Data source The eICU-CRD [19] provided the data between 2014 and 2015 for this retrospective analysis. The eICU program from Philips Healthcare automatically saved and obtained data from an eICU database that included over 200,000 admissions across different facilities in the United States [19]. We were granted access to this database, completed the test, and received certification from the PhysioNet Review Board (record ID: 40859994).
Study population Sepsis-diagnosed ICU participants were involved in the research upon admission. The possible or verified infection and a score increase of more than two points on the Sequential Organ Failure Assessment (SOFA) [1], as documented in the Acute Physiology and Chronic Health Evaluation (APACHE) IV dataset [20], were considered indicators of sepsis. Applying the International Classification of Diseases, 9th Edition (ICD-9) code, infections were detected by the eICU-CRD. ESRD is a final stage of CKD, diagnosed when renal replacement therapy (RRT) is needed for over 3 months, with a glomerular filtration rate (GFR) less than 15 mL/min [21]. ESRD in our study was verified by ICD-9 code. The criteria for exclusion were as follows:(1) Not a first-time admission to ICU; (2) ICU stay of shorter than 48 hours; (3) Age under 18 years; (4) Diagnosed as ESRD; (5) missing albumin level or system error; (6) missing AKI outcome. The research flowchart was referred to Fig. 1.
Variables The eICU database contained population data, bedside monitor readings, diagnoses using ICD-9 codes and lab data from regular medical treatment. Data of the participants in the eICU-CRD were collected within initial 24 hours after ICU admission. The apacheApsVar table was used to extract physiological information such as respiration rate, heart rate, mean arterial pressure (MAP)and temperature. Treatment details on vasopressor, mechanical ventilation, and RRT were also contained in the same table. Basic information was collected from the patient or apachePatientResult charts, including age, ethnicity, gender, height and weight. Weight divided by height squared (kg/m2) was used to compute BMI. The lab recorded levels of white blood cell (WBC), platelet, hemoglobin, total bilirubin (TBIL), aspartate transaminase (AST), alanine transaminase (ALT), blood urea nitrogen (BUN), albumin, serum creatinine (Scr), potassium, sodium, and lactate. Comorbidities like congestive heart failure (CHF), hypertension, chronic obstructive pulmonary disease (COPD), acute myocardial infarction (AMI), diabetes, hepatic failure, cirrhosis, and cancer were taken from the admissionDx table. Admission severity was evaluated using Glasgow Coma Scale ( GCS ) score, Acute Physiology Score III, SOFA score and Apache IV score.
Outcome The study’s outcome was the likelihood of AKI following ICU admission. AKI was defined as a increase in Scr of 0.3 mg/dL during 48 hours or at least 1.5 times the Scr baseline level during 7 days after entering the ICU, in accordance with kidney disease: Improving Global Outcomes (KDIGO) standards [22].
Missing data Among the 5459 patients in our study, The number of covariate missing values were 36 (0.66%) for ethnicity, 1 (0.02%) for gender, 103 (1.89%) for BMI, 264 (4.84%) for temperature,2 (0.04%) for heart rate, 20 (0.37%) for respiratory rate, 7 (0.13%) for MAP, 143 (2.62%) for WBC, 133(2.44%) for hemoglobin, 160 (2.93%) for platelet, 95 (1.74%) for ALT, 66 (1.21%) for AST, 569 (10.42%) for TBIL, 25 (0.46%) for BUN, 17 (0.31%) for Scr, 19 (0.35%) for potassium,19 (0.35%) for sodium, 1370 (25.10%) for lactate, 2 (0.04%) for SOFA score, 80 (1.47%) for GCS score, 595 (10.90%) for Acute Physiology Score III, 19(0.35%) for vasopressor, 595 (10.90%) for Apache IV score. Dummy variables indicated missing covariate values when continuous variables had over 1% missing values.
Statistical analysis For continuous variables, the standard deviation (SD) or median with interquartile range (IQR) are displayed, whereas counts and percentages are used for categorical data. For categorical variables, chi-squared tests were performed to examine differences in albumin quartiles. For continuous variables, the study employed one-way analysis of variance (ANOVA). The relationship between albumin and probability of AKI was investigated by both univariate and multivariate logistic regression. Results were given as OR with their 95% CI. Confounders were chosen if their impact estimates changed by more than 10% or if they showed a link with the outcomes [23]. After assessing clinical significance of included covariates, we established three models: (1) Crude model: include no variables. (2) Model I: adjust for demographics of BMI, age, ethnicity and gender. (3) Model II : adjust for BMI, age, ethnicity, gender, MAP, WBC, hemoglobin, platelet, BUN, Scr, potassium, sodium, site of infection, SOFA score, APACHE IV score, hypertension, AMI, CHF, COPD, diabetes, hepatic failure, cirrhosis, cancer, mechanical ventilation and RRT. Generalized additive model (GAM) was employed in smooth curve fitting to find any nonlinear correlations between blood albumin and AKI risk in sepsis patients. Bootstrap resampling, likelihood ratio tests, and piecewise regression were applied to examine the threshold effect between serum albumin and probability of AKI [24].
To verify the stability of the results, we completed sensitivity analysis. To investigate any unmeasured confounding between serum albumin and probability of AKI, we computed E-values [25]. The amount of an unmeasured confounder required to eliminate the link between serum albumin and probability of AKI was shown by the E-value. For statistical significance, a two-tailed p-value of less than 0.05 was judged appropriate. EmpowerStats (www.empowerstats.com, X&Y solutions, Inc. Boston, MA) and R statistical software tools (http://www.r-project.org, The R Foundation) were adopted for all data analyses.