Task-based fMRI brain activation in mild cognitive impairment and Alzheimer’s disease: an ALE meta-analysis

Abstract

Two decades of task-based fMRI studies have revealed atypical task-related activation and hypoactivation patterns in Alzheimer’s disease (AD) and mild cognitive impairment (MCI), implicating the impaired cognitive processing observed in these neurocognitive disorders. The current coordinate-based meta-analysis provides an updated picture of the pathophysiological neurocognitive mechanisms implicated in MCI and AD, which better controls for false positive findings. To pool and summarise these findings, we conducted activation likelihood estimation (ALE) meta-analyses on 90 eligible studies (N_total = 2824) with cluster-level family-wise error correction to compare AD/MCI and healthy controls (HC) on fMRI activity during cognitive tasks across four different domains. ALE assesses whether there is a spatial convergence of activation among experiments. We then conducted meta-analytic functional decoding on the ALE meta-analysis results to infer the functional relevance of the significant clusters. Significant activation clusters, mostly overlapping with the dorsal attention and frontoparietal networks, were observed in MCI and HC when comparing them across memory tasks and all tasks, regardless of cognitive domain. Functional decoding analyses suggest these clusters are linked to spatial and phonological processing. Significant activation converged in the superior temporal gyrus in AD and overlapped with the somatomotor network. Functional decoding indicates it is related to auditory functions. Our findings illustrated the spatial convergence of aberrant task-related activation and hypoactivation in AD and MCI, highlighting the atypical neurocognitive processing across a broad range of tasks.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11357-025-01850-z.

Introduction

Alzheimer’s disease (AD) is the most common type of dementia [1, 2]. Its clinical symptoms include cognitive, neurological, behavioural, and psychological symptoms, with memory loss as the primary feature [3–5]. Its cognitive symptoms include impairments in memory, language, executive functions, visuospatial functions, and global cognition [3, 4, 6]. The AD diagnosis criteria require the presence of biomarkers linked to amyloid-beta, tau, and neurodegeneration [7, 8]. Mild cognitive impairment (MCI) has been conceptualised as the intermediate state between normal cognitive decline associated with ageing and dementia [9–12]. The term MCI encompasses non-demented people with subjective cognitive complaints (self/informant-reported cognitive decline) and objective cognitive decline, whose functional independence is largely preserved [13]. MCI has been further specified into amnestic MCI (aMCI) and non-amnestic MCI (naMCI). aMCI is characterised by memory impairments, and naMCI is characterised by cognitive impairments that do not involve memory [14–16]. People with aMCI tend to progress to AD [14, 15, 17, 18].

The diagnoses of and screening for MCI and AD have improved [19–21]. However, more can be done to understand the pathophysiological changes associated with MCI and AD to improve early detection. Functional magnetic resonance imaging (fMRI) has been proposed and studied as a promising biomarker to identify MCI and AD [22–27]. However, most studies examined resting-state (task-free) fMRI as a biomarker [22–24, 27]. Individuals with aMCI and AD tend to present atypical task-based fMRI brain activation during various cognitive tasks. Hence, task-based fMRI is a potential biomarker for MCI and AD.

Past meta-analyses have identified hyperactivation and hypoactivation during task-based fMRI in MCI and AD [28–32]. Hyperactivation of brain regions may serve to compensate for other impaired brain regions to maintain cognitive performance or capacities [33–37]. However, hyperactivation has also been proposed to be induced by amyloid-beta [38, 39], driving AD pathogenesis by promoting the production and propagation of amyloid-beta and tau [40, 41]. Both mechanisms might also occur: hyperactivation in regions not typically implicated in early AD pathology, like the prefrontal lobes, might reflect compensation, and hyperactivation in regions implicated in early AD pathology, like the precuneus, might reflect disruption and drive AD pathology [42]. Meanwhile, hypoactivation of brain regions is attributed to neurodegeneration and amyloid-beta and tau accumulation [43–45].

Past research suggests that task-based fMRI brain activation changes in an inverted U-shape pattern as AD pathology progresses [42, 43, 46]. Corriveau-Lecavalier et al. [42] proposed a model of hyperactivation in AD spanning 5 phases (Phase 0 to Phase 4). Phase 0 corresponds to normal brain activation during tasks associated with normal ageing. Phase 1 captures the progression between normal and pathological ageing, with regions beginning to show hyperactivation. Phase 2 represents subjective cognitive decline and early MCI, and hyperactivation is expected to be at its maximum. In Phase 3, compensatory mechanisms fail, and pathology levels in regions susceptible to AD pathology plateau. Regions initially showing hyperactivation undergo hypoactivation due to neurodegeneration, which can temporarily lead to activation levels similar to Phase 0. Hyperactivation and hypoactivation can both be present. This phase corresponds to late MCI and mild dementia. Phase 4 corresponds to dementia, and it is linked to hypoactivation.

As mentioned, past meta-analyses have studied task-based fMRI [28–32, 47]. Schwindt and Black [29], Browndyke et al. [30], and Terry et al. [47] focused on episodic memory task-based fMRI, finding that medial temporal lobe (MTL) activity decreased, and the prefrontal cortex and parietal lobe activities increased during episodic memory tasks in people with AD compared to healthy controls (HC). While Schwindt and Black [29] only focused on people with AD, Browndyke et al. [30] and Terry et al. [47] looked at patients with MCI and AD. Browndyke et al. [30] and Terry et al. [47] found MTL hyperactivation and hypoactivation in people with MCI during episodic memory tasks compared to HC. H. Li et al. [28] compared whole-brain activations between HC and people with AD and between HC and people with MCI and conducted further analyses for task-based fMRI categorised under memory encoding, memory retrieval, executive function, and working memory. A portion of the analysis by Gu and Zhang [31] on multimodal MRI studies focused on brain activation associated with task-based fMRI in specific domains of memory encoding, memory retrieval, and working memory in MCI. These meta-analyses suggest that brain activations differ between people with AD and HC and between people with MCI and HC.

While past meta-analyses examined task-based fMRI studies on people with MCI and AD, their findings had several limitations. Firstly, past meta-analyses by Browndyke et al. [30], Schwindt and Black [29], Terry et al. [47], and H. Li et al. [28] used older GingerALE versions (2.3.1 or earlier) to conduct the activation likelihood estimation (ALE) meta-analysis. GingerALE uses the Benjamini–Hochberg procedure to implement the false discovery rate (FDR) correction, whereby p-values are sorted in ascending order before comparing them with a boundary criterion in a step-up manner [48]. However, the versions before 2.3.3 did not sort the p-values completely. This could lead to the threshold becoming more lenient, leading to increased false positives across multiple comparisons [48, 49]. This implementation error was flagged in May 2015, and the corrected version was released in the same month. Hence, these analyses should be rerun to provide a more accurate picture of task-based activation in AD and MCI. Secondly, past studies suggest that a voxel-wise FDR is inappropriate for inference on smoothed statistical maps and has low sensitivity and a high propensity for false positive findings [50, 51]. Eickhoff et al. [51] also concluded that cluster-level family-wise error provides higher power and lower propensity for false positive findings than voxel-wise FDR. While Gu and Zhang [31] utilised a version of GingerALE without the implementation errors, they carried out voxel-wise FDR thresholding. Hence, this present meta-analysis used cluster-level family-wise error to control for false positive results. Lastly, approximately 10 years have passed since the last meta-analysis on task-based fMRI in both MCI and AD. An updated meta-analysis with more recent studies will provide a current picture of the pathophysiological neurocognitive mechanisms implicated in MCI and AD.

H. Li et al. [28] conducted separate meta-analyses for the following three cognitive functions: executive function and working memory, memory encoding, and memory retrieval. Gu and Zhang [31] conducted separate meta-analyses for the following three cognitive functions: memory encoding, memory retrieval, and working memory. These cognitive functions are generally related to the executive function and memory domains. However, other cognitive domains are likely implicated. In MCI, neuropsychological testing comprises four or five cognitive domains, including memory, executive function, attention, language, and visuospatial domains (attention and executive function are sometimes combined) [9, 15, 52]. To reflect this, this meta-analysis grouped fMRI tasks into four cognitive domains: memory, language, visuospatial, and executive function (comprising working memory and attention). While H. Li et al. [28] studied memory encoding and retrieval, memory retrieval can be further split into recognition and recall [53]. Hence, our memory subdomains included memory encoding, recognition, and recall.

This review aims to evaluate the differences in brain activation in task-related fMRI between people with MCI and healthy controls and between people with AD and healthy controls using ALE meta-analyses. Based on the model of hyperactivation, we hypothesise that our coordinate-based meta-analyses would observe the spatial convergence of hyperactivation and hypoactivation clusters across experiments in MCI (corresponding to Phases 2 and 3), as well as that of hypoactivation clusters in AD (corresponding to Phase 4), during task-related fMRI.

Methods

This study followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) [54, 55].

Eligibility criteria

Two separate searches were done for the MCI and AD meta-analyses. The inclusion criteria for the studies include (1) an MCI group and a healthy control group for the MCI meta-analysis or an AD group and a healthy control group for the AD meta-analysis, (2) at least one task-based fMRI condition related to cognitive function that involves a comparison of task activation between MCI group and healthy controls or between AD group and healthy controls, (3) between-group comparison results in either Montreal Neurological Institute (MNI) or Talairach coordinates, (4) only English articles, and (5) group contrasts are based on comparisons between whole-brain activations (studies that only examined activations with regions of interest analyses or analyses with small volume correction were excluded).

Studies are excluded if the studies (1) only conducted group contrasts that utilised small volume corrections, (2) only used neuroimaging techniques other than fMRI, like positron emission tomography or electroencephalogram (which can cause variability due to different techniques), (3) have fMRI task that does not fall under the proposed cognitive domains (memory, language, visuospatial, and executive function, comprising working memory and attention), (4) were examining the effects of treatment or medication interventions in participants without task-based fMRI contrast results at baseline, (5) only included MCI or AD patients with a history of disease that could influence cognitions, (6) did not use a generalised linear model for group comparison, (7) do not provide any coordinates necessary for the meta-analysis, and (8) were not peer-reviewed research articles, like conference papers and book chapters.

Systematic literature search strategy

A systematic literature search was used to identify relevant research related to task-based fMRI. The search was conducted on four databases: Scopus, PubMed, Web of Science Core Collection, and APA PsychINFO. These databases cover the possible fields where relevant studies are likely to appear. Relevant research from the databases’ conception through 27 November 2023 was selected. The keywords associated with fMRI were “functional magnetic resonance imaging”, “fMRI”, “functional MRI”, “activity”, or “activation”. The keywords associated with MCI were “mild cognitive impairment”, “MCI”, or “aMCI”. The keywords associated with AD were “Alzheimer’s disease” or “AD”. Two separate searches were done for the two comparisons between people with MCI and healthy controls and between people with AD and healthy controls. For each database, the keywords associated with fMRI and MCI, as well as the keywords linked to fMRI and AD, were entered together to retrieve the relevant studies for the two comparisons. The keyword search was restricted to the title and abstract. The search strings can be found in Table S1. Studies included in past meta-analyses were included as relevant studies for screening [28–30, 47].

Selection process

All records were imported into Covidence [56]. Covidence then flagged duplicate studies, which were removed. The author (Z. L. L.) screened the titles and abstracts of all the records and removed records unrelated to the research aim. Records that were relevant or ambiguous were retained, and the full-text articles were retrieved. The author (Z. L. L.) then screened the full-text articles based on the eligibility criteria, and any ambiguous articles were referred to another author (J. Y.).

Data collection

Two authors (Z. L. L. and J. J. L.) collected the data independently and manually. The authors collected the year of publication, author names, sample size, subject characteristics (age, education, mini-mental state examination (MMSE) scores if available, and sex), and the criteria used to identify AD or MCI. The contrasts, cognitive domains involved in the fMRI task (memory, executive function, language, and visuospatial), and their subdomains, when the task is linked to the memory domain, were recorded. Lastly, the coordinates of significant clusters that emerged from the group contrasts were recorded. The two authors then resolved any differences related to data collection by referring back to the articles and discussing them with another author (J. Y.). In studies where significant brain coordinates were not reported or more information was needed to assess the study’s eligibility, we sent e-mails to the authors to request unreported data or information. A second e-mail was sent if we received no replies to the initial e-mail within two weeks. E-mails were sent to the authors of 12 studies; however, we only received the requested information from the author of 1 study.

To reduce within-group effects, coordinates from multiple contrasts based on the same subject groups were pooled together under one experiment [57, 58]. Such contrasts could come from different studies. If studies under the same first author had equal or similar sample sizes (differing by 1 or 2 participants) and equal or similar means and standard deviations for the descriptive statistics, they were pooled under one experiment. To be conservative, the sample size for the experiment would be recorded based on the study with the smallest sample size.

Some studies also split the MCI/AD or HC groups into two subgroups and conducted separate group contrasts for these groups with an HC or MCI/AD group, respectively. In cases where such a study contributes two group contrasts for a meta-analysis, the group contrasts will be reported as two separate experiments. To reduce within-group effects, we halved the sample size of the MCI/AD or control group that was split during the meta-analysis [59]. If the sample size of the split group was an odd number, the halved number was rounded down.

Meta-analysis

The meta-analysis was conducted using the ALE method implemented in GingerALE version 3.0.2, which aims to identify brain regions where activations among experiments converge [58, 60–62]. Studies that reported results using Talairach coordinates were transformed to their respective MNI coordinates using the method by Lancaster et al. [63] with GingerALE. A text file containing the Talairach coordinates was entered into the “Convert Foci” function with the “Talairach to MNI (SPM)” transform. The “Single Dataset” option was chosen under the data input because our foci were from individual studies. The more conservative mask and MNI coordinate system were selected. Reported coordinates from each study were compiled into a text file for each contrast and entered into GingerALE. For each study, the ALE method starts with an empty brain, where all voxels have a value of 0. For each reported coordinate, the value of 1 is assigned to that coordinate in the empty brain and then smoothed to neighbouring voxels with a Gaussian kernel [64]. The width of the Gaussian kernel is inversely related to the sample size of the study, with larger kernel sizes for studies with smaller sample sizes [61]. GingerALE carries out this process by default. These steps create a modelled activation (MA) map for each study. These MA maps are combined into an observed ALE map by calculating the union of activation probabilities [58]. Each voxel represents the probability that a significant cluster falls within that voxel. The observed ALE map is thresholded using a cluster-level family-wise error correction of p-value < 0.05 with a cluster-forming threshold of p-value < 0.001 at the voxel level, and 5000 permutations [51]. Each permutation involves randomising the MA values to create a permutated ALE map and thresholding it with the above voxel-level p-value. The largest cluster size from each permutated map is obtained. A distribution of largest cluster sizes is created across 5000 permutations to obtain a cluster-level null distribution for thresholding of the observed ALE map. The weighted mean centre of each cluster was obtained with GingerALE, and the anatomical labels and peak coordinates of the ALE clusters were obtained with the AtlasReader package version 0.3.2 [65] in Python version 3.11.7 using the automated anatomical labelling (AAL2) atlas [66]. The Nilearn package version 0.10.2 was used to generate brain plots of the ALE clusters [67]. Additional brain plots overlaid with the Yeo 7-network parcellation atlas were also created to relate our findings to resting-state research [68].

A few separate meta-analyses were planned. Firstly, ALE meta-analyses were conducted between people with MCI and healthy controls as well as between people with AD and healthy controls across all task-based fMRI studies for four contrasts (MCI > HC, HC > MCI, AD > HC, and HC > AD). Secondly, further analyses were conducted separately between people with MCI and healthy controls and between people with AD and healthy controls across task-based fMRI grouped according to the four cognitive domains of memory, executive function, language, and visuospatial. Thirdly, meta-analyses were also conducted for these contrasts, focusing on the specific functions of the memory cognitive domains (memory encoding, recognition, and recall). According to Eickhoff et al. [51], ALE meta-analysis requires at least 17 experiments to limit the effect of each experiment and have a sufficient power of 0.8 to detect effects present in about a third of the population. Hence, meta-analyses were carried out when there were at least 17 experiments for each contrast.

Fail-safe N analysis

We conducted the fail-safe N (FSN) analysis method proposed by Acar [64] to evaluate the robustness of our findings against the potential file drawer problem, which is the overestimation of the effects in meta-analyses, given that studies with null or unexpected findings might not be published. The FSN for each significant cluster is the largest number of generated experiments with null findings that can be included in the original list of real experiments before the ALE analyses no longer report the cluster as statistically significant. These null experiments were generated using the R code provided by Acar et al. [64]. Each of these null experiments was randomly generated to have a sample size and foci number that match one real experiment in the original list. The included foci were randomly sampled from within the grey matter mask used in the ALE meta-analysis. ALE analyses were conducted until the FSN for each significant cluster was identified. Following Enge et al. [69], the FSN analysis for each significant cluster was repeated five times, and the averaged FSN was reported to improve the reliability of the FSN number.

The lower and upper boundaries of the FSN for each cluster were also set before the FSN analysis. Based on the study that modelled the existence of the file drawer problem using data from BrainMap, Samartsidis et al. [70], Acar et al. [64], and Enge et al. [69] used a lower FSN boundary of 30 studies that reported null findings per 100 published studies to assess the file drawer problems in their ALE meta-analyses. This study adopts this lower boundary. The upper FSN boundary of the FSN was such that at least 5% of the experiments contributing to the significant cluster would be by real experiments. An FSN below the lower bound would suggest that the cluster is not robust against the potential file drawer problem. An FSN above the upper bound would suggest that the cluster is mainly driven by a small number of experiments. More information on the FSN analysis can be found in the Supplementary Information.

Functional decoding

We further conducted meta-analytic functional decoding on the ALE meta-analysis results to infer mental states associated with the significant clusters [71, 72]. We utilised the Neurosynth database (https://github.com/neurosynth/neurosynth-data version 0.7, July 2018) [73], which collates activation coordinates and key terms from task-based fMRI studies with an automated parser. Key terms from the Neurosynth database that were brain region names, unmeaningful descriptive words (e.g., studied, clips, and engaged), and broad categories (e.g., disorder, psychological, and cognitive) were removed to reduce computing time and irrelevant results (this modified database can be found on https://github.com/CogBrainHealthLab/VertexWiseR/blob/main/inst/extdata/neurosynth_dataset.pkl.gz). The meta-analytic functional decoding was then carried out using the ROIAssociationDecoder function from the NiMARE package version 0.1.1 [74] with the masks of the significant clusters obtained from GingerALE and the modified Neurosynth database. The function involves creating an MA map for each study in the Neurosynth database, averaging the MA values in the masks of the significant cluster, and correlating the averaged values with the term weights for the key terms [75]. The function returns the key terms correlated with the significant cluster and their correlation coefficients.

Results

Systematic literature search results

From the systematic literature search, we included 54 articles related to MCI and 36 articles related to AD. The selection process and exclusion of studies for MCI and AD patients can be found in the PRISMA flow charts in Fig. 1. There were five studies without significant clusters when comparing whole-brain activations between MCI and HC [76–80] and seven studies without significant clusters when comparing whole-brain activations between AD and HC [34, 77, 81–85]. These studies were not included since ALE assesses whether there is spatial convergence among published coordinates rather than the presence of an effect, and the inclusion of these studies would not affect the ALE results [86]. Information about the included studies related to MCI and AD can be found in Tables 1 and 2, respectively. Studies that likely used the same subject groups were grouped under the same experiment in Tables 1 and 2. Studies that split the MCI/AD or HC groups into two subgroups and conducted separate group contrasts for these groups with an HC or MCI/AD group can be identified in Tables 1 and 2 as having two separate rows for a single experiment.

Fig. 1.

PRISMA flow chart for studies related to MCI and AD. A MCI PRISMA flow chart and B AD PRISMA flow chart

Table 1.

Demographic information and characteristics of included MCI studies

Experiment	Country of study	N	Age (SD)	Years of education (SD)	MMSE (SD)	MCI composition	MCI diagnostic criteria	fMRI task domain	Specific memory domain	Group contrasts	Number of foci (cluster)
Alichniewicz et al. [87]	Germany	39 MCI 24 HC	62.31 (8.60) 60.67 (7.16)	13.10 (2.98) 13.67 (2.01)	28.59 (1.16) 29.21 (0.88)	aMCI	Artero et al. [88]	Executive Function	-	HC > MCI	5
Balardin et al. [89]	Brazil	18 MCI 17 HC	69.50 (1.91) 68.25 (1.54)	9.20 (1.13) 11.19 (1.35)	27.06 (0.53) 28.33 (0.37)	aMCI	Petersen et al. [90]	Memory	Encoding	MCI > HC	2
Belleville et al. [91]	Canada	15 MCI 15 HC	70.13 (7.34) 70.00 (7.26)	13.73 (4.33) 13.4 (3.10)	27.73 (1.87) 29.1 (0.74)	aMCI	Petersen [92]	Memory	Encoding	MCI > HC HC > MCI	1 3
									Recognition	MCI > HC HC > MCI	1 1
Bokde et al. [93]	Germany	16 MCI 19 HC	69.9 (7.8) 66.7 (4.2)	13.2 (3.3) 12.8 (2.9)	27.2 (1.5) 29.2 (1.0)	aMCI	Petersen et al. [90]	Visuospatial	-	MCI > HC	16
Bokde et al. [94]	Germany	8 MCI 8 HC	70.8 (5.3) 66.6 (3.9)		26.6 (1.3) 30 (0.0)	aMCI	Petersen et al. [95]	Executive Function	-	MCI > HC	16
								Memory	Encoding	MCI > HC HC > MCI	2 2
									Recognition	HC > MCI	6
Bosch et al. [96]	Spain	15 MCI 15 HC	74.63 (6.85) 72.20 (5.75)		25.50 (2.03) 27.67 (1.49)	aMCI-SD	Lopez et al. [97]	Language	-	MCI > HC HC > MCI	3 3
Catricalà et al. [98]	Italy	8 MCI 16 HC	71.62 (5.66) 63.44	12.75 (4.77) 8–18	27.75 (2.05)  >  = 25.44	aMCI	Petersen et al. [90]	Language	-	MCI > HC	9
Chiti et al. [99]	Italy	14 MCI 15 HC	75.8 (66–83) 71.9 (64–83)	7.6 (4.5) 7.6 (3.1)	28.3 (2.1) 29.6 (0.8)	VMCI	Winblad et al. [13] and moderate to severe degrees of leukoaraiosis	Memory	Encoding	MCI > HC HC > MCI	32 16
		21 MCI 15 HC	74.0 (58–87) 71.9 (64–83)	8.4 (4.4) 7.6 (3.1)	27.4 (2.2) 29.6 (0.8)	NVMCI	Winblad et al. [13] and without leukoaraiosis	Memory	Encoding	MCI > HC	61
Clément & Belleville [33, 100] & Clément et al. [101]	Canada	12 MCI 14 HC	68.50 (10.82) 67.21 (6.80)	14.92 (3.92) 14.57 (3.76)	28.92 (1.68) 29.29 (1.14)	Higher-Cognition MCI	Winblad et al. [13] and scored higher than the split-median of the MDRS Score [102]	Executive Function	-	MCI > HC HC > MCI	7 1
								Memory	Encoding	MCI > HC	5
									Recognition	MCI > HC HC > MCI	6 1
		12 MCI 14 HC	68.33 (6.91) 67.21 (6.80)	14.08 (4.23) 14.57 (3.76)	27.00 (1.81) 29.29 (1.14)	Lower-Cognition MCI	Winblad et al. [13] and scored lower than the split-median of the MDRS Score [102]	Executive Function	-	HC > MCI	2
								Memory	Encoding	MCI > HC HC > MCI	1 2
									Recognition	MCI > HC	4
Dannhauser et al. [103] & Dannhauser et al. [104]	UK	10 MCI 10 HC	72.0 (7.7) 68.0 (13.5)	10.3 (1.8) 10.1 (1.4)	24.5 (1.5) 28.3 (1.6)	aMCI	Petersen et al. [90]	Executive Function	-	HC > MCI	1
								Memory	Encoding	HC > MCI	1
Faraco et al. [105]	USA	16 MCI 24 HC	75.13 (6.53) 74.17 (5.50)	14.19 (3.47) 16.96 (2.33)		MCI	Global CDR of 0.05 [106]	Executive Function	-	MCI > HC	47
								Memory	Recognition	MCI > HC HC > MCI	13 34
								Visuospatial	-	HC > MCI	7
Frank et al. [107]	Germany	30 MCI 32 HC	60.80 (6.38) 59.77 (6.74)	9.41 (1.64) 9.59 (1.60)	28.02 (1.73) 29.51 (0.67)	aMCI	Artero et al. [88]	Memory	Recognition	HC > MCI	7
Giovanello et al. [108]	USA	12 MCI 12 HC	75.2 (4.3) 72.6 (5.9)	16.3 (2.9) 15.6 (3.1)	27.8 (1.7) 29.5 (0.9)	aMCI	Petersen et al. [95]	Memory	Recognition	MCI > HC HC > MCI	4 7
Hämäläinen et al. [109]	Finland	14 MCI 21 HC	72.4 (7.3) 71.2 (4.9)	8.1 (2.6) 7.9 (2.9)	25.6 (3.1) 27.7 (2.0)	aMCI-MD	Petersen et al. [110]	Memory	Encoding	MCI > HC HC > MCI	13 1
Hanseeuw et al. [111]	Belgium	16 MCI 16 HC	72.6 (7.9) 69.4 (4.8)	13.5 (2.7) 14.9 (2.4)	27.3 (1.6) 28.7 (1.5)	aMCI	Petersen [15]	Memory	Encoding	HC > MCI	5
Heun et al. [112]	Germany	20 MCI 28 HC	69.7 (7.1) 67.5 (5.4)		26.6 (1.5) 28.9 (1.1)	aMCI	Petersen et al. [90]	Memory	Recognition	MCI > HC	3
Jacobs et al. [113] & Jacobs et al. [114]	Netherland	18 MCI 18 HC	65.1 (4.5) 64.6 (3.4)		27.6 (2.3) 28.9 (1.0)	aMCI	Petersen and Negash [115] and CDR of 0.05 [106]	Visuospatial	-	MCI > HC	28
Jonin et al. [116]	France	17 MCI 19 HC	69.7 (4.1) 68.3 (4.4)	11.1 (3.1) 12.8 (3.3)	25.5 (2.0) 28.5 (1.3)	aMCI	Albert et al. [19]	Memory	Encoding	MCI > HC HC > MCI	4 7
Jin et al. [117]	USA	8 MCI 8 HC	60.9 (3.2) 60.6 (8.3)	16.9 (1.9) 16.9 (2.1)	28.1 (1.1) 29.6 (0.5)	aMCI	Petersen et al. [90]	Memory	Encoding	MCI > HC HC > MCI	1 2
									Recognition	MCI > HC HC > MCI	4 8
Jurick et al. [118]	USA	20 MCI 29 HC	73.4 (7.7) 75.3 (8.8)	16.2 (2.6) 16.5 (2.2)		MCI	Jak et al. [119]	Memory	Encoding	MCI > HC	8
Jurjanz et al. [120]	Germany	12 MCI 12 HC	66.6 (8.7) 62.1 (5.4)	10.8 (1.3) 11.1 (1.4)	28 (1.81) 29.58 (0.52)	aMCI	Petersen [15]	Memory	Recognition	HC > MCI	2
Kaufmann et al. [121]	Austria	6 MCI 9 HC	69.8 (5.3) 68.3 (7.5)		24.8 (1.2) 29.0 (1.2)	aMCI	Petersen [15]	Executive Function	-	MCI > HC	24
Kircher et al. [122]	Germany	21 MCI 29 HC	69.7 (7) 67.8 (5.4)		26.6 (1.4) 28.8 (1.2)	aMCI	Petersen et al. [123]	Memory	Encoding	MCI > HC	4
Kochan et al. [124]	Australia	35 MCI 22 HC	77.97 (3.88) 77.16 (3.31)	12.60 (3.86) 11.44 (3.74)	27.94 (1.56) 29.32 (0.95)	MCI	Winblad et al. [13]	Executive Function	-	HC > MCI	5
Lejko et al. [125]	Netherlands	25 MCI 15 HC	67.6 (5.2) 66.2 (3.8)		28.8 (1.1) 28.9 (1.2)	aMCI	Albert et al. [19]	Executive Function	-	HC > MCI	7
Leyhe et al. [126]	Germany	11 MCI 15 HC	75.0 (6.7) 70.6 (11.8)	13.4 (3.3) 13.7 (3.0)	27.6 (1.36) 29.7 (0.49)	aMCI	Petersen et al. [95]	Visuospatial	-	MCI > HC HC > MCI	10 1
Li X. et al. [127]	China	14 MCI 25 HC	63.95 (6.49) 62.52 (5.41)	10.88 (2.44)	26.50 (2.33) 28.64 (1.44)	aMCI-SD	Winblad et al. [13]	Memory	Encoding	HC > MCI	1
		20 MCI 25 HC	65.00 (8.06) 62.52 (5.41)	11.45 (2.33)	25.29 (1.77) 28.64 (1.44)	aMCI-MD	Winblad et al. [13]	Memory	Encoding	HC > MCI	8
Lou et al. [128]	Hong Kong	17 MCI 19 HC	72.18 (2.9) 71.11 (3.84)	5.94 (5.14) 8.68 (4.03)	26.41 (2.81) 28.74 (1.05)	aMCI	Petersen et al. [123]	Executive Function	-	MCI > HC HC > MCI	1 16
Machulda et al. [129]	USA	19 MCI 29 HC	75.0 (7.0) 73.0 (7.0)	16.0 (3.1) 14 1 (2.4)		aMCI	Petersen [15]	Memory	Encoding	HC > MCI	7
									Recognition	HC > MCI	4
		12 MCI 29 HC	79.0 (6.5) 73.0 (7.0)	13.1 (3.9) 14 1 (2.4)		naMCI	Petersen [15]	Memory	Encoding	HC > MCI	10
									Recognition	HC > MCI	12
Mandzia et al. [130]	Canada	14 MCI 14 HC	68.6 (7.4) 72.2 (6.4)	13.4 (2.8) 15.4 (2.8)	27.7 (1.1) 28.6 (1.1)	aMCI	Petersen et al. [95]	Memory	Encoding	HC > MCI	23
									Recognition	MCI > HC HC > MCI	4 13
Migo et al. [131] & Migo et al. [132]	UK	8 MCI 10 HC	69.6 (5.8) 70.3 (6.5)	17.0 (4.2) 16.0 (4.2)		aMCI	Petersen et al. [90]	Executive Function	-	MCI > HC	3
								Visuospatial	-	MCI > HC HC > MCI	2 4
Nemcova Elfmarkova et al. [133]	Czech Republic	21 MCI 55 HC	69.8 (7.4) 66.7 (7.3)	14.4 (2.6) 15.4 (2.5)	27.0 (1.4) 28.5 (1.2)	AD-MCI	Albert et al. [19]	Visuospatial	-	HC > MCI	2
		24 MCI 55 HC	65.1 (10) 66.7 (7.3)	14.0 (3.1) 15.4 (2.5)	26.8 (2.4) 28.5 (1.2)	PD-MCI	Litvan et al. [134]	Visuospatial	-	HC > MCI	1
Oedekoven et al. [135]	Germany	21 MCI 21 HC	65.6 (8.5) 24.5 (3.7)	14.1 (3.5) 16.4 (2.1)	26.7 (1.5) 29.2 (0.8)	aMCI	Albert et al. [19]	Memory	Recognition	MCI > HC	5
Petrella et al. [136]	USA	20 MCI 20 HC	75.0 (7.6) 71.2 (4.5)	15.0 (2.2) 15.9 (2.9)	26.7 (1.5) 28.4 (1.4)	aMCI	Petersen [137]	Memory	Encoding	HC > MCI	5
									Recognition	MCI > HC HC > MCI	2 6
Poettrich et al. [138]	USA	13 MCI 13 HC	60.5 (6.6) 59.8 (5.3)		28.3 (0.9) 29.1 (0.9)	aMCI	Petersen et al. [95]	Memory	Recall	MCI > HC	3
Risacher et al. [139]	USA	18 MCI 20 HC	72.3 (6.3) 71.4 (4.7)	16.3 (2.9) 17.1 (2.4)	26.6 (2.8) 29.1 (0.9)	aMCI	Albert et al. [19] and CDR	Memory	Encoding	MCI > HC HC > MCI	6 6
Rivas-Fernández et al. [140]	Spain	24 MCI 24 HC	67.96 (9.5) 66.0 (9.1)	10.88 (5.1) 13.96 (6.0)	27.33 (2.46) 28.58 (1.98)	aMCI-SD	Albert et al. [19]	Memory	Recognition	HC > MCI	2
Shu et al. [141]	China	26 MCI 29 HC	62.35 (6.34) 60.48 (6.22)	9.58 (1.60) 10.95 (1.78)	27.38 (1.53) 28.90 (1.01)	aMCI	Albert et al. [19]	Memory	Recognition	HC > MCI	1
Sinanaj et al. [142]	Switzerland	23 MCI 16 HC	75.69 (6.27) 70.12 (1.40)			aMCI	Petersen [15] and CDR of 0.5	Executive Function	-	MCI > HC HC > MCI	11 5
Staffen et al. [143]	Austria	12 MCI 13 HC	71.83 (5.17) 68.38 (7.94)		27.0 (1.76) 28.0 (1.12)	aMCI	Petersen [15]	Executive Function	-	HC > MCI	33
Threlkeld et al. [144]	USA	18 MCI 24 HC	77.1 (5.8) 77.9 (7.1)	15.83 (2.6) 16.21 (2.4)	27.1 (1.3) 28.4 (1.1)	aMCI	Petersen [15]	Language	-	MCI > HC	3
Trivedi et al. [145]	USA	16 MCI 23 HC	77.0 (8.4) 73.1 (5.5)	14.9 (3.3) 16.2 (3.0)	26.3 (2.3) 28.8 (1.2)	aMCI	Petersen and Morris [146]	Memory	Encoding	HC > MCI	6
Van Dam et al. [147]	USA	19 MCI 15 HC	77.6 (7.0) 74.6 (9.2)	14.6 (3.2) 16.9 (2.4)	27.1 (1.8) 28.8 (1.4)	aMCI	Sano et al. [148]	Executive Function	-	MCI > HC HC > MCI	41 20
van der Meulen et al. [149]	Switzerland	13 MCI 15 HC	69.2 (8.2) 68.1 (7.2)	13 (2.3) 14.3 (2.6)	26.7 (2.3) 29.5 (0.8)	aMCI	Petersen et al. [90]	Memory	Encoding	HC > MCI	12
									Recognition	HC > MCI	16
Weigard et al. [150]	USA	18 MCI 16 HC	71.2 (8.5) 72.1 (7.3)	17.1 (2.1) 16.1 (2.7)	26.7 (2.3) 27.8 (1.97)	aMCI	Petersen [15]	Memory	Encoding	HC > MCI	9
Yang et al. [151] & Yang et al. [152]	Australia	13 MCI 20 HC	70.38 (8.61) 64.00 (8.08)	12.31 (3.54) 14.66 (3.36)		PD-MCI	Litvan et al. [134]	Executive function	-	MCI > HC HC > MCI	4 1
Yetkin et al. [153]	USA	9 MCI 8 HC	72.0 (8.0) 65.0 (7.0)	16.0 (3.0) 13.0 (1.0)	28.4 (1.9) 30.0 (0.0)	aMCI	Petersen et al. [95]	Executive function	-	MCI > HC HC > MCI	12 11
Zanchi et al. [154]	Switzerland	21 MCI 22 Stable HC	73.5 (5.8) 68.7 (2.0)		27.0 (1.7) 28.6 (1.7)	aMCI	Petersen et al. [90]	Executive function	-	MCI > HC HC > MCI	3 3
		21 MCI 19 deteriorating HC	73.5 (5.8) 68.8 (4.1)		27.0 (1.7) 28.3 (1.4)	aMCI	Petersen et al. [90]	Executive function	-	MCI > HC HC > MCI	1 4

aMCI amnestic mild cognitive impairment, naMCI non-amnestic mild cognitive impairment, HC healthy controls, aMCI-SD amnestic mild cognitive impairment-single domain, aMCI-MD amnestic mild cognitive impairment-multiple domains, NVMCI non-vascular mild cognitive impairment, VMCI vascular mild cognitive impairment, PD-MCI Parkinson disease mild cognitive impairment, MDRS Mattis Dementia Rating Scale, CDR Clinical Dementia Rating

Studies that likely shared the same samples are grouped under the same experiments. The reported number of participants for these experiments that likely shared samples is based on the sample with the least number of participants

Table 2.

Demographic information and characteristics of included AD studies

Study	Country of study	AD diagnostic criteria	N	Age (SD)	Years of education (SD)	MMSE (SD)	fMRI task domain	Specific memory domain	Group contrasts	Number of foci (cluster)
Bentley et al. [155]	UK	NINCDS-ADRDA [156]	13 AD 18 HC	64.8 (4.4) 64.8 (4.2)	12.9 (1.0) 12.4 (0.9)	23.6 (1.3) 29.4 (0.4)	Memory	Encoding	HC > AD	1
Bokde et al. [157]	Germany	NINCDS-ADRDA [156]	12 AD 14 HC	71.2 (6.9) 67.1 (4.0)		25.3 (2.3) 29.2 (1.2)	Visuospatial	-	AD > HC	18
Bosch et al. [96]	Spain	NINCDS-ADRDA [156]	15 AD 15 HC	75.27 (5.66) 72.20 (5.75)		21.40 (3.06) 27.67 (1.49)	Language	-	AD > HC HC > AD	2 1
Chen et al. [158]	Taiwan	Unknown	15 AD 16 HC	79.92 (4.39) 68.33 (5.47)		13.64 (6.78) 28.53 (1.45)	Language	-	HC > AD	10
Donix et al. [159]	Germany	NINCDS-ADRDA [156]	12 AD 12 HC	69.6 (6.1) 62.1 (5.4)	14.5 (3.2) 15 (2.2)	24.5 (2.5) 29.6 (0.5)	Memory	Recognition	HC > AD	5
Gao et al. [160]	Hong Kong	NINCDS-ADRDA [156]	17 AD 13 HC	76.7 (2.4) 76.2 (4.1)	5 (3.8) 4.4 (2.6)	21.2 (2.9) 29.2 (0.6)	Executive Function	-	AD > HC	8
Golby et al. [161]	USA	NINCDS-ADRDA [156]	7 AD 7 HC	69 (8) 66 (11)		20.8 (2.0) 29.4 (0.5)	Memory	Encoding	HC > AD	7
Gould et al. [162]	UK	NINCDS-ADRDA [156]	12 AD 12 HC	77.3 (4.9) 77.3 (4.8)	11.3 (3.2) 11.4 (3.4)	36.33 (2.06) 29.08 (0.90)	Memory	Encoding	AD > HC HC > AD	2 3
								Recognition	AD > HC	1
Grön et al. [163]	Germany	NINCDS-ADRDA [156]	12 AD 12 HC	61.7 (5) 59.8 (2.6)		25.9 (3.5) 30 (0)	Memory	Recall	AD > HC HC > AD	6 17
Grossman et al. [164] & Grossman et al. [165]	USA	NINCDS-ADRDA [156]	11 AD 16 HC	73.0 (4.9) 73.9 (3.6)	15.3 (2.9) 13.8 (1.8)	20.2 (6.1) 29.7 (0.8)	Language	-	AD > HC HC > AD	7 14
Hämäläinen et al. [109]	Finland	NINCDS-ADRDA [156]	15 AD 21 HC	73.1 (6.7) 71.2 (4.9)	8.2 (2.7) 7.9 (2.9)	21.7 (2.7) 27.7 (2)	Memory	Encoding	AD > HC HC > AD	2 5
Hohenfeld et al. [166]	Germany	NIA-AA Research Framework [7]	9 AD 12 HC	64.67 (8.26) 65.25 (6.33)			Visuospatial	-	AD > HC HC > AD	7 3
Kato et al. [167]	USA	NINCDS-ADRDA [156]	7 AD 8 HC	73.6 (2.9) 65.1 (1.8)			Memory	Encoding	HC > AD	4
Kircher et al. [168]	Germany	DSM-IV Criteria [169]	10 AD 10 HC	71.8 (12) 67.2 (5.1)		22.3 (3.9) 29.3 (0.6)	Memory	Encoding	HC > AD	1
Leyhe et al. [126]	Germany	NINCDS-ADRDA [156]	11 AD 15 HC				Memory	Recognition	AD > HC HC > AD	11 12
Lim et al. [170]	South Korea	NINCDS-ADRDA [156]	12 AD 12 HC	69.5 (5.6) 68.6 (6.2)	10.8 (4.3) 11.3 (3.1)	20.3 (1.4) 29.1 (1.2)	Executive Function	-	AD > HC HC > AD	2 4
McGeown et al. [171] & McGeown et al. [172]	UK	NINCDS-ADRDA [156]	12 AD 9 HC	76.83 (8.50) 75.11 (1.62)	10.17 (2.33) 11.67 (2.29)	22.75 (2.30)	Executive Function	-	HC > AD	5
							Language	-	HC > AD	1
Meulenbroek et al. [173]	Netherlands	Dubois et al., 2007 [174]	21 AD 22 HC	72.4 (7.1) 69.6 (8.6)	16.1 (3.9) 16.5 (3.2)	25.3 (3.2) 29 (1.1)	Memory	Recall	AD > HC	4
Nelissen et al. [175]	Belgium	NINCDS-ADRDA [156]	15 AD 37 HC	73.2(6.8) 67 (8.6)	11.9 (2.4)	24.6(2.7)	Language	-	AD > HC	2
Olichney et al. [176]	USA	NINCDS-ADRDA [156]	15 AD 15 HC	72.9 (8.6) 68.7 (12.1)	14.7 (2.3) 15.5 (2.4)	24.4	Memory	Encoding	AD > HC HC > AD	7 20
Pariente et al. [177]	UK	NINCDS-ADRDA [156]	12 AD 17 HC	70.9 (6.4) 70.6 (5.6)	12.9 (2.3) 13.2 (3.8)	25.1 (1.8) 29 (1)	Memory	Encoding	AD > HC HC > AD	10 2
								Recognition	AD > HC HC > AD	3 5
Pihlajamäki et al. [178] & Pihlajamäki et al. [179]	USA	NINCDS-ADRDA [156]	15 AD 29 HC	78.3 (6.9) 74.2 (5.6)	13.3 (3.2) 15.6 (2.6)	23.3 (4.2) 29.7 (0.5)	Memory	Encoding	AD > HC HC > AD	23 1
Rémy et al. [180] & Rémy et al. [181]	Canada	NINCDS-ADRDA [156]	8 AD 11 HC	72.2 (10.8) 65.9 (5.7)	13.1 (2.8) 13.3 (2.6)	21.2 (6.4) 29.4 (0.5)	Memory	Encoding	AD > HC HC > AD	3 10
								Recognition	AD > HC HC > AD	4 8
							Executive Function	-	AD > HC HC > AD	1 8
Saykin et al. [182]	USA	NINCDS-ADRDA [156]	9 AD 6 HC	79 (5) 71 (4)	17 (2) 16 (2)		Language	-	AD > HC HC > AD	10 11
Shanks et al. [183]	UK	NINCDS-ADRDA [156]	9 AD 9 HC	74.89 (10.07) 75.11 (1.62)	12.22 (3.70) 11.67 (2.29)	23.11 (3.22) 28.89 (0.93)	Executive function	-	AD > HC HC > AD	3 3
Sperling et al. [184]	USA	NINCDS-ADRDA [156]	7 AD 10 HC	80.6 (6.9) 74.1 (7.3)		22.6 (2.2)	Memory	Encoding	AD > HC HC > AD	16 15
Thiyagesh et al. [185] & Thiyagesh et al. [186]	UK	NINCDS-ADRDA [156]	10 AD 11 HC	76 (6.5) 70.2 (4.4)	9.9 (1.4) 11.5 (2)	24.1 (3.5) 28.8 (0.75)	Visuospatial	-	AD > HC HC > AD	6 12
Vannini et al. [187]	Sweden	DSM-IV Criteria [169]	13 AD 13 HC	68.9 (6.9) 68.7 (7.8)	12.5 (3.6) 13.2 (3.9)	25.5 (2.33)	Visuospatial	-	AD > HC HC > AD	1 22
Venneri et al. [188]	UK	NINCDS-ADRDA [156]	9 AD responders 9 HC	79 (7.11) 75.11 (1.62)	10.11 (2.09) 11.67 (2.29)	21.89 (2.67)	Language	-	HC > AD	3
							Executive Function	-	HC > AD	5
		NINCDS-ADRDA [156]	17 AD non-responders 9 HC	75.35 (9.51) 75.11 (1.62)	11.88 (3.5) 11.67 (2.29)	23.65 (2.85)	Executive Function	-	HC > AD	3
Wierenga et al. [189]	USA	NINCDS-ADRDA [156]	10 AD 22 HC	77.7 (9.71) 78.09 (6.47)	15.4 (2.37) 16.05 (2.28)	24.3 (3.56)	Language	-	AD > HC HC > AD	2 1
Yetkin et al. [153]	USA	NINCDS-ADRDA [156]	9 AD 8 HC	68 (10) 65 (7)	14 (3) 13 (1)	23 (3.37) 30 (0.0)	Executive Function	-	AD > HC HC > AD	19 7

AD Alzheimer’s disease, HC healthy controls, AD non-responders AD patients who did not respond to cholinesterase inhibitor drugs, AD responders AD patients who responded to cholinesterase inhibitor drugs

Studies that likely shared the same samples are grouped under the same experiments. The reported number of participants for these experiments that likely shared samples is based on the sample with the least number of participants

ALE meta-analytic results

Based on the cutoff of 17 experiments, seven meta-analyses were conducted: comparison between MCI and HC across all cognitive domains (HC > MCI and MCI > HC), comparison between MCI and HC within the memory domain (HC > MCI and MCI > HC), comparison between MCI and HC within the memory encoding subdomain (HC > MCI), and comparison between AD and HC across all cognitive domains (HC > AD and AD > HC).

For the ALE meta-analyses across all cognitive domains, no significant cluster was found to converge for the HC > AD contrast, and a significant cluster converged for the remaining contrasts. The cluster for the HC > MCI contrast included the left precentral gyrus, opercular part of the inferior frontal gyrus, and middle frontal gyrus (predominantly frontoparietal network (FPN) and includes the dorsal attention network (DAN)); the cluster for the MCI > HC contrast included the right superior parietal gyrus, precuneus, and superior occipital gyrus (predominantly DAN and includes the visual network); the cluster for the AD > HC contrast included the right superior temporal gyrus, Rolandic operculum, supramarginal gyrus, and the postcentral gyrus (predominantly somatomotor network (SMN) and includes the ventral attention network). For the meta-analysis of differences in brain activations between MCI and HC within the memory domain, the significant cluster found for the MCI > HC contrast was similar to that found in the MCI > HC contrast across all cognitive domains. No significant cluster was found for the HC > MCI contrast. Lastly, no significant cluster converged for the HC > MCI contrast within the memory encoding subdomain. The ALE results can be found in Table 3. The brain plots for the significant clusters and the plots overlaid with Yeo’s 7 resting-state networks atlas can be found in Figs. 2 and 3, respectively. More information regarding the breakdown of the contributing experiments can be found in Table S2 in the Supplementary Information. The FSN of each significant cluster averaged over five runs, and the lower and upper FSN boundaries were also reported in Table 3. More information regarding the FSN for each run can be found in Table S3 in the Supplementary Information. The significant clusters in the MCI > HC contrast across all cognitive domains and in the memory domain had higher averaged FSNs than the lower FSN boundaries of their significant clusters, and the significant clusters in the HC > MCI and AD > HC contrasts across all cognitive domains had averaged FSNs lower than the lower FSN boundaries of their significant clusters. Table 4 shows the top five associated keywords for each of the significant contrasts based on functional decoding.

Table 3.

ALE results for the four contrasts

Cluster	Volume (mm3)	Weighted center (MNI)			Anatomical labela	Peak ALE value	Peak			Peak anatomical labela	Contributing studies (k)	Averaged FSNb
		X	Y	Z			X	Y	Z
HC > MCI
All Domains (42 experiments, total sample size of 1408, 378 foci)
1	832	− 38.5	6.9	31.1	61.54% Precentral_L 29.81% Frontal_Inf_Oper_L 7.69% Frontal_Mid_2_L	0.0259	− 38	6	32	Precentral_L	5	< 12.6 (12.6–58)
Memory Domain (26 experiments, total sample size of 829, 246 foci)
-	No sig. cluster
Memory Encoding Domain (19 experiments, total sample size of 557, 126 foci)
-	No sig. cluster
MCI > HC
All Domains (35 experiments, total sample size of 1087, 430 foci)
1	896	19	− 68.5	49.4	62.73% Parietal_Sup_R 29.09% Precuneus_R 7.27% Occipital_Sup_R	0.0235	18	− 68	48	Parietal_Sup_R	4	12.6 (10.5–45)
Memory Domain (19 experiments, total sample size of 600, 189 foci)
1	856	20.3	− 69	50.2	74.77% Parietal_Sup_R 17.76% Precuneus_R 7.48% Occipital_Sup_R	0.0208	20	− 68	52	Parietal_Sup_R	3	5.8 (5.7–41)
HC > AD
All Domains (28 experiments, total sample size of 668, 229 foci)
-	No sig. cluster
AD > HC
All Domains (23 experiments, total sample size of 630, 180 foci)
1	800	63.1	− 21.4	11.3	73.00% Temporal_Sup_R 16.00% Rolandic_Oper_R 6.00% SupraMarginal_R 5.00% Postcentral_R	0.0171	64	−20	14	Temporal_Sup_R	4	< 6.9 (6.9–57)

AD Alzheimer’s disease, MCI mild cognitive impairment, HC healthy controls, L left, R right, precentral precentral gyrus, Frontal_Inf_Oper inferior frontal gyrus (opercular part), Frontal_Mid middle frontal gyrus, Parietal_Sup superior parietal gyrus, Occipital_Sup Superior occipital gyrus, Temporal_Sup superior temporal gyrus, Rolandic_Oper Rolandic operculum, SupraMarginal Supramarginal gyrus, Postcentral postcentral gyrus, FSN fail-safe N

^aThe automated anatomical labelling (AAL2) atlas was used to identify the anatomic labels of the ALE clusters [66]

^bThe averaged FSN is reported, and numbers in the brackets represent the lower and upper FSN boundaries

Fig. 2.

Brain plots of the significant clusters

Fig. 3.

Brain plots of the significant clusters with Yeo’s 7 resting-state networks. Contours of the significant clusters are in black

Table 4.

Functional decoding of each contrast with significant ALE meta-analysis results

Contrasts	Features	r
HC > MCI All Cognitive Domains	Language	0.092
	Phonological	0.090
	Working Memory	0.082
	Word	0.082
	Words	0.077
MCI > HC All Cognitive Domains	Spatial	0.123
	Visual	0.076
	Attention	0.075
	Target	0.075
	Frontoparietal Network	0.072
MCI > HC Memory Domain	Spatial	0.122
	Visual	0.078
	Attention	0.076
	Eye Fields	0.074
	Target	0.073
AD > HC All Cognitive Domains	Auditory	0.216
	Speech	0.125
	Sounds	0.117
	Pitch	0.113
	Sound	0.111

r is the correlation between the activation clusters from the significant ALE meta-analysis results of each contrast and the feature from the database

Supplementary ALE meta-analytic results

Eight supplementary ALE meta-analyses were conducted for contrasts that have 10 or more experiments but do not meet the cutoff of 17 experiments [51]. The supplementary ALE results are reported in Table S4, and the accompanying ALE plots can be found in Fig. S1. Additionally, the ALE plots overlaid with Yeo’s 7 resting-state networks atlas can be found in Fig. S2, and the functional decoding of the significant clusters can be found in Table S5. Briefly, no significant clusters converged for fMRI experiments in the HC > MCI executive function domain and memory recognition subdomain contrasts and in the HC > AD memory encoding subdomain contrast. Significant clusters converged in three MCI > HC contrasts that span experiments in the executive function domain, memory encoding subdomain, and memory recognition subdomain. Additionally, we observed significant clusters in the memory domain of both HC > AD and AD > HC contrasts.

Discussion

This meta-analysis found convergence among the results from different experiments examining task-based fMRI brain activation differences between MCI and HC and between AD and HC. The meta-analysis found hyperactivation and hypoactivation clusters in MCI compared to HC. In the HC > MCI contrast, a significant cluster converged across all cognitive domains, primarily containing the left precentral gyrus and the opercular part of the left inferior frontal gyrus. In the MCI > HC contrast across all cognitive domains and the memory domain, there was a significant cluster at the right superior parietal gyrus and precuneus. This provides some support for the model of hyperactivation, given that the model proposes that hypoactivation and hyperactivation can both be present [42], although MCI phases are not distinguished here.

This meta-analysis found a hyperactivation cluster in AD. In the AD > HC contrast across all cognitive domains, a significant cluster converged at the superior temporal gyrus and the Rolandic operculum. No clusters were identified in the HC > AD contrasts across all cognitive domains. These findings countered the proposal by the model of hyperactivation that people in Phase 4 (dementia) experience hypoactivation during tasks [42]. There are some possible explanations. First, this meta-analysis included several studies that primarily examined participants with mild or early dementia. These participants might be better represented by Phase 3, where hyperactivation and hypoactivation might be present, rather than by Phase 4. Second, brain regions later affected by AD pathology might still show hyperactivation associated with AD pathology during dementia. The transition from hyperactivation to hypoactivation in brain regions might be longer than proposed by the model of activation. Regions later affected by AD pathology might still show hyperactivation associated with AD pathology during dementia, and hypoactivation might only occur in later stages of dementia. Hence, Phase 3 might correspond to mild and moderate dementia instead, and Phase 4 reflects severe dementia.

Interpretation of activation clusters

The significant cluster for the HC > MCI contrast across all cognitive domains included the left precentral gyrus and the left inferior frontal gyrus. Past meta-analyses did not report significant clusters that converged at the left precentral gyrus in this contrast [28, 30, 47]. The precentral gyrus contains the primary motor cortex, and it is critical for voluntary motor movements [190]. The left precentral gyrus is also activated when engaging in working memory tasks [191, 192]. Grey matter volume atrophy at the left precentral gyrus is associated with poorer working memory in controls, AD patients, and patients with behavioural variant frontotemporal dementia [193]. This significant cluster also included the left inferior frontal gyrus, replicating findings from past meta-analyses. H. Li et al. [28] and Browndyke et al. [30] reported significant clusters in this area across all task-based fMRI and episodic memory encoding, respectively. The left inferior frontal gyrus has multiple functions; it is activated during memory encoding, retrieval, and verbal working memory tasks [191, 194], and lesions at the left inferior frontal gyrus, including the opercular part, were found to lead to worse inhibitory control in the Go/NoGo task [195]. The opercular part of the left inferior frontal gyrus is also involved in language processing, being more activated when processing complex words compared to simple ones [196]. Functional decoding results also support that this cluster is related to language processing and working memory. This cluster lies predominantly in the FPN, which is involved in executive functions, like goal-directed behaviours, inhibition, and working memory [197]. Past research reported lower regional homogeneity in the left inferior frontal gyrus, decreased amplitude of low-frequency fluctuation in the left precentral gyrus, and lower resting-state functional connectivity between these two regions in aMCI compared to HC [198]. Additionally, another study reported decreased grey matter volume in the left precentral gyrus and altered functional connectivity between the left inferior frontal gyrus and brain regions with grey matter atrophy in aMCI compared to HC [199]. Consolidating this information, lower activation in this region in MCI might be linked to poorer working memory functions in MCI [200]. One plausible explanation for finding this cluster based on fMRI tasks across all cognitive domains might be because working memory is important to other cognitive domains and activities of daily living [201]. Greater activation in this region might differentiate healthy controls from MCI.

The significant clusters in the MCI > HC contrast across all cognitive domains and within the memory domain are similar. They primarily contain the right superior parietal gyrus and precuneus. The activation of the right superior parietal gyrus is associated with visuospatial attention and higher working memory load, especially in spatial working memory tasks [202]. A lesion study found that superior parietal gyrus lesions were associated with poorer performance in manipulating information in working memory [203]. This highlights the role of the right superior parietal gyrus in working memory. Meanwhile, the precuneus is involved in higher-order cognitive functions, like visuospatial imagery, episodic memory retrieval, and self-processing [204]. This is corroborated by the functional decoding results, which found that this cluster is related to spatial processing and attention and that the cluster predominantly lies in the DAN, which is involved in visuospatial attention [197]. MCI shows attentional and spatial processing deficits compared to HC [205, 206], and this cluster might be linked to these deficits.

Resting-state functional connectivity between the DAN (including the superior parietal gyrus) and the default mode network was anticorrelated in young adults, but this anticorrelation decreased in older adults [207]. Such anticorrelations were found to be lower in elderly individuals with MCI compared to healthy elderly individuals, though the included seed region did not contain the superior parietal gyrus [208]. While the functional decoding results did not include memory in the top five features (which we might expect since this cluster is related to fMRI memory tasks), the results contained attention and spatial processing, which are also important for memory [209, 210]. Hyperactivation in this cluster may serve as successful or failing compensatory mechanisms to maintain memory-related functions.

The significant cluster for AD > HC contrast across all cognitive domains is located at the right superior temporal gyrus and the Rolandic operculum. In line with previous research, our functional decoding results revealed that this region is involved in language processing [211, 212]. However, magnetic stimulation-induced virtual lesions at this region did not significantly decrease task performance during speech perception tasks, though the right superior temporal gyrus was activated during sub-lexical speech perception fMRI tasks [213]. This cluster also overlaps with the SMN, which is involved in somatosensory processing [197]. Resting-state and task-based functional connectivity identified that the temporal lobe can be divided into their anterior and posterior regions [214]. The superior temporal gyrus was part of the posterior region, which is linked to auditory functions. People with AD experience deficits in the language domain. These deficits include impairments in semantic and lexical aspects of language as well as changes in spontaneous speech, like the amount of hesitation and articulation rates [215]. Brain activation differences in AD during tasks are likely linked to such deficits. Aside from auditory functions, the right superior temporal gyrus is also linked to memory. The activation of the right superior temporal gyrus is linked to short-term auditory memory [216] and some types of visual search tasks [217]. The right superior temporal gyrus also appeared as a significant cluster in the AD > HC contrast for episodic memory in two meta-analyses [29, 30]. The higher activation of the right superior temporal gyrus across different task-based fMRI paradigms in AD, compared to HC, possibly suggests a compensatory mechanism to augment language processing and verbal short-term memory.

The FSN analysis evaluated the robustness of the significant clusters against the file drawer problem, which is the overestimation of the effects in meta-analysis due to null studies not being published. We adopted a lower FSN bound of 30 null studies per 100 published studies for our clusters—a conservative estimate based on past studies [69, 70]. Our results suggest that the significant clusters for the MCI > HC contrast across all cognitive domains and in the memory domain are generally robust against the file drawer problem. However, the significant clusters in the HC > MCI and AD > HC contrast across all cognitive domains might not be robust against the file drawer problem. Hence, caution is needed when interpreting the latter two clusters.

Surprisingly, regions linked to AD progression based on past studies [218–220], like the medial temporal lobe, were not significant in this meta-analysis on task-based fMRI. Additionally, MCI is the intermediate state between cognitive decline associated with normal aging and dementia [9, 12], and the disease progression and cognitive decline are much worse in AD than in the HC group [221–223]. Hence, we would expect more and larger activation clusters when comparing AD and HC. There are some potential explanations for these puzzling results. First, it is likely that clusters were not identified because there is too much heterogeneity among the included AD studies, in terms of the within-cognitive domain tasks. Consequently, the patterns of task-related activation and hypoactivation in AD can be highly varied even within a single cognitive domain. Thus, the amount of spatial convergence in the activated clusters across these studies is fairly minimal. Second, some studies suggested that functional brain networks become more disorganised as AD progresses [224–226], possibly explaining the heterogeneous task-related activation patterns. Lastly, it might be because regions showing hyperactivation are undergoing hypoactivation, which leads to brain activation during AD being temporarily similar to those in HC, as proposed by the model of hyperactivation [42].

Differences from past meta-analyses

The results from this meta-analysis also differed from past meta-analyses. The current meta-analyses found much fewer significant clusters compared to past meta-analyses in this area [28–31, 47]. The issues related to the GingerALE implementation and the use of voxel-wise FDR mentioned in the introduction are likely to contribute to some differences in the results. Additionally, the current meta-analysis adopted different inclusion criteria compared to H. Li et al. [28] and Gu and Zhang [31]. For instance, this present meta-analysis only included task-based fMRI paradigms that fell within the domains of memory, language, visuospatial, and executive function. Hence, studies that use other task-based fMRI paradigms involving simple motor tasks [227], pain processing [228], and self-awareness [80] were excluded. Furthermore, we excluded studies that applied small volume corrections and those that did not use the general linear model approach to model task-related activation across groups. Gu and Zhang [31] published their article in 2019, and 7 MCI studies included in this meta-analysis were published after 2019. Lastly, we were careful to avoid double counting any studied samples that were likely to be analysed multiple times across different studies. For instance, studies under the same first author with similar sample sizes (differing by ± 2 participants) and similar descriptive statistics are pooled under one experiment. These procedures were not specified in other meta-analyses. Hence, such differences might explain the result differences. A summary of the differences among the meta-analyses can be found in Table S6 in the Supplementary Information.

Limitations and future research

This study has several limitations. A single reviewer was in charge of abstract screening and full-text review. This can lead to relevant research articles being excluded [229, 230], contributing to fewer included studies. Moreover, caution should be taken when interpreting the results for the MCI > HC contrast across all cognitive domains and within the memory domain. This is because the two highest contributing experiments in both contrasts to the significant clusters identified by the ALE meta-analysis were from the same study by Chiti et al. [99]. The study examined two different MCI groups, contrasting each MCI group with the same HC group to obtain brain activation differences; hence, these two contrasts were recorded as two experiments. While we partially accounted for the unit of error analysis by halving the sample size of the HC group, the resulting comparisons between the two experiments remain correlated. Hence, the results might not be robust, especially for the latter contrast within the memory domain, since two of the three experiments contributing to the cluster were from the same study. Future studies are needed to confirm these findings.

There is also heterogeneity among studies. For instance, in the memory domain and encoding subdomain, there were heterogeneous tasks. For example, studies included under the encoding subdomain analysed different forms of encoding, like the encoding of related and unrelated words, successful/unsuccessful encoding, and contrasting the encoding of new words against old words. Such heterogeneity among the tasks and contrasts can influence the results. While the results need to be interpreted with caution, such a meta-analysis does help identify consistent findings across these heterogeneous studies.

Compared to past ALE meta-analyses [28, 47], the number of included AD studies did not increase much. This might be because recent studies examined fMRI task activation with regions of interest analyses instead. However, ALE meta-analyses recommend including only studies that utilised whole-brain comparison methods [57], and recent studies using ROI analyses were not included. Future studies can consider sharing their uncorrected statistical images and imaging data. This can allow future meta-analyses to use image-based meta-analysis and mega-analysis instead, which preserves more information from the original analyses than only using the peak coordinates in coordinate-based meta-analysis [57, 231].

Lastly, most of the included studies in this meta-analysis are from countries that are Western, educated, industrialised, rich, and democratic (WEIRD) [232]; hence, there are significant concerns about the generalisability of these results to non-WEIRD populations. Future studies can attempt to replicate such analyses in other non-WEIRD countries.

Conclusions

The present meta-analyses revealed the spatial convergence of abnormal task-related activations and hypoactivations during tasks in MCI and AD. For MCI, domain-general task-related hypoactivation converged in the left precentral and inferior frontal gyri, and hyperactivation was observed in the right superior parietal gyrus and precuneus. For AD, domain-general task-related hyperactivation converged in the right superior temporal gyrus. During memory tasks, we observed significantly greater task-related activity in the right superior parietal gyrus and precuneus among MCI subjects relative to HC subjects.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contribution

Zheng Long Lee: conceptualisation, methodology, software, investigation, formal analysis, writing—original draft, writing—review and editing, visualisation. Li Jianyi: investigation, writing—review and editing. Savannah Kiah Hui Siew: methodology, writing—review and editing. Charly Billaud: software, writing—review and editing. Junhong Yu: methodology, software, investigation, formal analysis, writing—review and editing, supervision, funding acquisition.

Funding

This work was supported by the Nanyang Assistant Professorship (Award no. 021080–00001). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data availability

The data, code, and results can be found at the OSF link: https://osf.io/6wqz4/?view_only=b435684380f448308201def06d323e52.

Declarations

Ethical approval

This is a systematic review and meta-analysis of published data. No ethical approval is required.

Registration and protocol

The review was not registered, and a protocol was not prepared.

Competing interests

The authors declare no competing interests.