Statin Use Is Associated With Reduced Mortality On Mechanically Ventilated Patients : A Retrospective Propensity-Matched Analysis Of MIMIC-III Database

Background: The patients on mechanical ventilation were always critically ill and ventilation may cause injury and inammatory disorders in the lungs of such patients. We sought to determine whether statin treatment has a protective effect on the outcome of the ventilated patients. Methods: A retrospective observational study of mechanically ventilated patients from the Medical Information Mart for Intensive Care III (MIMIC-III) database. To ensure the robustness of the ndings, we utilized the gradient boosted model, propensity score analysis, doubly robust estimation and an inverse probability ‐ weighting model in the statistical procedure. The propensity score of statin use was calculated by age, sex, severity scores and 32 other variables. Results: Compared with nonusers, the use of statins was associated with improved 28-day survival in the unmatched cohort (HR 0.85 95% CI 0.80~0.90) and matched cohort (HR 0.79 95% CI 0.73~0.85). Statin use was also associated with improved 60-day (HR 0.81 95%CI 0.75~0.85), 90-day (HR 0.81 95%CI 0.76~0.86) and in-hospital (HR 0.81 95%CI 0.76~0.86) survival in the matched cohort. Statins use was associated with longer ventilator free days (VFD28, 14.93±13.11 vs 12.06±13.26) and longer ICU free days (13.41±12.14 vs 10.86 ±12.19) in the matched cohort. Different types of statins were associated with improved 28-day survival. The subgroup analyses results showed improved 28-day survival in most subgroups, but this improvement was not observed in patients with pneumonia, septicemia or acute respiratory failure. Conclusions: In a population of mechanically ventilated patients, the use of statins may be associated with reduced mortality. This protective effect was observed only in mild-to-moderate patients instead of critical inammatory ones.

The differences in outcomes between the studies might be attributed to the characteristics of different ICU populations. One recent retrospective study enrolled all the ICU patients who had been prescribed statins and compared the outcome with a matching cohort. The result showed a bene cial association of statin use and 90-day mortality improvement [21]. However, the ICU population in this study may have been healthier or more self-limiting disease processes than other ICU populations, such as patients with sepsis or acute lung injury/ARDS; this could have resulted in the relatively lower 90-day mortality observed in this study compared to that reported in other studies. To avoid such patient selection bias and to, moreover, research on the mortality bene t of statin use in the critically ill, we enrolled ventilated patients who had taken statin before or during ventilation as our study population. Ventilation is a key life-saving treatment measure for critically ill patients, including patients with acute respiratory distress syndrome (ARDS), trauma, shock and other life-threatening conditions. Epidemiological data have shown that there were an estimated 790,257 hospitalizations involving mechanical ventilation in 2005 in the USA, representing 2.7 cases of mechanical ventilation per 1000 people. The estimated national associated cost was $27 billion, representing 12% of hospital costs, which accounts for a large amount of resources in the critical care department [22]. Furthermore, ventilation can cause lung injury and lead to ventilator-associated pneumonia (VAP) and other severe complications, which may increase the mortality of critically ill patients [23]. Lung injury remains one of the major complications of mechanical ventilation in the intensive care unit (ICU).
Whether statin use is associated with a lower mortality in critically ill patients requiring ventilation in the intensive care unit (ICU) remains unclear. Therefore, we designed this observational study to research the potential bene cial effect of statin use among critically ill ventilated patients.

Study design and data source
This is a retrospective observational study. We analyzed data from a large database: Medical Information Mart for Intensive Care (MIMIC-III). The MIMIC-III database is an openly available dataset developed by the MIT Laboratory for Computational Physiology, comprising deidenti ed health data associated with nearly 54,000 intensive care unit admissions [24]. The data in the MIMIC-III database are composed of comprehensive clinical datasets from patients admitted to the ICUs of Beth Israel Deaconess Medical Center in Boston, MA, from June 1, 2001 to October 31, 2012. The requirement for institutional review board (IRB) approval from our institution was exempted because MIMIC-III is a third-party anonymized publicly available database with pre-existing IRB approval.

Participants
Patients who underwent mechanical ventilation were selected from the MIMIC-III database. The ventilation data were extracted from the chartvcents table. Statin usage information was extracted from the prescription table. We selected those who had taken statins before or during ventilation as the statin cohort and those who underwent ventilation without statins as the control cohort. Those who took statin medicine after extubation were excluded from this study. We included only adults in this study, so those under 18 years old were excluded. A total of 3999 patients were selected for the statin cohort. Then, we performed a propensity matching by age, sex, Simpli ed Acute Physiology Score II (SAPSII), and 32 other variables. The detail propensity score calculation was explained in the statistics part later. Each statinexposed patient was matched with the closest corresponding non-exposed patient (that is, a patient who was not exposed to statins) at a 1:1 xed ratio (nearest match cohort). Finally, 3363 patients were matched and included in each cohort.

Medication exposure
Patients who had taken statins before or during ventilation were selected as the statin cohort, and those who underwent ventilation without statins were selected as the control cohort. Patients taking statin medicine after extubation were excluded from this study. Atorvastatin, pravastatin, rosuvastatin and simvastatin were the 4 most common statin types. A total of 187 patients received atorvastatin plus simvastatin.

Outcomes
The primary outcome was 28-day and in-hospital all-cause mortality. The primary statistical method of comparison for the time-to-event end points was expressed by the Kaplan-Meier curves and tested by the log-rank test. A Cox proportional hazards model was used to estimate the hazard ratio (HR) of 28-day mortality and its associated 95% con dence interval (CI). We included several variables in the model to adjust the 28-day survival (age, SAPSII score, sex, liver disease, diabetes, obesity, hypertension, etc.). The APACHE III score was not recorded for every patient, so we could not extract all APACHE III scores. The SAPSII was chosen to represent the severity of illness. The secondary outcome analyses also included ICU free days in 28-day, ventilator free days in 28 days, 60-day survival, 90-day survival and in hospital survival. Eight plasma biomarkers were extracted and studied to measure the host responses during the rst 10 days after intubation. The HRs of each type of statins were studied by COX models. The effect of statin in different subgroup population were analyzed in the end.

Statistical methods
The doubly robust estimation method was applied to infer the sensitivity analysis of the primary outcome. "Doubly robust estimation combines a multivariate regression model with a propensity score model to estimate the association and causal effect of an exposure on an outcome" [25,26]. Usually, the regression model or the propensity score model was applied individually to estimate a causal effect.
When the two approaches were built in one estimation model, only one of the two models needs to be correctly speci ed to obtain an unbiased effect estimator, thus the term doubly robust analysis. For propensity scores of the statin use, a machine learning algorithm named gradient boosted model (GBM) was employed to maximally correlated with the negative gradient of the prede ned loss function. Regression tree was used and a total of 35 covariates were used in the model so that covariate imbalance between the statin and non-statin groups was minimized. Using the estimated propensity scores calculated by GBM model as weights, an inverse probability weighting (IPW) model was used to generate a weighted cohort [26,27]. An IPW weighted COX regression was then built adjusting for the variables that remained unbalanced between the groups. The propensity score was also employed to build the 1:1 matching cohort of the non-statin cohort.
Some studies have reported different potencies between statins; for example, simvastatin exerted better antibacterial effects than rosuvastatin, and the latter was found to have a more potent lipid-lowering capacity [28,29]. Therefore, we analyzed the difference in 28-day survival among statin types in the Cox regression model. To test the e ciency of statins on patients with different pro les, we performed a subgroup analysis with age and SAPSII as covariates. Patient categorical data are presented as percentages, and continuous data are listed as means with standard deviations (SDs). We used Student's t tests for continuous variables and chi-square or Fisher's exact tests for dichotomous variables.

Baseline results
All 53432 cases in the MIMIC-III database were screened. A total of 24,769 individuals who had undergone ventilation met the inclusion criteria. A total of 17065 patients who had not received statin treatment were included in the non-statin cohort. We constructed a propensity score model by employing the 39 covariates with the GBM. The contributions of individual covariates to the nal propensity score are illustrated in (Supplement Fig. 1). The top covariates include diagnosis, age, history of hypertension, presence of CHF and peripheral vascular disease: such factors would commonly in uence the decision regarding whether to prescribe statin. Based on the estimated propensity scores, IPW was applied to standardize the differences between the statin and non-statin cohorts. As shown in Table 1, most of the covariates of the weighted cohorts were balanced between the two cohorts with or without statin. There were 3999 patients who received statin treatment before or during the ventilation time. After matching, there were 3363 patients in each cohort (Table 1). They were similarities in age, sex, SAPSII and 36 more variables. The characteristics of the patients are presented in Table 1

The primary outcome and doubly robust analysis
The Kaplan-Meier analysis showed that statin users had a better 28-day and in hospital survival curve than the non-users in both unmatched and matched cohorts (Fig. 1). The univariate Cox model showed that the use of statins before or during ventilation was associated with an improved 28-day survival in both the overall cohort (HR 0.85 95% CI 0.80 ~ 0.90 Fig. 1A) and the matched cohort (HR 0.79 95% CI 0.73 ~ 0.85 Fig. 1B). In the multivariate Cox model, the use of statins was associated with a bene cial effect on 28-day survival in the unmatched cohort (HR 0.73 95% CI 0.68 ~ 0.77 Table 2) and the matched cohort (HR 0.73 95% CI 0.68 ~ 0.78 Table 2). Under the doubly robust estimation framework, an univariate and a multivariate regression model was developed to adjust for these unbalanced covariates on the weighted cohort (   (Table 3). Regarding the biomarkers, we found no large differences in the matched cohort. Additionally, we found that statin use was not associated with higher aspartate aminotransferase (AST) or alanine aminotransferase (ALT) levels after matching (Fig. 2). The WBC and neutrophils were signi cantly lower in the statin cohorts. However, the creatinine was slightly higher in the statin group.

Sensitivity and subgroup analysis
We further analyzed the different statins and their effect on patients' 28-day survival. All the different kinds of statins were found to be associated improved 28-day survival of the patients.  (Fig. 3A).
In the subgroup analysis, we found that statins had a protective effect in most of the subgroups but not in individuals with obesity. More importantly, statin use showed no effect in in ammation-related groups, such as the septicemia, respiratory failure and pneumonia subgroups (Fig. 3B). In the current study, patients with coronary disease accounted for a large proportion of the sample. However, in the noncoronary artery disease subgroup, we also found that statin use was associated with reduced mortality.

Discussion
This study revealed that statin use was associated with improved 28-day survival and in-hospital survival. All kinds of statins showed reduced HRs. This evidence might indicate the protective effect of statins in ventilation patients. In the subgroup analyses, we found that statins had a protective effect in most of the subgroups but not in the septicemia, acute respiratory failure or pneumonia subgroups. All these results might explain the protective effect of statins.
In addition to lowering cholesterol, statins exert pleiotropic effects [3][4][5][6] such as anti-in ammatory, antioxidant, and immunomodulatory effects, especially in the context of pulmonary disorders [30]. Statins may reduce COPD exacerbation [31]. Statins have also been suggested to be effective in patients with acute lung injury or acute respiratory distress syndrome in some observational studies [32,33].
These clinical effects may be mediated by a reduction in pulmonary and systemic in ammation.
Simvastatin decreased bronchoalveolar lavage IL-8 by 2.5-fold (P = 0.04) [34]. Statins also showed a protective effect against sepsis. When compared with nonusers, simvastatin (HR, 0.72; 95% CI, 0.58-0.90) and atorvastatin (HR, 0.78; 95% CI, 0.68-0.90) users had improved 30-day survival [10]. The current study examined the anti-in ammatory effect of statins from a ventilation patient cohort, which has not been previously reported. A recent study showed that infections in older adults were associated with prolonged, impaired neutrophil migration. Simvastatin improves neutrophil migration in vivo in healthy individuals and in vitro in milder infective events but not in severe sepsis, supporting its potential utility as an early intervention during pulmonary infections [35,36]. Lung injury remains one of the major complications of mechanical ventilation in the ICU. This injury could result from an altered host immune response after mechanical stretch [37]. Using statin therapy to protect patients with lung injury could therefore be a reasonable strategy, as these drugs could abate the host in ammatory response to infection [13], especially within the lungs [3].
Animal models may have further explained the protective effect of statins against lung injury. Statins increase glucocorticoid receptor expression in alveolar macrophages and downregulate NF-κB activation associated with the increased number of alveolar macrophages [12]. Two other animal models of lung injury induced by mechanical ventilation also support these ndings [38,39]. The protective effect may be due to the anti-in ammatory effect of statins. Prior statin use was associated with a lower baseline IL-6 concentration, and continuation of atorvastatin treatment in this cohort was associated with improved survival [40]. Statins have been shown to reduce vascular leakage and in ammation in animal models of lung injury [41]. Statins may also attenuate lung injury by downregulating the expression of in ammatory cytokines [42,43]. Our previous study revealed the lung-protective effect of statins caused by the reduction of in ammatory cell in ltration [44]. Additionally, statins may have direct antibacterial effects and modulate bacterial virulence [45][46][47]. However, some randomized controlled studies have come to the opposite conclusion [14][15][16][17][18]. In a randomized controlled trial (RCT) with a follow-up of over one year, there was no signi cant difference in cumulative survival between the rosuvastatin and placebo groups (58% vs 61%; p = 0.377) [49]. Simvastatin therapy was not signi cantly associated with the difference between the study groups in mortality at 28 days (

Declarations
Ethic approval and consent to participate: The requirement for institutional review board (IRB) approval from our institution was exempted because MIMIC-III is a third-party anonymized publicly available database with pre-existing IRB approval. The consent to participate was waived because the retrospective design of the study.
Consent for publication: Not applicable.
Availability of data and material: The data can be available by contacting the corresponding author Jiang Du by email gowindj@163.com. The source codes for statistical analyses can be found at https://github.com/gowindj/vent_statin [50]. The source codes for postgreSQL can be found at https://github.com/gowindj/public [51].
Competing interest: All the authors declare that they have no potential nancial or ethical con icts of interest regarding the contents of this manuscript.  The plot of relative in uence factor of covariates. /The relative in uence factor measures how discriminative the 35 covariates of the propensity score model are when predicting the likelihood of statin prescription.

Figure 2
Kaplan-Meier curves for all-cause mortality in statin recipients vs. nonrecipients in the unmatched cohorts and propensity score-matched cohorts. / A: Statin use was associated with improved 28-day survival in the unmatched cohort (HR 0.85 95% CI 0.80~0.90); B: Statin use was associated with improved 28-day survival in the matched cohort (HR 0.79 95% CI 0.73~0.85).

Figure 4
The plots of subtype analysis and subgroup analysis. /A:The different types of statin use and survival; B: The forest plot of statins in the 28-day survival subgroup analyses.