Magnesium and Cognitive Health in Adults: A Systematic Review and Meta-Analysis

Abstract

Magnesium (Mg) plays a key role in neurological functioning and manifestations. However, the evidence from randomized controlled trials (RCTs) and cohorts on Mg and cognitive health among adults has not been systematically reviewed. We aimed to examine the associations of various Mg forms (supplements, dietary intake, and biomarkers) with cognitive outcomes by summarizing evidence from RCTs and cohorts. PubMed, Embase, PsycINFO, and the Cochrane Central Register of Controlled Trials were searched for relevant peer-reviewed articles published up to May 3, 2024. Three random-effects models were performed, when appropriate, to evaluate the relationship between Mg and cognitive outcomes: 1) linear meta-regression, 2) nonlinear (quadratic) meta-regression, and 3) meta-analysis using Mg variables categorized based on pre-existing recommendations. Three RCTs and 12 cohort studies were included in this systematic review. Evidence from the limited number of RCTs was insufficient to draw conclusions on the effects of Mg supplements. Cohort studies showed inconsistent dose–response relationships between dietary Mg and cognitive disorders, with high heterogeneity across populations. However, consistent U-shape associations of serum Mg with all-cause dementia and cognitive impairment were found in cohorts, suggesting an optimal serum Mg concentration of ∼0.85 mmol/L. This nonlinear association was detected in meta-regression (P_quadratic = 0.003) and in meta-analysis based on the reference interval of serum Mg (0.75–0.95 mmol/L) [<0.75 compared with 0.85 mmol/L: pooled hazard ratio (HR) = 1.43; 95% confidence interval (CI) = 1.05, 1.93; >0.95 compared with 0.85 mmol/L: pooled HR = 1.30; 95% CI = 1.03, 1.64]. More evidence from RCTs and cohorts is warranted. Future cohort studies should evaluate various Mg biomarkers and collect repeated measurements of Mg intake over time, considering different sources (diet or supplements) and factors affecting absorption (for example, calcium-to-Mg intake ratio). This systematic review was preregistered in PROSPERO (CRD42023423663).

Statement of Significance.

Dementia is a global leading cause of disability and death among older adults, and there is a need to carefully evaluate existing evidence on related risk factors to guide future research. This review summarized and evaluated the most updated evidence on associations of various magnesium forms (supplements, dietary intake, and biomarkers) with cognitive outcomes in adults, based on evidence from randomized controlled trials and cohort studies.

Introduction

Dementia, a global and national leading cause of disability and death among older adults [[1], [2], [3]], is a group of symptoms affecting cognitive abilities, memory, behaviors, and abilities to perform daily activities [1]. It currently affects >55 million people worldwide, and the prevalence is predicted to rise to 78 million, with a cost of caregiving estimated to exceed US$ 2 trillion by 2030, because of the global population aging [1,4]. To relieve the substantial social and economic impact of dementia, evidence-based strategies are urgently needed to support prevention. Well-established modifiable risk factors for cognitive decline and dementia include lifestyle factors (for example, smoking, unbalanced diet, misuse of alcohol, and physical inactivity), vascular risk factors, depression, and psychosocial stress [5]. Electrolytes, including magnesium (Mg), have also drawn attention as potential risk factors, because of their critical roles in neurological functioning and manifestations [[6], [7], [8], [9], [10]].

Mg is the fourth most abundant cation in the body and is an essential electrolyte for maintaining cellular and organ function [11,12]. It passes the blood–brain barrier and plays a key role in neuronal maturation and functioning of the central nervous system [13,14]. The exact pathways linking Mg and brain health remain unclear. Several putative mechanisms include maintaining the integrity of the blood–brain barrier through the prevention of oxidative stress and inflammation [15,16], maintaining myelin sheaths and myelinated axons [17], and alleviating excitotoxic damage via inhibiting N-methyl-D-aspartate receptor activity and excessive calcium (Ca) influx [18,19]. In animal models, dietary Mg deficiency impaired hippocampus-dependent memories [20], whereas administration of Mg ions orally and directedly in the brain has been shown to reduce neuroinflammation [21], prevent synaptic loss [22], and improve learning and memory functions [22,23].

Evidence from human studies on Mg and cognitive health has been controversial. In a cross-sectional study of 2466 older adults aged ≥60 y from the NHANES 2011–2014, higher total Mg intake (from both diet and supplements) was associated with higher Digit Symbol Substitution Test score [highest quartile (>412 mg/d) compared with lowest quartile (<232 mg/d): β = 4.34; 95% confidence interval (CI) = 1.14, 7.54; P_trend <0.01] [24]. In other cross-sectional analyses, higher dietary Mg intake has been associated with lower odds of mild cognitive impairment (MCI). For example, a study with 612 Chinese adults (aged ≥55 y) with mean dietary Mg intake of 222 ± 69 mg/d found that the highest, compared with lowest, quartile was protective [odds ratio (OR) = 0.39, 95% CI = 0.19, 0.79], especially among females [25]. However, results from prospective long-term cohorts were mixed, with some studies showing no clear associations between dietary Mg and dementia or even opposite trends [[26], [27], [28], [29], [30], [31]].

A recent meta-analysis of 21 case-control studies demonstrated that Mg concentration in serum and plasma was lower in patients with Alzheimer’s disease (AD), compared with healthy controls [standardized mean difference (SMD) = –0.89, P < 0.001] [32], but the difference was not clear for Mg in cerebrospinal fluid (CSF) (SMD = –0.16, P = 0.36). It differs from the conclusion of an earlier meta-analysis of 13 cross-sectional studies, which found lower Mg concentrations in CSF and hair, but not in serum, among AD patients, relative to healthy controls [33]. These 2 reviews were limited to cross-sectional and case-control studies and did not examine other important forms of Mg (for example, Mg intake from diet and supplements) in relation to cognitive outcomes.

To our knowledge, there have been no comprehensive reviews of randomized controlled trials (RCTs) and cohort studies on Mg in relation to cognitive health among adults. Given that Mg deficiency is a modifiable risk factor for numerous health conditions [12], there is a critical need to understand its association with cognitive decline to inform prevention strategies. The goal of this study was to assess associations of Mg (supplements, dietary intake, and biomarkers) with various cognitive outcomes in adults by systematically reviewing the evidence from RCTs and cohort studies.

Methods

This systematic review was conducted and reported complying with the PRISMA guidelines (http://www.prisma-statement.org/) [34]. The protocol of this review was registered with PROSPERO, an international database of prospectively registered systematic reviews, on May 28, 2023 (ID: CRD42023423663). No modifications were made to the protocol other than adding a database (the Cochrane Central Register of Controlled Trials) and updating the study status.

Data sources and searches

An electronic search was conducted using PubMed, Embase, PsycINFO, and the Cochrane Central Register of Controlled Trials (via Ovid) databases for peer-reviewed journal articles published starting from the earliest possible date in the record up to May 3, 2024. A comprehensive search strategy was developed using key search terms of Mg and cognitive health as shown in Supplemental Tables 1–4. Mg exposures were searched using terms including, but not limited to, “magnesium”, “magnesium deficiency”, “magnesium ion”, and “magnesium intake.” Cognitive outcomes were searched using terms regarding global cognitive function (for example, “cognition,” “neuropsychological tests,” and “cognition assessment”), specific cognitive domains most affected by aging (for example, “attention,” “memory,” and “executive function”), and cognitive disorders (for example, “dementia,” “Alzheimer’s disease,” and “MCI”). A combination of relevant indexing terms (that is, Medical Subject Heading or MeSH for PubMed, Emtree for Embase, and Thesaurus for PsycINFO) and text words in abstracts, titles, and author-provided keywords were used to identify articles that investigated the association between Mg and cognitive health. A manual literature search was also conducted for capturing articles not identified by electronic search, using reference lists of 1) included studies and 2) relevant systematic reviews and meta-analyses. In addition, the clinical trial registry (https://clinicaltrials.gov/) was searched on May 22, 2024, for relevant ongoing and completed clinical studies.

Study selection

All identified studies were downloaded from databases and checked for duplicates in Rayyan [35]. Each abstract was screened by 2 reviewers independently in Rayyan, based on the pre-established criteria. Any disagreement between reviewers was discussed as a group until a consensus was reached. The same double-independent process was adopted for full-text evaluation. Studies were considered for inclusion if 1) they were human studies with participants aged ≥18 y, 2) exposure or intervention was Mg (for example, supplements, dietary intake, and biomarkers), and 3) outcome was cognitive function, measured by cognitive tests or diagnosed cognitive disorders (for example, all-cause dementia, AD, and MCI). Studies were excluded if 1) they were not original studies published in a peer-reviewed journal (for example, conference abstracts, preprints, and protocols); 2) they were published in languages other than English; or 3) they included participants with medical conditions that substantially affect Mg or cognition status, including dialysis, traumatic brain injury, or patients receiving intensive care. Studies on Mg-containing supplements or drugs were included in this review only if Mg was the only active component. As described in the protocol, any human study, except case reports, was included in the stages of abstract and full-text evaluation, but only cohort studies and RCTs were included in subsequent steps when evidence was identified from enough RCTs or cohorts (≥6). Details of inclusion and exclusion criteria are summarized in Supplemental Table 5.

Data extraction and risk-of-bias assessment

Each study was assigned to a reviewer for extracting qualitative data on study characteristics and quantitative data on association measures, using structured tables developed by the team. The extracted data was fully reviewed by a second reviewer. Qualitative data includes the location of the study population, funding source, study design, sample size, description of participants (age, sex, and health status), Mg exposure (form of Mg, assessment method, and unit), cognitive outcome (outcome type, assessment tool, validity of the tool, baseline status, and number of follow-up assessments), and duration of follow-up. Quantitative data for measures of association [for example, the difference in means, hazard ratio (HR), and OR] were collected along with 95% CIs. If a series of statistical models were performed in a study, the results from the most fully adjusted models were extracted. Study restrictions (for example, females only), stratification (for example, by race), and matching methods (for example, participants matched by age and sex) were extracted and considered as strategies to control for confounders.

Data for risk-of-bias assessment was extracted from each study independently by 2 reviewers. Grading results were compared and discussed to resolve any disagreement. RCTs were graded as “low risk,” “some concerns,” and “high risk” using version 2 of the Cochrane risk-of-bias tool based on the adherence to the criteria in 5 domains: randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result [36]. Cohorts were graded from 0 to 9 using the Newcastle–Ottawa Scale, which assesses 3 domains: selection, comparability, and outcome, with higher scores indicating lower risk [37]. The risk of bias was assessed and reported at the outcome level.

Data synthesis

For cohort studies, the characteristics and statistical outcomes were summarized by exposure type. Dose–response curves were charted for each study, based on reported HRs or ORs from the most fully adjusted models. In the figure for dietary Mg, the study-specific lowest categories were assigned as the reference group, as originally reported. There is a lack of a consistent reference group in studies on circulating Mg in serum and plasma. Therefore, we set the category that contains 0.85 mmol/L (the midpoint of the reference interval: 0.75–0.95 mmol/L) [38,39] as the uniform referent and recalculated risk estimates and 95% CIs for each study, using the method by Hamling et al. [40]. This method estimates cell counts, namely the “effective counts,” in the 2 × 2 table, adjusted for confounding, from which the asymptotic correlation between the adjusted log-risk estimates can be calculated for each exposure level compared with the referent. The Hamling method was implemented in SAS (accessible at www.pnlee.co.uk/Software.htm, retrieved on December 30, 2023). The calculations were conducted on a per-study basis.

Three models were performed to evaluate the associations between Mg and cognitive outcomes: 1) linear meta-regression, 2) nonlinear (quadratic) meta-regression, and 3) meta-analysis using Mg variables categorized based on existing recommendations. The first 2 models were designed to explore the trend of the association, whereas the third model was to evaluate whether pre-existing recommendations or reference intervals were related to cognitive outcomes. The details of the models are described below:

When 4 or more studies reported similar exposure-outcome associations, linear and nonlinear dose–response meta-regressions were performed using a 2-stage random-effects hierarchical regression model, implemented in the dosresmeta R package [41,42]. This method, initially outlined by Greenland and Longnecker [43], employs covariance matrix estimates to account for within-study correlations across exposure categories. These correlations are then integrated into linear and nonlinear trend estimations using the method of moments.

In addition, the Hamling method was used to combine Mg categories and to recalculate risk estimates based on predefined Mg cutoffs within each study. For example, all dietary Mg categories ≥ the Estimated Average Requirement (EAR) were combined into 1 group (“≥EAR”), and the rest were combined into the “<EAR” group. Then the risk estimate was recalculated for comparing Mg intake ≥EAR with <EAR. Given that the EAR for Mg varies by sex and age, we adopted 308 mg/d as the cutoff in the analysis, which is the average of the EAR for males (350 mg/d) and females (265 mg/d) aged over 30 y [44]. On the basis of the recalculated risk estimates, a random-effects meta-analysis was conducted to evaluate the weighted average association between dietary Mg (≥ compared with < the EAR) and cognitive outcomes with the metafor R package. Using the same method, meta-analyses were performed for serum Mg based on the reference interval, comparing the category containing serum Mg of 0.85 mmol/L with <0.75, 0.75–0.84, 0.86–0.95, and >0.95 mmol/L.

Several approximations were made in the analyses: The reported mean or midpoint of the Mg measure within each category was assigned to the corresponding risk estimate. If categories were open-ended, a mean Mg measure was estimated by multiplying the threshold by 0.8 for the lowest category and 1.2 for the highest category (Supplemental Method 1). When person-years for each Mg category were not reported, they were estimated by multiplying the number of participants by the mean or median follow-up years. When mean or median follow-up time was not provided, the entire follow-up duration was considered as a proxy. If Mg intake was reported as mg/1000 kcal, Mg intake/d was estimated by multiplying mg/1000 kcal with the reported average energy intake/d.

Sensitivity analyses were performed after excluding cohort studies 1) reporting ORs instead of HRs, 2) risk-of-bias score <7, and 3) not in generally healthy populations, as well as replacing number of participants at risk (n) with person-years in each Mg category. Statistical heterogeneity across the studies was evaluated using the Cochran Q statistic and the I² index, with a higher index indicating a greater extent of heterogeneity. Analyses were performed using R version 4.1.1 (The R Foundation for Statistical Computing) and SAS version 9.4 (SAS Institute). Meta-analysis was not performed for RCTs because of the limited number of studies (n = 3) and high heterogeneity between study populations.

Evaluation of evidence strength

The strength of evidence was assessed using the Grading of Recommendations Assessment, Development, and Evaluation approach [45] when there are >2 studies reporting similar exposure-outcome associations. Evidence strength is considered “insufficient” without assessment when there are ≤2 studies. Strength was graded as high, moderate, low, very low, or insufficient, based on 5 basic factors (risk of bias, inconsistency, indirectness, imprecision, and publication bias) and 3 additional criteria assessing the effect size, dose–response gradient, and plausible confounding that may change the effect. The publication bias assessment was not feasible because of the limited number of studies on a similar exposure-outcome association. Therefore, we assumed the presence of publication bias and graded it as “likely.” The grades for RCTs start as “high” and observational studies as “low,” with subsequent upgrades or downgrades based on the criteria. Given the strong correlation between serum and plasma Mg (r = 0.96), with minimal difference (<0.004 mmol/L) per a previous study [46], the term "serum Mg" was used to represent both biomarkers when presenting combined results in data synthesis and evaluation of evidence strength.

Results

Search results

A total of 3812 articles were identified in the databases, among which 3 RCTs and 12 cohort studies were included in this systematic review. The process of literature search and study selection is described in Figure 1. Studies excluded during full-text evaluation (n = 145) were summarized in Supplemental Table 6 with exclusion reasons. No additional eligible publications were identified by searching reference lists from included studies and relevant reviews. Our search using ClinicalTrials.gov and the Cochrane Central Register of Controlled Trials additionally identified 13 registered clinical trial protocols that aimed to investigate the association between Mg supplements and cognitive function. However, most of these trials were either recently completed or are currently ongoing, and no additional related publications were found. A summary of these trials is provided in Supplemental Table 7 to serve as a follow-up list for future systematic reviews.

FIGURE 1.

PRISMA flow diagram showing the process of literature search and study selection.

Randomized, controlled trials

Effects of Mg supplementation on cognitive function

Three double-blind RCTs were identified investigating the effects of Mg supplements on cognitive test scores [[47], [48], [49]]. In the study by Ni et al. [47], with female patients aged 18–70 y who underwent breast cancer procedures, 12-wk supplementation with Mg-L-threonate (1.2 g/d, containing 91.8 mg of Mg) in 34 participants did not improve cognitive function, as measured by the Telephone Interview for Cognitive Status, compared with placebo (n = 54) at 6-mo follow-up (P > 0.05). However, they used a relatively low dose of Mg and did not report baseline Mg status or other Mg sources (for example, Mg from food and beverages). This study was ranked as “high risk” due mainly to missing outcome data in 29% of the participants in the Mg group and 12% in the placebo group, as well as a potential risk in the selection of the reported results (Supplemental Figure 1). The Zhu et al.’s [48] study was nested within the Personalized Prevention of Colorectal Cancer Trial, including 123 adults aged 40–85 y. Participants in the Mg glycinate group received a personalized dose of Mg (mean dose = 216.5 mg/d) to reduce the Ca:Mg intake ratio to 2.3. After 12-wk treatment, the Mg group had a greater increase in the Montreal Cognitive Assessment (MoCA) score, compared with the placebo group (Mg group: 2.3 ± 2.7; placebo group: 0.5 ± 1.9; P = 0.01) in participants aged >65 y, though this effect was weak in participants aged ≤65 y. Further mediation analysis found that 86% of this beneficial effect among those >65 y could be explained by reduction in 5-methylcytosine (5-mC) levels at cg13496662 and cg06750524 CpG sites in apolipoprotein E (APOE). Although cg13496662 was a novel finding of the Zhu et al.’s [48] study, cg06750524 methylation had been earlier found to be associated with the APOE ε4 variant [50], the allele strongly related to a higher risk of AD [51]. The overall risk of bias in this study was low, except for some concern regarding potential risks in selection of the reported results. In another RCT, among 22 adults diagnosed with compensated liver cirrhosis, participants who received 520 mg/d Mg oxide (MgO) orally for 8 wk had higher average long-term memory score than the placebo group (Mg group: 4.17; placebo group: 2.75; P = 0.031) at the end of the study. Both Mg and placebo groups showed improvement in MoCA score compared with the baseline (P < 0.05), but not for recall memory score. This study also measured cognitive function using the clock completion test, digit span, and Lowenstein Occupational Cognitive Assessment, but the statistical results were not completely reported. This study was ranked as “high risk” because of potential deviation from the intended intervention (incomplete information on outcome measurement and statistical methods), missing outcome data, and potential risk in the selection of the reported results.

Cohort studies

Associations between dietary Mg intake and risk for cognitive disorders

Six prospective cohort studies investigated associations between dietary Mg intake and risk of cognitive disorders [[26], [27], [28], [29], [30], [31]], with only 1 study additionally reporting results for total Mg intake, from both diet and supplements [29]. The risk of bias in these studies was moderate, with 4 studies losing 1 point because of the limited representativeness of the exposed cohort [26], limited justification for the validity of the food frequency questionnaire [28,31], or unclear description of loss to follow-up [29] (Supplemental Table 8). Study characteristics are summarized in Table 1 [[26], [27], [28], [29], [30], [31],[52], [53], [54], [55], [56], [57]] and results are plotted in Figure 2A. Reported dietary Mg intake ranged from 116 to 508 mg/d, and was relatively lower in populations from China and Japan [27,28,31], and higher in populations from the United Kingdom and United States [26,29]. Only 1 study, from the UK Biobank, reported dietary Mg extending beyond 400 mg/d [26]. The Cherbuin et al.’s [30] study was not shown in Figure 2A because it reported linear associations without providing the observed range of Mg intake.

TABLE 1.

Characteristics of cohort studies examining the associations of dietary Mg (n = 6), Mg-containing drug (n = 1), and circulating Mg (n = 5) with cognitive disorders

First author, publication year, (reference) location	Source of population (study year)	Sample size analyzed (n)	Male (%)	Baseline mean age (SD),1 range	Baseline health status	Exposure type, unit, categories (method)	Outcome type (method)	Effect measure, HR (95% CI)2	ROB score3	Follow-up duration (y)4	Funding source
Dietary Mg (n = 6)
Takeuchi, 2023 [26] UK	The UK Biobank (from 2009–2012 to September 2021)	161,376	46.7	58.6 (8.0), NR	Assumed generally healthy	Dietary Mg, mg/d Q1 (≤263) Q2 (263–314) Q3 (314–360) Q4 (360–423) Q5 (>423) (24 h dietary assessment)	All-cause dementia (hospital inpatient records and death registry data, ICD-9/10)	Q1: 1 (REF) Q2: 0.97 (0.80, 1.18) Q3: 0.90 (0.73, 1.12) Q4: 0.80 (0.64, 1.01) Q5: 1.07 (0.84, 1.37)	8	11.3 y (median or mean NR)	Government, nonprofit
Luo, 2022 [27], China	The Shanghai Aging Study (from 2010–2011 to 2014–2016)	1565	46.7	71.1 (7.2), >60	Assumed generally healthy	Dietary Mg, mg/d T1 (<202) T2 (202–268) T3 (>268) (FFQ)	All-cause dementia (DSM-IV)	T1: 1 (REF) T2: 1.11 (0.65, 1.90) T3: 2.26 (1.02, 5.00)	9	5.2 y (mean)	Government
Kimura, 2022 [28], Japan	The Hisayama Study (from December 1988 to November 2012)	1060	42.2	69.5 (6.5), ≥60	Assumed generally healthy	Dietary Mg, mg/1000 kcal Q1 (M: ≤85; W: ≤94) Q2 (M: 86–98; W: 95–107) Q3 (M: 99–111; W: 108–124) Q4 (M: ≥112; W: ≥125) (FFQ)	All-cause dementia (DSM-III-R)	Q1: 1 (REF) Q2: 0.77 (0.58, 1.01) Q3: 0.61 (0.46, 0.83) Q4: 0.69 (0.50, 0.95)	8	24 y (median or mean NR)	Government
							AD (NINCDS-ADRDA)	Q1: 1 (REF) Q2: 0.84 (0.59, 1.19) Q3: 0.63 (0.43, 0.93) Q4: 0.79 (0.52, 1.19)	8
							VaD (NINDS-AIREN)	Q1: 1 (REF) Q2: 0.78 (0.48, 1.26) Q3: 0.64 (0.38, 1.07) Q4: 0.63 (0.35, 1.13)	8
Lo, 2019 [29], United States	The WHIMS and the WHIMS-ECHO5 (from 1995 to December 2012)	6473	0	70.1 (3.8), 65–79	Assumed generally healthy	1) Total Mg intake, mg/d Q1 (<197) Q2 (197–257) Q3 (257–318) Q4 (318–399) Q5 (>399) (FFQ + a dietary supplement questionnaire) 2) Dietary Mg, mg/d Q1 (<170) Q2 (170–216) Q3 (216–263) Q4 (263–323) Q5 (>323) (FFQ)	Probable Dementia (DSM-IV)	Total Mg: Q1: 1 (REF) Q2: 1.02 (0.73, 1.44) Q3: 0.87 (0.60, 1.28) Q4: 1.06 (0.69, 1.63) Q5: 1.25 (0.75, 2.08) Dietary Mg: Q1: 1 (REF) Q2: 1.00 (0.70, 1.42) Q3: 0.97 (0.66, 1.41) Q4: 1.11 (0.74, 1.67) Q5: 1.08 (0.64, 1.81)	8	9.6 y (mean)	Government, profit
							MCI (the Petersen's criteria)	Total Mg: Q1: 1 (REF) Q2: 0.77 (0.58, 1.03) Q3: 0.63 (0.45, 0.87) Q4: 0.67 (0.46, 0.97) Q5: 0.61 (0.39, 0.96) Dietary Mg: Q1: 1 (REF) Q2: 0.97 (0.73, 1.29) Q3: 0.85 (0.62, 1.17) Q4: 0.78 (0.55, 1.13) Q5: 0.71 (0.45, 1.14)	8
Cherbuin, 2014 [30], Australia	The PATH through life study (from 2001–2002 to 2009–2010)	1406	48.0	62.5 (1.5), 60–64	Assumed generally healthy	Dietary Mg, per 100 mg increment normalized to a daily energy intake of 10 MJ (FFQ)	MCI, amnestic (the Petersen’s criteria and Winblad’s criteria)	0.07 (0.01, 0.56)	9	8 y (median or mean NR)	Government
							Other mild cognitive disorders (DSM-IV)	0.47 (0.22, 0.99)	9
Ozawa, 2012 [31], Japan	The Hisayama Study (from December 1988 to November 2005)	1081	42.3	69 (NR), ≥60	Assumed generally healthy	Dietary Mg, mg/d Q1 (≤147) Q2 (148–169) Q3 (170–195) Q4 (≥196) (FFQ)	All-cause dementia (DSM-III-R)	Q1: 1 (REF) Q2: 0.61 (0.43, 0.86) Q3: 0.50 (0.34, 0.75) Q4: 0.63 (0.40, 1.01)	8	17 y (median or mean NR)	Government
							AD (NINCDS-ADRDA)	Q1: 1 (REF) Q2: 0.58 (0.35, 0.95) Q3: 0.53 (0.31, 0.92) Q4: 0.72 (0.38, 1.37)	8
							VaD (NINDS-AIREN)	Q1: 1 (REF) Q2: 0.44 (0.25, 0.79) Q3: 0.34 (0.17, 0.67) Q4: 0.26 (0.11, 0.61)	8
Mg-containing drug (n = 1)
Tzeng, 2018 [52], Taiwan	the LHID (from 2000 to December 2010)	6188	59.9	NR (NR), ≥50	General patients recorded in the LHID	Prescribed MgO, MgO users vs. nonusers (insurance claims data)	All-cause dementia (ICD-9-CM, DSM-IV, and DSM-IV-TR)	Nonusers: 1 (REF) Users: 0.52 (0.41, 0.79) Nonusers: 1 (REF) Users (use 1–365 d): 0.59 (0.42, 0.82) Users (use ≥365 d): 0.42 (0.14, 0.93)	6	10 y (median or mean NR)	Nonprofit
							Degenerative dementia: AD + nonVaD (ICD-9-CM)	Nonusers: 1 (REF) Users: 0.36 (0.28, 0.55)	6
							AD (ICD-9-CM)	Nonusers: 1 (REF) Users: 1.32 (0.90, 2.28)	6
							NonVaD (ICD-9-CM)	Nonusers: 1 (REF) Users: 0.20 (0.15, 0.36)	6
							VaD (ICD-9-CM)	Nonusers: 1 (REF) Users: 1.24 (0.80, 1.99)	6
Circulating Mg (n = 5)
Chen, 2021 [53], United States	The REGARDS (from December 2003 to April 2015)	2063	44.9	64.1 (9.0), ≥45	Assumed generally healthy	Serum Mg, mmol/L Level 1 (<0.75) Level 2 (0.75–0.81) Level 3 (0.81–0.87) Level 4 (≥0.87) (ICP-MS)	Cognitive impairment (SIS score ≤4)	Level 1: 1 (REF) Level 2: 0.59 (0.37, 0.94) Level 3: 0.54 (0.34, 0.88) Level 4: 0.59 (0.36, 0.96)	9	11.3 y (median or mean NR)	Government
Thomassen, 2021 [54], Denmark	The CGPS (from 2003–2015 to December 2018)	102,648	44.9	58.0 (NR), IQR: 48–67	Assumed generally healthy	Plasma Mg, mmol/L Q1 (0.72, IQR: 0.70–0.74) Q2 (0.78, IQR: 0.77–0.79) Q3 (0.81, IQR: 0.81–0.82) Q4 (0.85, IQR: 0.84–0.86) Q5 (0.90, IQR: 0.88–0.92) (NR, standard hospital assays)	All-cause dementia (national patient registry and death registry)	Q1: 1.15 (1.02, 1.31) Q2: 1.12 (0.98, 1.28) Q3: 1.08 (0.95, 1.24) Q4: 1 (REF) Q5: 1.12 (0.98, 1.28)	7	9.5 y (median)	Nonprofit, private donor
							VaD (ICD-8/10)	Q1: 1.49 (1.19, 1.85) Q2: 1.32 (1.05, 1.66) Q3: 1.21 (0.96, 1.53) Q4: 1 (REF) Q5: 1.32 (1.05, 1.66)	7
							AD (ICD-8/10)	Q1: 1.01 (0.86, 1.18) Q2: 1.04 (0.88, 1.21) Q3: 1.03 (0.87, 1.20) Q4: 1 (REF) Q5: 1.03 (0.88, 1.22)	7
Alam, 2020 [55], United States	The ARIC study (from 1990–1992 to 2018–2019)	12,040	43.7	56.9 (5.7), NR	Assumed generally healthy	Serum Mg, mg/dL Q1 (≤1.4) Q2 (1.5) Q3 (1.6) Q4 (1.7) Q5 (≥1.8) (colorimetric method)	All-cause dementia (expert adjudication, death certificate, medical history, ICD-9/10-CM)	Q1: 1.24 (1.07, 1.44) Q2: 1.08 (0.95, 1.24) Q3: 1.03 (0.91, 1.16) Q4: 1.08 (0.95, 1.22) Q5: 1 (REF) per 1 SD (0.009 mmol/L) decrease in serum Mg: 1.07 (1.02, 1.11)	9	24.2 y (median)	Government
Tu, 2018 [56], China	NR (from October 2013 to June 2015)	327	63.3	62.5 (10.2), 18–80	All with a history of acute ischemic stroke	Serum Mg, mmol/L T1 (≤0.82) T2 (0.83–0.88) T3 (≥0.89) (colorimetric method)	Poststroke Cognitive Impairment (MMSE ≤19 if illiterate, ≤22 if a primary school, and ≤26 points if secondary school and above)	T1: 2.24 (1.23, 4.06) T2: 0.77 (0.40, 1.48) T3: 1 (REF)	5	1 mo (mean)	Government
Kieboom, 2017 [57], The Netherlands	The Rotterdam Study (from 1997–2008 to January 2015)	9569	43.4	64.9 (9.7), ≥45	Assumed generally healthy	Serum Mg, mmol/L Q1 (≤0.79) Q2 (0.80–0.83) Q3 (0.84–0.85) Q4 (0.86–0.89) Q5 (≥0.90) (colorimetric method)	All-cause dementia (DSM-III-R)	Q1: 1.32 (1.02, 1.69) Q2: 1.04 (0.82, 1.32) Q3: 1 (REF) Q4: 1.22 (0.96, 1.55) Q5: 1.30 (1.02, 1.67)	8	7.8 y (median)	Government, nonprofit
							AD (NINCDS-ADRDA)	Similar associations were found in all-cause dementia, but the statistical results were nonsignificant and not fully reported.	8

Abbreviations: AD, Alzheimer's disease; ARIC, Atherosclerosis Risk in Communities study; CGPS, Copenhagen General Population Study; CI, confidence interval; DSM, Diagnostic and Statistical Manual of Mental Disorders; FFQ, food frequency questionnaire; HR, hazard ratio; ICD, International Classification of Diseases; ICD-9-CM, International Classification of Diseases-9-Clinical Modification; ICD-10-CM, International Classification of Diseases-10-Clinical Modification; ICP-MS, inductively coupled plasma mass spectrometry; IQR, inter quartile range; LHID, Longitudinal Health Insurance Database; M, Men; MCI, mild cognitive impairment; Mg, magnesium; MgO, magnesium oxide; MMSE, Mini-Mental State Examination; NINCDS-ADRDA, National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer’s Disease and Related Disorders Association; NINDS-AIREN, National Institute of Neurological Disorders and Stroke and the Association Internationale pour la Recherche et l’Enseignement en Neurosciences; NR, not reported; Q, quartile or quintile; REF, reference group; REGARDS, Reasons for Geographic and Racial Differences in Stroke Study; ROB, risk-of-bias; SIS, Six-Item Screener; T, tertile; VaD, vascular dementia; W, women; WHIMS, Women’s Health Initiative Memory Study; WHIMS-ECHO, Women’s Health Initiative Memory Study-Epidemiology of Cognitive Health Outcomes.

¹Mean age (SD) was estimated for overall participants based on Cochrane's formula (details of the method can be found at: https://www.statstodo.com/CombineMeansSDs.php) when it was only reported in stratified samples in original studies.

²All cohort studies reported HR (95% CI), except for Chen et al. [53] and Tu et al. [56], which reported OR (95% CI).

³ROB was assessed at the outcome level with the highest possible score of 9. Details can be found in Supplemental Table 8.

⁴The follow-up duration was estimated by counting the number of years between the start date (baseline) and the end date when the mean or median follow-up year was not reported.

⁵Participants were postmenopausal women who were initially enrolled for hormonal therapy.

FIGURE 2.

Results of cohort studies examining the associations of dietary Mg (A) and circulating Mg (B) with cognitive disorders. AD, Alzheimer's disease; ARIC, Atherosclerosis Risk in Communities study; CGPS, Copenhagen General Population Study; MCI, mild cognitive impairment; Mg, magnesium; NA, not available; REGARDS, Reasons for Geographic and Racial Differences in Stroke Study; WHIMS, Women’s Health Initiative Memory Study.

Five out of 6 studies evaluated the association between dietary Mg intake and all-cause dementia [[26], [27], [28], [29],31], reporting inconsistent results across cohorts. No clear association was found in 161,376 adults (mean age = 58.6 ± 8.0 y) from the UK Biobank followed for >11 y, or in 6473 female adults (mean age = 70.1 ± 3.8 y) from the Women’s Health Initiative Memory Study (WHIMS), with a mean follow-up of 9.6 y [26,29]. However, in 2 analyses using data from the Hisayama Study, with >1000 Japanese adults aged ≥60 y, higher dietary Mg intake was associated with lower dementia risk after 17 and 24-y follow-ups, respectively [28,31]. Although they categorized dietary Mg in different ways (sex-specific quartiles of Mg intake per 1000 kcal or study-specific quartiles of Mg intake/d), both studies observed the lowest dementia risk at the third quartile of Mg intake ∼180 mg/d [HR (95% CI) = 0.61 (0.46, 0.83) and 0.50 (0.34, 0.75), respectively], compared with the lowest quartile, below 150 mg/d. In contrast, higher dietary Mg was associated with greater dementia risk in 1516 older Chinese adults (mean age = 71.1 ± 7.2 y) from the Shanghai Aging Study [highest (>268 mg/d) compared with lowest tertile (<202 mg/d): HR (95%CI) = 2.26 (1.02, 5.00)] with mean follow-up of 5.2 y, but this association was confined to participants with Ca:Mg intake ratio ≤1.69 [27].

Only 2 analyses reported results on AD and vascular dementia (VaD), also using data from the Hisayama Study [28,31]. Both observed a U-shape association between dietary Mg intake and AD, with the lowest risk detected at the third quartile, ∼180 mg/d [HR (95% CI) = 0.63 (0.43, 0.93) and 0.53 (0.31, 0.92), respectively], compared with the lowest quartile. In addition, both studies found that higher dietary Mg tended to be associated with lower VaD risk, but it was stronger in the Ozawa et al.’s [31] study [highest (≥196 mg/L) compared with lowest quartile (≤147 mg/L): HR (95% CI) = 0.26 (0.11, 0.61)], and weak in the Kimura et al.’s [28] study, with a longer follow-up. MCI was investigated in 2 studies. The Cherbuin et al.’s [30] study found a linear trend, where higher dietary Mg was associated with lower MCI risk [HR (95% CI) = 0.07 (0.01, 0.56) per 100 mg/d increase in dietary Mg] in 1406 Australian adults, aged 60–64 y, and followed for 8 y. The Lo et al.’s [29] study observed a similar association in 6473 females from the WHIMS, but it was only strong for total Mg intake when combining diet and supplements [highest (>323 mg/d) compared with the lowest quintile (<170 mg/d): HR (95% CI) = 0.61 (0.39, 0.96)].

Dose–response meta-regression was performed using data from 4 studies that investigated dietary Mg intake and risk of all-cause dementia [[26], [27], [28], [29]]. The Ozawa et al.’s [31] study was excluded to avoid duplication of the study population with the Kimura et al.’s [28] study. The meta-regression did not detect a clear linear association [pooled adjusted HR (95% CI) = 0.98 (0.80, 1.21), P = 0.87] (Table 2 [53,56]). A possible U-shape association was detected in the nonlinear model (P_quadratic = 0.077), and the lowest risk of all-cause dementia was found at 240 mg/d [pooled adjusted HR (95% CI) = 0.76 (0.59, 0.98)], when compared with the lowest observed intake (116 mg/d). Results remained consistent in the sensitivity analysis after substituting the number of participants at risk (n) with person-years in the models.

TABLE 2.

Results of random-effects dose–response meta-regressions of prospective cohort studies

Models	Observed dose range	Studies (n)	Number of observed unique doses (n)	Follow-up (y)	Mg variable	P	Pooled adjusted, HR (95% CI)	I2 (%)	P for Cochran Q test
Exposure: dietary Mg intake (per 100 mg/d increase); outcome: all-cause dementia
Linear model	116–508 mg/d	4	17	5.2–24	Mg intake	0.872	0.98 (0.80, 1.21)	73.6	0.010
Sensitivity analysis1	116–508 mg/d	4	17	5.2–24	Mg intake	0.888	0.99 (0.81, 1.21)	72.9	0.011
Nonlinear (quadratic) model	116–508 mg/d	4	17	5.2–24	Mg intake	0.053	0.76 (0.59, 0.98)2	58.6	0.025
					Mg intake squared	0.077
Sensitivity analysis1	116–508 mg/d	4	17	5.2–24	Mg intake	0.051	0.76 (0.60, 0.98)2	58.7	0.024
					Mg intake squared	0.074
Exposure: serum Mg (per 0.1 mmol/L increase); outcome: all-cause dementia or cognitive impairment3
Linear model	0.464–1.080 mmol/L	5	21	0.08–24.2	Serum Mg	0.029	0.93 (0.88, 0.99)	76.0	0.002
Sensitivity analysis4	0.464–1.080 mmol/L	3	15	7.8–24.2	Serum Mg	0.098	0.97 (0.94, 1.01)	31.0	0.234
Nonlinear (quadratic) model	0.464–1.080 mmol/L	5	21	0.08–24.2	Serum Mg	0.008	0.44 (0.18, 1.09)5	71.8	<0.001
					Serum Mg squared	0.003
Sensitivity analysis4	0.464–1.080 mmol/L	3	15	7.8–24.2	Serum Mg	0.025	0.70 (0.47, 1.03)6	0.0	0.480
					Serum Mg squared	0.020

Abbreviations: CI, confidence interval; HR, hazard ratio; Mg, magnesium; MMSE, Mini-Mental State Examination; SIS, Six-Item Screener.

¹Number of participants at risk (n) in each Mg intake category was substituted with person-years in sensitivity analysis.

²The reported HR was the lowest HR observed at a dietary Mg of 240 mg/d, compared with the lowest intake (116 mg/d).

³All studies reported HRs for all-cause dementia as outcome, except for Chen et al. [53] reporting ORs for cognitive impairment defined by SIS score ≤4 and Tu et al. [56] reporting ORs for cognitive impairment 1 mo after stroke defined by MMSE score.

⁴In the sensitivity analysis, 2 studies (Chen et al. [53] and Tu et al. [56]) were excluded because of that 1) study reported ORs instead of HRs, 2) risk-of-bias score <7, 3) not a generally healthy population, or 4) outcome was not all-cause dementia. In addition, number of participants at risk (n) in each dose category was substituted with person-years.

⁵The reported HR was the lowest HR observed at serum Mg of 0.91 mmol/L, compared with the lowest concentration (0.464 mmol/L).

⁶The reported HR was the lowest HR observed at serum Mg of 0.845 mmol/L, compared with the lowest concentration (0.464 mmol/L).

In the analysis using cutoffs based on the EAR defined as 308 mg/d, no clear association was found between dietary Mg intake and all-cause dementia [≥EAR compared with <EAR: pooled adjusted HR (95% CI) = 1.14 (0.76, 1.70)] (Figure 3). Results from the Kimura et al.’s [28] and Ozawa et al.’s [31] studies were not included in this meta-analysis because their highest Mg intake categories were below 308 mg/d.

FIGURE 3.

Re-analysis of the associations between Mg exposures and all-cause dementia based on predefined cutoffs for Mg. Given that the EAR for Mg varies with sex and age, we adopted 308 mg/d as the cutoff in the analysis, which is the average of the EARs for males (350 mg/d) and females (265 mg/d) aged over 30 y. All studies reported HRs for all-cause dementia as outcome, except for Chen et al. [53] reporting ORs for cognitive impairment defined by SIS score ≤4 and Tu et al. [56] reporting ORs for cognitive impairment 1 mo after stroke defined by MMSE score. CI, confidence interval; df, degree of freedom; EAR: Estimated Average Requirement; HR, hazard ratio; Mg, magnesium; MMSE, Mini-Mental State Examination; OR, odds ratio; RE, random-effects; REF, reference group; SIS, 6-Item Screener.

Associations between Mg-containing drug intake and risk for cognitive disorders

One retrospective cohort study was identified, which investigated the association of prescribed MgO with risks of all-cause dementia and subtypes among 6188 Taiwan adults aged ≥50 y, using 10-y insurance claim data [52] (Table 1 [[26], [27], [28], [29], [30], [31],[52], [53], [54], [55], [56], [57]]). MgO users had a lower risk of all-cause dementia [HR (95% CI) = 0.52 (0.41, 0.79)] compared with nonusers, and this association was stronger among long-term users (≥365 d). A clear association was also found for nonVaD dementia [MgO users compared with nonusers: HR (95% CI) = 0.20 (0.15, 0.36)], but not for AD or VaD. The results did not account for Mg intake from diet, MgO dose, or medication compliance. The risk of bias was scored as 6, because of inadequate justification of final adjusted models and limited information on loss to follow-up.

Associations between circulating Mg concentration and risk for cognitive disorders

Five prospective cohort studies investigated the associations between circulating Mg and cognitive disorders, including 4 with serum Mg and 1 with plasma Mg [[53], [54], [55], [56], [57]]. The overall risk of bias in these 5 studies was moderate (Supplemental Table 8); only 1 study scored <7, because of limited representativeness of the exposed cohort, short follow-up (<5 y), and inadequate justification of final adjusted models [56]. Study characteristics are summarized in Table 1 [[26], [27], [28], [29], [30], [31],[52], [53], [54], [55], [56], [57]] and results are plotted in Figure 2B. The term "serum Mg" was used to represent both serum and plasma Mg when presenting combined results. Observed serum Mg concentrations ranged from 0.46 to 1.08 mmol/L, with 1 study reporting values below 0.60 mmol/L [55], and 2 extending beyond 0.95 mmol/L [56,57]. All 5 studies reported results on cognitive impairment, including 3 on diagnosed all-cause dementia [54,55,57], 1 on cognitive impairment defined by Six-Item Screener (SIS) score ≤4 [53], and 1 on cognitive impairment after stroke, defined by Mini-Mental State Examination score, with varying cutoffs based on education level [56]. Results were consistent across the 5 studies, suggesting a U-shape relationship, where the lowest risk of all-cause dementia or cognitive impairment was detected in the serum Mg categories containing 0.85 mmol/L. Stratified analyses were reported in 2 United States cohorts, suggesting that the association was not modified by age, sex, race (Black or White), or serum Ca concentration (all P_interaction ≥0.1) [53,55]. Two studies reported results on AD, including Thomassen et al. [54], conducted among 102,648 Danish adults (mean age = 58 y) with a median follow-up of 9.5 y, and Kieboom et al. [57], among 9569 Dutch adults (mean age = 64.9 y) with a median follow-up of 7.8 y. Both observed a U-shape association similar to the results for all-cause dementia, but the association with AD was weaker. On the other hand, Thomassen et al. [54], the only study that reported results on VaD, found that plasma Mg had a stronger U-shape association with VaD, relative to AD. Higher VaD risk was observed at the lowest quartile (0.72 mmol/L) [HR (95% CI) = 1.49 (1.19, 1.85)] and the highest quartile (0.90 mmol/L) [HR (95% CI) = 1.32 (1.05, 1.66)] when compared with 0.85 mmol/L.

Dose–response meta-regression, including all 5 studies, detected a linear relationship where every 0.1 mmol/L increment in serum Mg was associated with a 7% lower risk of all-cause dementia or cognitive impairment [pooled adjusted HR (95% CI) = 0.93 (0.88, 0.99), P = 0.029] (Table 2 [53,56]). This association was attenuated in sensitivity analysis [pooled adjusted HR (95% CI) = 0.97 (0.94, 1.01), P = 0.098], when number of participants at risk (n) were replaced by person-years, and when 2 studies [53,56] were excluded because of 1) reporting ORs, 2) risk-of-bias score <7, 3) not a generally healthy population, or 4) outcome was not diagnosed all-cause dementia. A nonlinear relationship was detected by the quadratic model (P_quadratic = 0.003), again suggesting a U-shape relationship, where the lowest risk of dementia or cognitive impairment was detected at serum Mg of 0.91 mmol/L [pooled adjusted HR (95% CI) = 0.44 (0.18, 1.09)], compared with the lowest observed concentration (0.464 mmol/L). In sensitivity analysis, the nonlinear association remained strong (P_quadratic = 0.020), and the lowest risk was observed at 0.845 mmol/L [pooled adjusted HR (95% CI) = 0.70 (0.47, 1.03)], compared with 0.464 mmol/L.

In analysis using cutoffs based on the reference interval (0.75–0.95 mmol/L), we found that serum Mg either lower or higher than the interval was associated with higher risk of dementia or cognitive impairment, compared with 0.85 mmol/L [<0.75 compared with 0.85 mmol/L: pooled adjusted HR (95% CI) = 1.43 (1.05, 1.93); >0.95 compared with 0.85 mmol/L: pooled adjusted HR (95% CI) = 1.30 (1.03, 1.64)] (Figure 3). Even within the reference interval, serum Mg below or above 0.85 mmol/L tended to be associated with higher risk, compared with the midpoint [0.75–0.84 compared with 0.85 mmol/L: pooled adjusted HR (95% CI) = 1.09 (0.98, 1.20); 0.86–0.95 compared with 0.85 mmol/L: pooled adjusted HR (95% CI) = 1.14 (1.02, 1.28)]. This U-shape relationship was slightly attenuated in sensitivity analysis but remained clear (Supplemental Figure 2). I² was high (90.6%, that is, great heterogeneity) when all 5 studies were included in the analysis comparing serum Mg <0.75 with 0.85 mmol/L, but it decreased to 0.0% in the sensitivity analysis where Chen et al. [53] and Tu et al. [56] were excluded. This suggests that heterogeneity may result from the difference in outcome definition, effect matrix (HR or OR), population health status, or study quality.

Discussion

On the basis of our assessment of evidence strength, we conclude with moderate confidence that serum Mg is associated with all-cause dementia and cognitive impairment, characterized by a U-shape pattern with the lowest dementia risk detected around serum Mg of 0.85 mmol/L (Supplemental Table 9). However, the evidence on dietary Mg and dementia was graded as “Low,” mainly because of inconsistent results across studies, and lack of clear dose–response gradient. Evidence on other Mg types or cognitive outcomes was insufficient to draw a conclusion. Most of the included studies were published in recent years, indicating a notable surge in attention to this topic. However, more evidence is needed from RCTs and cohort studies that investigate the effects of different Mg forms on various cognitive outcomes.

The pathways linking Mg and cognitive health remain unclear. Several putative mechanisms include supporting the integrity of blood–brain barrier by preventing oxidative stress and inflammation [15,16], maintaining myelin sheaths and myelinated axons [17], and alleviating excitotoxic damage via inhibiting N-methyl-D-aspartate receptor activity and excessive Ca influx [18,19]. In animal models, dietary Mg deficiency impaired hippocampus-dependent memories [20], whereas administration of Mg ions orally and directedly in the brain has been shown to reduce neuroinflammation [21], prevent synaptic loss [22], and improve learning and memory functions [22,23]. In addition, Mg is involved in vascular tonus and blood pressure regulation, and Mg deficiency has been associated with elevated angiotensin II-mediated aldosterone, thromboxane, vasoconstrictor prostaglandins, as well as a higher risk of hypertension [12], which may partially explain the association between Mg and VaD. Two recent human studies evaluated the relationship between Mg and brain morphology using neuroimaging data [58,59]. A UK Biobank cohort study with 6001 adults, aged 40–73 y, found that every 1 mg/d increase in baseline dietary Mg intake was associated with 0.001% (SE = 0.0003) increase in gray matter, 0.0013% (SE = 0.0006) in left hippocampal volume, and 0.0023% (SE = 0.0006) in right hippocampal volume, in both men and women [59]. In another study among 1466 older adults (mean age = 76.2 ± 5.3 y) free of prevalent stroke, higher serum Mg was cross-sectionally associated with greater total brain volume and frontal, temporal, and parietal lobe volumes, as well as lower odds of subcortical infarcts [highest (>0.88 mmol/L) compared with the lowest quintile (<0.76 mmol/L): OR (95% CI) = 0.44 (0.25, 0.77)] and lacunar infarcts [highest compared with lowest quintile: OR (95% CI) = 0.40 (0.22, 0.71)] [58]. These studies provide additional clues to plausible pathways between Mg deficiency and cognitive disorders that involve neurodegeneration and cerebrovascular damage in particular brain regions. On the other hand, the mechanism underlying hypermagnesemia and cognitive health has been scarcely explored in previous studies. Hypermagnesemia has been associated with hypocalcemia [60] and cardiovascular conditions characterized by bradycardia, hypotension, and conduction defects [61], which are potent risk factors for stroke [[62], [63], [64]]. Given that stroke is a well-known cause of VaD [65], it may mediate the association between hypermagnesemia and dementia.

The inconsistent results observed in cohort studies on dietary Mg and cognitive disorders may be partially explained by different Ca:Mg intake ratios across study populations. Higher dietary Mg was associated with elevated dementia risk in the Shanghai Aging Study among participants with Ca:Mg intake ratio ≤1.69 (that is, low Ca and high Mg diet) [27], while it was associated with lower dementia risk among participants from Hisayama Study, with Ca:Mg intake ratio ∼3 (that is, high Ca and low Mg diet) at population level [28,31]. Ionized Mg acts as a physiologic antagonist to ionized Ca [66]. Studies found that Ca competes with Mg for intestinal absorption and transport [67], highlighting the importance of maintaining a balanced intake of these 2 nutrients. This is supported by results from the RCT by Zhu et al. [48], which showed that maintaining a Ca:Mg ratio at 2.3 by taking Mg supplements optimized plasma vitamin D concentration, reduced methylation in the APOE gene, and improved MoCA score [48,68]. Future studies should take the Ca:Mg intake ratio into consideration when they evaluate the effects of Mg. The inconsistent results may also be explained by differences in the range of dietary Mg across study populations, making direct comparisons challenging. In this review, Mg intake was measured only once in most cohorts, thus, the results may not represent long-term dietary intake over years of follow-up. In addition, Mg intake from supplements was not accounted for in most cohort studies, which may lead to an underestimation of Mg intake and distort its true association with cognitive outcomes. These limitations should be avoided in future studies on Mg intake.

Cohort studies on serum and plasma Mg consistently supported a U-shape association with cognitive disorders, suggesting that blood Mg biomarkers may be a useful predictor for brain aging. Although serum Mg only accounts for <1% body Mg and may not represent the total Mg content, it reflects Mg homeostasis and correlates well with the free Mg-ion, a physiologically active form of the element Mg, in erythrocytes and serum [69,70]. Thus, it is the predominant measure of Mg status used in clinical settings and research [39]. In this review, we adopted the reference interval (0.75–0.95 mmol/L) derived from the distribution (95th percentile) in a United States nationally representative sample from the NHANES I in 1974 [38]. However, there is no international consensus for a normal range of serum Mg. The reference interval is under debate and more studies are needed to inform an evidence-based reference range. Our review summarized the most updated evidence and tested multiple models to evaluate the association between serum Mg and cognitive outcomes, which may contribute to the establishment of a new reference range from the view of cognitive health. In addition, more studies using other important Mg biomarkers, such as free Mg-ion in CSF and erythrocytes, are needed to help identify the best form of Mg that predicts long-term cognitive outcomes.

Our literature search identified 2 systematic reviews with meta-analyses on Mg and AD, but they were limited to cross-sectional and case-control studies [32,33]. They investigated AD as the only outcome and Mg biomarkers as exposures of interest, with no evidence reported on dietary or supplemental Mg. The review conducted by Veronese et al. [33] identified 13 studies published up to May 4, 2015, finding that patients with AD had lower Mg in CSF (2 studies, SMD = –0.35, P = 0.02) and in hair (2 studies, SMD = -0.75, P = 0.0001), compared with healthy controls, but no clear difference was detected in serum, plasma, and ionized/blood cell Mg. With more updated studies published by 2021 included in the review by Du et al. [32], the difference in serum/plasma Mg between AD and controls was strong (18 studies, SMD = –0.89, P < 0.001), but the difference in CSF Mg was attenuated (3 studies, SMD = –0.16, P = 0.36). These 2 reviews provide evidence for various Mg biomarkers (for example, CSF, hair, and blood cell Mg) that were not available in our review of RCTs and cohorts, but their results should be interpreted with caution, because of potential selection bias in case-control studies and uncertainty in temporal precedence in cross-sectional studies.

This systematic review and meta-analysis has several limitations. First, we included only publications in English; thus, language and publication bias cannot be ruled out. Second, our meta-analyses of cohort studies may be affected by the limited number and risk of bias of original studies, potential residual confounding, ecological bias, and measurement precision of Mg exposures, which may limit the interpretation of data. Third, sub-group analysis (for example, by age, sex, or race) was not conducted, because of the limited number of included studies. Fourth, covariate adjustment varied across cohort studies. Our analyses combined risk estimates from the most complete model in each study, assuming that the difference in adjustments would not substantially affect the results. This assumption cannot be easily verified and, thus, warrants cautious interpretation of our meta-analytic results. Fifth, possible reverse causality cannot be completely ruled out based on data from prospective cohorts. In this review, 92% of cohort studies demonstrated that participants were excluded from the analysis if they had cognitive outcomes of interest at baseline, which may alleviate the possibility of reverse causality. In addition, the follow-up durations of 92% of cohorts were >5 y, and 50% were >10 y. The long-term follow-up also enhanced the ability to establish temporal precedence and improved the control over reverse causality. However, more evidence from RCTs is still needed to help determine the direction of the association. Lastly, the outcomes were reported as dichotomous variables (yes or no) in cohort studies, which cannot capture the severity of the cognitive disorder and may underestimate the variability. In addition, Mg intake beyond the Tolerable Upper Intake Level is lacking; thus, the risk for cognitive disorders at very high Mg intake levels cannot be tested.

Our systematic review summarized evidence on relationships between Mg supplements and short-term cognitive test scores from RCTs, as well as prospective associations of Mg intake and serum Mg with long-term cognitive disorders from cohort studies. Evidence from RCTs should be re-evaluated once the results are published from additional ongoing and recently completed trials. Given the limited evidence from trials so far, large population-based cohort studies improved the strength of evidence by providing complementary data for future clinical or policy decision-making. Future cohort studies are recommended to evaluate various Mg biomarkers and to conduct repeated measurements of Mg intake over time, taking sources (diet or supplements) and other nutrients (for example, Ca:Mg ratio) into consideration.

Author contributions

The authors’ contributions were as follows – FC, TS, KLT: designed research; FC, JW, YC, RL, YW, YC: conducted abstract and full-text screening, extracted data, and performed risk-of-bias assessment; FC: created the protocol, performed statistical analysis, and wrote the article; KLT: edited the manuscript; all authors: read and approved the final manuscript; and KLT: had primary responsibility for final content.

Conflict of interest

No potential conflicts of interest relevant to this article were reported.

Funding

This study was funded by NIH P50HL105185 and P01AG023394 (KLT). The funding source had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication. JW, YC, RL, YW, and YC contributed to this systematic review without receiving funding or salary support.

Data availability

Data described in the manuscript will be made available upon request pending (for example, by e-mailing the corresponding author).

Appendix A. Supplementary data

The following is the Supplementary data to this article: