The Model for End-stage Liver Disease (MELD) as a predictor of short-term mortality in Staphylococcus aureus bloodstream infection: A single-centre observational study

Background Automated laboratory-based prediction models may support clinical decisions in Staphylococcus aureus bloodstream infections (BSIs), which carry a particularly high mortality. Small studies indicated that the laboratory-based Model for End-stage Liver Disease (MELD) score is a risk factor for mortality in critically ill patients with infections. For S. aureus BSIs, we therefore aimed to assess a potential association of the MELD score with mortality. Methods In this single-centre observational study, all consecutive patients with a first episode of methicillin-susceptible S. aureus BSI occurring between 2001 and 2013 were eligible. Relevant patient data were retrieved from our prospective in-house BSI database. We assessed the association of the MELD score at day of BSI onset (range ± two days) with 30-day all-cause mortality using uni- and multivariable logistic regression analysis. Results 561 patients were included in the final analysis. The MELD score at BSI onset was associated with 30-day mortality in S. aureus BSIs (odds ratio per 1-point increase, 1.06; 95% confidence interval, 1.03‒1.09; P < 0.001). After adjustment for relevant patient and infection characteristics, an increased MELD score remained a predictor of 30-day mortality (adjusted odds ratio per 1-point increase, 1.05; 95% confidence interval, 1.01‒1.08; P = 0.005). Conclusions In our study population, the MELD score at BSI onset was an independent predictor of mortality in S. aureus BSIs. We therefore suggest to prospectively validate the MELD score as part of clinical decision support systems in inpatients with suspected or confirmed BSI.

Introduction Staphylococcus aureus is a leading cause of bloodstream infection (BSI) carrying a high mortality, even if treated with adequate antimicrobial and supportive measures [1,2]. The considerable mortality stresses the importance of prediction models to support clinical decisions in S. aureus BSIs.
The Model for End-stage Liver Disease (MELD) is a widely-used risk model, which was initially created to predict mortality in patients with portal hypertension undergoing placement of transjugular intrahepatic portosystemic shunts [3]. Subsequently, it was thoroughly validated as a predictor of mortality among different patient populations across a broad spectrum of liver diseases-primarily to allocate organs for liver transplantation [4][5][6]. The MELD score incorporates three routine laboratory parameters, i.e. serum creatinine, serum bilirubin, and the International Normalized Ratio (INR), which could enable rapid automated clinical decision support without depending on complex clinical variables.
Small studies indicated that the MELD score is a predictor of mortality in various patient populations with infections or infectious complications [7][8][9]. We therefore hypothesized that the MELD score is an independent predictor of mortality in S. aureus BSIs. The overall goal of our study was achieved in terms of establishing that the automatically calculated MELD score might be an independent predictor of mortality in patients with BSIs caused by S. aureus.

Study design and setting
This was an observational study performed at the University Hospital Basel, an 800-bed tertiary referral centre in Northwestern Switzerland with >35,000 hospitalisations per year; treatment modalities cover all surgical and medical specialties including kidney and bone marrow transplantations. The Ethics Committee of Northwestern and Central Switzerland (EKNZ) approved this study with a waiver on written/oral informed consent (study number 2016-01515).

Patient selection
Patients aged !18 years with a first methicillin-susceptible S. aureus BSI episode occurring between January 2001 and December 2013 were eligible for the study. We excluded recurrent BSI episodes (i), patients with a missing 30-day follow-up (ii), patients with missing MELD parameters at BSI onset (iii), patients with missing medication data on vitamin K antagonists and novel oral anticoagulants (iv), and patients treated at BSI onset with a vitamin K antagonist (v) or a novel oral anticoagulant (vi) because of their effect on the INR. culture >7 days after the last identification from blood culture. An episode of BSI without a definite source of infection was defined as BSI of primary origin. The time stamp of the first positive blood culture by the in-house microbiological laboratory was defined as the day of BSI onset. Onset of a BSI after more than two days of hospitalisation was interpreted as hospitalacquired BSI. The diagnosis of liver cirrhosis was collected retrospectively from medical records and verified by reviewing ultrasonographic, laboratory, endoscopic, and pathological reports [12]. Immunosuppression was defined as described previously [13]. In brief, the definition included the presence of conditions leading to immunosuppression (e.g. end-stage renal failure, haematologic malignancies), immunological/inflammatory diseases requiring immunosuppressive therapy, or an absolute neutrophil count <500/μl. An adequate empirical antimicrobial therapy was defined in retrospect as providing coverage against methicillinsusceptible S. aureus.

Microbiology
During the entire study period, rapid processing and incubation of the blood cultures were ensured 24 hours per day using the same blood culture system with at least one pair of aerobic/ anaerobic bottles (BacT/ALERT; bioMérieux, Durham, North Carolina, Unites States of America).

Statistical analysis
Patient and infection characteristics were presented as proportions, means with standard deviation and as medians with interquartile ranges (IQR) as appropriate. For 30-day survivors and non-survivors, we assessed unadjusted differences in the patient and infection characteristics with the use of the Chi-square test or Fisher's exact test for categorical variables and the Student's t-test or the Wilcoxon test for continuous variables, depending on the data distribution.
We assessed the association of the MELD score at BSI onset with 30-day all-cause mortality by means of uni-and multivariable logistic regression analysis; The multivariable logistic regression model was built using a forward selection procedure (at P < 0.10) for the following co-variables: Age (continuous), gender, Charlson Comorbidity Index [14] at BSI onset, days of hospitalisation prior to BSI onset, BSI origin (primary or secondary), and presence or absence of liver cirrhosis, immunosuppression, intravenous drug use, and any surgery in the last 30 days prior to BSI onset [15][16][17]. The dichotomous origin of BSIs was used in the multivariable model, as S. aureus BSIs of primary origin have been described as independent predictor of mortality relative to other BSI sources [17].
Subsequently, we stratified the patients into the following MELD categories at BSI onset: scores <10, 10-15, and/or >15. The association of these MELD categories with 30-day allcause mortality was analysed by means of a univariable Cox proportional hazards model and comparison of the corresponding Kaplan-Meier curves using the Log-rank test.
For patients with and without liver cirrhosis, we compared laboratory and outcome parameters with the use of the Chi-square test or the Fisher's exact test for categorical variables and the Student's t-test or the Wilcoxon test for continuous variables, depending on the data distribution. We analysed all data using SPSS, version 22 (SPSS; IBM, Chicago, Illinois, United States of America); All P-values are two-sided with values 0.05 being considered as significant.

Patient selection
Within the 13-year study period, 788 patients had a first S. aureus BSI episode caused by methicillin-resistant S. aureus (MRSA; n = 27) or methicillin-susceptible S. aureus (n = 761). Seven hundred sixty-one patients with a first episode of methicillin-susceptible S. aureus BSI were eligible for the study (Fig 1). Of these, 561 were included in the final analysis after exclusion of 11 patients with a missing 30-day follow-up, 76 patients with missing MELD parameters at BSI onset, 61 patients with missing medication data, and 52 patients treated with a vitamin K antagonist or a novel oral anticoagulant.
In univariable analysis, the MELD score at BSI onset was positively associated with 30-day all-cause mortality (OR per 1-point increment, 1.06; 95% CI, 1.03-1.09; P < 0.001) ( Table 2). After adjustment for relevant patient and infection characteristics, the MELD score at BSI onset remained positively associated with 30-day all-cause mortality (adjusted OR per 1-point   increment, 1.05; 95% CI, 1.01-1.08; P = 0.005). The positive association of the MELD score at BSI onset was also observed with in-hospital mortality in uni-and multivariable analysis (S4 Table).

Discussion
In our study population consisting of patients with methicillin-susceptible S. aureus BSI, the laboratory-based MELD score was an independent predictor of short-term mortality. We focused on a leading pathogen in BSIs and excluded MRSA cases since empirical antimicrobial therapy did not cover MRSA. The results of this study are in line with several small observational studies indicating that the MELD score might be a predictor of mortality in patients with serious infections-with or without concomitant liver disease [7][8][9]: In a recent study by Juntermanns et al., liver transplant recipients with a pretransplant MELD of !31 points were at higher risk for dying due to septicaemia than individuals with a lower pretransplant MELD score (P < 0.05) [9]. In a retrospective study, patients with liver cirrhosis and bacterial infections (in 69.6% of patients [369/ 530] due to a spontaneous bacterial peritonitis) had an eight percent increase in the mortality rate with every one point rise in the MELD score [8]. Another study demonstrated a good performance of the MELD for mortality prediction in patients with Vibrio vulnificus skin and soft tissue infections (n = 39) [7]. If carefully validated and modified to accurately predict mortality in BSIs, the MELD could offer an original opportunity for automated clinical decision support in BSIs without relying on complex clinical variables, which is a major barrier to the application of most mortality prediction scores in daily clinical routine [16,18,19]. Regarding clinical decision support, the MELD fulfils important preconditions for a successful prediction model: It incorporates three readily available laboratory variables, which are routinely and repeatedly measured in most inpatients, and which have been demonstrated to be good predictive markers of mortality in patients with and without infections [20][21][22][23][24].
For the observed association of the MELD parameters with mortality, we did not investigate potential underlying pathophysiological mechanisms, as has partly been elucidated for other routine laboratory parameters in BSIs and sepsis (e.g. thrombocyte count) [25]. We speculate that the MELD is a surrogate marker for the presence and severity of a variety of diseases (e.g. kidney injury, liver diseases, malignancies), which may be used for risk stratification and risk adjustment.
In our study, the INR at BSI onset was-in contrast to creatinine and bilirubin-not associated with mortality in uni-and multivariable analysis, although we excluded patients on vitamin K antagonists and novel oral anticoagulants, which could have biased the results. A positive association of the INR with mortality was demonstrated in high-risk patients with sepsis or trauma [26][27][28][29]. Conversely, in our study population, only about 30% of patients were admitted to an intensive care unit during the initial hospitalisation period and the INR at BSI onset was rather low with a median of 1.1 (IQR, 1.0-1.3).
Our study has several limitations. First, we had to exclude 26% of patients which might have led to a selection bias possibly overestimating the effect of the MELD score on mortality, as patients with unmeasured MELD parameters on hospital admission may be healthier than patients, in whom the serum creatinine, bilirubin and INR had been obtained as part of the diagnostic workup. However, only eleven patients had to be excluded due to a missing 30-day follow-up in this hospital-wide observational study. Second, although we have adjusted for important patient and infection characteristics in our analysis, we cannot exclude that unmeasured confounders might have remained, which could have biased our estimates (e.g. time to adequate antimicrobial therapy). Third, our study was performed at a single centre, which may limit the generalizability of our results. In this context, it is important to note that the results of our study do not apply to patients on vitamin K antagonists and novel oral anticoagulants, as these individuals had been excluded from the analysis. Fourth, we did not assess patients with BSIs caused by pathogens other than S. aureus-particularly microorganisms with less pathogenic properties such as Escherichia coli. Fifth, there are studies suggesting an interlaboratory variability in all three components of the MELD score with a mean interlaboratory difference of about five MELD points [30], which may have further reduced the generalizability of our results.
In conclusion, in our study population, the MELD score at BSI onset was an independent predictor of mortality in S. aureus BSIs. We therefore suggest to prospectively validate the MELD score as part of clinical decision support systems in inpatients with suspected or confirmed BSI.
Supporting information S1
Data curation: JAR MD RF.