Dose–Response Association Between High-Density Lipoprotein Cholesterol and Stroke: A Systematic Review and Meta-Analysis of Prospective Cohort Studies

Introduction Studies investigating the effect of high-density lipoprotein cholesterol (HDL-C) on stroke and stroke subtypes have reached inconsistent conclusions. The purpose of our study was to clarify the dose–response association between HDL-C level and risk of total stroke and stroke subtypes by a systematic review and meta-analysis. Methods We performed a systematic search of PubMed, Embase, and Web of Science databases through July 30, 2020, for prospective cohort studies that reported the HDL-C–stroke association and extracted the estimate that was adjusted for the greatest number of confounding factors. Restricted cubic splines were used to evaluate the linear and nonlinear dose–response associations. Results We included 29 articles, which reported on 62 prospective cohort studies including 900,501 study participants and 25,678 with stroke. The summary relative risk per 1-mmol/L increase in HDL-C level for total stroke was 0.82 (95% CI, 0.76–0.89; I 2 = 42.9%; n = 18); ischemic stroke (IS), 0.75 (95% CI, 0.69–0.82; I 2 = 50.1%; n = 22); intracerebral hemorrhage (ICH), 1.21 (95% CI, 1.04–1.42; I 2 = 33.4%; n = 10); and subarachnoid hemorrhage (SAH), 0.98 (95% CI, 0.96–1.00; I 2 = 0%; n = 7). We found a linear inverse association between HDL-C level and risk of total stroke and SAH, a nonlinear inverse association for IS risk, but a linear positive association for ICH risk. The strength and the direction of the effect size estimate for total stroke, IS, ICH, and SAH remained stable for most subgroups. We found no publication bias with Begg’s test and Egger’s test for the association of HDL-C level with risk of total stroke, IS, and ICH. Conclusion A high HDL-C level is associated with reduced risk of total stroke and IS and an increased risk of ICH.


Introduction
Studies investigating the effect of high-density lipoprotein cholesterol (HDL-C) on stroke and stroke subtypes have reached inconsistent conclusions. The purpose of our study was to clarify the dose-response association between HDL-C level and risk of total stroke and stroke subtypes by a systematic review and metaanalysis.

Methods
We performed a systematic search of PubMed, Embase, and Web of Science databases through July 30, 2020, for prospective cohort studies that reported the HDL-C-stroke association and extracted the estimate that was adjusted for the greatest number of confounding factors. Restricted cubic splines were used to evaluate the linear and nonlinear dose-response associations.
Many prospective cohort studies reported that a high-density lipoprotein cholesterol (HDL-C) level protected against the development of stroke (5)(6)(7)(8)(9)(10)(11). However, the "good cholesterol" label for HDL-C has been challenged by several recent randomized controlled trials demonstrating that HDL-C-elevating therapy increased the risk of cardiovascular diseases (12,13). Thus, a full understanding of the effect of HDL-C level on stroke and stroke subtypes is warranted. Only one systematic review, conducted in 2008, examined the association between HDL-C level and risk of total stroke (14). Another meta-analysis in 2013 investigated the association between HDL-C level and risk of hemorrhagic stroke (15). However, up to 10 more cohort studies have been published recently on the association of HDL-C level with total stroke, ICH, and SAH, showing inconsistent results (9-11,16-24). No metaanalysis has been performed on the association of HDL-C level with IS, and a dose-response meta-analysis on the association of HDL-C level with total stroke and IS is lacking. We therefore performed this systematic review and dose-response meta-analysis of prospective cohort studies to quantitatively evaluate possible linear or nonlinear associations between baseline HDL-C level and risk of total stroke, IS, ICH, and SAH.

Data sources and searches
We followed the protocol for the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) Statement for our meta-analysis (25). We conducted a systematic literature search of PubMed, Embase, and Web of Science databases for all reports of prospective cohort studies that examined the association between HDL-C level and stroke and were published through July 30, 2020, with no restriction on language. We also searched the reference lists of all related articles and reviews.

Study selection
Two authors (R.Q. and M.H.) independently searched articles, selected relevant studies based on their title and abstract, then evaluated these articles by reviewing the full text. Inclusion criteria for prospective cohort studies were as follows: 1) study participants were aged ≥18 years; 2) the study investigated the association between HDL-C level and risk of stroke or stroke subtypes; 3) the study reported the effect estimates, relative risks (RRs), or hazard ratios (HRs), with 95% CIs for ≥3 HDL-C categories or per-unit increase in HDL-C level; and 4) the study reported the number of cases, exposed person-years, or participant numbers in each category of HDL-C level. We excluded cross-sectional and casecontrol studies, commentaries, letters, reviews, meta-analyses, and studies with unusable data. If data from the same study were reported more than once, only the most recent and complete data were included.
Data extraction and quality assessment R.Q. and L.L. independently extracted the following information from each study: first author, publication year, study name, study location, follow-up period, age range, sex, stroke and HDL-C assessment method, baseline levels of HDL-C, case number of percategory HDL-C exposure, total persons or person-years of percategory HDL-C exposure, reported RRs or HRs and 95% CIs for each HDL-C category, and adjusted covariates. Included studies were assessed for quality according to the 9-point Newcastle-Ottawa Quality Assessment Scale (NOS) (26). Any discrepancy was resolved by discussion with a senior investigator (D.H.).
When HDL-C levels were reported in milligrams per deciliter (mg/dL), we used the scaling factor of 38.67 to translate 1-mg/dL HDL-C to 1-mmol/L HDL-C. Studies that provided results separately for men and women or reported multiple stroke subtypes within an article were treated as independent studies. For studies reporting results separately for fatal and nonfatal stroke, we combined the RRs and then included the pooled RR in the metaanalysis.

Data synthesis and analysis
We considered the RR and 95% CI of the effect size for all studies. The reported HRs in the primary studies were considered equal to RRs (28). We first used the DerSimonian and Laird random-effects model, which considers both within-study and between-study variation, to calculate summary RRs and 95% CIs for high versus low HDL-C level (29). Studies reporting only a continuous risk estimate of stroke were excluded from our analysis. We then pooled the study-specific dose-response RRs and 95% CIs per 1-mmol/L increase in HDL-C level (29).
We used generalized least squares regression to estimate the study-specific dose-response association (30). The natural RRs and CIs across categories of HDL-C level were used to compute study-specific slopes (linear trends) and 95% CIs. A generalized least squares regression model estimates the linear dose-response coefficients and considers the covariance for each exposure category within each study because they are estimated relative to a common referent HDL-C level category. In this method, the distribution of cases and person-years, or cases and noncases, with the RRs and estimates of uncertainty (eg, CIs) for ≥3 quantitative categories of exposure were required. If studies reported only the total number of cases or person-years, the number of person-years or cases in each category was obtained from the total number of person-years or cases divided by the number of reported categories. We assigned the mean, median, or midpoint of HDL-C level in each category to the corresponding risk estimate. When the lowest or highest categories were open-ended, we assumed the width of the category to be the same as the closest category when estimating the midpoint (31). For the studies already reporting a linear dose-response trend for per n-mmol/L increase in HDL-C level, we calculated the dose-response RRs per 1-mmol/L increase in HDL-C level with this formula: RR 1 = EXP (LN (RR n )/n*1), where RR 1 represents the dose-response RRs for each 1-mmol/L increase in HDL-C level and RR n represents the dose-response RRs for each n-mmol/L increase in HDL-C level (EXP: exponential function; LN: log base e) (32). All study-specific dose-response RR estimates were then pooled by using the DerSimonian and Laird random effects model (29). With heterogeneity (I 2 ) ≥50%, a random-effects model was used to calculate the summary RRs and 95% CIs; otherwise a fixed-effects model was used, which considered both within-and between-study variation. The Hartung-Knapp-Sidik-Jonkman method was used to evaluate the stability of results for N <10 (33). A potential nonlinear association was examined by modeling HDL-C level by using restricted cubic splines with 3 knots located at the 25th, 50th, and 75th percentiles of the distribution (34). The P for nonlinearity was calculated by testing the null hypothesis that the coefficient of the second spline is equal to zero (35).
Heterogeneity was assessed by Cochran Q and I 2 statistics (36). For the Q statistic, P < .10 was considered significant. For the I 2 statistic, I 2 values of 0%, 25%, 50%, and 75% were considered to reflect no, low, moderate, and high heterogeneity, respectively. We also performed subgroup analyses by sex, region, follow-up period, publication year, sample size, and the covariates (alcohol drinking, education, body mass index, systolic blood pressure, physical activity, lipid-lowering medication use, and other lipid profile parameters) adjusted in the analysis.
A sensitivity analysis was performed to assess the influence of each individual study by omitting 1 study at a time and calculating a pooled estimate for the remainder of the studies (37). Potential publication bias was assessed with Egger's and Begg's tests (38,39). Conversion from DerSimonian-Laird results to Hartung-Knapp-Sidik-Jonkman results involved using Microsoft Excel software (Microsoft Corp). Other analyses were conducted with Stata 12.1 (Stata Corp), and all tests were 2-sided with a significance level of P < .05.

Results
Literature search and study characteristics. Our literature search identified 7,366 articles; 1,113 were duplicates, leaving 6,253. After screening the titles and abstracts, we selected 201 potentially eligible articles. After detailed evaluation, we included 29 articles describing 62 prospective cohort studies in our metaanalysis with a total of 900,501 study participants of which 25,678 had stroke (5-11,16-24,40-52).
HDL-C level and risk of total stroke. To explore the association between HDL-C level and risk of total stroke, we examined 18 studies that included 256,427 participants overall and 12,328 PREVENTING CHRONIC DISEASE www.cdc.gov/pcd/issues/2021/20_0278.htm • Centers for Disease Control and Prevention people with stroke. We excluded 8 studies in comparing the highest versus lowest category of HDL-C because they provided only a continuous risk estimate. The pooled RR was 0.79 (95% CI, 0.72-0.87; I 2 = 46.4%; P heterogeneity = .05) ( Table 2). The 18 studies were included in the dose-response analysis; the pooled RR for total stroke was 0.82 (95% CI, 0.76-0.89) per 1-mmol/L increase in HDL-C level, with low heterogeneity (I 2 = 42.9%; P heterogeneity = .03) ( Table 3) . We found a linear dose-response association between HDL-C level and risk of total stroke (P nonlinearity = .96) ( Figure). No evidence of heterogeneity was detected between subgroups (Table 4). We observed an inverse association for most subgroups, except a nonsignificant association in studies of women, with a follow-up period of less than 10 years, without adjustment for physical activity or without adjustment for other lipid profile parameters (Table 4). HDL-C level and risk of IS. We included 10 studies consisting of a total of 706,482 participants and 19,047 people with stroke in the binary analysis of the association of IS risk with HDL-C level. The pooled RR was 0.75 (95% CI, 0.68-0.82; I 2 = 44.3%; P heterogeneity = .06; Table 2). Another 12 studies provided only a continuous risk estimate, so 22 studies were included in the dose-response analysis of IS risk. The pooled RR for IS was 0.75 (95% CI, 0.69-0.82) per 1-mmol/L increase in HDL-C level, with low heterogeneity (I 2 = 50.1%; P heterogeneity = .004) ( Table 3). We found a nonlinear dose-response association between HDL-C level and IS risk (P nonlinearity = .13) ( Figure). No evidence of heterogeneity was detected between subgroups (Table 4). Subgroup analyses showed a nonsignificant association in studies with a sample size of less than 10,000.
HDL-C level and risk of ICH. Ten studies consisting of 246,607 participants overall and 1,467 people with ICH were included in the analysis of HDL-C level and risk of ICH. The summary RR was 1.13 (95% CI, 0.93-1.36; I 2 = 29.9%; P heterogeneity = 0.17) in the binary analysis ( Table 2). The pooled results showed that risk of ICH was increased 26% per 1-mmol/L increase in HDL-C level (RR 1.21; 95% CI, 1.04-1.42), with low heterogeneity (I 2 = 33.4%, P heterogeneity = 0.14) ( Table 3). We found a linear dose-response association between HDL-C level and risk of ICH (P nonlinearity = 0.28) ( Figure). The effect size and direction of the pooled estimates were robust for most subgroups.
HDL-C level and risk of SAH. Data from 7 studies that included a total of 127,935 participants of which 551 had SAH provided information on the association between HDL-C level and risk of SAH. The pooled RR was 0.69 (95% CI, 0.50-0.95; I 2 = 30.7%; P heterogeneity = 0.19) ( Table 2) in the binary analysis. With a per-1-mmol/L increase in HDL-C level, the pooled RR was 0.98 (95% CI, 0.96-1.00; I 2 = 0%; P heterogeneity = 0.61) ( Table 3). Hartung-Knapp-Sidik-Jonkman results showed that risk of SAH was decreased 14% per 1-mmol/L increase in HDL-C level (RR 0.86; 95% CI, 0.75-0.98). We found a linear dose-response association between HDL-C level and risk of SAH (P nonlinearity = 0.94) ( Figure). The pooled estimates remained relatively stable on subgroup analyses.
Sensitivity analyses and publication bias. In sensitivity analyses, the results were robust when excluding one study at a time in the analysis of total stroke, IS, ICH, and SAH. We found no publication bias with Begg's test for risk of total stroke (P = 0.10), IS (P = .15), and ICH (P = .86), and Egger's test for risk of total stroke (P = .10), IS (P = .31), and ICH (P = .63). Publication bias was not assessed for the association between HDL-C level and SAH because of limited studies.

Discussion
We aimed to clarify the association between HDL-C level and risk of total stroke and stroke subtypes and found an inverse linear association between HDL-C level and risk of total stroke and IS. For each 1-mmol/L increase in HDL-C level, the risk of total stroke decreased by 18% and that of IS decreased by 24%. For ICH, we found a positive linear association, with the risk of ICH increased 21% per 1-mmol/L increase in HDL-C level. In addition, we found a marginal inverse linear association between HDL-C level and risk of SAH. Results of previous reviews and meta-analyses evaluating the association between HDL-C level and total stroke, ICH, and SAH were consistent with our study (14,15). However, previous research suggesting a negative association between HDL-C level and total stroke was based on a review of 8 cohort studies and 3 case-control studies (14). Our review did not report the association between HDL-C level and stroke subtypes because of the limited data on that relationship (14). In the current meta-analysis, we quantitatively evaluated the possible linear or nonlinear association of HDL-C level with total stroke, IS, ICH, and SAH.

PREVENTING CHRONIC DISEASE
We found an inverse linear association between HDL-C level and risk of total stroke. The reduced risk of total stroke may be due to the anti-atherosclerotic effects of HDL-C (42). The oxidation of LDL is thought to play an important role in the development of atherogenesis. HDL is a powerful antioxidant that exists in the subintimal space of the artery at a concentration 20 times greater than that of LDL and thus plays an important role in preventing atherosclerosis by inhibiting LDL oxidation in the artery wall (53). Additionally, HDL-C may play a central role in the reverse transport of cholesterol, thereby preventing the accumulation of excess cholesterol in peripheral tissues and the processes that initiate atherogenesis (54). However, subgroup analyses by sex showed significantly decreased risk of total stroke in men but not in women. The reason behind such inference remains unknown, and future experimental studies are needed to explore the potential mechanism.
Compared with a previous meta-analysis of HDL-C level and hemorrhagic stroke (15), 5 cohort studies were additionally included in our meta-analysis of the association of HDL-C level and ICH risk. We found a positive linear association of HDL-C level and ICH risk, which agreed with the previous meta-analyses. The possible mechanisms are as follows. First, HDL also has an antithrombotic function. A high HDL-C level can increase the risk of ICH by promoting fibrinolysis (10), which was found to be associated with the inhibition of coagulation cascade and the stimulation of blood clot fibrinolysis (58). In addition, HDL attenuates platelet function by stimulating endothelial cells to produce nitric oxide and prostacyclin (58,59).
Results of a previous meta-analysis reported a significant positive association between HDL-C level and SAH based on 2 cohort studies (15). Five cohort studies were additionally included in our meta-analysis of HDL-C level and SAH risk. We found a marginal inverse linear association between HDL-C level and SAH risk. More large-sample cohort studies are needed to firmly establish this association.
Our meta-analysis has several strengths. To our knowledge, this is the first meta-analysis to systematically examine the association between HDL-C level and risk of major stroke subtypes by using both binary and dose-response analyses. Also, all included studies had a prospective design, large sample size, and long followup. In addition, the high mean NOS score, 8.24, indicated a relatively high quality of the articles included.
Our meta-analysis also had several limitations. First, IS is a mixed term, including lacunar infarction, large-artery occlusive infarction, and embolic infarction. Only 1 study explored the distinction between IS subtypes, so we could not explore the association between HDL-C level and each IS subtype (10). Second, most included studies did not exclude participants using medication, which may have confounded the association of HDL-C level with risk of total stroke and stroke subtypes. Third, HDL-C level was measured only at baseline, so we could not consider the effect of HDL-C changes during follow-up. Finally, all included studies were observational, and we need further analyses based on randomized clinical trials for assessing the causality of HDL-C level on stroke.
The effects of HDL cholesterol levels on stroke risk vary by type of stroke. A high HDL-C level was associated with reduced risk of total stroke and IS, but an increased risk of ICH. Reasonable control of HDL-C level will prevent and control incident stroke. However, because the HDL particle is so complex, we do not know whether the particle size, number, HDL-C content, or functionality is the best marker of stroke risk. Future studies with information on potential mechanisms are needed.