We searched PubMed, Cochrane Library, and Embase databases for studies examining the association between LDL-C and risk of HS. The following search terms were used: (“hemorrhagic stroke” [MeSH Terms] OR “intracerebral hemorrhage” [Title/Abstract] OR “subarachnoid hemorrhage” [Title/Abstract]) AND (“cholesterol, ldl” [MeSH Terms] OR “low density lipoprotein cholesterol” [Title/Abstract]). The search was limited to studies published before October 2024. The language was restricted to English publications. A detailed search strategy is provided in the Supplementary Material. This systematic review has not been registered.

Included studies had to meet the following criteria: a prospective design (prospective cohort studies (PCs) or nested prospective case–control study) or RCT; investigate the association between LDL-C level and the risk of HS (intracerebral hemorrhage (ICH), subarachnoid hemorrhage (SAH), or both); report effect estimates (risk ratio (RR), hazard ratio (HR), or odds ratio (OR)) and 95% confidence intervals (CIs) for comparisons between different concentration levels, or sample number and cases in each group to be able to calculate the RR; provide a clear definition of hypertension. Duplicate publications from the same study were excluded. Publications with two or more categories containing zero cases were also excluded.

Two investigators reviewed the included studies and completed standard data extraction forms separately. This form included the following information: author, publication year, study design, study name, location, number of participants, hypertension (%), mean/median age (range), female sex (%), mean/median follow-up duration, endpoint, first occurrence/recurrence, time of LDL-C measurement, details of each LDL-C category such as LDL-C concentration, sample size, cases, effect estimates, 95% CI, and adjusted covariates. The one with the largest number of adjusted variables was extracted for studies reporting several effect estimates.

The quality of the included cohort studies was comprehensively assessed using the Newcastle–Ottawa Scale (NOS) [14]. Studies with more than six stars were regarded as high quality. The quality of the included RCTs was comprehensively assessed using the Cochrane Collaboration’s risk of bias tool, with reference to the Cochrane Handbook [15]. The third investigator resolved disagreements between investigators in data extraction and quality assessment.

Our study regarded RRs as the effect size, with HRs considered equivalent to RRs. For studies reporting results for men and women separately, we combined the estimates using a fixed-effects model to obtain an overall RR of HS for an individual study. A random-effects model was used with I² $>$ 50% [15] or p_{heterogeneity} 0.05; otherwise, a fixed-effects model was used. We first calculated the pooled RRs and 95% CIs for median versus low levels of LDL-C. We defined the median group as follows: if the trend in RRs was a line with groups of LDL-C concentration, the middle LDL-C concentration group was considered the median group; if the trend in RRs was a curve, the group with the minimum RR was considered the median group. The low LDL-C concentration group, also defined as the reference group, was the lowest LDL-C concentration group in each study. Subgroup analysis was stratified by reference groups, age, proportion of sex, and follow-up time to find other potential factors that affect the relationship between LDL-C concentration and the risk of HS. Then, we focused on whether there was a difference in RRs among populations with and without hypertension. Hypertension was defined as SBP $\ge$140 mm Hg and/or diastolic blood pressure (DBP) $\ge$90 mm Hg or current antihypertensive treatment. The hypertension group was defined as the proportion of hypertensive participants over 60%, while the non-hypertension group was defined as the proportion of hypertensive participants less than 40%. When we pooled the RRs, we also conducted a sensitivity analysis in which we calculated pooled RRs for studies with proportions of hypertensive participants over 60%, 65%, and 75%, respectively. Meta-regression was also conducted to eliminate the effect of potential factors on the relationship between LDL-C concentration and risk of HS, such as age, sex, LDL-C concentration of reference group, and follow-up time. We then investigated the shape of the relationship between LDL-C and HS in a dose-response analysis using a two-stage generalized least squares model for trend estimation (GLST) model, using restricted cubic splines with three knots at flexible percentiles [16, 17] to obtain an optimal fitted curve. For each study, we extracted the mean or median of the categories of LDL-C. If neither the mean nor the median was reported, we calculated the midpoint of each category. For studies with an open-ended highest or lowest LDL-C category, we assumed that the interval was the same as that of the nearest category. We also extracted the number of cases, sample numbers, RRs, and 95% CIs for each category of LDL-C [18]. If the reference category was not the lowest LDL-C, we used the method described by Orsini to translate it to the lowest group [19]. If the unit of LDL-C was mmol/L, the unit was converted to mg/dL by multiplying by 38.67 [20, 21]. Considering studies using different lowest LDL-C concentrations as reference are unsuitable for the GLST model [22], we divided the original studies into three groups according to the LDL-C concentration of the reference group 50 mg/dL, 70 mg/dL, and 100 mg/dL. Finally, we defined the group included LDL-C level with the lowest pooled RR as the optimal group and compared the lower group adjacent to the optimal group and the optimal group to find the effect of very low LDL-C level on the risk of HS.

Heterogeneity was mainly assessed using the I² statistic. We considered low, moderate, and high I² values of more than 25%, 50%, and 75%, respectively [23]. Potential publication bias was visualized using a funnel plot and estimated using Egger’s and Begg’s tests. Sensitivity analysis was performed by removing one study at a time and then evaluating the remaining studies. All statistical analyses were performed with Stata 16.0 (StataCorp LLC, College Station, TX, USA). A threshold of p 0.1 was used to determine whether heterogeneity or publication bias was present [15]. Otherwise, p values were two-sided, with a significance level of 0.05.

Fig. 1 shows the results of the study selection process. The initial search yielded 1574 studies from PubMed, 1840 from the Cochrane Library, and 1770 from Embase. After excluding duplicates, non-original articles, and irrelevant studies, 178 potentially eligible studies were screened. We excluded studies without information on the studied variables and ultimately included 16 studies in the final meta-analysis [9, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38]. A manual search of the reference lists of these studies did not yield any new eligible studies.

Fig. 1.

A flowchart of the selection process in the study. RR, risk ratio; HR, hazard ratio; LDL-C, lowering low-density lipoprotein cholesterol; LDL, lowering low-density lipoprotein.

Sixteen studies were included, with 7 RCTs [25, 26, 27, 29, 31, 35, 37] and 9 PCs [9, 24, 28, 30, 32, 33, 34, 36, 38] (Supplementary Table 1), involving 304,763 participants with 2125 (0.70%) HS events. Nine studies ascertained HS as the endpoint [26, 27, 29, 31, 32, 33, 35, 36, 38], and seven studies ascertained ICH as the endpoint [9, 24, 25, 29, 33, 36, 38]. Seven studies were conducted in Asia [9, 24, 28, 29, 30, 34, 38], two in Europe [32, 33], three in North America [26, 36, 37], and four in countries on more than two continents [25, 27, 31, 35]. One study only included women [26], whereas all others included men and women. Only one study [37] focused on recurrence; the remaining studies defined the first occurrence as the outcome. Twelve studies [9, 24, 26, 28, 29, 30, 32, 33, 34, 35, 36, 38] estimated the relationship between HS and LDL-C measured at baseline; four studies [25, 27, 31, 37] used the LDL-C value after taking lipid-lowering medicine. Seven studies [25, 28, 29, 30, 31, 33, 35] reported RRs among hypertensive participants or the study population comprising more than 60% of participants with hypertension. Six studies [9, 24, 30, 32, 34, 35] adjusted the RRs with at least one other type of lipid, including high-density lipoprotein (HDL), total cholesterol, or triglyceride (TG).

PCs were assessed using the NOS (Supplementary Table 2), and RCTs were evaluated via the Cochrane Collaboration’s risk of bias tool (Supplementary Fig. 1). The scores of the nine PCs were seven or more; thus, all studies were considered high quality. For the RCTs, 100% had a low risk of reporting bias, and 71% had a low risk of attrition bias, which are important risks for our analysis. Furthermore, 85% of the RCTs had a low risk of performance bias and detection bias, and most RCTs had an unclear risk of selection bias and other biases.

We pooled 16 studies that provided the RR between categories of LDL-C levels (median versus low) and the risk of HS and found a significant relationship between LDL-C and the risk of HS (pooled RR = 0.81, 95% CI: 0.69–0.97, p = 0.020; Fig. 2), with low heterogeneity (I² = 28.4%, p = 0.138; Fig. 2). The funnel plot (Supplementary Fig. 2), Egger’s test (p = 0.938), and Begg’s test (p = 0.893) showed no significant publication bias. The sensitivity analysis results suggested that the pooled RR was not influenced by any single study (RR range: 0.76–0.86; Supplementary Fig. 3).

Fig. 2.

Forest plots of LDL-C and risk of HS. CI, confidence interval; IV, inverse variance method.

We also conducted subgroup analysis stratified by reference groups, age, proportion of sex, and follow-up time to find other potential factors affecting the relationship between LDL-C concentration and the risk of HS. The results presented a significant relationship between LDL-C concentration and risk of HS when the reference group was below 70 mg/dL (RR = 0.63, 95% CI: 0.46–0.87, p = 0.005; Fig. 3A), when age was below 60 years old (RR = 0.79, 95% CI: 0.65–0.95, p = 0.013; Fig. 3B), when the proportion of females was less than males (RR = 0.72, 95% CI: 0.54–0.96, p = 0.024; Fig. 3C), and when follow-up time was more than 10 years (RR = 0.63, 95% CI: 0.48–0.82, p = 0.001; Fig. 3D).

Fig. 3.

Subgroup analysis of LDL-C concentration and risk of HS. (A) Reference groups of LDL-C concentration stratified subgroup analysis. (B) Subgroup analysis stratified by age. (C) Subgroup analysis stratified by proportion of sex. (D) Subgroup analysis stratified by follow-up time.

The results for proportions of the study population with hypertension of 60%, 65%, and 75% were consistent. The pooled RRs for the risk of HS between the median and low LDL-C level groups in the hypertensive population were 0.84 (95% CI: 0.53–1.34; Fig. 4A), 0.68 (95% CI: 0.39–1.18; Fig. 4B), and 0.53 (95% CI: 0.27–1.06; Fig. 4C) for the above three proportions, respectively; the result of pooled RR among non-hypertensive populations was 0.49 (95% CI: 0.33–0.75; Fig. 4A–C).

Fig. 4.

Forest plots of LDL-C and risk of HS stratified by blood pressure status. (A) Proportion of hypertension: 60%. (B) Proportion of hypertension: 65%. (C) Proportion of hypertension: 75%. DL, DerSimonian-Laird.

We also conducted a meta-regression analysis to eliminate the effect of potential factors on the relationship between LDL-C concentration and the risk of HS. The results showed that hypertension remained an independent factor affecting this relationship (RR = 1.486, 95% CI: 1.039–2.216, p = 0.033; Table 1).

To obtain actual, precise dose-response relationships, we divided studies with hypertensive populations with reference LDL-C levels in the original studies into three groups: 50 mg/dL, 70 mg/dL, and 100 mg/dL, respectively. We detected a significant non-linear relationship and linear relationship (except reference group 100 mg/dL) between LDL-C and HS among the hypertensive populations of the above three groups (p for non-linear relationship: 0.016, 0.041, 0.002; p for linear relationship: 0.005, 0.019, 0.854, respectively) and for non-hypertensive populations (p for non-linear relationship 0.0001; p for linear relationship 0.019).

The non-linear trend among hypertensive populations suggested that an LDL-C level of approximately 80 mg/dL could be associated with the lowest risk of HS (Fig. 5B); the risk of HS rose as the LDL-C level increased (Fig. 5C). The risk of HS also noticeably rose when the LDL-C level was very low, below 60 mg/dL (Fig. 5A).

Fig. 5.

Relative risk (solid line) with 95% CI (long dashed lines) for the association of LDL-C with the risk of HS in a hypertensive population. (A) Hypertensive population: The reference group was LDL-C 50 mg/dL. (B) Hypertensive population: The reference group was LDL-C 70 mg/dL. (C) Hypertensive population: The reference group was LDL-C 100 mg/dL. (D) Non-hypertensive population: The reference group was LDL-C 70 mg/dL.

We detected that the lowest risk of HS among non-hypertensive populations was an LDL-C level of approximately 115 mg/dL. In comparison, the lowest risk of HS among hypertensive populations was in an LDL-C level range of 60–80 mg/dL. In addition, the degree of risk reduction of HS was greater; the effective lowering LDL-C range (95% CI of RR did not conclude 1) was wider among non-hypertensive populations than among hypertensive populations (Fig. 5B,D).

Finally, we calculated RRs to compare the lower group adjacent to the optimal group and the optimal group among hypertensive populations (50–69 mg/dL vs. 70–99 mg/dL) and non-hypertensive populations (70–99 mg/dL vs. 100–129 mg/dL). The results showed that when the proportion of the study population with hypertension was more than 65%, the RRs were significant (RR = 1.84, 95% CI: 1.36–2.50; RR = 1.94, 95% CI: 1.41–2.66, respectively; Fig. 6A,B); the RRs were not significant among non-hypertensive populations (RR = 1.15, 95% CI: 0.97–1.36; RR = 1.15, 95% CI: 0.97–1.36, respectively; Fig. 6A,B).

Fig. 6.

Forest plots of LDL-C and risk of HS stratified by blood pressure status in lower versus optimal. (A) Proportion of hypertension: 65%. (B) Proportion of hypertension: 75%.