Detailed Description of Data Source

Detailed Description of Data Source

The database includes basic admission and demographic information (not all of which were used in the present study), vital signs, laboratory and radiology results, medications, discharge diagnoses, nursing notes, physician discharge summaries and dates of death. MIMIC-II contains patients from five ICU types: medical (MICU), surgical (SICU), cardiac (CCU), cardiac surgery recovery (CSRU) and neonatal (NICU). Other clinical data were added to the database including pharmacy, provider order entry records, admission and death records, discharge summaries and 9th Revision (ICD-9) codes.

All data were extracted from the MIMIC-II database (v2.6), including demographic information (e.g., age and gender), clinical information from admission notes (diuretic use), laboratory results (e.g., blood gas analysis, minimum serum chloride, minimum serum potassium, maximum serum sodium, serum creatinine and urine chloride levels), pharmacy information (e.g., diuretic and alkali administration) and discharge diagnoses according to ICD-9 codes. In addition, cumulative gastric output fluids during the ICU stay were collected, and patients above the 90th percentile were considered to have high gastric output.

The MIMIC-II database also includes the severity of illness as assessed through the simplified acute physiology score (SAPS I) and Sequential Organ Failure Assessment (SOFA) score [1],[2] .

Complete Statistical Analysis

First, we used descriptive statistics, including the mean with standard deviation, median and interquartile range or frequencies, to describe the population as appropriate. Next, we investigated the association between the maximum serum bicarbonate level as a continuous variable and in-hospital mortality. For the univariate analysis, we used chi-square tests,t-tests, Kruskal-Wallis tests, one-way analysis of variance and Mann-WhitneyUtests to evaluate statistical significance. All tests were two-sided, and ap-value of < 0.05 was considered significant.

To obtain multivariable models, we first used multiple imputations (5 replications) to create complete data sets because up to 3% of observations were missing values for Na+, K+ and Cl-. We performed several logistic regression analyses to determine the association between increased serum bicarbonate and its duration and hospital mortality. The robustness of the association between serum bicarbonate and clinical outcome was tested by progressively adding variables to the models that might have an effect on serum bicarbonate or the clinical outcome considered. The following covariates were included in the models: age, gender, SOFA and SAPS-I scores, type of ICU admission (medical or surgical), minimum and maximum serum bicarbonate levels, acute kidney injury (AKI) severity, hyponatremia and the main comorbidity groups obtained from ICD-9–Clinical Modification codes using the Elixhauser comorbidity index. Other factors associated with high serum bicarbonate levels in univariate analysis also includedhypokalemia, hypernatremia, gastric output and renal replacement therapy (RRT), all of which were considered during ICU stays. Based on the initial exploration of maximum serum bicarbonate levels for hospital mortality, which revealed a U-shaped association, we assessed a cubic spline regression model with serum bicarbonate as a continuous variable with two knots at 24 and 31 mEq/L. The values were then categorized as less than 25 or greater than 30 vs. 25-30 mEq/L. This same approach was made using SBE with one knot at -1mEq/L and 5mEq/L. After testing for colinearity, multiple covariate analyses were applied. For interpretation, we provided odds ratio (OR) values for each 5-mEq/L increases or decrease. We also performed bootstrap validation of the full-adjusted models using the entire cohort with 1,000 resampling runs to assess bias due to over-fitting. The specific covariates used in each logistic regression model are specified in the text and/or tables. Serum albumin was not included in the multivariate analysis because less than 50% of patients had these measurements available. We performeda subgroup analysis to search for interactions by each known cause of metabolic alkalosis to investigate the association between high serum bicarbonate levels and in-hospital mortality.

Because ICU discharge and mortality represent competing risks, we used the cumulative incidence function to analyze time to hospital discharge and mortality over 28 days.To investigate whether high serum bicarbonate had any influence on the impact on long-term survival, we performed a multivariate Cox regression analysis in all patients discharged alive from the hospital. In Cox regression analysis, we also adjusted for age, gender, simplified acute physiology score (SAPS-I), sequential organ failure assessment (SOFA), main comorbidities, type of admission (clinical or surgical), diuretic use, either previous to or during ICU stay, alkali administration, hypo/hyperkalemia, hypo/hypernatremia, renal replacement therapy (RRT) and gastric output, minimum serum bicarbonate, partial CO2 pressure > 45mmHg, acute kidney injury severity and mechanical ventilation during ICU stay. All covariates were tested to maintain proportional hazard assumption3.Statistical analyses were performed using SPSS 19.0 for Windows.

Arterial Blood Sample Analysis – Sensitivity Analysis

In a sensitivity analysis, we attempted to separately evaluate patients with arterial blood gas measurements (n = 12,696). In these patients a standard base excess (SBE) greater than 5 mEq/L (n = 3,273) was used to define metabolic alkalosis instead of serum bicarbonate levels. The majority of patients with high SBE (72.2%) also had alkalemia (pH > 7.45). In these subset analyses, metabolic alkalosis remained independently associated with the death rate in uni- and multivariate analyses (see supplementary table 4).

References

1. Le Gall JR, Loirat P, Alperovitch A, Glaser P, Granthil C, Mathieu D, Mercier P, Thomas R, Villers D. (1984) A simplified acute physiology score for ICU patients. Crit Care Med 12:975-977.

2. Vincent JL, de Mendonça A, Cantraine F, Moreno R, Takala J, Suter PM, Sprung CL, Colardyn F, Blecher S. (1998) Use of the SOFA score to assess the incidence of organ dysfunction/failure in intensive care units: results of a multicenter, prospective study. Working group on "sepsis-related problems" of the European Society of Intensive Care Medicine. Crit Care Med 26:1793-1800.