We conducted a retrospective propensity score matched cohort study using the Medical Information Mart for Intensive Care-IV (MIMIC-IV), a large, freely available, de-identified, comprehensive database that includes patients admitted to the BIDMC ICUs from 2008 to 2019 (22). The database contains non-identifiable bedside health data, including demographics, vital signs, laboratory data, prescriptions, fluid balance, caregivers notes, procedural and diagnostic codes (22). A member of our team (LCF) passed the Examination of Protection of Human Research Participants and was granted access to the database (record ID: 33047414). This study was conducted and reported following the STrengthening the Reporting of OBservational Studies in Epidemiology (STROBE) statement (23).

All consecutive patients were considered for inclusion. The inclusion criteria were as follows: (1) Diagnosed with sepsis (Sepsis-3)(1) upon hospital admission. (2) Aged 18 years or older. (3) For patients with multiple sepsis episodes and ICU stay records, only the first sepsis episode was evaluated. Patients with an ICU stay of less than 24 hours were excluded from the study.

Medication prescriptions were identified from the prescription drug file based on both generic and brand names. The medication exposure was defined simply as any statin use or no statin use during ICU stay, regardless of the statin type. The primary outcome was 28-day all-cause mortality. The secondary outcomes were ICU mortality, in-hospital mortality, length of ICU stay, length of hospital stay, duration of mechanical ventilation (MV) and continuous renal replacement therapy (CRRT). To assess the impact of statin use on clinical outcomes, eligible patients were allocated into either the statin group or the no statin group based on whether they received statins or not during their ICU stay.

All data were extracted from the MIMIC-IV database using Structured Query Language (SQL). The SQL script codes for data extraction were available on GitHub (https://github.com/MIT-LCP/mimic-iv). The following data were collected: demographics, including age, gender, race and body mass index (BMI); vital signs, including temperature, heart rate, respiratory rate (RR) and mean blood pressure (MBP); comorbidities, including cerebrovascular disease, congestive heart failure, chronic pulmonary disease, renal disease, severe liver disease, cancer and diabetes; severity scores, including acute physiology score III (APS III), charlson comorbidity index (CCI), glasgow coma scale (GCS), logistic organ dysfunction system (LODS), oxford acute severity of illness score (OASIS) and sequential organ failure assessment (SOFA); laboratory tests, including serum vitamin D, hemoglobin, white blood cells (WBC), platelets, creatinine, blood urea nitrogen (BUN), alanine aminotransferase (ALT), aspartate aminotransferase (AST), total bilirubin, glucose, potential of hydrogen (pH), partial pressure of oxygen (pO2), partial pressure of carbon dioxide (pCO2), partial pressure of arterial oxygen to fraction of inspired oxygen ratio (PaO2/FiO2 ratio), base excess, lactate, sodium, potassium, calcium, chloride, anion gap and international normalized ratio (INR); clinical measures, including first-day vasopressor, antibiotic lag, duration of MV and duration of CRRT. Comorbidities were assessed upon admission. Initial vital signs and clinical indices acquired within 24 hours of ICU admission were used as baseline characteristics. Variables with missing values of more than 50% were excluded from the analysis, while those with less than 50% were included. The missing rate for each variable is presented in the Supplementary Material : Supplementary Table S1 , Supplementary Figure S1 . Missing values of the included variables were imputed using the missForest method to decrease bias and avoid participant exclusion (24).

No sample size calculation was performed as this was a retrospective exploratory study (25), and the sample size was determined by the number of patients available in the MIMIC-IV database (over 364,627 patients). Continuous variables were presented as mean (standard deviation [SD]) or median (interquartile range [IQR]) and were analyzed using either the Student’s t-test or the Mann-Whitney U-test depending on their distribution (26). Categorical variables were expressed as numbers (percentages) and compared using either the chi-square test or Fisher’s exact test (27). For the primary outcome, the Cox proportional hazards model was employed to estimate the hazard ratio (HR) and 95% confidence interval (CI). The Kaplan-Meier method and the log-rank test were used to calculate and compare the cumulative incidence of 28-day all-cause mortality. For dichotomous secondary outcomes, the logistic regression model was applied to compute the odds ratio (OR) and 95% CI. The Hodgese-Lehmann method was utilized to determine the median difference (MD) and 95% CI for continuous secondary outcomes. Multicollinearity between variables was assessed using the variance inflation factor (VIF), with VIF values of less than 5 indicating no multicollinearity ( Supplementary Material : Supplementary Table S2 , Supplementary Figure S2 ; Supplementary Table S3 , Supplementary Figure S3 ). A two-tailed p < 0.05 was considered statistically significant for all analyses. All statistical analyses were performed using R software (version 4.2.3; R Foundation for Statistical Computing, Vienna, Austria).

To address potential confounding factors and selection bias inherent in observational studies, we performed PSM following the methodological guidelines proposed by Lonjon and colleagues (28). According to a consensus statement (29), the following variables were included in the propensity score model for matching: age, gender, congestive heart failure, cerebrovascular disease, diabetes, malignant cancer, severe liver disease, APS III, CCI, heart rate, first care unit, ALT, total bilirubin, base excess, calcium, anion gap and INR. The propensity score, which represents the predicted probability of receiving statins, was calculated using baseline covariates in a logistic regression model. Patients were matched using the 1:1 nearest neighbor method without replacement and with a caliper width of 0.05. After PSM, a matched cohort of patients with comparable baseline characteristics was assembled. The covariate balance between groups was evaluated using standardized mean differences (SMD) before and after matching, with a SMD < 0.1 indicating negligible differences (30). Furthermore, Stepwise Cox regression analyses were employed to build adjusted models while adequately considering possible confounders in the matched cohort. Variables with a p-value < 0.1 in the univariate analysis were selected for further stepwise multivariate analysis. The independent variables included in the final model were age, gender, race, BMI, APS-III, CCI, LODS, OASIS, SOFA, GCS, respiratory rate, temperature, hemoglobin, WBC, creatinine, ALT, total bilirubin, pH, pCO2, lactate, calcium, potassium, anion gap, INR, antibiotic lag, first-day vasopressor and statin use ( Supplementary Material : Supplementary Table S4 ).

To evaluate the impact of different variables on 28-day all-cause mortality in patients with sepsis, we conducted subgroup analyses in the matched cohort based on the following variables: age (>60 versus <=60 years), gender (female versus male), race (white, black, unknown, other), BMI (obesity, overweight, normal, underweight) and CCI (<6 versus >=6).

To validate the robustness of the findings in the matched cohort, sensitivity analyses were conducted in the unmatched cohort. Stepwise Cox regression analyses were employed to identify independent prognostic factors and adjust for potential confounders. Variables with a p-value < 0.1 in the univariable analysis were entered into the multivariable analysis by stepwise selection. The independent variables incorporated into the final model were: age, gender, race, BMI, APS-III, CCI, LODS, OASIS, SOFA, GCS, respiratory rate, temperature, hemoglobin, WBC, creatinine, ALT, total bilirubin, pH, lactate, Sodium, potassium, chloride, anion gap, INR, antibiotic lag, first-day vasopressor and statin use ( Supplementary Material : Supplementary Table S5 ).

A total of 30133 adult patients with sepsis were identified from the database. After excluding ineligible records, 20230 patients were included in the unmatched cohort, with 8972 (44.34%) in statin group and 11258 (55.66%) in no statin group. After PSM, 12140 patients were included in the matched cohort, with 6070 in the statin group and 6070 in the no statin group. The process of patient selection is illustrated in Figure 1 .

In the unmatched cohort, patients who received statin tended to be older and more likely to be male, exhibited lower APS-III and OASIS, as well as a shorter antibiotic lag. The baseline characteristics of both the unmatched and matched cohorts are shown in Table 1 . After PSM, all variables were well-balanced in the matched cohort (SMD < 0.10) ( Figure 2 ). The distribution of propensity scores of the two groups before and after matching are depicted in Figure 3 .

In the unmatched cohort, approximately 44.34% (8972/20230) patients received statins, while in the matched cohort, approximately 50% (6070/12140) patients received statins. Various forms of statins were used during ICU stay, including atorvastatin, fluvastatin, lovastatin, pitavastatin, pravastatin, rosuvastatin, simvastatin and other statins. Clinical indications for the initiation and discontinuation of statins were not available in the database.

While randomized controlled trials are widely regarded as the gold standard of evidence-based medicine, conducting prospective randomized controlled trials to assess the effect of statin use on sepsis prognosis is challenging due to the large number of patients required to achieve a sufficient cohort of patients who actually develop sepsis. We believe that the best alternative to a prospective randomized controlled trial is exactly what we have done: identify a cohort, follow them over time, even if not concurrently, and match cases to controls by propensity matching on important clinical characteristics.

To account for the selection bias and unmeasured confounders inherent in observational studies, we employed the PSM approach (32) to ensure that all patients were pseudo-randomized to the treatment and control groups as in a typical RCT. PSM enables the generation of an unbiased average treatment effect of statin on clinical outcomes among patients with sepsis admitted to the ICU. PSM allows simultaneous modeling of the propensity for unbiased group allocation and modeling of the outcomes using multivariate regression adjustment, thereby obtaining double robust and unbiased estimates of the average treatment effect of statins (32).

After adjusting for various biases inherent in observational studies using PSM, we observed a beneficial effect of statin use on the outcome of sepsis, which is contrary to the findings of several RCTs. Though the methodological differences between RCTs and observational studies are frequently cited as the primary source of such discrepancy, our study employed a pseudo-randomized quasi-experimental approach that successfully adjusted for selection biases and supported the results of most observational studies (16, 33, 34).

Our findings challenge the previous assertion made by Majumda that a healthy user effect explains why observational studies demonstrated the beneficial effects of statins on sepsis patients (35). Because the use of the PSM approach allows individuals to be assigned randomly to different groups, thus eliminating the possibility of a healthy user effect (20).

There are several possible reasons why most RCTs failed to detect a beneficial effect of statins in patients with sepsis. A comprehensive review of these RCTs revealed that sepsis diagnoses were often underreported, and many trials could not provide additional data upon request, increasing the risk that a non-representative sample of statin-treated patients was enrolled and assessed for sepsis outcomes (20). It is noteworthy that the PSM approach used in this study is not inherently superior to large-scale RCTs with complete data reporting, but it helps mitigate biases in observational studies and address noncompliance issues in RCTs.

Subgroup analyses revealed consistent beneficial effects of statin therapy in sepsis patients irrespective of pre-existing cerebrovascular diseases and chronic heart failure. Notably, while the plaque-stabilizing properties of statins constitute the primary mechanism underlying their cardiovascular protective effects, this observed sepsis-associated mortality reduction in both subgroups suggest potential pleiotropic mechanisms independent of atherosclerotic plaque modulation.

The pleiotropic effects of statins have been well documented in the literature. However, despite this, the underlying mechanism by which statins confer benefit in sepsis remains unclear (36). Potential explanations for this beneficial effect include: First, statins may attenuate the severity of sepsis through their anti-inflammatory, immunomodulatory, antioxidative, and antithrombotic effects (37–40). In animal models of sepsis, statins have been demonstrated to inhibit the elevation of inflammatory mediators (41, 42), resulting in improved survival rates (43, 44). Previous clinical studies have shown that statins may exert potential antioxidant properties in models of sepsis, which could help mitigate tissue damage and organ dysfunction (45). Statins have been reported to inhibit the expression of toll-like receptors (TLR) 4 and 2 on monocytes in human endotoxemia models, leading to a decrease in inflammatory cytokine production (45). Statins may interfere with transcription factors such as nuclear factor kappaB (NF-kappaB) and activation protein-1 (AP-1), which could result in a reduction in the synthesis of proinflammatory cytokines, including interleukin-1 (IL-1) and IL-6 (46). Similarly, an association between statin treatment and reduced levels of tumor necrosis factor (TNF) and IL-6 has been observed in patients experiencing acute bacterial infections (47). Statins have been demonstrated to inhibit adhesion molecules in both neutrophils/monocytes and endothelial cells, resulting in a decreased migration of polynuclear neutrophils into tissues (48–50). Statins may assist in restoring the balance between endothelial nitric oxide synthase (eNOS) and inducible nitric oxide synthase (iNOS), which is disrupted in sepsis (51). By substantially boosting eNOS expression while downregulating iNOS, statins have the potential to prevent or reverse sepsis-related endothelial dysfunction (51). Furthermore, statins may play a crucial role in mitigating the negative effects of sepsis on the coagulation system by inhibiting the expression of tissue factor and plasminogen activator inhibitor-1, improving protein C function (52), lowering prothrombin fragment levels, and significantly upregulating the expression of thrombomodulin (53, 54). Second, Statin use was associated with a lower risk of bacterial infection. Statins may have direct antimicrobial properties (37, 55), as the enzymes in the mevalonate pathway, which are potentially modified by statin therapy, are also involved in the development of Gram-positive bacterial infections (55). It is noteworthy that statins may also exhibit antifungal properties due to the similarities between the ergosterol biosynthetic pathway in fungi and the cholesterol synthesis in humans, implying a direct effect on Candida species (56, 57). The immunomodulatory, antioxidative, anti-inflammatory, antithrombotic, and direct antimicrobial effects of statins may account for the beneficial effects against sepsis observed in our study.

The main strength of this study lies in the utilization of the PSM analytical approach, which allows for the generation of doubly robust unbiased estimates of the average treatment effects of statins in patients with sepsis. However, this study also has several limitations. First, the observational design inherently precludes definitive causal inferences, as unmeasured confounding factors may influence the observed associations despite our rigorous propensity score matching and multivariable adjustment approaches. Second, the study may be subject to potential residual confounders that are not recorded in the MIMIC-IV database. Although PSM is a robust method for addressing multiple baseline differences between groups, variables included in this study are confined to relevant variables available in the MIMIC-IV database, potentially introducing bias from unmeasured confounders. Third, the study did not identify the specific effects of individual statins on sepsis. In this study, the exposure was simply defined as either the use of any statin or no statin during the ICU stay. Previous studies have demonstrated that simvastatin, atorvastatin and rosuvastatin exhibit antibacterial properties, while other statins do not (15, 58). As a result, studies conducted without distinguishing the effects of different statins are prone to underestimate their effects, and future studies should be conducted to compare the clinical outcomes associated with individual statins. Fourth, the impact of prior statin use on clinical outcomes was not investigated. The study focused only on statin use during the ICU stay. However, pretreatment with simvastatin has been demonstrated to improve sepsis survival in mouse models by preserving cardiac function, lowering circulatory inflammatory cytokines, decreasing neutrophil migration to the lung, and enhancing T-cell function (42, 43, 59). Therefore, studies conducted without considering the impact of prior statin exposure may overestimate the beneficial effects of statins.