Antimicrobial Resistance Patterns and Protective Effects of Statins in Bacteremic Patients

The significant increase in antimicrobial resistance over the past few years is a serious global public health concern, particularly as the development of new antimicrobial agents had been slow for many years. Infections with resistant organisms are associated with poor clinical outcomes and higher cost burdens. Determining antimicrobial resistance patterns can help identify problem areas and modify treatment practices to improve clinical outcomes. Additionally, identifying adjunctive therapies can also help improve clinical outcomes among infected patients. The objectives, hypotheses, methods and results of this dissertation are threefold: Manuscript 1: The objective was to analyze antimicrobial resistance trends in E. faecalis and E. faecium between 2003 and 2015 in five acute care facilities of the Veterans Affairs New England Healthcare System as antimicrobial resistance patterns among Enterococcus have changed over the past decade. Using a multicenter ecologic study design, we evaluated antimicrobial resistance patterns for blood and urine cultures of enterococci. In E. faecium urine cultures, a decline in gentamicin resistance, as well as a small decease in vancomycin resistance were observed. Enterococcus resistance towards ampicillin, linezolid, and tetracycline was stable over the study period. Daptomycin resistance did not emerge over the study period. Manuscript 2: The objective was to evaluate the impact of statin exposure on clinical outcomes, including inpatient mortality and length of inpatient stay, among bacteremic patients. The hypothesis was that statin use would be associated with positive clinical outcomes compared to non-statin use. We conducted a retrospective cohort study using the deidentified Optum Clinformatics (OptumInsight, Eden Prairie, MN) with matched Premier Hospital data (October 2009-March 2013). Our retrospective cohort study observed lower mortality for incident users and prevalent users continuing statin use during admission. Though non-significant in incident users, the point estimate was similar to that observed in other studies. Manuscript 3: The objective was to identify a statin therapy duration among pre-defined baseline statin users at which use of statins minimizes the risk of inpatient mortality among bacteremic patients. The hypothesis was that a certain minimum duration of statin use during the hospitalization would improve survival. A casecontrol design was used to test this hypothesis using the Optum Clinformatics with matched Premier Hospital data (October 2009-March 2013). Classification and regression tree analysis was conducted among cases and controls matched on disease risk scores. Among matched pairs of cases and controls with at least 90 days of pre-admission statin use, the continuation of statin use during admission for at least 2 days provided a better survival benefit among bacteremic patients.


Introduction
Antimicrobial resistance among Enterococcus has been increasing in the United States (U.S.) over the past several years. 1,2 These bacteria have developed resistance to nearly every antibiotic used for treatment. 2,3 Further, the occurrence of resistance in enterococci to comparatively new antibiotics, such as daptomycin 4 and tigecycline, 5 as well as developing resistance to adjunctive therapies, such as gentamicin, is a substantial public health concern. 2 Infections with resistant organisms are associated with poor clinical outcomes, 6 and increased health and cost burdens. 7 As antibiotic resistance changes, enterococcal infections are becoming more difficult to treat, potentially leading to the administration of inappropriate empiric therapies 8,9 that could increase mortality risk. 10 Vancomycin resistance in enterococci (VRE) has increased extensively in the past few decades, and is considered a "serious threat" as each year, 20,000 (or 30%) Enterococcus healthcare-associated infections are vancomycin-resistant causing an estimated 1,300 deaths in the U.S. 11 With the changing epidemiology of infections, 12,13 frequent review of resistance, with respect to historical patterns, is crucial to identifying problems with resistance and response to such changes, such as modifications to treatment practices. 14 This study aimed to describe antimicrobial resistance trends of E. faecalis and E. faecium in a large, regional healthcare system.

Methods
A multi-center ecologic study design was used to evaluate annual antibiotic Standards Institute (CLSI) guidance 15,16 . Changes in the percent resistance over the study period were assessed with generalized linear mixed models (GLMM). GLMM accounts for the clustered nature of the study design by incorporating correlations among responses through the inclusions of random effects in the linear predictor and/or by modeling the correlations among the data directly. 16 Generalized models are commonly used when response variables have error distribution models other than a normal distribution. 16 The general form of the model is: y The events/trial syntax (number resistant/number of isolates tested for a given antibiotic) served as the response variable and the 'year' functioned as the independent categorical variable. The binomial distribution and logit link were used for all GLMM models due to the event/trial syntax of a dependent variable. To account for the interdependence of samples within the five facilities (clusters), 'facility' was included as a random effect in the GLMM models. Significance was defined as an α (alpha) of 0.05 and all models were run in SAS 9.4 (SAS Institute, Cary, NC, USA).

Results
Over the thirteen-year study period, 10 Vancomycin resistance in E. faecium was high in our study, however, we observed a small, but significant decrease in urine isolates which could be a result of successful infection control strategies and antimicrobial stewardship activities. 21 In E. faecalis, vancomycin resistance was low and decreased non-significantly in both culture sites.
This findings are positive considering the spread of VRE over the past two decades. 22 Other encouraging results besides decreasing vancomycin resistance in E.
faecalis, were the stable resistance rates or small decreases in resistance rates to conventional therapies, 17 30 We also observed low ampicillin resistance in E.
faecalis, which remained stable over the study period.
Tetracycline resistance in E. faecium increased in urine and blood cultures.
While tetracycline exhibits clinically significant anti-enterococcal activity, it is generally considered a second-line agent and is seldom used for enterococci treatment. 18 As a result, susceptibility testing against this antibiotic declined, and therefore, the tetracycline resistance was only tested in few blood isolates after 2010 (n<3), resulting in larger differences in percent resistance year to year. Tetracycline resistance increases were not as large in years when more isolates were tested.
Linezolid resistance was observed in 4 of 5 facilities in our study, but the rates remained stable and low (<3%) over the study period. Linezolid is approved by FDA for the treatment of VRE faecium infections 31    Objectives: To evaluate the impact of statin exposure on clinical outcomes in bacteremic patients.
Methods: A retrospective cohort study was conducted using Optum Clinformatics TM with matched Premier Hospital data to assess inpatient mortality and length of stay (LOS) among statin-exposed vs. non-exposed bacteremic patients hospitalized between April 2010-March 2013. Patients who received at least two consecutive days of antibiotic therapy within the first three days of hospital admission were included. In the primary analysis, only incident statin users were included to avoid the "healthy user" bias. Non-users were defined as patients without any pharmacy records for statins. Cox proportional hazards regression models, adjusted by propensity scores, were developed to evaluate the effect of statins on clinical outcomes. Secondary analyses were conducted among existing statin users.

Introduction
Bloodstream infections are the sixth most common principal reason for The specific mechanism by which mortality is reduced among patients with bacteremic infections remains undefined. A proposed mechanism has been the moderation of the overall inflammatory response. 11 Other previously observed antiinflammatory effects with statins have included lowering of C-reactive protein (CRP), chemokine release (MCP-1, RANTES), cytokines (IL-1β, TNF α, IL-6, IL-8), and adhesion molecules (P-selectin, VLA 4, CD11a, CD11b, CD18). 12,13 Statins may also have a direct antimicrobial effect 14 , and possible antibacterial activity of statins against a variety of pathogens may be attributed to their ability to suppress cell growth, and to promote apoptosis. [15][16][17] In murine models, statin treatment inhibits apoptosis in sepsis 18 , reduces nitric oxide overproduction 19 , regresses the endotoxic shock induced damaged vascular responsiveness 19 , and also improves survival as it maintains cardiac function and hemodynamic status after an onset of sepsis. 20 A randomized double-blind placebo-controlled trial among patients with acute bacterial infections found a significant reduction in the levels of inflammatory cytokines among statin users. 21 A number of meta-analyses 10,22 and observational studies 4-6,23 have reported survival benefits among bacteremic patients exposed to statins compared to those not exposed to statins. However, published research has not reached a consensus on this association as several studies failed to observe significant results 24,25 and/or result estimates varied considerably. 4,5,23,25 Optimal statin use duration required to provide mortality benefits is still unknown, but the continuation of statin use during hospital admission has been found to offer pronounced effects on survival. 4 Differences between statin users and non-users in previous studies have varied by data source 6,22 , study designs 10,[22][23][24][25] and sample size of the statin user group. 4,5,23 Observational studies evaluating this association have evident differences in statin user and non-user group, potentially causing confounding of the exposureoutcome relationship. 4,5,23 Many hospital based studies evaluating protective effects of statins did not have information about medical history or medication use prior to the admission. 23,25,26 Therefore, this study aimed to determine whether the association between statins and better clinical outcomes was observed among a privately insured population with administrative data linked to hospital data. The primary objective of this study was to evaluate the impact of incident statin use, and secondarily existing statin exposures, on clinical outcomes, including inpatient mortality and length of hospital stay, in bacteremic patients exposed to statins versus those not exposed to statins in a large real-world clinical setting.

Research Design and Methodology
A retrospective cohort study design was used to assess two different outcomes, inpatient mortality and length of hospital stay, among statin-exposed vs. non-exposed patients. A retrospective cohort study design was used because it allows the comparison of individuals with differing exposures, which can be observed in order to determine the health effects of the exposure over a period of time. 27

Study Population
Included in the analysis were adult patients (>18 years) having a primary diagnosis for bacteremia or septicemia (International Classification of Diseases, 9th  28 We only included patients with hospital admissions between 04/1/2010 and 03/31/2013, to allow for a continuous enrollment period of 6 months prior to admission ( Figure 1). Antibiotic therapy for each patient during the hospital stay was assessed. Patients who received at least two consecutive days of at least one antibiotic therapy for bacteremia [29][30][31][32] within the first three days of the admission were included. For patients with multiple admissions for bacteremia, only the first admission was included. Medication use was identified from both outpatient prescriptions and medications given during the hospital stay.

Definition of Statin Use
For the primary analysis, we identified incident statin users, which was defined as those initiating a statin (i.e., atorvastatin, cerivastatin, lovastatin, pravastatin, rosuvastatin and simvastatin) within 90 days of hospital admission, or during the hospitalization, after having not used statins in the three months prior to the initial pharmacy record. A one-day gap in therapy was allowed, including separate one-day gaps on several different occasions. Non-users were patients without any pharmacy records for statins from the study start period through hospital discharge. The date of the hospital admission was defined as the index date.
Secondary analyses were conducted among existing statins users. These analyses were conducted separately among patients who, irrespective of their statin initiation time, had at least a continuous 90-day exposure for statins prior to hospitalization and did not continue during admission (existing, outpatient-only users), and among patients who had a continuous statin exposure for at least 90 days prior to hospitalization and continued statins for at least the first 5 days after hospitalization (existing-continuous users). These existing statin users were compared with nonusers to assess differences in the outcomes.

Outcomes
The primary outcome of interest in this analysis was inpatient mortality. The secondary endpoint that we evaluated was hospital length of stay. Inpatient mortality was defined as death occurring during the hospital stay. The length of hospital stay was calculated as the number of days between hospital admission and the hospital discharge date. For the length of stay, patients who died during the admission were excluded from the analysis.

Statistical Analysis
To identify baseline differences between the exposed and non-exposed groups, we analyzed demographic and clinical data including current and prior comorbidities. 28 For categorical variables, if the assumptions for the chi-squared test were not met (expected count of 75% of cells >5), the Fisher's exact test was utilized.
For continuous variables, the t-test was used for normally distributed data, and the non-parametric Wilcoxon Rank Sum test was used if the normality assumption of ttest was violated as assessed graphically and the Shapiro-Wilk test for normality.
The propensity score is the probability of treatment assignment conditional on observed baseline characteristics, derived from the inclusion of various demographic, hospitalization-related, and clinical characteristics in a logistic regression model. The propensity score attempts to mimic a randomized controlled trial (RCT) by balancing the exposure groups on observed baseline characteristics. 33,34 In this study, the propensity score was the predicted probability of statin use, as calculated from the baseline covariates included in an unconditional logistic regression model which was built with manual backward elimination. 34 39 and plotted propensity scores to review the overlap of propensity score between groups. Patients from the statin user and nonuser groups were stratified by propensity score quintile to achieve homogeneity between exposure groups within quintiles of the predicted probability of statin use. 39 Covariate balance within propensity score quintiles was reviewed. 35

Results
For the primary analysis, we identified 1,709 patients who met our inclusion and exclusion criteria (see Figure 2).

Discussion
In the primary analysis of this retrospective cohort study among privately insured patients with bacteremia, the difference in inpatient mortality among incident statin users compared to non-users was not significant, although the point estimate The continuation of statin use during hospital admission was found to offer a greater benefit, in terms of inpatient mortality. Similar results have been observed in a retrospective cohort study. 4 Additionally, a recent RCT reported a significantly lower 28-day mortality rate (5% vs. 28%; P=0.01) in the subgroup of existing-continuous statin users. 43 These results support continuing statins through the period of inflammation, as the inflammatory response has been found to be lower among patients on statins at the same time as they developed an infection. 44,45 Additional limitations of a number of the studies included in the aforementioned meta-analyses were (a) control for few confounders 23,5,8,11,25 , (b) lack of information about pre-hospitalization medication use 23,25,26 , (c) combined incident and existing statin use 8,11,26 , and (d) combined pre-hospital and post-hospital use. 26,46 These limitations may explain the conflicting findings between studies in regards to the impact of statin use on mortality among patients with infections.
Well-designed RCTs can overcome these limitations. Several RCTs have evaluated the anti-inflammatory or immunomodulatory effects of statins, including the ASEPSIS trial (EUCTR2005-004636-52), which investigated the difference in rates of sepsis converting to severe sepsis and of critical care admissions between statintreatment and placebo groups, and found the acute administration of atorvastatin in patients with sepsis may prevent sepsis progression. 47 The "Statin Therapy in the Treatment of Sepsis" trial (NCT00676897), found significantly lower coenzyme Q10 levels, which may be associated with the inflammatory cascade in septic shock, in septic shock patients compared to healthy controls. 48 The "Statin for Immunomudulation in Sepsis" trial (NCT00452608) and "Effect of Atorvastatin on the conditions. 49 To our knowledge, only one study has explored the association of statin use with LOS, and found non-significant results for hospital (β = -0.8 days, 95% CI -2.2-1.7 days) or intensive care unit LOS (β = -0.1 days, 95% CI -3.7 to 3.8 days) length of stay. 25 The present study has several strengths. Firstly, we used administrative data from a major private insurer linked to hospital data which allowed us to evaluate medical history, previous medication use, as well as conditions present during the admission and all medication exposures during the hospitalization. Secondly, we attempted to account for "healthy user bias" by including incident statin users in our primary analysis since patients taking preventive medications, such as statins, are more likely to engage in healthy behaviors leading to favorable health outcomes compared to non-statin users. 50,51 Additionally, patients taking preventive medications have a higher probability of being up-to-date with immunizations and having quit smoking, and are less likely to have been admitted to a nursing home or need advanced medical care. 52 Third, we balanced baseline characteristics between statin users and non-users that were significantly different using PS methods in an effort to control for confounding. Lastly, to account for possible biases in socioeconomic and health behaviors, 53 we included demographic and clinical characteristics in the PS models.

Limitations
The results of this study have potential limitations. We could not study the protective effects of each statin separately due to small numbers. The effect of statins on inpatient mortality in patients with sepsis may be different for individual statins. 54 We also could not assess dose-dependent effects, changes in statin therapy (drug or dose) prior to admission, at admission, or during the admission, or the effects of adherence due low sample sizes. In our review of statin doses, dispensing quantity in incident users mostly reflected moderate to high doses. As we used an administrative claims database for our analysis, we assumed outpatient statin exposure to be equivalent to filling a prescription. In the primary analysis, our definition of incident statin use was broad due to small numbers and included patients initiating prior to admission or after admission, and also included those not continuing statins during the admission (38%). As such, we could not evaluate the association using more specific definitions of incident statin use.
Furthermore, there is a possibility of statins having a different impact on clinical outcomes based on the causative pathogen, since the mechanism of action is not exactly known and it may vary for different pathogens. Microbiology data was not available for potential causative pathogen, but we identified organisms using ICD-9 codes, where available. Bacteremic treatment varies by organism type and we were only able to use general inclusion criteria of having received an antibiotic which may be used for bacteremia. [29][30][31][32] Since we only evaluated a general bacteremic population, our results may not be generalized to pathogen-specific bacteremias. As such, patients without appropriate initial antibiotic treatment may have been included.
Despite using propensity scores to control for confounding, we could not control for unmeasured confounding. We also could not differentiate bacteremic severity, although we included ventilation status and sepsis proxies using diagnosis-related groups (DRG).

Conclusions
In conclusion, our retrospective cohort study quantified the effect of both incident and existing statin use on clinical outcomes such as inpatient mortality and hospital length of stay among bacteremic patients in a real-world clinical population.
Result estimates for incident and existing-continuous statin use, although nonsignificant for incident users, were similar to previous meta-analyses that observed reductions in inpatient mortality after statin use among bacteremic patients. Further   1) Adjusted by propensity score quintiles (reference quintile I).
2) The final PS model equation for predicting incident statin exposure.
3) The final PS model equation for predicting existing, outpatient-only statin exposure. 4) The final PS model equation for predicting existing-continuous statin exposure.
In the above equations, y is the probability of receiving a statin, α is the intercept, β's are the coefficients on the independent variables and ε is standard error.

Figure 3: Adjusted proportional hazards among incident statin users vs. nonusers
Note-On the x-axis, 'tmdc' represents "time to death".

Abstract
Background: There is no consensus as to whether statin therapy should be continued among patients presenting to the hospital with bacteremia, and if so, what duration would be associated with better survival.
Objectives: To identify a statin therapy duration that would decrease mortality in bacteremic statin users. While numerous studies have found reduced mortality with statins in bacteremic patients, statin duration and measurement of outcomes differ across these studies. [3][4][5] As a result, rates of survival vary, particularly as statin exposure varies. 4,10 Several studies not only observed a decline in inpatient mortality after continuing statin use during admission, 4,10 but also an increase in mortality after cessation of statin therapy. 4,11 Since the length of statin treatment time varies between studies, there is no consensus on the duration of statin continuation that would provide the maximum advantage in terms of clinical outcomes.

Methods
While several meta-analyses 12,13 and observational 4-6,10 studies observed protective effects with statins in bacteremia, one meta-analysis 14 did not observe improvements in clinical outcomes after statin use. However, this meta-analysis was conducted among critically-ill patients with severe sepsis, and some of the included studies only had short durations of statin use. 4,5,10 Other studies with shorter statin durations also did not demonstrate a statistically significant association between statin use and mortality. 10,14,15 A recent RCT evaluating benefits of continued statin therapy on inflammatory parameters and sepsis among patients with pre-existing statin use 16 did not find clinical benefits of continuation. As such, there is a lack of evidence regarding the appropriate exposure duration needed for statins to provide the utmost protective effects in bacteremic patients. The main objective of this study was to identify a time breakpoint of statin continuation which minimized inpatient mortality among bacteremic patients.

Research Design and Methodology
A case-control study design was used to estimate a time breakpoint in statin continuation at which the highest clinical benefit would be seen in terms of survival (i.e., lowest inpatient mortality). A case-control study is an analytical study that compares individuals who have a specific outcome (cases) with a group of individuals that do not have the outcome (controls). A case-control design was utilized because it is the most effective study design for evaluating multiple exposures when an outcome is rare. 17

Data Sources
This study was conducted using deidentified Optum Clinformatics TM   18 by any causative organism. We excluded patients who, on the first three days after hospital admission, did not receive a minimum of two successive days of at least one bacteremic antibiotic therapy. [19][20][21][22] The index date was defined as the date of the first hospital admission during the study period, and subsequent multiple hospital admissions were not considered for the analysis. From this cohort, only patients with a minimum of 3 months of continuous statin use in the 3 months prior to admission were selected for inclusion ( Figure 1).

Cases and controls
Cases included those who died during the admission. In a case-control study, controls should be drawn from the same population from which cases are derived, in order to reduce the chance that group differences account for the difference in the exposure being evaluated. 23 Thus, controls were selected from the same cohort of adult patients who had a primary diagnosis of bacteremia on hospital admission and received an antibiotic therapy, but experienced a different outcome (i.e., no inpatient mortality). Controls were matched to cases on disease risk score (DRS). 24 DRS is a confounder summary method, commonly used in case-control studies to control for confounding by calculating the predicated probability of an outcome in the absence of exposure. 25,26 A recent simulation study 27 suggested the DRS model could cause higher bias due to misspecification at higher outcome incidences, however, when the outcome is rare, DRS matching would increase the statistical efficiency of casecontrol studies. The DRS is considered a useful method in case-control studies, 25 especially when the association between covariates and exposure is modest (squared multiple correlation coefficient amid exposure and confounders <90%). 28 The stratified DRS is a retrospective balancing score and therefore it works in a similar manner in case-control studies as the propensity score works in cohort studies. 26

Statin duration
Among

Statistical Analysis
Disease risk scores (DRS) were calculated to control for confounding. A DRS is the probability of a patient having a particular outcome in the absence of the exposure. 24,25 Therefore, we calculated DRS as the probability of inpatient mortality among unexposed patients, that is, statin users not continuing use post-admission.
Using likelihood ratio tests, we compared each independent variable to the null model. Variables with a p-value <0.25 in likelihood ratio tests were included in an initial multivariate model and removed using a backward elimination approach, if the p-value of the parameter estimate was less than 0.05 to arrive at the final DRS model.
The model was checked for multicollinearity using variance inflation factor (VIF) and correlation matrices, and goodness of fit was checked using the Hosmer-Lemeshow goodness-of-fit test. 29,30 The final DRS model c-statistics was 0.91. The full DRS model equation can be found in the footnote of the main results. Using nearest neighbor matching within a caliper of 0.25 distance, a single control was selected for each case. 31 We checked DRS balance between cases and controls using graphical displays (see figure 2).
To partition statin continuation days associated with the lowest risk of death (i.e., highest survival), we conducted a classification and regression tree analysis (CART). 32,33 The CART analysis, which includes an optimal tree selection based on CART models are useful because of their non-parametric, non-linear structure. 33 As a result, they do not make any distribution assumptions, they treat the data generation process as unknown, and they do not require a functional form for the predictors.
Indeed, tree methods are probably one of the most easily interpreted statistical techniques; they are conceptually simple yet analytically powerful. 33,34 In the CART analysis, the primary dependent variable was inpatient mortality (case/control) and the independent variable was days of continued statin exposure during admission. The trees were automatically developed to forecast inpatient mortality by considering every possible cut-point on statin continuation duration at every node in the classification tree. We checked the fitness of the tree by plotting cross-validated error rate vs. size of a tree, and identified an appropriate complexity parameter (CP). 33

Results
During the three-year study period, 61 (

Discussion
In this DRS matched case-control study, we identified a specific continued statin use duration threshold required to provide the maximum survival benefit among bacteremic patients, which was the continuation of statin therapy for at least 2 days. Our study is a step forward in the direction of identifying an optimal statin duration for inflammatory conditions. To the best of our knowledge, this is the first study in a large, privately insured population in the U.S. evaluating an optimal continued statin therapy duration among patients with inflammatory conditions. We utilized a machine-learning analysis method, CART, that allowed us to identify an exact statin therapy duration at which inpatient mortality was lowest among patients with bacteremia. Other factors may affect the impact of statin therapy on mortality, including baseline differences in patient characteristics, differences in bacteremic severity, regional differences in infections, statin prescription patterns, pre-admission and post-admission comorbid conditions as well as medication use. However, we used DRS to match controls to cases in order to account for confounding. The strength of an observational study generally depends on the quality of the data source. However, we utilized administrative data from a large, national insurer, which is not affected by recall or surveillance bias. Further, used administrative and hospital linked data from a real-world clinical population with health-coverage from a major private payer.
Our study offers evidence regarding continuation of statin therapy in existing statin users presenting to the hospital with bacteremia. Although our findings indicate benefits with continuation of statins during admission, greater information is needed regarding the risks of continuation, in terms of adverse events, to enable a clear benefit-risk assessment. There is an ongoing controversy about the benefit-risk assessment of statins in general ("statin wars") 45

Limitations
Our study has a few limitations. First, we were unable to assess adherence or dose-dependent effects of statins that might affect bacteremic mortality. Depending on the severity of the infection, different statin doses and different statins may be used. While statin therapy may been continued more frequently in lower-risk patients, as clinicians hear about the potential protective effects of statins, there may channeling bias in the opposite direction, where more severe patients are kept on their statin therapy. 46 Second, our study relied on a claims database, which raises the concern of misclassification due to coding errors throughout medical claims processing. Further, use of this database assumed actual exposure from prescription claims and hospital charges for medications. Third, we could not study differences in mortality with statin continuation duration in bacteremia caused by specific pathogens. A previous study 26 observed greater protection with statins in S. aureus bacteremia compared to bacteremia caused by Gram-negative bacilli, while also suggesting greater survival in nosocomial versus community-associated bacteremia. 26 Our study could not evaluate these differences. Moreover, the sample size of our study was small. Lastly, the limitations of CART analysis include an inability to fully describe the observed data due to uncertainty that remains in the prediction of the model and potential existence of multiple threshold values despite a single "optimal" split. 47

Conclusions
In conclusion, this disease risk score matched case-control study conducted in   Note: On the y-axis, 0 represent controls, while 1 represent cases.
On the x-axis, estimated probability is the disease risk score. Note: "in.days" represents "continued-during-admission statin therapy duration".
The study included an equal number (n=58) of cases and controls, producing a 50% survival rate at the root node. Here y is the probability of inpatient mortality, α is the intercept, β's are the coefficients on the independent variables and ε is standard error.