Modiable Risk Factors for Incident Dementia and Cognitive Impairment: An Umbrella Review of Evidence

Dementia and cognitive impairment can be attributed to both genetic and modiable risk factors. Recently, considerable evidence emerged and urgently require standardized evaluation. To address it, we conducted an umbrella review of prospective studies regarding the associations of dementia and cognitive impairment with modiable factors to evaluate the strength and validity of the existing evidence. We searched PubMed, Embase, CINAHL and Cochrane Database of Systematic Reviews to identify relevant systematic reviews and meta-analyses of prospective studies. Mendelian randomization studies were reviewed to assess the causality for these associations. For each association, we analyzed the summary effect size, 95% condence interval, 95% prediction interval, heterogeneity, small study effect and excess signicance bias. Based on these estimates, the evidence was graded into levels of convincing, highly suggestive, suggestive, or weak. In total, 12015 articles were identied, of which 118 eligible studies yielded 243 unique associations. Convincing evidence was found for associations of dementia and cognitive impairment with early-life education, midlife to late-life plasma glucose, body mass index, atrial brillation, benzodiazepine use, and gait speed. Suggestive to highly suggestive evidence was found for associations of dementia and cognitive impairment with midlife to late-life blood pressure, homocysteine, cerebrovascular diseases, hearing impairment, respiratory illness, anemia, smoking, alcohol consumption, diet, sleep, physical activity and social engagement. Among convincing evidence, Mendelian randomization studies veried genetic predicted causal relationships for education and plasma glucose with Alzheimer's disease. Modiable factors identied in this study, especially those with high-level evidence, should be considered in dementia prevention. (FPG)


Introduction
As population aging and life expectancy extends, over 50 million people worldwide suffer with dementia and the number will triple to 152 million by 2050 [1]. The high prevalence of dementia brings heavy nancial and caring burden to families and society. The estimated global cost of dementia in 2018 is US$1 trillion and the gure will rise to US$ 2 trillion by 2030 if it's not properly prevented or treated [1]. 75% of caregivers feel stressed between caring dementia patients and meeting other responsibilities [2].
There is only one controversial disease-modifying drug, aducanumab [3], for Alzheimer's dementia (AD) while no therapies to delay the onset or progression of other types of dementia yet, largely because of complicated gene-environment interactions and unclear pathophysiological mechanisms which make it di cult to perform targeted drug development. The dilemma of high disease burden while no effective treatment emphasizes the necessity for primary prevention.
Promisingly, some studies have reported downwards trend of dementia [4], probably results from higher education and/or better vascular risk factors management [5,6]. The Lancet Commission indicated that modi able factors accounting for around 40% dementia worldwide, including education, hypertension, diabetes, obesity, depression, hearing impairment, brain injury, physical inactivity, social isolation, smoking, alcohol consumption and air pollution [7,8]. Modi cation of the factors has been suggested in dementia prevention. In earlier report, we proposed 21 suggestions for AD prevention targeting modi able factors [9]. Nevertheless, hierarchies of evidence have not been determined across various factors, and problems of different study designs, between-study heterogeneity may bias the results.
To address the issue, we performed an umbrella review of existing evidence to identify the compelling associations of modi able factors throughout the life course with incident dementia and cognitive impairment. We focused on evaluating the strength and validity of available associations then identi ed the factors with high-level evidence, which should provide perspectives for dementia prevention.

Methods
The umbrella review was conducted following the Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) guidance [10] (S1 Table), as well as standardized methods and principles of umbrella reviews [11,12]. It was registered with PROSPERO, number CRD42020195729.

Search strategy
We searched PubMed, Embase, CINAHL, and Cochrane Database of Systematic Reviews from inception to July 31, 2020 using the terms (dementia OR Alzheimer OR Alzheimer's OR cognitive impairment OR cognitive decline OR cognitive dysfunction) AND (systematic review OR meta-analysis OR Mendelian randomization) (S2 Text). No restrictions or lters were applied in the search strategy. We also manually searched the cited references of the included articles.

Selection criteria
Eligible articles were included if they were systematic reviews and meta-analyses of randomized controlled trials (RCTs) or observational prospective studies (OPSs), Mendelian randomization (MR) studies examining associations of modi able factors with incident dementia and cognitive impairment. All-cause dementia (ACD) including AD, vascular dementia (VD) and any other types of dementia, cognitive impairment (CI) and cognitive decline (CD) were all considered. The articles speci cally recruited participants with known dementia at baseline were excluded. When multiple reviews on the same association were identi ed we chose the one with the most recent and the largest number of primary studies to avoid duplication. If the most recent review is not at the same time the largest one, we explore the reason for the discrepancy. If the most recent review includes more prospective studies while the largest one not, we kept the most recent one, otherwise we kept the largest one. Articles containing non-OPSs studies but conducting the subgroup analysis of OPSs were included. Articles including less than three component studies were excluded. Only English written articles permitting easy access to the source information were included.

Data extraction
Three investigators (YN Ou, Y Qu, KM Wu) independently retained eligible articles by scanning the titles and abstracts then reviewing the full-text. Any disagreement was resolved by consensus and arbitration. For each eligible meta-analysis, we extracted PMID/DOI, rst author, publication year, journal name, sources of funding, number of included studies, number of total cases and participants, exposure, outcome, summary effect sizes along with their corresponding 95% con dence interval (CI), any measure of heterogeneity and publication bias. For each primary study, we extracted rst author, publication year, number of cases and participants, effect sizes and CI. For each MR study, we extracted data on PMID/DOI, rst author, publication year, journal name, number of total cases and participants, exposure, outcome, genetic instruments, MR causal effect estimates and CI.

Quality appraisal
For included systematic reviews and meta-analyses, the methodological quality was assessed with A Measurement Tool to Assess Systematic Reviews 2 (AMSTAR 2) [13]. AMSTAR 2 includes 16 items, seven of which can critically affect the validity of a review and its conclusions so that regarded as critical domains. The methodological quality is rated as high, moderate, low and critically low depending on weaknesses in critical and/or non-critical items.

Statistical analysis
For each association, we calculated the summary relative risk (RR) and 95% CI using random effects methods [14]. For the articles that presented results with odds ratio (OR), we transformed primary ORs to RRs using the algorithm: RR = OR / [(1 − p 0 ) + (p 0 × OR)] (p 0 indicates the incidence of endpoint in nonexposed group of the cohort) [15]. A 95% prediction interval (PI) was calculated to better evaluate the precision of the result. The heterogeneity was assessed by I² statistic [16] and it is considered large when I 2 > 50%.
We examined whether smaller studies gave greater side effects estimates than larger studies by Egger's regression asymmetry test [17]. p<0.10 was taken as statistical evidence of the small study effects.
We evaluated whether there was a relative excess bias of signi cant ndings in the published literature by excess statistical signi cance test [18,19], which assessed whether the observed (O) number of statistically signi cant studies (p<0.05) is too large compared with the expected (E) number using the χ 2 test. The expected number of studies was calculated by summing the statistical power estimated of each component study (assuming the true effect size the same as that of the largest component study).
p<0.10 was taken as statistical evidence of the presence of excess signi cance.
Gradation of the evidence strength followed the rules 11 : convincing (class I) when random effects p<10 −6 , number of cases>1000, I 2 <50%, 95% PI excluding the null, no small-study effects and no excess signi cance bias, highly suggestive (class II) when p<10 −6 , number of cases>1000, largest study with a statistically signi cant effect and class I criteria not met, suggestive (class III) when p<10 −3 , number of cases>1000 and class I-II criteria not met, weak (class IV) when p<0.05 and class I-III criteria not met, non-signi cant when p>0.05.
For MR studies, we did not conduct quantitative syntheses and just present here a descriptive analysis of the individual studies. p<0.05 was taken as statistical evidence of the presence of causal relationship.
All statistical analyses were done using Stata version 12.0.

Study selection and characteristics
Overall, 12015 articles were retrieved from systemic literature search and 611 articles were eligible for fulltext screening (Fig 1). A total of 493 articles not meeting the inclusion criteria were excluded (S3 Table). Seventy systematic reviews and meta-analyses eligible for umbrella review (S4 Text) were comprised of 5 articles of RCTs, 64 articles of OPSs and 1 article of both RCTs and OPSs. Forty-eight MR studies containing 157 causal analyses were included.
Most of the eligible articles didn't report a prior established protocol (item 2), a list of excluded studies (item 7) or the sources of funding for primary studies (item 10), which unfortunately downgraded the initial AMSTAR 2 rating. Therefore, we performed an adjusted analysis of the reporting quality by ruling out the above three domains. The adjusted reporting quality were high (39%) in 27 articles, moderate (24%) in 17 articles, low (31%) in 22 articles and critically low (6%) in 4 articles (S5 Table). Evidence from RCTs studies Only 9 associations were available from RCTs (S6 Table). Two interventions (22%), antihypertensive medications in midlife to late life and nonpharmacological interventions in late life, were found signi cantly (p<0.05) protective of ACD and CI while the evidence was weak. Of 9 associations, none had a 95% PI excluding the null value, 4 (44%) showed large heterogeneity (I²>50%), 2 (22%) presented smallstudy effects (Egger's p<0.10) and 3 (33%) were observed with excess signi cance bias.
Eighty-ve associations were examined in OPSs for ACD (S8 Table) (Fig 3). Seven (8%) associations were graded as highly suggestive evidence covering risk factors diabetes mellitus, high homocysteine level, prevalent stroke, incident stroke, hearing impairment, respiratory illness and current smoking in midlife to late life. Eleven (13%) associations were graded as suggestive evidence covering risk factors high systolic blood pressure (SBP) in midlife, antihypertensive medications, orthostatic hypotension, cerebral small vessel disease (CSVD) -white matter hyperintensities (WMHs), apathy, living alone and feeling loneliness in late life, high fasting plasma glucose (FPG) level, physical activity, wine consumption and ever smoking in midlife to late life. Thirty (35%) associations were graded as weak evidence. Of 85 associations, 23 (27%) had a 95% PI excluding the null value, 25 (29%) showed large heterogeneity, 11 (13%) presented small-study effects and 12 (14%) were observed with excess signi cance bias.
A total of 69 associations were extracted from OPSs for AD (S9 Table). Three (4%) associations were graded as convincing evidence covering risk factors low education ( (Fig 4). The association of one (1%) risk factor, diabetes mellitus in midlife to late life, was graded as highly suggestive evidence. Seven (10%) associations were graded as suggestive evidence covering risk factors hypertension, obesity, overweight and high BMI in midlife, antihypertensive medications in late life, statins and hearing impairment in midlife to late life. Twenty-three (33%) associations were graded as weak evidence. Of 69 associations, 4 (6%) had a 95% PI excluding the null value, 20 (29%) showed large heterogeneity, 7 (10%) presented small-study effects and 17 (25%) were observed with excess signi cance bias.
For VD, 21 associations were evaluated in OPSs (S10 Table). No association was graded as convincing evidence. The association of one (5%) risk factor, diabetes mellitus in midlife to late life, was graded as highly suggestive evidence. The association of one (5%) risk factor, current smoking in midlife to late life, was graded as suggestive evidence. Ten (48%) associations were graded as weak evidence. Of 21 associations, 3 (14%) had a 95% PI excluding the null value, 9 (43%) showed large heterogeneity, 5 (24%) presented small-study effects and 5 (24%) were observed with excess signi cance bias.
Eight associations for PDD were explored in only one OPS (S11 Table). No association was graded as convincing, highly suggestive or suggestive evidence. Four (50%) associations were graded as weak evidence. none association had a 95% PI excluding the null value, 5 (63%) showed large heterogeneity, 1 (13%) presented small-study effects and 5 (63%) were observed with excess signi cance bias.
Finally, 21 associations were assessed in OPSs for CI/CD (S12 Table). No association was graded as convincing evidence. The associations of two (10%) risk factors, low gait speed in late life and physical activity in midlife to late life, were graded as highly suggestive evidence. The associations of two (10%) risk factors, diabetes mellitus and hearing impairment in midlife to late life, were graded as suggestive evidence. Seven (33%) associations were graded as weak evidence. Of 21 associations, 3 (14%) had a 95% PI excluding the null value, 5 (24%) showed large heterogeneity, 3 (14%) presented small-study effects and 2 (10%) were observed with excess signi cance bias.
On the whole, the evidence gradation was convincing for 12 associations, highly suggestive for 12 associations and suggestive for 27 associations throughout the life course (Fig 5). In early life, education is the only, but determining, protective factor for dementia. While in midlife, blood pressure and BMI critically in uence the risk of dementia and cognitive impairment. When it moves into late life, blood pressure and BMI still matter and antihypertensive medications will bene t for prevention of dementia.
Late-life depression, apathy, social isolation, low gait speed, WMHs and anemia would affect the risk of dementia and cognitive impairment as well. Additionally, the incidence of dementia and cognitive impairment are related to benzodiazepine use, plasma glucose, homocysteine, atrial brillation, stroke, statins, hearing impairment, respiratory illness, smoking, alcohol consumption, sleep, healthy diet and physical activity in midlife to late life.

Evidence from MR studies
As for MR studies (S13 Table), genetically predicted increment of intelligence, education especially college/university completion, height, DBP (diastolic blood pressure), lipoprotein-a, cancer, rheumatoid arthritis, gut blautia and downstream product γ-aminobutyric acid (GABA), Vitamin D and Vitamin D binding protein play protective roles in AD. In contrast, elevated fasting glucose, white matter hyperintensities, periodontitis, alcohol consumption, branched-chain amino acids (isoleucine), decreased β cell function and complement C3 may increase the risk of AD.

Discussion
We provide a comprehensive overview of 243 reported associations of modi able risk factors with incident dementia and cognitive impairment by incorporating evidence from 70 systematic reviews and meta-analyses. 51 associations were graded as suggestive to convincing evidence and 87 were graded as weak evidence. The main factors bringing down the grade of evidence were between-study heterogeneity indicated in one-third of the meta-analyses, the limited cases number especially for VD and PDD, the presence of small-study effects or excess signi cance bias. Heterogeneity was mainly derived from inconsistent diagnostic criteria for dementia and cognitive impairment as well as various exposure assessment tools among primary studies.
Education is the only signi cant modi able risk factor in early life and it is convincing enough for application in the practice of dementia prevention. It is thought that education enhances the tolerance to age-related brain changes, which is implicit in the concept of cognitive reserve [20]. In contrast, social isolation in late life, either living alone or feeling loneliness, would weaken the cognitive reserve [21].
Vascular risk factors in midlife to late life including atrial brillation, stroke, CSVD, blood pressure, plasma glucose, BMI, homocysteine, smoking and alcohol consumption were supported by suggestive to convincing evidence increasing the risk of dementia and cognitive impairment. Hypoxia and hypoperfusion due to vascular diseases can accelerate the progression of neurodegenerative pathology [22]. In contrast, late-life high BMI is convincingly protective of dementia and cognitive impairment. Such a paradox could be partly explained by increased leptin hormone secreted by the adipose tissue, which may modulate hippocampal synaptic plasticity [23]. Furthermore, late-life BMI could in uence the levels of cerebrospinal uid (CSF) core AD biomarkers, amyloid-β (Aβ) and t-tau, as well as the volumes of hippocampus and cortex [24]. Antihypertensive medications and statins can be effective in dementia prevention based on suggestive evidence, and no particular antihypertensive class is signi cantly related to dementia and cognitive decline. Our analysis of OPSs reveals that healthy lifestyles including physical activity and fruit and vegetables intake contribute to decrease the incidence of dementia and cognitive impairment. However, the protective effects of physical activity and unsaturated fat supplementation are not signi cant in meta-analyses of RCTs.
The evidence for association of depression in late life and dementia is convincing. Neuroendocrine changes, hippocampal atrophy and vascular depression hypothesis are prominent mechanisms that may link depression and dementia [25]. Nevertheless, reverse causation may exist because depression is also a prodromal symptom of dementia. The uncertainty is similar for another neuropsychiatric symptom, apathy. Is it a sign of an early stage, or an independent risk factor for dementia? The lack of studies on early life apathy and dementia leaves this question open.
Motor dysfunction is predictive of dementia occurrence especially for PDD. Low gait speed in late life convincingly increases the incidence risk of dementia although the general evidence level of motor factors is weak partially due to the small sample size and large publication bias. The shared neural networks of cognition and motor function which are modulated by both dopaminergic and nondopaminergic transmitter systems [26] have been suggested to explain the underpinning mechanism.
Midlife to late life sleep disturbances including insomnia and long nocturnal sleep duration are responsible for the increased incidence of dementia and cognitive decline. A U-shaped association was reported between sleep duration and cognition [27]. Possible mechanisms of insu cient sleep time include inadequate glymphatic clearance pathways and vulnerable deposition of Aβ [28]. While excessive sleep time leading to in ammation activation may also result in cognitive deterioration [29]. Benzodiazepines are widely applied for treatment of sleep disturbances, but their use is convincingly linked to increased risk of dementia. Thus, it is necessary to nd alternative pharmacological or nonpharmacological interventions in sleep management for effective dementia prevention.
Additionally, comorbidities in midlife to late life including hearing impairment, respiratory illness and anemia generated a higher risk of dementia and cognitive impairment. Detailed mechanisms are not yet known. These comorbidities may impair cognition by mediating other risk factors such as hearing impairment causing social isolation, and respiratory illness and anemia inducing hypoxic ischemic brain vascular changes.
Compared with previous work of Bellou [30], we signi cantly updated the evidence with advanced prospective studies and further summarized MR studies to nd more solid associations. In addition, cognitive impairment with no dementia was also included to achieve an early-stage prevention. The evidence hierarchy of some factors was elevated, including education, gait speed, atrial brillation, BMI, homocysteine, stroke, WMHs, hearing impairment, respiratory illness, sleep and smoking, etc. While it is less suggestive for associations of cancer, aluminum, infectious diseases, low-frequency electromagnetic elds and NSAIDs. We also added evidence from MR studies to identify the risk factors not affected by confounding or reverse causation.
Our study has several limitations. First, meta-analyses of RCTs are inadequate. Bene cial factors supported by e OPSs need to be further veri ed by RCTs. Second, studies for dementia other than AD or VD are scarce. In our umbrella review, the dementia endpoints of the included articles are mainly AD or VD even if for articles of all-cause dementia. Only one article reported the risk factors for PDD and no studies speci cally focus on frontotemporal dementia, dementia with Lewy bodies or other types of dementia. Therefore, the prevention strategies based on this study may not be applicable to all types of dementia.
Third, although between-study heterogeneity, small-study effects and excess signi cance bias were evaluated, we cannot be sure that we excluded all of the biases that are inherent to individual studies. The inconsistence of baseline participants characteristics, exposure assessment, outcome diagnostic criteria, follow-up duration and adjustment model couldn't be fully controlled. Assessing the quality of the primary studies included in the meta-analyses is beyond the scope of an umbrella review. Fourth, we excluded the systemic reviews without quantitative analysis, so we missed some reported risk factors. Fifth, we did not establish whether the MR reports included assessment of the phenotype in relation to the polymorphisms studied in the cohorts, which might limit their validity, for example in the case of homocysteine [31]. Finally, our study was conducted on published English articles so the results can be updated by additional epidemiological studies.

Conclusions
Our study mapped the associations of incident dementia and cognitive impairment with various modi able risk factors. We identi ed factors with convincing evidence including education in early life, overweight, depression and low gait speed in late life, severe hypoglycemia, high FPG level, atrial brillation and benzodiazepine use in midlife to late life, among which, education and FPG are supported by MR causal analysis. In whole, our study provides evidence for modi able risk factors that can be considered in dementia prevention strategies. However, some of the evidence is limited by study quality and heterogeneity which need to be validated in further research.     Characteristics and quantitative synthesis of the eligible meta-analyses of OPSs for ACD by life course.

Declarations
Only the risk factors graded as suggestive to convincing evidence were listed. Convincing evidence existed in early-life education, late-life depression and low gait speed, midlife to late-life severe hypoglycemia, atrial brillation and benzodiazepine current or ever use. Highly suggestive evidence existed in midlife to late-life diabetes mellitus, high homocysteine level, prevalent or incident stroke, hearing impairment, respiratory illness and current smoking. Suggestive evidence existed in midlife high SBP, late-life antihypertensive medications, orthostatic hypotension, CSVD-WMHs, apathy and social isolation, midlife to late-life high FPG level, physical activity, wine consumption and ever smoking. ACD = All Cause dementia, CSVD = Cerebral Small Vessel Disease, FPG = Fasting Plasma Glucose, OPSs =