Study design
This retrospective cohort study was conducted on surgical cases performed from November 2008 to 2019 using the MIMIC-IV database (version 2.0). MIMIC-IV is an available and real-world database that consists of 76,943 ICU cases and provides 30-day follow-ups. This database facilitates comprehensive healthcare exploratory research. Additionally, this extensive database contains a wide range of information, including laboratory measurements, documented vital signs, medication administration, procedure records, outcome events, and other medical records related to patients admitted to the ICU or emergency department at the two medical centers [11].Database exploration clearance was obtained from author JY (certification ID:11088373). The adherence of this study to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines has been ensured [12].
Participants
Approval for the establishment of the MIMIC-IV database has been granted by the Institutional Review Boards (IRBs) of Massachusetts Institute of Technology (0403000206) and the Beth Israel Deaconess Medical Center (2001-P-001699/14). Due to the anonymized and accessible nature of the data utilized in this investigation, specific informed consent from individuals was not necessary [13, 14]. In this research, adult patients who were admitted to the ICU for the first time post-CABG procedures at Harvard Medical School and MIT, were included. To gather information about the cohort, the diagnosis according to the International Classification of Diseases, Ninth and 10th Revision (ICD-9, ICD-10) was utilized. Exclusion criteria involved individuals with missing data on survival time (<0 days), stage of acute kidney injury (AKI) or ventilation time, ICU stay (<1 day), and incomplete or unavailable vital medical records. Eventually, the final analysis included a total of 5224 patients (Fig 1).
Data sources and measurement
The cohort study analyzed relative data derived from the MIMIC-IV database in America. In order to extract data, Navicat Premium software (version 16.0.11) was utilized, employing Structured Query Language. The code for data extraction can be found on GitHub [15]. Vital sign data were collected, and the mean values during the ICU stay were calculated. The following data were extracted: basic demographic information (sex, age, weight, height, marital status, race, and admission type, et al.); comorbidities factors (myocardial infarct [MI], peripheral vascular disease [PVD], cerebrovascular disease [CD], mild liver disease [MLD] congestive heart failure [CHD], chronic pulmonary disease [CPD], diabetes with chronic complications [DWCC], hypertension, hyperlipidemia, and atrial fibrillation [AA], etc.); vital signs (heart rate [HR], diastolic blood pressure [DBP], respiration rate [RR], etc.); severity of illness (Sequential Organ Failure Assessment (i.e., “SOFA”) score, and Charlson Comorbidity Index); laboratory investigations (hemoglobin, lactate, pH, bicarbonate, sodium, potassium, and chloride etc.); treatment (ventilation use, continuous renal replacement therapy, norepinephrine, epinephrine etc.); and calculated survival time. During the initial week in the ICU following the CABG procedure, the criteria for AKI were determined by employing both the Kidney Disease Improving Global Outcomes (KDIGO) criteria for serum creatinine and urine output (UOP) [16].
Variable definition
Blood chloride concentration was presented as millimoles per liter (mmol/L). Body mass index (BMI) was determined by dividing an individual’s weight (in kilograms) by the square of their height (in meters). Calcium (mg/dl) × 0.2495, magnesium (mg/dl) × 0.4114, and glucose × 0.0555 were converted to standardized units (mmol/L). CABG was defined as a history of MI requiring open-heart surgery. Thirty-day death was calculated by subtracting the ICU admission time from the death time as the primary outcome. We defined AKI by using both the Kidney Disease Improving Global Outcomes (KDIGO) serum creatinine and urine output (UOP) criteria[16] during the first week in ICU after the CABG procedure.
Potential confounding factors
Significant variables (P < 0.2 from the generalized linear model (GLM) and univariable logistic regression models) and those with P < 0.1 in the final multivariable logistic regression models [17-19]. The selection details of potential confounding factors are summarized in Supplementary File 1. Six models (1, 2, 3, 4, 5, and 6 [Table 2]) were explored sequentially, including covariates to adjust for relative confounders. These factors included patients’ demographic variables (age, sex) and comorbidity variables (MI, PVD, CD, and MLD). The analyses were further adjusted for admission type. Finally, analyses were adjusted for sodium, glucose, lactate, bicarbonate, blue light, and norepinephrine use.
Quantitative variables
Serum Cl– was selected as thequantitativevariable. The cohort was divided into Q1 (90–103.7 mEq/L), Q2 (103.8–106.6 mEq/L), and Q3 (106.7–117.5 mEq/L) groups according to tertiles of blood-Cl–-levels.The Q1 and Q2 groups were the targets of interest, as abnormal blood-Cl–-levels were hypothesized to increase 30-day mortality.
Secondary analysis
The models included all factors chosen from the analysis of univariable logistic regression to control for any potential known confounders. The objective was to evaluate the correlation between levels of blood- Cl– and mortality within 30 days in this particular cohort. K–M curves were plotted to illustrate differences in 30-day mortality among the three groups (i.e., Q1–Q3). We also examined the association between the fluctuation of blood-Cl–-levels and mortality within 30 days. Subsequently, the regression model was adjusted for the threshold of Cl– level and the association between each 1mmol/L change in blood-Cl–-levels and 30-day mortality around the threshold was tested.
Exploratory and sensitivity analyses
Additional sensitivity analyses were performed in this study. With exploratory intent, the OR increases for 30-day mortality per 1 mmol/L change in Cl– were evaluated. The subgroup variables were sex (female, male), age (<65, ≥65 years), MI, PVD, CD, MLD, BMI (< 25, ≥ 25 kg/m2), and elective surgery.
Statistical analysis
Details regarding the missing variables were summarized in Supplementary File 2. Missing data were handled with multiple imputations using chained equations with three imputed datasets before statistical analysis. Variables pertaining to continuous data distribution (both normal and nonnormal) and categorical data were presented using various descriptive statistics. The mean with standard deviation (SD), median with interquartile range (IQR), and frequency with percentage (%) were utilized for continuous, nonnormally distributed, and categorical variables, respectively. We tested for statistical differences in percentage, median, and mean among the three groups using the chi-squared (χ2) test, Kruskal–Wallis test, and one-way analysis of variance (ANOVA), respectively. To investigate the nonlinear correlation between blood chloride levels and mortality within 30 days in ICU patients after CABG procedures, we utilized RCS and GAM techniques.
The K–M survival curve revealed statistical differences in 30-day mortality among the three groups. Potential known factors were introduced in the univariate logistic analyses to investigate the impact of changes in blood-Cl–-levels on 30-day mortality, and then with the included confounders into the multivariate logistic analyses. Six models were applied in the multivariate logistic regression analysis by gradually adding variables to the models, as follows: Model 1 was crude, without any adjustment; sex was added for Model 2; additionally, adjusted age for Model 3; next, MI, PVD, CD, and MLD were introduced to Model 4; then, plus admission type for Model 5; at last, sodium, glucose, lactate, bicarbonate, blood urea nitrogen, and norepinephrine were added to adjust for Model 6. The analysis reference was a Cl– level of 107.0 mmol/L, and adjustments were made for the aforementioned (model 6) covariates.Furthermore, blood-Cl–-levels were measured at specific percentiles (5th, 35th, 65th, and 95th) and marked with four knots. To assess the correlation between Cl– and the 30-day mortality, we performed a threshold analysis. This entailed comparing a model containing solely a linear term to a model including both linear and cubic spline terms through a likelihood ratio test. Additionally, we examined the potential influence of confounding factors on the relationship between blood-Cl–-levels and mortality within a 30-day period. Therefore, subgroup analysis with stratified multivariable logistic regression models was conducted, and the interactions between subgroups were examined using likelihood ratio testing.
We introduced unadjusted and adjusted ORs of the logistic regression models to report the results. Two-tailed tests and corresponding 95% CIs were applied for all analyses. p ≤ 0.05 and p < .017 were considered significant differences in the pairwise and multiple comparison analyses. We conducted all analyses utilizing R (http://www.R-project.org, The R Foundation for Statistical Computing, Vienna, Austria) and the Free Statistics software, version 1.8.