Two independent reviewers (A.P. and K.P.) systematically screened the Medline (through PubMed), Scopus and Cochrane databases from inception to December 28, 2019 to identify studies investigating the association between LVH and CSVD neuroimaging markers. The detailed search strategy is available in the Supplement (Supplementary methods). All reference lists of the derived eligible articles were hand-searched for potential eligible studies not identified through the database search (“snowball” procedure). No language or publication year restrictions were applied. Eligible studies were evaluated for possible population overlap according to geographical setting, chronological period, sample size, outcome under study, and type of statistical analysis. In case of overlap, we opted for the most recent study. We further contacted the corresponding authors of articles presenting evidence that relevant data were available but not quantifying the associations under study, in order to request supplementary analyses. Differences between the two reviewers were solved through team consensus.

We considered as eligible cohort, cross-sectional, and case-control studies, as well as secondary analyses of randomized controlled trials exploring the association between LVH and neuroimaging markers of CSVD. Cases series, case reports, systematic or narrative reviews, animal and in vitro studies were excluded. We included studies of the general population or studies focused on specific high-risk populations, such as patients with stroke, hypertension, cardiovascular disease, diabetes mellitus, and chronic kidney disease. All analyses were performed separately for the general population and high-risk population studies. We excluded studies examining populations with genetic diseases predisposing to CSVD (e.g., Cerebral Autosomal Dominant/Recessive Arteriopathy with Subcortical Infarcts and Leukoencephalopathy [CADASIL, CARASIL], Fabry disease), autoimmune diseases and vasculitis, primary cardiomyopathies (e.g., dilatative or hypertrophic) and those including solely dementia individuals. Studies without a non-LVH comparison group were also excluded.

The exposure variables of interest included: (1) dichotomously defined LVH, diagnosed by electrocardiography (ECG), transthoracic echocardiography (TTE), or cardiac magnetic resonance imaging (MRI), and (2) continuous LVM measures, indexed (LVMI) or not to body surface area, assessed by TTE or cardiac MRI. For TTE-assessed LVH, we preferably included studies defining LVH as LVM ≥115 g/m² in males and ≥95 g/m² in females [13], but other cut-off points were also considered. ECG-assessed LVH should be defined by validated (e.g., Sokolow-Lyon indices or Cornell voltage criteria) [24,25] methods in standard 7 or 12-lead ECG. LVM should be calculated by TTE parameters according to the method of Devereux et al. [26].

The primary outcomes of our study included the following neuroimaging markers of CSVD, in accordance with the STandards for ReportIng Vascular changes on nEuroimaging (STRIVE [11]): lacunes, WMHs, CMBs, and EPVSs. Eligible were considered MRI or computed tomography (CT) studies assessing lacunes and WMHs, as previous literature has described compliance validity between the two methods [27,28], and MRI studies evaluating CMBs and EPVSs. We included studies defining lacunes as round or ovoid, subcortical, fluid-filled cavities, measuring between 3 and 15-mm in maximal diameter, consistent with a previous acute small deep brain infarct or hemorrhage in the territory of one perforating arteriole. Studies exploring lacunar strokes, defined as lacunes with acute clinical manifestations, were also included. We further post hoc decided to include studies examining “silent infarcts” provided that >80% of the included events were lacunes. WMHs should be identified as hyperintense areas on T2-weighted MRI sequences, isointense or hypointense on fluid-attenuated inversion recovery (FLAIR) imaging or as CT hypodensities. The studies should assess WMHs presence or severity through semiquantitative visual rating scales (e.g., Fazekas) or WMH volume via automated or semi-automated methods. Due to the high prevalence of WMHs in the elderly [12], the individual studies dichotomized WMH outcome based on specific burden levels (either based on a scale or a volumetric measurement) instead of mere presence. For simplicity, we use the term “extensive WMHs” to refer to this outcome although the individual studies used different methods for its assessment. CMBs had to be visualized as small (≤10 mm) areas of signal void with associated blooming on T2*-weighted MRI sequences. EPVSs should be defined as fluid-filled spaces following the course of a vessel with cerebrospinal fluid-like signal intensity.

A predefined spreadsheet was used to extract the following data: publication details (authors, year), study information (geographical region, recruitment period, design, population under study, sample size, follow-up parameters), study sample characteristics (age, gender, smoking, body mass index, hypertension history, diabetes mellitus, stroke, coronary artery disease), LVH/LVMI ascertainment (assessment method, definition, method/scale of quantification), CSVD assessment (marker under study, imaging modality, definition, method/scale of quantification, number of cases), and statistical analysis details (analysis type, effect estimates, 95% confidence intervals [CIs], adjusting variables). The corresponding author was contacted in case of missing data.

We evaluated studies for risk of bias using the Newcastle-Ottawa scale [29]. As the vast majority of eligible studies were of cross-sectional or cohort design, we applied the nine-item cohort subscale to all studies. The following criteria were assessed: representativeness of the exposed population; selection of the non-exposed group; LVH ascertainment; outcome absence at study onset; comparability of the exposed and non-exposed group for age and hypertension; CSVD markers assessment; follow-up period length and completion. Cross-sectional studies, by definition, did not receive any points for longitudinal assessment items (outcome absence at study onset, follow-up period length and completion). The detailed pre-defined handling of each criterion for the purposes of this systematic review is outlined in the Supplementary Table 1.

For each eligible study, we extracted association estimates and 95% CIs between presence of LVH and presence or incidence of neuroimaging CSVD markers. In 21 out of the 27 studies in our meta-analysis, the association estimates were odds ratios (ORs) derived from logistic regression analyses. Two prospective studies presented relative risks (RRs), but as the prevalence of the examined outcome was <10% in their population we considered RRs to be comparable to ORs [30] and pooled them with the other studies. Where ORs were not presented, we hand-calculated unadjusted ORs using 2×2 tables, based on data from the published articles. In studies presenting only ORs stratified by LVMI increments, we obtained the OR for the presence or absence of LVH by applying the method described by Hamling et al. [31]. For studies examining WMH volume or WMH severity measured as continuous outcomes, we transformed the presented beta coefficients to standardized mean differences and then used the latter to estimate the OR for a dichotomized WMH measure, based on validated formulae with the use of an online tool (https://campbellcollaboration.org/research-resources/effect-size-calculator.html ) [32].

We then performed random-effects meta-analyses of the derived association estimates to obtain pooled ORs and 95% CIs for each outcome. The method described by DerSimonian and Laird [33] was our primary meta-analytical approach. For our main analyses we also performed alternative random-effects meta-analytical approaches (ORs calculated via the Paule-Mandel between-study variance estimator [34], 95% CIs with the Hartung-Knapp [35] and modified Hartung-Knapp [36] methods), as detailed in the supplement (Supplementary methods) [34,36-46]. All analyses were performed separately for the general population and high-risk population studies. The presence of heterogeneity was evaluated by the I² and the Cochran Q statistics. We defined low, moderate and high heterogeneity as an I² of <25%, 25% to 75%, and >75%, respectively (significance threshold: P<0.10) [47]. To explore potential sources of heterogeneity, we performed sensitivity and subgroup analyses stratified by study design (cross-sectional, cohort), LVH assessment method (TTE, ECG), LVH definition criteria (ECG: only ↑QRS voltage-based criteria; TTE: LVMI ≥115 g/m² in males and ≥95 g/m² in females, body surface indexed), CSVD assessment method (MRI, CT), level of adjustment (studies adjusted for age, sex, hypertension, and other vascular risk factors), and fulfilment of the quality criteria of the Newcastle-Ottawa scale. Where possible, we further performed analyses for different left ventricular (LV) morphology patterns: normal LV geometry, concentric remodeling, eccentric hypertrophy, and concentric hypertrophy [13]. In order to explore the effect of each individual study in the overall estimate we conducted “leave-one-out” sensitivity analyses.

The results of our main analyses were graphically presented with funnel plots. The effect of potential publication bias (small-study effects) was explored in cases of ≥10 pooled studies using the Egger’s test (significance threshold: P<0.10) [48]. In case of statistically significant small-study effects, we adjusted the pooled effect estimates for publication bias using a “trim and fill” analysis [49].

Statistical significance for the main analyses was set at a two-sided P<0.05. All analyses were conducted with the STATA Software version 13.0 (Stata Corporation, College Station, TX, USA).

Figure 1 summarizes the study selection process. Following screening of 1,456 articles yielded by the literature search, we identified 34 articles meeting our eligibility criteria (60 studies were excluded after full-text screening as described in Supplementary Table 2). Three of them were excluded due to population overlap (Supplementary Table 3) [50-52]. Of the 31 studies [20-22,53-80] (n=25,562) included in our systematic review, only 27 [21,22,53-60,62-69,72-80] (n=21,010) provided appropriate data to also be used in the meta-analysis. Quantitative synthesis of articles examining associations between LVMI and WMH severity or volume as continuous variables was not possible because of the highly heterogeneous statistical methodologies. Of the studies included in meta-analysis, 13 examined presence of lacunes, 16 assessed extensive WMHs and three examined presence of CMBs. No eligible articles investigating EPVSs were identified.

The descriptive characteristics of the included studies are presented in Table 1.

The most commonly used definition for ECG-based diagnosis of LVH was the Sokolo-Lyon and/or Cornell definition, but some older studies [56,57,65-67,77] considered isolated QRS changes “normal” and defined LVH only if additional ST-segment and/or T-wave changes were present. Regarding TTE-defined LVH, the diagnosis was usually made according to internationally accepted standards, i.e., LVMI ≥115 g/m² for men and ≥95 g/m² for women (indexed to body surface area). Some studies, however, used different cut-off points [55,68,69,78], no indexing (g) [71], or height-based indexes (e.g., g/m².7) [59,72,76]. Six studies (n=1,279) [22,53,68,72,76,78] also evaluated LV morphology, classifying it as normal LV geometry, concentric remodeling, eccentric hypertrophy, or concentric hypertrophy.

Of the 13 studies [21,22,56,57,64,65,67,68,73,74,76,79,80] examining lacunes (n=13,529), nine were of cross-sectional (n=6,272), three of cohort (n=7,020), and one of case-control (n=237) design. Overall, lacunes were identified in 1,588 individuals. Five of the studies (n=6,650) were based on the general population and the remaining eight (n=6,879) on high-risk population subsets. Mean age of all the individuals was 67 years (range, 57 to 76). Lacunes were assessed by MRI in eight studies (n=6,091), by CT in two (n=844), whereas the three studies examining clinically manifest lacunar stroke utilized either MRI or CT (n=6,594).

Twenty-two studies [20-22,53-55,58,59,61-64,66,68-72,75,77,78,80] investigated WMHs (n=15,636). In 14 of these studies (n=8,540) the outcome was presence of extensive WMHs, whereas six studies (n=6,319) examined WMHs severity or volume as continuous outcomes. Two studies (n=777) presented both types of data. Eighteen studies were of cross-sectional (n=13,494) and three of cohort design (n=2,011), whereas one study presented both a cross-sectional (n=131) and a cohort analysis (n=113). Overall, nine of the studies (n=10,432) were based on the general population and the remaining 13 (n=5,204) on high-risk population subsets. Mean age of all individuals was 65 years (range, 30 to 76). WMHs were assessed by MRI in 21 studies (n=15,026) and CT in one (n=610). Presence of extensive WMHs was defined by the Fazekas-scale in seven studies (n=816). There was, however, heterogeneity regarding the cutoff point used to define extensive WMHs as well as the location of WMHs assessment (periventricular, deep, or both). Regarding continuous WMH data, either semi-quantitative scales were used to assess WMH severity, or semi-automated and automated computer-based algorithms calculated WMH volume. Two recent studies [20,54] examined the association between LVMI and diffusion tensor imaging parameters of WM integrity.

All three studies [60,63,69] (n=493) examining presence of CMBs were of cross-sectional design and based on high-risk population subsets. CMBs were identified in 151 individuals. Mean age of all the individuals was 63 years (range, 52 to 72). All studies utilized the same MRI-based definition.

The overall study quality was moderate. Only one study [79] (3%) fulfilled all nine criteria of the Newcastle-Ottawa scale (Supplementary Table 4). The median quality score was 5/9 for studies examining lacunes and WMHs and 4/9 for those examining CMBs. This could be explained by the cross-sectional design employed by 25 studies (81%), thus inherently limiting their maximum score to 6/9. Furthermore, only 10 studies (32%) were based on the general population, thus fulfilling the representativeness of the exposed cohort criterion. Most studies fulfilled the criteria for exposure and outcome assessment methods (87% and 94%, respectively), despite the between-study heterogeneity. Regarding the comparability criteria for age and hypertension, 15 (49%) studies controlled for both, 10 (32%) controlled for age but not hypertension, while only six studies (19%) presented unadjusted results. Lastly, concerning the cohort-specific criteria, three of the six cohort studies assessed CSVD markers at study onset, all six had follow-ups longer than 3 years, and attrition rates were <20% for three studies.

In studies of the general population we found presence of LVH to be associated with the odds of lacunes (OR, 1.49; 95% CI, 1.12 to 2.00; five studies; 6,650 individuals; 1,037 cases) and extensive WMHs (OR, 1.73; 95% CI, 1.38 to 2.17; five studies; 4,432 individuals) (Figure 2). Similar results were also obtained from studies in high-risk populations (lacunes: OR, 2.39; 95% CI, 1.32 to 4.32; eight studies; 6,879 individuals; 551 cases) (extensive WMHs: OR, 2.01; 95% CI, 1.45 to 2.80; 11 studies; 4,885 individuals) (Figure 2). A meta-analysis of the three high-risk population studies with data on CMBs also showed a significant association between LVH and presence of CMBs (OR, 2.54; 95% CI, 1.04 to 6.22; three studies; 493 individuals; 151 cases) (Supplementary Figure 1). When using various alternative meta-analytical approaches the associations for lacunes and extensive WMHs remained statistically significant, indicating the robustness of our findings (Supplementary Table 5).

Of note, the results for lacunes and extensive WMHs in the general population were also stable across studies adjusting their analyses for hypertension and other vascular risk factors on top of age and sex (lacunes: adjusted OR, 1.50; 95% CI, 1.09 to 2.06) (extensive WMHs: adjusted OR, 1.74; 95% CI, 1.34 to 2.25) (Figure 3).

When exploring LV geometry patterns and LVH subtypes, we documented different magnitudes of associations with lacunes and extensive WMHs (Figure 4). Specifically, in both studies of the general and high-risk-populations, we found gradually increasing associations estimates for concentric remodeling, eccentric hypertrophy, and concentric hypertrophy with the odds of lacunes and extensive WMHs.

Table 2 summarizes the results derived from eight studies (five in the general, three in high-risk populations) exploring associations between LVM or LVMI and heterogeneous methods for a continuous or ordinal assessment of WMHs severity or volume, which could not be included in the meta-analysis. In accordance with our main results, five of the eight studies showed statistically significant associations between higher LVM or LVMI and higher WMH severity or volume, while the association estimates were directionally consistent across all studies.

In studies of the general population, meta-analyses of WMHs studies showed no heterogeneity (I²=0%, P=0.77), while those of lacunes only moderate heterogeneity (I²=46%, P=0.12) (Figure 2). In subgroup analyses the results were stable for both cross-sectional and cohort studies, as well as for studies assessing LVH by either ECG or TTE (Table 3). Additionally, when restricting our analyses to studies defining LVH with the currently considered most optimal approaches (ECG: only ↑QRS voltage-based criteria; TTE: LVMI ≥95 g/m² for women and ≥115 g/m² for men, body surface indexed) the results remained stable (Table 3). Yet, we found moderate heterogeneity in studies of high-risk populations for both lacunes and WMHs (I²=71%, P=0.001 and I²=53%, P=0.02, respectively) (Figure 2). Although none of the sub-analyses entirely resolved the heterogeneity, the results were stable across the examined subgroups (Table 3). Overall, sensitivity analyses restricted to studies fulfilling each one of the Newcastle-Ottawa criteria showed consistent associations of LVH with both lacunes and WMHs (Supplementary Table 6). In “leave-one-out” sensitivity metaanalyses, we found no evidence that any single study significantly influenced the results of the main analyses (Supplementary Figure 2).

Funnel plots for the main analyses are presented in the supplement (Supplementary Figure 3). We did not perform the Egger’s test for meta-analyses of studies of the general population, or for high risk populations for the outcome lacunes due to <10 pooled studies. The Egger’s test showed statistically significant small-study effects for the 11 high risk WMHs studies (P=0.01). After adjusting this analysis for publication bias with the “trim and fill” method [49] the association remained statistically significant (OR, 1.45; 95% CI, 1.03 to 2.04) (Supplementary Figure 4).