We searched PubMed, Web of Science, Embase, and Cochrane Library databases from database inception to June 1, 2020. The search strategy was [(“Mild cognitive impairment” OR “Alzheimer's Disease”) AND (“magnetic resonance spectroscopy” OR “MRS” OR “MR Spectroscopy”)]. The search was limited to English language studies only. Regardless of the primary outcome or the type of study, we have considered all possible eligible studies for review.

Studies meeting the following criteria were included: (1) the proton MRS was performed to compare MCI patients, AD patients, and healthy controls. (2) NINCSD-ADRDA criteria were chosen as the diagnostic standard for AD, and criteria used for diagnosis of MCI were clearly reported. (3) At least one single metabolites ratio or concentration in a specific brain region was reported. (4) Specifications for spectrum acquisition were reported.

The exclusion criteria were as follows: (1) studies were published in languages other than English. (2) Original data could not be extracted, or the full text could not be obtained. (3) Duplicate or similar data published research. (4) The subjects were animals. (5) The subjects were taking drugs, had other significant medical conditions or substance abuse that could interfere with cognitive functioning.

After applying the inclusion and exclusion criteria, we finally identified 63 articles and extracted the following characteristics for meta-analysis: the interested brain regions and the corresponding metabolites ratios and concentrations, the field strength, repetition time/echo time (TR/TE), and other characteristics which are shown in Table 1. Meanwhile, we also extracted standard deviations (SD) or standard error of mean (SEM) or median, as our main results.

Stata 16.0 (Stata Corp) software was used to perform all statistical analysis. The sample size, mean value, and SD were used to generate the effective sizes, and when the mean ± SEM or median was provided, we converted it into mean ± SD for meta-analysis. Then, we calculated the standardized mean difference (SMD) and 95% CI and drew a forest map to compare the relationship between the metabolites' concentrations or metabolites' ratios between the healthy control group and AD patients, the healthy control group and MCI patients, and AD patients and MCI patients. We used the Q-test and I² index to evaluate heterogeneity. The statistical significance of the Q-test was set as P < 0.1, and heterogeneity was assessed by I² index, with 25, 50, and 75%, indicating that the heterogeneity was low, medium, and high (Higgins et al., 2003). For the statistical model, we first chose fixed effects model with the method of inverse-variance for analysis. If the heterogeneity of the results is greater, we used the random effect model with the method of Dersimonian and Laird for analysis. All statistical significances were set at P < 0.05 and P < 0.1 was regarded as a trend.

We identified 4,436 articles from database searching and 2,812 remained after removal of duplicates. Then, 133 full-text articles were assessed for eligibility. Finally, according to the inclusion and exclusion criteria, 63 studies were included in the systematic review and meta-analysis (MacKay et al., 1996; Ernst et al., 1997; Parnetti et al., 1997; Lazeyras et al., 1998; Schuff et al., 1998, 2002; Jessen et al., 2000, 2005, 2009; Kantarci et al., 2000, 2002, 2007; Catani et al., 2001, 2002; Block et al., 2002; Chantal et al., 2002; Huang et al., 2002; Herminghaus et al., 2003; Frederick et al., 2004; Ackl et al., 2005; Chao et al., 2005, 2010; Fernández et al., 2005; Metastasio et al., 2006; Zhu et al., 2006, 2015; Franczak et al., 2007; Rami et al., 2007; Azevedo et al., 2008; Ding et al., 2008; García et al., 2008; Pilatus et al., 2009; Siger et al., 2009; Wang et al., 2009, 2012; Zhang et al., 2009; Griffith et al., 2010; Li et al., 2010; Liu et al., 2010, 2013; Watanabe et al., 2010; de Souza et al., 2011; Fayed et al., 2011, 2014; Foy et al., 2011; Modrego et al., 2011; Zimny et al., 2011; Lim et al., 2012; Seo et al., 2012; Shiino et al., 2012; Yang et al., 2012; Targosz-Gajniak et al., 2013; Graff-Radford et al., 2014; Bai et al., 2015; Delli et al., 2015; Mandal et al., 2015; Riese et al., 2015; Guo et al., 2016; Zeydan et al., 2017; Marjańska et al., 2019; Mitolo et al., 2019; Oeltzschner et al., 2019; Figure 1).

The meta-analysis comprised of a total of 3,271 subjects, with 1,086 MCI patients, 1,256 AD patients, and 1,907 healthy controls. The following regions were investigated: hippocampus (including MTL), PC, temporal lobe, occipital regions, paratrigonal white matter area, temporo-parietal lobe, parietal lobe, frontal lobe (gray and white matter area), and anterior cingulate. The key characteristics of the studies are shown in Table 1 among which 6 studies were classified as high quality and 57 studies were classified as medium quality (Table 2).

Nineteen studies (Jessen et al., 2000, 2005, 2009; Block et al., 2002; Schuff et al., 2002; Ackl et al., 2005; Chao et al., 2005; Franczak et al., 2007; Wang et al., 2009, 2012; Zhang et al., 2009; Watanabe et al., 2010; Foy et al., 2011; Seo et al., 2012; Shiino et al., 2012; Liu et al., 2013; Targosz-Gajniak et al., 2013; Zhu et al., 2015; Huang et al., 2017) investigated the ratios of metabolites in hippocampus from 358 MCI patients, 890 AD patients, and 787 healthy control subjects. Specifically, 12 studies (Ackl et al., 2005; Franczak et al., 2007; Wang et al., 2009, 2012; Zhang et al., 2009; Watanabe et al., 2010; Foy et al., 2011; Seo et al., 2012; Liu et al., 2013; Targosz-Gajniak et al., 2013; Zhu et al., 2015; Huang et al., 2017) performed a comparison of the changes between 358 MCI patients and 425 healthy control subjects, 14 studies (Jessen et al., 2000, 2005, 2009; Block et al., 2002; Schuff et al., 2002; Ackl et al., 2005; Chao et al., 2005; Wang et al., 2009, 2012; Zhang et al., 2009; Watanabe et al., 2010; Foy et al., 2011; Shiino et al., 2012; Huang et al., 2017) compared the differences in metabolites between 890 AD patients and 679 healthy control subjects, and 5 studies (Ackl et al., 2005; Wang et al., 2009, 2012; Zhang et al., 2009; Huang et al., 2017) were conducted to observe the differences of metabolites in 155 AD patients and 130 MCI patients. Moreover, there were another two articles (Modrego et al., 2011; Seo et al., 2012) longitudinally tracking the metabolite differences in the hippocampus between MCI-converter and MCI-stable patients.

A total of 29 studies (Kantarci et al., 2000, 2002, 2007; Chao et al., 2005; Rami et al., 2007; García et al., 2008; Olson et al., 2008; Wang et al., 2009, 2012; Griffith et al., 2010; Watanabe et al., 2010; de Souza et al., 2011; Fayed et al., 2011, 2014; Zimny et al., 2011; Lim et al., 2012; Seo et al., 2012; Shiino et al., 2012; Yang et al., 2012; Targosz-Gajniak et al., 2013; Graff-Radford et al., 2014; Liu et al., 2014; Riese et al., 2015; Guo et al., 2016; Zeydan et al., 2017; Marjańska et al., 2019; Mitolo et al., 2019; Oeltzschner et al., 2019) investigated the ratio of metabolites in posterior cingulate with a sample size of 770 MCI patients, 585 AD patients, and 1,378 healthy controls. To be specific, 25 studies (Kantarci et al., 2000, 2002, 2007; Chao et al., 2005; Rami et al., 2007; García et al., 2008; Olson et al., 2008; Wang et al., 2009, 2012; Griffith et al., 2010; Watanabe et al., 2010; de Souza et al., 2011; Fayed et al., 2011, 2014; Zimny et al., 2011; Lim et al., 2012; Seo et al., 2012; Yang et al., 2012; Targosz-Gajniak et al., 2013; Liu et al., 2014; Riese et al., 2015; Guo et al., 2016; Zeydan et al., 2017; Mitolo et al., 2019; Oeltzschner et al., 2019) compared the differences in metabolites between 770 MCI patients and 1,132 healthy control subjects, 16 studies (Kantarci et al., 2007; Rami et al., 2007; Ding et al., 2008; Wang et al., 2009, 2012; Watanabe et al., 2010; de Souza et al., 2011; Fayed et al., 2011, 2014; Zimny et al., 2011; Lim et al., 2012; Shiino et al., 2012; Graff-Radford et al., 2014; Guo et al., 2016; Marjańska et al., 2019; Mitolo et al., 2019) compared the differences in metabolites between 610 AD patients and 822 healthy control subjects, and 12 studies (Kantarci et al., 2007; Rami et al., 2007; Wang et al., 2009, 2012; Watanabe et al., 2010; de Souza et al., 2011; Fayed et al., 2011, 2014; Zimny et al., 2011; Lim et al., 2012; Guo et al., 2016; Mitolo et al., 2019) made a comparison between 440 AD patients and 421 MCI patients. Moreover, there were another two studies (Kantarci et al., 2007; Seo et al., 2012) longitudinally tracking the metabolite differences between MCI-converter and MCI-stable patients.

There were 7 studies (Kantarci et al., 2000; Block et al., 2002; Herminghaus et al., 2003; Frederick et al., 2004; Rami et al., 2007; Azevedo et al., 2008; Li et al., 2010) investigating the ratio of metabolites in the temporal lobe and included 82 MCI patients, 157 AD patients, and 207 healthy controls. Of these studies, 3 (Kantarci et al., 2000; Rami et al., 2007; Li et al., 2010) compared the differences in metabolites between 82 MCI patients and 124 healthy control subjects, and 5 (Block et al., 2002; Herminghaus et al., 2003; Frederick et al., 2004; Rami et al., 2007; Azevedo et al., 2008) compared the differences between 157 AD patients and 110 healthy control subjects.

Eight studies (Herminghaus et al., 2003; Ackl et al., 2005; Chao et al., 2005; Zhu et al., 2006; Siger et al., 2009; Li et al., 2010; Modrego et al., 2011; Targosz-Gajniak et al., 2013) with a total sample size of 639 (162 AD patients, 187 MCI patients, and 290 healthy controls) were included for meta-analysis to investigate the ratio of metabolites in the parietal lobe. Specifically, 3 studies (Herminghaus et al., 2003; Ackl et al., 2005; Zhu et al., 2006) compared the differences in metabolites between 80 AD patients and 71 healthy control subjects in parietal WM, and 5 studies (Herminghaus et al., 2003; Ackl et al., 2005; Chao et al., 2005; Zhu et al., 2006; Siger et al., 2009) compared the differences in metabolites between 162 AD patients and 151 healthy control subjects in parietal GM. Moreover, there were 2 studies (Modrego et al., 2005, 2011) longitudinally tracking the metabolite differences between MCI-converter and MCI-stable patients.

There were 5 studies (Block et al., 2002; Azevedo et al., 2008; Watanabe et al., 2010; Graff-Radford et al., 2014; Marjańska et al., 2019) with a total sample size of 481 (195 AD patients and 286 healthy controls) included for meta-analysis to investigate the ratio of metabolites in the occipital lobe. Moreover, there were 3 more studies (Modrego et al., 2005, 2011; Seo et al., 2012) longitudinally tracking the metabolite differences between MCI-converter and MCI-stable patients.

Three studies (Lim et al., 2012; Guo et al., 2016; Huang et al., 2017) investigated the anterior cingulate including 66 MCI patients, 83 patients with AD, and 70 healthy control subjects. Specifically, 3 studies (Lim et al., 2012; Guo et al., 2016; Huang et al., 2017) performed a comparison in the changes between 66 MCI patients and 70 healthy control subjects, 3 studies (Lim et al., 2012; Guo et al., 2016; Huang et al., 2017) performed a comparison in the changes between 83 AD patients and 70 healthy control subjects, and 3 studies (Lim et al., 2012; Guo et al., 2016; Huang et al., 2017) were conducted to observe the differences of metabolites in 83 AD patients and 66 MCI patients.

Four studies (Ernst et al., 1997; Fernández et al., 2005; Rami et al., 2007; Zhang et al., 2009) investigated the temporo-parietal lobe including 80 AD patients and 71 healthy control subjects to compare the metabolites between the two groups.

Four studies (Parnetti et al., 1997; Chao et al., 2005; Zhu et al., 2006; Siger et al., 2009) with a total sample size of 218 (109 AD patients and 109 healthy controls) were included for meta-analysis to investigate the ratio of metabolites in the frontal region. Specifically, 3 studies (Parnetti et al., 1997; Zhu et al., 2006; Siger et al., 2009) compared the differences in metabolites between 61 AD patients and 61 healthy control subjects in the frontal WM, and 3 studies (Chao et al., 2005; Zhu et al., 2006; Siger et al., 2009) compared the differences in metabolites between 96 AD patients and 102 healthy control subjects in the frontal GM.

Three studies (Catani et al., 2001; Metastasio et al., 2006; Yang et al., 2012) reported data from paratrigonal white matter including 89 MCI patients and 177 healthy control subjects to compare the metabolites between the two groups.

NAA is an important indicator of neuronal metabolism and plays a number of roles, which includes maintaining fluid balance in the brain, providing acetic acid salt for the synthesis of myelin in oligodendrocytes, and providing energy for the glutamylation of amino acid in neuronal mitochondria. Also, NAA is considered as a biomarker of neuronal function and density, as it can reflect the density and functional state of neurons and axons. Moreover, the concentration of NAA in the cortex can provide information about the growth of neurons. In addition, its concentration in white matter can reflect the development of axons. Due to NAA being located at the synaptic terminal, dendrites, and neuronal somata, its concentration may reflect the level of synapses and the ratio of NAA/Cr reflects the integrity of synapses (Onwordi et al., 2021). Our meta-analysis showed that during the pathological development from MCI to AD, the ratio of NAA/Cr in the hippocampus and PC, and the concentration of NAA in PC decreased dramatically. Meanwhile, the level of NAA/Cr in AD patients was decreased than that in the healthy subjects in the other brain lobes, such as the parietal lobe, the frontal lobe, the temporal lobe, and the temporo-parietal lobe. Previous studies showed that the hippocampus and PC were crucial brain regions that are differentially affected by neuropathological changes in AD patients (Silverman et al., 1997). Hippocampal atrophy is one of the pathological and radiological signs of AD, and the resting state functional magnetic resonance imaging and ¹⁸F-deoxyglucose PET demonstrated that hippocampal dysfunction is closely related to the cognitive impairment in AD patients (Yan et al., 2020). The PC plays a crucial role in the default mode network of the brain, and previous studies demonstrated that the functional connections between the PC and the hippocampus have also been weakened during cognitive impairment (Huang et al., 2002). Studies have reported that the atrophy of the hippocampus was closely related to the loss of neurons and the number of neurons had a close connection with NAA (Shiino et al., 2012). The results of this meta-analysis suggested that with the pathological development of AD, the neurons in the hippocampus and the PC were gradually damaged. And the formation of Aβ and NFT in the brain of AD patients may lead to the decrease of NAA and the gradual loss of synapses, which is consistent with the progress of cognitive dysfunction in AD. Similarly, there are studies showing that the sensitivity and specificity of NAA/Cr of PC in predicting the conversion of MCI to AD were 82% and 72%, respectively (Fayed et al., 2008). Consequently, NAA may be considered as a biomarker for monitoring the progression of MCI to AD.

mI is a good indicator of the proliferation of glial cells, as it is mainly expressed in glial cells. Previous studies have reported the increased levels of mI and mI/Cr in the PC were found in MCI and AD patients (Yang et al., 2012), which was consistent with our research results. In this meta-analysis, the results showed that in the pathological development from MCI to AD, in addition to the decrease in the ratios of NAA/Cr and the concentration of NAA in PC, there was also an increase in the ratios of mI/Cr and mI/NAA; and, compared with the HC, the mI concentration in the PC of AD and MCI patients was significantly increased. Meanwhile, the level of mI in MCI and AD patients was significantly higher than that in the healthy subjects in the other brain regions, such as the temporo-parietal lobe and PWM. The increase in mI concentration may be caused by the activation of astrocytes or microglia, which is related to the neuroinflammation process, and has been considered as one of the core pathological features of AD. In addition, the increased mI may affect the phosphorylation of membrane proteins or cause changes in phospholipid metabolism, affecting the formation of Aβ, and thus leading to the deposition of amyloid plaques. In addition, the increased deposition of Aβ also induced the formation of nutritionally impaired synapses, and the astrocytes wrapped and phagocytosed the diseased synapses to remove the aberrations in the synapses. However, with the development of the disease, the deposition intensified, and this pathological change promoted the increase of inflammatory response, which would disrupt the normal form of synapses (Gomez-Arboledas et al., 2018). Studies have found that the change in the ratio of mI/Cr in PC was closely related to the early decline of cerebrospinal fluid Aβ42, and the decrease in the level of CSF Aβ42 can be detected 10–20 years before the onset of cognitive impairment (Bateman et al., 2012). Studies have found in the brain of Down's syndrome and other dementia patients, the ratio of mI/Cr was also significantly increased before significant manifestations of cognitive dysfunction (Voevodskaya et al., 2016). Therefore, these results suggested that the change in the level of mI may precede the onset of cognitive impairment, which had the potential to be applied to early diagnosis of AD.

Moreover, it is worth noting that this study showed that the concentration of Cr was significantly lower in the hippocampus of AD and MCI patients than that of healthy people, but there was no significant change in other brain regions such as PC. Interestingly, it is generally believed that the concentration of Cr is basically constant and uniformly distributed throughout the brain and is not changed with age or various diseases. Therefore, the level of Cr is often used as a reference value to indicate the level of other neurochemical substances. Some studies have also found that the concentration of Cr was relatively reduced in the late stage of AD as well as subcortical ischemic vascular dementia, which may be due to the fact that Cr existed in neurons and glial cells at the same time, and was affected by the density of brain tissue (Shiino et al., 2012).

Studies have regarded the ratio of mI/NAA a standard method to determine the severity of AD, as the sensitivity and specificity of the ratio of mI/NAA in the diagnosis for AD patients were 83 and 98%, respectively, and was consistent with the MMSE score (Shiino et al., 2012). This meta-analysis found that the ratio of mI/NAA increased during the progression from MCI to AD in PC, and the same result was also observed in the hippocampus. But since only 3 studies were included and a large heterogeneity was observed, this result should be interpreted cautiously. In addition, there was no meta-analysis results in the procession from MCI to AD in the hippocampus. Therefore, the change of mI/NAA is consistent in the progression from MCI to AD, but whether it can be used as markers in early diagnosis of AD is still questionable.

The findings with respect to changes in the levels of Cho and Glx in MCI and AD were less consistent. As discussed above, the ratio of Cho/Cr was significantly higher in the PC of AD patients than that of HC, but there was a downward trend in MCI subjects. In addition, the concentration of Cho was found to be raised in the PC of AD patients compared with healthy controls, and no differences were seen in the MCI subjects. In contrast, the concentration of Cho was found to be reduced in the hippocampus of AD patients and MCI patients compared with healthy controls. It had been reported that a cholinergic lesion emerged as early as the MCI state and primarily in the presynaptic membrane, which may affect the long-term potentiation (Nordberg and Winblad, 1986). Cho plays an important role in the formation of cell membranes, and the change of concentration directly reflects the synthesis and degradation of membranes. Cho could be converted into acetylcholine (AchE) by choline acetyltransferase (ChAT) to play a neuroregulatory role, and the Cho signal may be closely related to the activity of ChAT (Klein, 2000). A clinical autopsy study found increased ChAT activity in the hippocampus of patients with MCI, which could explain that the decrease of Cho in the MCI stage is due to the increased activity of ChAT and the utilization of more choline substrates, thus resisting the damaging effect of cholinergic neurons (Ikonomovic et al., 2003). This compensatory activity may increase with the progression of the disease. Meanwhile, neuronal death will lead to an increase in membrane turnover, which will increase the ratio of Cho/Cr in AD patients. Recent investigations reported that the increase in the level of Cho in PC of AD patients may be the result of cell membrane rupture providing free Cho, which was in response to a decrease in the release of acetylcholine from cholinergic neurons in the brain of AD patients (Watanabe et al., 2010). In frontotemporal dementia and dementia with Lewy bodies, the increase in the ratio of Cho/Cr in the PC can also be detected. Interestingly, there was no significant change in the early stage of AD disease. In MCI patients, the concentration of Cho in PC remained basically the same with healthy controls, while the ratio of Cho/Cr had an upward trend, which may be due to the gradual aggravation of cholinergic neuron damage with the progression of the disease. Currently, cholinergic inhibitors such as donepezil, rivastigmine, and galanthamine are clinically used to treat AD. A meta-analysis showed that these drugs had modest but clinically significant overall benefits in stabilizing cognition, function, behavior, and overall clinical changes (Tan et al., 2014). Therefore, the change of Cho may reflect the severity of AD and was considered as a potential target for early detection and interventions.

In recent years, more research has focused on the change of Glx in patients with MCI and AD. Glx is a class of excitatory amino acid, including glutamate (Glu) and glutamine (Gln) (Bleich et al., 2003). In the brain, Glu and Gln are in dynamic equilibrium, and they can maintain and regulate synaptic information transmission through mutual transformation. In addition, Glu plays a crucial role in mitochondrial metabolism, neurotransmission of pyramidal cells, cerebral cortex function, and glutamate/GABA-glutamine cycle. And Glu-mediated synaptic transmission is critical for brain functions. However, excessive and continuous excitatory glutamatergic stimulation can lead to the death of neurons (Fayed et al., 2011). Interestingly, our results showed that compared with HC, the concentrations of Glx and Glu in PC of AD patients were lower, while the ratio of Glx/Cr was relatively higher, and the ratio of Glx/Cr in the PC of MCI patients was higher, while the concentration of Glu had a downward trend. In animal experiments, it was also found that the Glu/Cr was decreased in AD model mice (Liang et al., 2017). Studies have reported that the Aβ can induce several changes in nerve cells including the loss of neuronal viability and synaptic activity, leading to the reduction in glutamate levels. Meanwhile, the decrease of Glu content will affect the A-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptors (AMPARs), which play a key role in synaptic function and cognition. In addition, this decrease in AMPARs may be the reason for the loss of synaptic and the decrease of cognitive function in AD (Liu et al., 2010). Consequently, Glx/Cr and Glu may be seen as signs of cognitive deterioration in AD.

Several limitations to the current meta-analysis should be pointed out. First, the number of longitudinal studies to investigate the changes of metabolites between MCI-converter and MCI-stable patients was limited, so the sample size for analysis was relatively small. Therefore, more longitudinal studies are required to observe and explain the metabolite changes during the progress of MCI to AD. In addition, a significant effect of heterogeneity was found in many studies, and we were temporarily unable to do any moderating analysis to detect systematic influence on heterogeneity. Additionally, the detection results of MRS are affected by multiple parameters such as TR, TE, and ROI. This meta-analysis did not unify these parameters, which may lead to heterogeneity and affect the results.