Participants and Measurements
This retrospective, observational cohort study was performed using two publicly available ICU datasets, the MIMIC-III 29 and the eICU 30 Collaborative Research Database (eICU-CRD). The MIMIC-III database was released in 2016 by the Massachusetts Institute of Technology Laboratory for Computational Physiology (MIT-LCP) and contained data from a single tertiary care hospital (Beth Israel Deaconess Medical Center). The eICU-CRD is a multicenter critical care database containing data from rural/nonacademic hospitals across the US and was made available in 2018 by Philips Healthcare with the help of researchers from MIT-LCP. There is no overlap in the patients included in these two databases.
We included adult patients 18 years of age or older who received RRT (either intermittent HD or continuous RRT [CRRT]) in the ICU for AKI. The AKI in this study was defined according to the Kidney Disease Improving Global Outcomes clinical practice guidelines 31 and diagnosis codes.. We only used the creatinine criteria because of unreliable urine data in the retrospective databases. For patients who did not have more than one creatinine value to make a comparison, but who had RRT records, we included patients who were diagnosed as having AKI based on their ICD-9 diagnosis codes (Table S1). If a patient had been admitted to the ICU multiple times in one hospitalization course, data from the ICU admission that included the initial dialysis treatment was extracted for the study. Patients with a history of end-stage kidney disease who underwent chronic peritoneal dialysis or HD (Table S1) were excluded from the study. We also excluded patients who had chronic kidney disease (CKD) stage 4 and 5 based on ICD-9 codes (Table S1), because we were interested in patients who did not have advanced CKD at baseline. Patients with a history of any organ transplant were also excluded as they may have other confounding risk variables that affect mortality. We excluded patients who did not have complete records of vital signs and creatinine data one day before RRT start.
The variables collected consisted of demographics, medical history, mechanical ventilation usage, AKI stage, vital signs, laboratory test results, dialysis modality (intermittent HD or CRRT), and medications (diuretics and vasopressors, see Table S2). The time window of mechanical ventilation, vital signs, laboratory tests, and medications were recorded one day before RRT initiation. Past medical history was extracted from database records using ICD-9 codes (Table S1). Relevant past medical history included in the study were diabetes mellitus (DM), CKD, hypertension (HTN), congestive heart failure (CHF), liver cirrhosis (LC), and cancer. Vital signs in this study included the mean values of the following variables: shock index (SI), Glasgow Coma Scale (GCS), mean arterial pressure (MAP), respiratory rate (RR), and heart rate (HR). The mean SI was calculated by the formula: SI = mean HR/mean systolic blood pressure. For laboratory tests, we used the mean value of all variables recorded one day before the date of the first dialysis therapy because some laboratory data values would have been influenced by dialysis. We excluded the variables with >25% missing values, except for albumin level, because we thought that albumin was an important factor for mortality prediction. Table S3 reveals the percentages of missing data in the laboratory tests. Multiple imputation by chained equations (MICE) with five imputed datasets was used to impute the missing values of the laboratory tests and vital signs and the results were pooled using the MICE package 32.
We modified the codes from https://github.com/nus-mornin-lab/oxygenation_kc and https://github.com/MIT-LCP/mimic-code/tree/master/concepts/severityscores to calculate the SOFA score using variables collected one day before RRT start in the eICU and MIMIC datasets based on methods used in the original study 33. For patients with missing variables, the
The primary aims of the investigation were to assess whether the predialysis creatinine level was associated with 30-day mortality independent of other risk factors and to explore the association between the predialysis creatinine level and the SOFA score.
Statistical Analyses
The study cohort was stratified into two groups according to the median creatinine value. Categorical variables were presented as counts, proportions, and frequencies; continuous variables were expressed as mean with standard deviation. Numeric variables of clinical characteristics between the two groups were compared using the Student’s t test. The chi-square test was used to compare the differences of the categorical variables. Logistic regression was used to calculate odds ratios (ORs) and 95% confidence intervals (95% CIs) for the analyses of predictors of mortality. The comparison of survival status between the two groups was done using the Kaplan–Meier curve with significance levels determined by the log rank test. We implemented four models for the adjustments of the covariates: model 1, adjusted for age, sex, and ethnicity; model 2, adjusted for all variables in model 1 plus DM, HTN, CKD, malignancy, and LC; model 3, adjusted for all variables in model 2 plus GCS, HR, MAP, RR, SI, ICU days before dialysis, CRRT, diuretics, vasopressors, and mechanical ventilation; model 4, adjusted for all variables in model 3 plus laboratory parameters. We used the Kruskal–Wallis test to compare the SOFA score difference between the groups.
Each analysis was repeated using the MIMIC dataset to explore the heterogeneity using sensitivity analysis. Analyses were performed using R version 3.6.1 (R Foundation for Statistical Computing).