Mapping of impaired functional connectivity during memory phases in Alzheimer's disease and its association with cortical β-amyloid

Background: Amnesia in Alzheimer's disease (AD) could be due to disrupted encoding, consolidation dysfunction, or an impairment in the retrieval of stored memory information. The different memory phases relate with different parts of functional brain systems. Methods: We combine task functional magnetic resonance imaging and amyloid positron emission tomography in 72 participants (36 AD and 36 controls), to investigate the relationship between memory performance, memory phase-locked functional connectivity, and cortical β-amyloid deposition. Results: We found that AD was mainly characterized by decreased functional connectivity in a new data-driven Network composed of regions from default mode network, limbic network and frontoparietal network during the memory maintenance and retrieval phase. Within the Network, AD had more regions with reduced connectivity during the retrieval phase than other phases, locating mainly in the medial prefrontal cortex, posterior cingulate cortex, middle temporal and inferior parietal cortex of left hemisphere. Furthermore, functional connectivity in the Network related to memory performance. Crucially, the magnitude of the Network connectivity reduction during retrieval negatively correlated with mean cortical β-amyloid, and this relationship mediated the relationship between cortical β-amyloid and memory performance. Conclusions: Our ndings show that memory deciency in AD relates with decreased connectivity in specic network and cortical β-amyloid only during retrieval phase. These ndings help to map impaired functional connectivity during memory phases and explain the relationship between memory deciency and cortical β-amyloid.

cingulate cortex and medial prefrontal cortex. Furthermore, some cortex neurons, which are not normally the engram cells, can have engram function during memory retrieval (11).
Together, these ndings suggest that neural interactions might exert an in uence on memory retrieval, and thus affect memory performance in AD. However, this hypothesis has not been tested in humans. Besides, as the Aβ is regarded as the key marker of the "Alzheimer continuum" (12), it is necessary to clarify the adverse effect of cortical deposition on memory-related neural function along the disease continuum.
To address this caveat in the literature, we characterized detailed alterations in cortical connectivity networks during different memory phases (encoding, maintenance and retrieval) in 36 AD and 36 nondemented participants by task functional magnetic resonance imaging (fMRI). In the same participants, we measured cortical Aβ deposition using positron emission tomography (PET). In addition, we investigated how increasing levels of Aβ pathology in cortex affect cortical connectivity networks. To assess the clinical relevance of the observed functional alterations, we related them to measures of memory performance.
Our primary hypothesis was that AD functional connectivity changes within speci c cortical networks during memory retrieval. The cortical Aβ deposition would correlate with the functional connectivity, and this relationship would mediate the in uence of cortical Aβ on memory performance. The ndings further our understanding of the mechanisms underlying memory de ciency in AD by Aβ pathology on critical memory systems and suggest mapping of functional connectivity as a therapeutic target.

Participants
Participants with clinically diagnosed AD were recruited from the Memory Clinic at Shanghai Jiao Tong University School of Medicine a liated Ruijin Hospital. AD was diagnosed using the syndromal categorical cognitive staging scheme and positive amyloid deposition by PET, based on the National Institute on Aging-Alzheimer's Association (NIA-AA) workgroups (12). Based on cognitive status, neuropsychologic test performance, and Aβ positivity determined by agreement between one nuclear medicine specialist and one memory-disorder specialist, 36 AD participants were included in this study.
The Aβ-normal controls were recruited from one community without dementia. They were assessed by the identical neuropsychological tests as AD. Structural, functional MRI and Aβ PET data were also acquired from these participants.
All participants were initially screened by the Mini-Mental State Examination (MMSE, Chinese Version) (13), global clinical dementia rating (CDR > 0.5 for AD diagnosis), Zung Self-rating Anxiety Scale, Selfrating Depression scale and activity of daily living questionnaire. Demographics included sex, age, education level, occupation, concomitant diseases, and medications. After inclusion, each participant underwent neuropsychological tests that included the Beijing version of the Montreal Cognitive Assessment (MoCA) and the Chinese version of Addenbrooke's Cognitive Examination-Revised (ACER) with subtests of memory, language, attention, uency and visual-spatial processing (14).
The present study was approved by the ethics committee, Shanghai Jiao Tong University a liated Ruijin Hospital, China. All participants in the study or their caregivers signed written informed consent after fully understanding the procedure involved.

Task Design In Fmri
The task was an event-related design, with a total of 34 encoding objects and 56 retrieval objects (Fig. 1).
During the encoding phase, the participants studied 34 unique objects [3 × 3 degrees of visual angle], which were presented for 2 s and had 2-6 s randomized interstimulus intervals. In the following 2-min maintenance phase, the screen was blank, and no stimulus was presented. Then in the retrieval phase, there were 56 test trials presenting objects with the same visual angle. The objects were presented sequentially, followed by 2-6 s interstimulus intervals. The participants were asked to judge whether the objects were "old" (studied) or "new" (unstudied). The participants had up to 4 s to respond. Among these trials, half presented old (studied) objects, and half presented new (unstudied) objects. Task accuracy was de ned as the percentage of the correct answer in 56 recognition objects.
The participants viewed an MRI-compatible liquid crystal display monitor via a mirror mounted to the head coil. All responses were recorded by an MRI-compatible optical mouse. The order of object presentation was randomized across all participants. The duration of the task fMRI slightly varied from 11 to 13 min based on interstimulus interval randomization and reaction times.

Acquisition Of Mr And Pet Images
MRI data were acquired on a whole-body PET/MR scanner (Biograph mMR; Siemens Healthcare, Erlangen, Germany) with a standard 8-channel head coil. Whole-brain functional images were acquired during the tasks, with a 3000-ms repetition time (TR), a 30-ms echo time (TE), a 192 × 192 mm eld of view (FOV), a 90° ip angle, 35 slices, and voxel size of 3.0 × 3.0 × 3.0 mm 3 (each measurement consisted of 260 acquisitions in interleaved mode with a total scan time of 13 min and 8 s).
PET scans using an 18 F orbetapir (AV45) tracer to image Aβ were performed on the same day of fMRI.
The participants received an IV injection of 18 F orbetapir at a mean dose of 3.7 MBq/kg body weight after nishing the task fMRI. Static AV45-PET data were acquired in sinogram mode for 15 min using the following parameters: 128 slices (gap, 0.5 mm) covering the whole brain; FOV, 500 mm; matrix size, 344 × 344; voxel size, 2.6 × 2.6 × 3.1 mm 3 , reconstructed with high-de nition (HD) PET (21 subsets, 4 iterations) and post-ltered with an isotropic full-width half-maximum (FWHM) Gaussian kernel of 2 mm. The T1weighted 3-dimensional structural image was simultaneously acquired with TR = 1900 ms, TE = 2.44 ms and 192 slices covering the whole brain. Attenuation correction for PET was performed using MR-based attenuation maps derived from a dual-echo Dixon-based sequence.

Fmri Activation Analysis: Standard Univariate Analysis
Functional and T1 MRI data were preprocessed by Freesurfer 6.0 and FsFast version 5.0 (surfer.nmr.mgh.harvard.edu). The preprocessing pipeline for fMRI included templating from the middle time point of raw functional data, masking, functional-anatomical registration, motion correction, slicetiming, and resampling the raw time series to the left and right hemispheres. Quality control of functionalanatomical registration was performed by both automatic rating from Fressurfer and visual inspection. Two-dimensional spatial smoothing was performed for surface data with a Gaussian kernel of 5-mm in FWHM.
The encoding and retrieval phases were evaluated separately using the surface-based stream. For the rst-level general linear model (GLM) analysis, contrasts (encoding stimulus against interstimulus rest; recognition stimulus against interstimulus rest) were calculated by the Gamma hemodynamic response function within the cortical surface at each voxel. We inspected each case after the rst-level analysis by visualization to ensure that they registered well with the FreeSurfer average surface (common space). All cases were concatenated for further analysis.
In the second-level analysis, task accuracy-related activation maps were generated in all participants by GLM with accuracy as a regressor. Signi cant clusters were computed after permutation resampling by bootstrapped Monte Carlo simulations (10,000 iterations) at p = 0.001 to correct for multiple comparisons across all brain voxels. We nally had clusters with that size or larger during the simulation and corrected the threshold to p < 0.05 (15). The coe cients of activation were extracted after mapping on the native surface using an inverse a ne transform for further analysis.
This pre-processing included slice timing correction, realignment of functional scans and normalization to MNI space and spatial smoothing (Gaussian kernel of 8 mm FWHM). In the denoising step, linear regression was used to remove the in uence from: (1) Blood oxygen level-dependent (BOLD) signal from the white matter and CSF voxels ( ve components each, derived using the anatomical component-based correction implemented using the ART toolbox), (2) six residual head motion parameters and their rst order temporal derivatives, (3) scrubbing of artifact/outlier scans, and (4) effect of task-condition using event regressors (encoding and retrieval stimulus) convolved with the hemodynamic response function. Finally, the denoising step included temporal bandpass ltering (0.008-0.09 Hz), and linear detrending of the functional time course.
Following pre-processing, we extracted individual phase-locked (encoding, maintenance and retrieval) averaged time-series from 200 cortical nodes de ned by the Schaefer fMRI atlas, which is based on a data-driven fMRI brain parcellation (17). The 200 nodes can be assigned to seven priori functional networks, including visual network, somatomotor network, dorsal attention network (DAN), ventral attention network (VAN), limbic network, frontoparietal network (FPN) and default mode network (DMN). This atlas is well-suited for joint analyses of pathological PET in the cortex and fMRI, since the nodes cover the neocortex and are adaptive from volume to surface. As described before, we performed the connectivity analysis on the time series after removing event-related effects (16).

Data-driven Detection Of Empirical Networks From The Maintenance Phase
The functional connectivity between two brain regions (nodes) was de ned as the Fisher z-transformed Pearson's correlation coe cient between the BOLD time-series in each region. We used the functional connectivity measures in the maintenance phase from the control group to generate an undirected weighted covariance matrix, C, where c ij represents the Fisher z-transformed Pearson's correlation coe cient between node i and j (i.e. edge strength between nodes). After setting negative weights to 0, we used the Louvain algorithm (implemented in the MATLAB Brain Connectivity Toolbox, https://sites.google.com/site/bctnet/) to get the control group-level empirical networks (10) in the maintenance phase (Supplementary material).
We restricted the community detection analysis to the 136 nodes allocated to the limbic network, DMN, DAN, VAN and FPN (16), as de ned by the Schaefer atlas. Given the stochastic nature of the Louvain algorithm, we used a consensus clustering approach to ensure the robustness of the nal community structure by agreement.m and consensus_und.m functions in the Brain Connectivity Toolbox (18)(19)(20) (see Supplementary material for details). Finally, these nodes were clustered into three new empirical networks from controls.

Functional Connectivity Statistical Analysis In Each Memory Phase
The new empirical networks from controls during the maintenance phase were used to group nodes for all participants in each memory phase. The mean functional connectivity within/between the new networks was compared between groups in each phase, respectively. We also measured the relationship between cognition performance and mean functional connectivity within/between each empirical network.
Furthermore, the network component analysis was performed using node-wise connectivity maps within the network that showed group differences in mean functional connectivity. For functional connectivity in each phase, it entered into a group analysis gauging differences in AD and controls using codes similar to the NBS (v1.2) algorithm (21). Different from the traditional NBS, all the links' t-values were calculated in a general linear model after regressing out age and sex (22). Applying a two-sided suprathreshold, the differences in node-node functional connectivity between two groups would be visualized as binary outcomes in each memory phase (Supplementary material).
We also gauged the BOLD series' information content using an entropy measure for discrete time series called sample entropy (23) (setting embedding dimension m = 2 and the range in SD from all time series). The measure should be higher for time series with lower predictability and random disorder, and conversely reduced for more ordered and predictable time series (24).

Pet Image Analysis
For 18 F-orbetapir (AV45) PET, we employed an automatic pipeline to extract cortical standardized uptake value ratios (SUVRs) implemented in PETSurfer FreeSurfer 6.0 (https://surfer.nmr.mgh.harvard.edu/fswiki/PetSurfer). In detail, structural T1 images were used to creates a high-resolution segmentation to run the following partial volume correction (PVC) methods. The PET/anatomical image registration was then performed and visually checked. To minimize partial volume effect from cortical atrophy in AD, we applied the Symmetric Geometric Transfer Matrix as PVC methods for the following ROI analysis (25), using the cerebellum cortex as the reference region. Surfacebased SUVR maps were smoothed on the 2-dimensional surface by a Gaussian kernel of 5 mm in full width at half maximum. Again, we applied the 200-parcellation from Schaefer fMRI atlas that had been adapted from volume to cortex surface and extracted cortical SUVRs.

Statistical analysis
Statistical analyses were performed with R (Version 3.6.2). AD and controls were compared using independent t-test or Chi-square test for continuous or nominal variables, respectively. We used Pearson's correlation coe cient to test for relationships between phase-locked connectivity within/between empirical networks, cortical amyloid and cognitive performance. All signi cant results were doublechecked by the RVAideMemoire package in R, performing a permutation Pearson's product-moment correlation test or Student's t-test with 10,000 permutations.
To control for age, education and sex as covariates, we further used linear regressions with network connectivity and cortical amyloid as predictors of cognitive performance (variables were normalized into z scores). We tested the hypothesis that cortical amyloid in uences memory performance by modulating cortical network connectivity with a mediation analysis using the 'mediation' package, implementing a nonparametric bootstrap method with bias-corrected and accelerated con dence intervals and 10,000 simulation draws. All tests were 2 tailed, and values of corrected p < 0.05 were considered statistically signi cant unless speci ed otherwisea. Supplementary Fig. 1 summarizes our analysis approach.

Participants and task-related activation
A total of 36 AD and 36 cognitively unimpaired controls satis ed the aforementioned inclusion criteria. The two groups were matched in age, education, and sex (Table 1), and AD had worse cognitive performance in all cognitive domains assessed by the MMSE, MoCA, and ACER as expected. Additionally, in the fMRI task, the AD group had lower accuracy than the control group. Furthermore, task accuracy had signi cant correlations with the MoCA (R = 0.610, p < 0.001), the ACER (R = 0.652, p < 0.001), and more strongly with ACER memory scores (R = 0.696, p < 0.001). As expected, in a standard fMRI GLM analysis for activation, there were activations in widespread visual, frontoparietal and attention networks of the left hemisphere (voxel-wise p < 0.001, cluster-wise corrected p < 0.05 after correction by randomized permutation simulator, Table 2). Besides visual network, the most prominent activation was in the left superior frontal cortex (Talairach Montreal Neurological Institute (MNI305) coordinates: -5.8, -7.8, 53.5; p = 0.001, activation β = 5.429, lower right in Fig. 2). The activation was located in the ventral attention network of the Schaefer functional atlas, and mainly correlated with ACER memory (R = 0.435, p < 0.001), followed by ACER (R = 0.348, p < 0.001) and MoCA (R = 0.373, p < 0.001). In the encoding phase, there was no signi cant linear or quadratic relationship between neural activation and task accuracy when this variable was used as a regressor in a second-level voxel-wise analysis.

Empirical Networks During Maintenance From Controls
To investigate memory phased-locked functional connectivity changes within task-relevant cortical networks, we rst used a data-driven community detection algorithm (18,26) to partition the cortex into 3 group-level networks (Network 1, 56 nodes; Network 2, 50 nodes; Network 3, 30 nodes), using the time series during the maintenance phase in the control group (see Fig. 2 for the resulting partition). All three empirical networks were visually symmetrical in the cortex. The 'empirical' Network 1 node assignments overlapped with the a priori network assignments from the Schaefer functional atlas in DMN, limbic network and FPN. The Network 2 mainly overlapped with dorsal and ventral attention networks, as well as the spatial activation maps from the standard fMRI activation analysis. The Network 3 generally included regions assigned to FPN and its adjacent DMN regions. For each participant, we de ned the phase-locked (encoding, maintenance, retrieval) functional connectivity within/between the Network 1-3 by functional connectivity strength (Fisher z-transformed rvalue) in the edges connecting nodes in the three Networks. Thus each participant had three phasespeci c functional connectivity matrices, and their mean matrices for AD and control group were shown in Fig. 3.
Reduced Network 1-Network 1 functional connectivity in AD only during the retrieval phase. The rst and second columns: Mean functional connectivity in each task phase of network edge (node-node connection) in the AD and controls. Each edge (cell of the matrix) represents the functional connectivity strength (Fisher z-transformed r-value) within and between the Network 1-3. The third column: The difference of functional connectivity in each task phase of each network edge (node-node connection) between the AD and controls. Black lines indicate the boundary separating nodes allocated to the empirical Network 1-3.

Network 1 Connectivity Is Reduced In Ad
We rst investigated group differences between the AD and control in the mean functional connectivity within/between the three empirical networks. The AD group had signi cantly reduced connectivity within Network 1 during maintenance (0.205 vs 0.236, p = 0.040) and retrieval (0.159 vs 0.183, p = 0.017). No difference in functional connectivity within Network 1 was observed in the encoding phase. We investigated the speci city of group differences in phase-locked network connectivity in two additional analyses. First, there was no solid group difference in network functional connectivity averaged over each memory phase in the Network 2 (p = 0.081, 0.088, 0.449 for encoding, maintenance and retrieval) and Network 3 (p = 0.073, 0.914, 0.934 for encoding, maintenance and retrieval). Second, there was also no group difference in functional connectivity during each memory phase in the networks' interaction (i.e. Network 1-Network 2, Network 3-Network 2, all p > 0.250).

Relationship Between Network 1 Connectivity And Memory
There were signi cant positive correlations between the Network 1-Network 1 connectivity and ACER memory performance in encoding, maintenance and retrieval phases (R = 0.249, 0.318, 0.304; p = 0.032, 0.004, 0.008; df = 70), indicating that participants who showed the greatest reduction of functional connectivity within the Network 1 also showed the worst memory performance (Fig. 4). We also observed similar positive relationships between memory performance and connectivity within the Network 2 during maintenance and retrieval phase, as well as Network 3 during maintenance. These results were robust to the permutation test at the level of p < 0.05. There was no similar relationship for connectivity between the networks in any phase.
Memory scores in ACER were associated with Network 1-Network 1 functional connectivity in the encoding, maintenance and retrieval phase, while only associated with Network 2-Network 2 functional connectivity in the maintenance and retrieval phase. No other associations were observed. R and P corr represented correlation coe cient and signi cance across two groups. P corr in the correlation analysis was corrected by 10,000 permutations. The grey zone around blue lines represents the 95% con dence Mapping Of Impaired Network Connectivity In Network 1 As we found decreased connectivity in Network 1 of AD group during retrieval, we further explored nodewise functional connectivity comparison between AD and controls during encoding, maintenance and retrieval respectively. During retrieval, 29 in 56 nodes were found having signi cantly reduced links (32 edges), and they located in the medial prefrontal cortex, posterior cingulate cortex, middle temporal and inferior parietal cortex of left hemisphere, as well as inferior temporal and medial prefrontal cortex of right hemisphere. The majority of nodes with reduced connectivity were assigned to DMN (yellow balls in Fig. 5), followed by the limbic network and FPN (blue and green balls in Fig. 5). In the maintenance, only 16 in 56 nodes were found with reduced links (21 edges) in AD. The reduced connectivity had a similar pattern as that in the retrieval phase, predominantly in the DMN of the left hemisphere. During encoding, we observed only 10 nodes with 9 reduced links in AD (Fig. 5). To elucidate the nature of the reduced connectivity in AD, we investigated entropy as a measure of BOLD time series predictability (24). To that end, we computed the entropy-based measure (SampEn) in AD and controls. For completeness, we extract time series in all nodes (encoding, maintenance and retrieval phases respectively) and conducted node-wise comparisons between two groups. The p values were corrected by Benjamini/Hochberg false discovery rate at the level of 0.05 in each phase. The AD showed reduced entropy in 5 nodes (temporal pole, olfactory cortex, medial prefrontal cortex and inferior prefrontal cortex in the left and cingulate cortex in the right hemisphere), which were all in the Network 1 during retrieval phase. Except left temporal pole, four nodes among them had reduced functional connectivity in AD (showed as larger nodes in Fig. 5). The results indicated that decreased functional connectivity appears to be partially characterized by more regular and predictable BOLD time series in AD during information retrieval.

Relationship Between Cortical Aβ And Functional Connectivity
There  (Fig. 7). The mediation effect remains signi cant at p = 0.032 for ACER. As expected, there was no signi cant mediation effect when the mediator variable was connectivity within the Network 2, Network 3 or between Network 1-3.
Greater cortical Aβ deposition has a signi cant but indirect association with worse memory performance, mediated via connectivity during retrieval within the Network1. Mediation effects were computed by 10,000 bootstrapped samples and the 95% con dence intervalw were reported. N1_1 = Network 1-Network 1; R = Retrieval.

Relationship Between Cortical Aβ And Task-related Activation
Our results support the conclusion that global cortical amyloid is related to functional connectivity reduction in Network 1 for AD during memory retrieval, as well as memory performance. For completeness, we conducted an additional analysis to test relationship between Aβ and inter-individual differences in accuracy-related BOLD activation during retrieval derived from standard fMRI GLM analysis. We found strong evidence for BOLD modulation by global cortical amyloid deposition as

Discussion
We found that AD participants were mainly characterized by decreased functional connectivity in the Network 1 community (dominantly composed of DMN, limbic network and FPN) during the maintenance and retrieval phase in the memory task. Within Network 1, AD had more nodes with reduced connectivity during the retrieval phase than other phases, and the nodes located in the medial prefrontal cortex, posterior cingulate cortex, middle temporal and inferior parietal cortex of left hemisphere, as well as inferior temporal and medial prefrontal cortex of right hemisphere. The novel contribution of the study is the demonstration that global cortical Aβ was associated with reduction in Network 1 connectivity only during retrieval. The mediation analysis indicated that cortical Aβ might affect memory performance by modulating retrieval relevant functional connectivity within the Network 1.

Network 1 Connectivity Is Predominantly Impaired In Ad
As memory maintenance is believed to require the coordination of more neocortex regions (27-29), we clustered regions with assignments to priori networks based on functional connectivity during the maintenance phase. Although Network 1 was composed of four priori networks, and its regions were spatially well-organized in the bilateral temporal cortex, medial prefrontal cortex, posterior cingulate cortex and occipitotemporal regions (Fig. 2). The pattern of Network 1 is like the joint of the anteriortemporal and posterior-medial system (30), which work interactively for encoding and retrieval memory (31,32). The reduced functional connectivity within Network 1 is consistent with evidence found in resting-state fMRI, both in AD and cognitively unimpaired amyloid-β + individuals (22,33). Importantly, we found the reduction speci cally in the maintenance and retrieval phase of memory. As in our previous work, retrieval de ciencies could be the cause of poor performance (5). An intervention study provided additional evidence for the importance of the retrieval phase in recognition (34). Taken together, these suggested that Network 1 successfully clustered regions that were working together for memory maintenance and retrieval, and it was impaired in AD.
Functional connectivity within empirical networks is related to memory performance We report signi cant positive correlations between the Network 1 connectivity and ACER memory performance in encoding, maintenance and retrieval phases. Compared to similar positive relationships between memory and the Network 2 and Network 3 in some phases, Network 1 connectivity correlated with memory more steadily and strongly during all phases. As the dominant role of DMN in Network 1, these results are consistent with studies that highlight the importance of DMN activation (35) and cortical network reorganization for better cognitive performance (36,37), especially when large shifts of attention are required during cognitive operations.
Note that the task accuracy-related activation during retrieval was located in the left superior frontal cortex, within VAN/Network 2. In the working memory model, retrieval requires phonological loop and central executive, which is generally de ned as attention allocation e ciency (38,39). In our study, the task accuracy also correlated with memory performance, suggesting the failure to recruit the VAN may also hamper the cognitively demanding selection.

Impaired Node-wise Connectivity During Memory Phases
As we found the hallmark of the AD group had decreased connectivity in Network 1 during retrieval, we further explored node-wise functional connectivity comparison. As expected from the above results, more nodes had different connectivity during the retrieval phase than other phases. Some nodes were clustered in the left posterior cingulate cortex, originally assigned to left DMN. During a memory process, the functional connectivity between DMN sub-networks increases in the retrieval of episodic memories (40). In AD, reduced resting-state functional connectivity was observed in posterior DMN nodes as precuneus/posterior cingulate cortex (41). Besides, we observed another hub in the DMN, medial prefrontal cortex, with reduced connectivity. Similarly, previous research on the schema memories showed the activation of the mPFC associated with successful memory for schema items (42). Importantly, our results appeared to be strongly left-lateralized in the retrieval phase. This ts with earlier studies showing similar left-lateralized effects in early Alzheimer's disease using resting-state fMRI (22), volumetric grey matter measurements (43), FDG-PET (44) or amyloid PET (45).
To elucidate the nature of decreased functional connectivity in AD, we calculated node-wise entropy in Network 1. Nodes with reduced entropy (i.e. reduced BOLD signal predictability) showed clear spatial correspondence with nodes of reduced connectivity relative to CN, only during the retrieval phase. This indicates that some regions in Network 1 with decreased functional connectivity are more predictable and less random with reduced complexity. The results revealed retrieval related connectivity-entropy coupling and showed difference in AD from normal aging (24).

The Relationship Between β-amyloid, Memory And Cortical Activity
We nd a negative correlation between cortical amyloid and ACER memory performance. Our ndings were consistent with previous amyloid studies, which found that higher amyloid deposition correlated with lower immediate memory and delayed recall scores in MCI participants and amyloid-positive healthy controls (46,47).
It was once argued that amyloid pathology and neurodegeneration have adverse, in part synergistic, effects on prospective cognition (48). In addition to the linear relationship found between Aβ and cognition (49), Aβ load was revealed for its nonlinear role in moderating the BOLD activation effect on behavioral task performance (50, 51). Taken together, these ndings suggested the involvement but lack of direct and consistent effect of Aβ on the cognitive de ciency. Several of our results are in accordance with this hypothesis. First, the correlation between cortical amyloid and memory performance/task activation was not regionally speci c. Second, the negative correlation between cortical Aβ and Network 1 connectivity was signi cant only in the retrieval phase. Third, the mediation analyses indicated a signi cant effect of Aβ on memory performance, mediated through a direct effect on Network 1 connectivity during the retrieval phase. Furthermore, we report no signi cant relationships between Aβ and Network 2 or 3 connectivity change, or Network 1 during other phases.
These results also suggested that global cortical amyloid might exert an in uence on memory performance through an effect on functional connectivity during memory retrieval within the Network 1, which mainly included DMN, limbic network and FPN. Participants with higher cortical amyloid deposition exhibited the pronounced Network 1 connectivity decrease, which predicted worse memory performance.

Limitations
Our study has some strengths that extend the current literature, including measurement of cortical amyloid, connectivity, entropy in the same participants, and examination of memory phase-locked functional connectivity. One key limitation is that it is not possible to make strong inferences regarding the direction of causality from purely observational studies. Related to this, there may be (unmeasured) biomarkers (e.g tau) and neuronal variables, which exert a more direct causal in uence on memory performance. These concerns can only be addressed in fully randomized interventional experimental designs. Second, as we did not nd regionally speci c effects from amyloid, our ndings are the result of the global effects of Alzheimer's disease pathology in the cortex. We cannot be sure which regions are more vulnerable to local pathological change.

Conclusion
We employ a multimodal neuroimaging approach to test the relationship between cortical amyloid and memory performance in AD and controls. Taken together, our results suggest that mean cortical Aβ deposition is directly related to Network 1 functional connectivity decrease during memory retrieval, and that greater reductions are associated with worse memory performance. In a mediation analysis, we show that cortical Aβ may impair memory performance through its relationship with Network 1 connectivity during memory retrieval. These ndings help mapping impaired functional connectivity during memory phases and explain memory de ciency due to cortical Aβ in AD patients.  Figure 1 Task design in fMRI. The memory task in the fMRI study, including encoding, blank (memory maintenance), and retrieval phases. In the encoding phase, 34 objects were sequentially presented for 2 seconds respectively. After 2-min blank, the participants indicated, as quickly as possible, whether the object was present or absent in the encoding phase, using two buttons of an MR-compatible button box.

List Of Abbreviations
During retrieval, each object appeared on the screen until the response or up to 4 s before the next letter was shown. The inter-object interval was randomly set between 2 to 4 s.   Relationship of memory scores and Network 1-3 connectivity estimates Memory scores in ACER were associated with Network 1-Network 1 functional connectivity in the encoding, maintenance and retrieval phase, while only associated with Network 2-Network 2 functional connectivity in the maintenance and retrieval phase. No other associations were observed. R and Pcorr represented correlation coe cient and signi cance across two groups. Pcorr in the correlation analysis was corrected by 10,000 permutations.
The grey zone around blue lines represents the 95% con dence interval for predictions from the linear    Mediation analysis between cortical amyloid, task related Network1 connectivity, and memory performance Greater cortical amyloid deposition has a signi cant but indirect association with worse memory performance, mediated via connectivity during retrieval within the Network1. Mediation effects were computed by 10,000 bootstrapped samples and the 95% con dence intervalw were reported. N1_1 = Network 1-Network 1; R = Retrieval.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.