AD and MCI groups showed significantly decreased FC and FDG scores than the CN group. Besides, the FC and FDG scores of the AD group were also significantly lower than the MCI group.

FIGURE 4.

Group comparison of the FDG and FC scores and its relationship with the MMSE scores. (a) Group comparisons of the FDG and FC scores in CN, MCI and AD groups. (b) The FDG and FC scores of the third canonical dimension were positively correlated with the MMSE scores. AD, Alzheimer's disease; CN, cognitively normal; MCI, mild cognitive impairment; MMSE, Mini‐Mental State Exam. *p < .05, **p < .01.

Furthermore, for the third canonical dimension, both the FDG scores (r = .280, p = .008) and FC scores (r = .221, p = .036) were significantly correlated with the MMSE scores in all subjects, as illustrated in Figure 4b. No significant correlations were found between other FDG or FC scores and the MMSE scores.

4. DISCUSSION

In this study, we evaluated the relationship between glucose metabolism and functional connectivity using simultaneously acquired resting‐state ¹⁸F‐FDG‐PET/fMRI and examined their associations with cognitive decline in CN, MCI and AD subjects. The hybrid PET/MRI system allowed us to capture concurrent metabolic and functional changes in relation to AD pathology, providing optimal spatial registration and enhanced patient comfort in a timely‐efficient manner (Chen et al., 2018; Prakken et al., 2023). Furthermore, considering the rapid changes in mental states over minutes and the swift variations in physiological and metabolic conditions over seconds (Heiss, 2016), the combined measurements of FDG uptake and FC were crucial to an accurate evaluation of their interwoven relationship, eliminating the interference caused by intra‐subject variations in physiological and cognitive conditions. In comparison to the conventional region‐ and correlation‐based analyses, we utilized a multivariate approach (i.e., sCCA) to focus on the overall topological pattern. This approach was robust and well‐suited for uncovering the complex and potentially subtle effects that might be concealed within high‐dimensional datasets (Calhoun & Sui, 2016). We observed three prominent patterns of glucose metabolism linked to functional connectivity in the AD continuum. Among them, we found that the metabolic‐functional associations were not limited to colocalized regions, but also occurred over long‐range connections. Furthermore, the covariation of glucose metabolism and functional connectivity were disrupted in MCI and AD patients, which contributed to cognitive decline.

Overall, we identified three multivariate metabolic‐functional patterns that encompassed widespread brain regions including the default‐mode, executive and control, salience, attention, and subcortical networks. One of the interesting observations was that the first pair of canonical variates featured the subcortical regions. Specifically, glucose hypometabolism of the SBC (including thalamus and putamen) mostly covaried with the functional connectivity of the SBC, as well as between the DMN and the COP (Figure 3). In agreement with our findings, both glucose hypometabolism (Choo et al., 2007; Nestor et al., 2003) and altered functional connectivity (Allen et al., 2007; Kenny et al., 2013; Zhou et al., 2013) have been reported in subcortical regions of patients at various stages of AD, and further suggested to be correlated with measurements of cognitive decline and disease severity (Choo et al., 2007; Zhou et al., 2013). As the thalamus and putamen are involved in multiple cognitive processes, such as declarative memory, working memory, and directing attention (Dahlin et al., 2008; Newman, 1995; van der Werf et al., 2000), metabolic and functional disruptions in these regions were considered to be an early clinically significant event during the progression of AD.

As expected, the second and third pairs of canonical variates highlighted the covariation between glucose metabolism and functional connectivity in the COP/SAL, the FPT, and the DMN, which belong to the association system. The association system is responsible for a vast majority of cognitive functions. To name a few, the DMN is involved in the self‐referential process and episodic memory, and was the first large‐scale network found to be disrupted in AD (Greicius et al., 2004; Palmqvist et al., 2017). The COP/SAL and FPT are organized in parallel and involved in various cognitive processes, including attention, salience detection, and cognitive control (Dosenbach et al., 2008; Elton & Gao, 2014; Hausman et al., 2022). In order to support effective communications between functional hubs, the association cortices are highly energy‐demanded, which also makes them vulnerable to neurodegeneration and AD‐related pathologies (Badhwar et al., 2017; Manza et al., 2020; Tomasi et al., 2013). As a result, both metabolic and functional alterations have been extensively reported in the DMN (Badhwar et al., 2017; Buckner et al., 2005; Greicius et al., 2004; Sheline & Raichle, 2013), the FPT (Badhwar et al., 2017; Grothe, Teipel, & Initiative, 2016), and the COP/SAL in patients with AD (Brier et al., 2012; Thomas et al., 2014; Zhang et al., 2022). Some of these disruptions were partially attributed to, and further promoted the deposition of amyloid plaques in cortical hubs (Buckner et al., 2005; Drzezga et al., 2011; Grothe et al., 2016).

In addition to independent observations of glucose metabolism and functional connectivity, their interwoven relationship has long been discussed. From a biological perspective, glucose plays a fundamental role in the generation of adenosine triphosphate (ATP) through oxidative phosphorylation, providing a vital source of energy for the brain (Mergenthaler et al., 2013). A substantial proportion of this energy is dedicated to supporting synaptic transmission, a process closely linked to the spontaneous oscillations that can be measured by fMRI (Logothetis et al., 2001; Niessing et al., 2005; Tomasi et al., 2013). Advances in brain imaging techniques have allowed us to further explore the spatial pattern of metabolic‐functional association. Using hybrid PET/MRI, strong correlations between metabolic and functional metrics have been reported in specific brain regions, such as the SAL and the DMN, in healthy subjects (Aiello et al., 2015; Riedl et al., 2014). In patients with MCI and AD, covariations have also been found in the DMN, as well as the frontal gyri and subcortical regions (Drzezga et al., 2011; Marchitelli et al., 2018). This direct spatial correspondence is in accordance with the hypothesis for the energy efficiency of the connectivity hubs, that local high energy cost is causally linked to high demand of functional communications in the same region (Tomasi et al., 2013). On the other hand, compelling evidences have suggested mismatches between the spatial distributions of metabolic and functional activities. For example, Manza et al. did not find a significant correlation between glucose metabolism and network segregation (an integration of intra‐ and inter‐network functional connectivity) in the association or the sensorimotor networks in healthy subjects (Manza et al., 2020), while Shokri‐Kojori et al. argued for a deviation between glucose utilization and neuroglial activity particularly in the FPT (Shokri‐Kojori et al., 2019). Moreover, although focal correlations existed, the spatial overlaps between glucose hypometabolism and functional disconnection were reported to be 42%–71% in patients with MCI and AD, indicating distinct patterns of metabolic and functional disruptions (Drzezga et al., 2011; Marchitelli et al., 2018). In this study, we observed both one‐to‐one and distant correspondences between glucose metabolism and functional connectivity, which partially explain the discrepancy between the spatially matched and mismatched results. Specifically, we found colocalized metabolic‐functional correlations in the DMN, the SBC and the COP/SAL regions, which were in line with previous results. In addition, we found long‐term correlations between the SBC FDG and the COP‐DMN connectivity, as well as between the DMN FDG and the SBC‐VAT connectivity. In fact, significant correlations between two imaging factors at distant locations are not seldomly seen. Several multimodal studies have reported that glucose metabolism was linked to whole‐brain connectivity (Drzezga et al., 2011), gray matter volume (Strom et al., 2022; Villain et al., 2008), and even amyloid deposition (Pascoal et al., 2019) in distant and non‐overlapping brain regions. Authors of these studies suggested that synaptic activity reflected by FDG uptake may affect ATP‐requiring functions of remote or down‐stream regions and foster structural and pathological changes. Our study has provided in vivo evidence to support their findings and suggest a diversified but still intimate relationship between metabolism and function in the brain.

Finally, knowing how the co‐varying patterns of FDG and FC change across the spectrum of AD and relate to cognitive decline could help understand the pathogenesis and progression of the disease. In this regard, we found a general trend of decreased FDG and FC scores in all three canonical variates in the MCI and AD groups (Figure 4a). Similarly, impaired metabolic‐functional coupling has been observed in widespread regions including the subcortical areas, the DMN, the SAL and the FPT in patients at various stages of AD in both resting (Maleki Balajoo et al., 2022; Marchitelli et al., 2018; Zhang et al., 2022) and active states (Y. Zhang, Du, et al., 2021). Scherr et al. further suggested that local amyloid beta might contribute to the decoupling of metabolic‐functional associations (M. Scherr et al., 2018). And our previous study showed that disrupted metabolic‐functional coupling might play a crucial role in cognitive decline (Zhang et al., 2022). Indeed, we found that the FDG and FC scores of the third canonical dimension (characterized by the DMN and FPT regions) were positively correlated with the MMSE scores (Figure 4b), which were consistent with previous studies (Mosconi et al., 2008; Xiong et al., 2022; Yao et al., 2013; Zhou et al., 2013). Decreases in glucose uptake in the important areas of the brain can't sustain the necessary support of neuronal activity and lead to reduced cognitive function (Lin et al., 2014; Mergenthaler et al., 2013; Neth & Craft, 2017). Taken together, our multivariate analysis suggests that both glucose metabolism and functional connectivity in the co‐varying regions are impaired in patients with MCI and AD, which may contribute to their cognitive impairment and provide a useful multimodal biomarker to the overall cognitive performance.

In conclusion, we deployed simultaneous PET/MRI and multivariate analyses to found a widespread and heterogeneous association between glucose metabolism and functional connectivity in the continuum of AD. Importantly, the spatial pattern of correlation exhibited both one‐to‐one and distributed correspondences, indicating the presence of local energy‐efficiency strategy and interaction between synaptic disruption and functional disconnection in remote regions, respectively. We further showed that this co‐varying pattern of FDG and FC were impaired in patients with MCI and AD, which were related to their cognitive decline.

AUTHOR CONTRIBUTIONS

Yaoyu Zhang, Wenli Li and Yao Li conceived and designed the study. Miao Zhang, Lijun Wang, Guanyu Ye, Hongping Meng, Xiaozhu Lin, Jun Liu and Biao Li recruited subjects and acquired the data. Wenli Li, Ruodong Huang, Jialin Hu and Yaoyu Zhang analyzed the data and interpreted the results. Yaoyu Zhang, Wenli Li, and Yao Li drafted the manuscript. All authors contributed to the article and approved the submitted version.

CONFLICT OF INTEREST STATEMENT

The authors have no relevant financial or non‐financial interests to disclose.

Li, W. , Zhang, M. , Huang, R. , Hu, J. , Wang, L. , Ye, G. , Meng, H. , Lin, X. , Liu, J. , Li, B. , Zhang, Y. , & Li, Y. (2024). Topographic metabolism‐function relationships in Alzheimer's disease: A simultaneous PET/MRI study. Human Brain Mapping, 45(2), e26604. 10.1002/hbm.26604

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.