Data source
This retrospective observational study relied on data from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) dataset16,17. MIMIC-IV is an openly accessible critical care medicine dataset. The most recent version (2.2) was released on January 6, 2023. This database contains over 50,000 de-identified patient records from individuals admitted to the Intensive Care Unit (ICU) at Beth Israel Deaconess Medical Center in Boston, Massachusetts between 2008 and 2019. The Institutional Review Board of Beth Israel Deaconess Medical Center approved the waiver of informed consent and the sharing of research resources. The authors obtained permission to access the database, as indicated by Certificate Number 56161429. This study was conducted in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines18.
Study population
We obtained a cohort of 9,645 adult patients (aged 18 years or older) from the MIMIC IV database who were diagnosed with sepsis and admitted to the ICU. Patients who were missing SI data on the first day of hospitalization were excluded from the analysis. The diagnosis of sepsis was based on the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3), which required either suspected or documented infection and a minimum two-point increase in the Sequential Organ Failure Assessment (SOFA) score. Because pre-hospital organ-specific data are unavailable in MIMIC-IV, we assumed a baseline SOFA of 0 for all patients and defined organ dysfunction as a first-day ICU SOFA ≥ 2. For patients with multiple admissions, only their first ICU admission was considered. The final cohort was divided into high and low SI groups based on the optimal SI cut-off value (100 µg/dL)19,20. The flow chart depicting the patient screening process is presented in Fig. 1.
The flowchart of extracting patients in this study.
Data extraction
To extract the data, structured query language (SQL) script codes were obtained from the GitHub website (https://github.com/MIT-LCP/mimic-iv) 21 Various characteristics of the patients, including their age, sex, body mass index, and Charlson comorbidity index, were collected for analysis. Using the International Classification of Diseases coding systems 9 and 10, we extracted information on comorbidities such as hypertension, cerebrovascular disease, myocardial infarction, congestive heart failure, chronic lung disease, liver disease, diabetes mellitus, renal disease, cancer, and AIDS.
To operationalize the Sepsis-3 criteria, we defined the onset of sepsis as the first co-occurrence of suspected infection (antibiotic administration and culture sampling) and acute organ dysfunction (SOFA score ≥ 2) within a 24-h window of ICU admission.,Baseline data were collected, including the severity of illness (Sequential Organ Failure Assessment [SOFA] score, Oxford Acute Severity of Illness Score [OASIS], Acute Physiological Score III [APSIII]), vital signs (systolic blood pressure (mmHg), diastolic blood pressure (mmHg), mean arterial pressure (mmHg), heart rate (beats min-1), respiratory rate (min-1), pulse oximetry (%), and body temperature (℃)), and laboratory results (SI level(ug/dl), hemoglobin level(g/dl), platelet count(K/ul), white blood cell count(K/ul), prothrombin time(S), partial thromboplastin time(S), international normalized ratio, blood urea nitrogen level(mg/dl), creatinine level(mg/dl), electrolytes, and blood gas analysis.
Data on different types of microorganisms and therapies, including mechanical ventilation and vasopressors, were also collected. If a variable was recorded more than once during the first 24 h, the mean value was used.
Primary outcome and secondary outcomes
The primary outcome of this study was 28-DACM. Secondary outcomes included in-hospital mortality, 90-day all-cause mortality, length of stay in days in the intensive care unit (ICU), and total hospital length of stay. For the assessment of outcomes, including in-hospital and 28-day mortality, we utilized the MIMIC-IV database, which captures both in-hospital death records and death dates from state records for up to one-year post-discharge. This comprehensive data collection allowed us to accurately determine the mortality status of patients for a continuous 28-day period following their hospitalization, regardless of whether the death occurred within the hospital or after discharge.
Statistical analysis
The optimal cut-off value for the SI in a receiver operating characteristic (ROC) curve was determined as the value with the highest sum of sensitivity and specificity. This approach was selected to pinpoint the SI level that best balances sensitivity and specificity in predicting 28-DACM, offering a potentially more insightful threshold compared to quartiles or the median. Given that deleting data with missing values would result in the loss of a significant amount of useful information, we adopted a multiple imputation approach for data with less than 30% missing values (Supplementary Figure S1). This allows us to make full use of the existing data information to enhance the accuracy of our predictions. For variables with missing values less than 30%, the MICE package for multiple imputations was used22. Categorical variables were expressed as frequencies or percentages and analyzed using the chi-squared test or Fisher’s exact test. The Shapiro–Wilk test was performed on continuous variables to determine the normality assumption. Continuous variables were described by their mean ± standard deviation or median and interquartile range, depending on the normality or skewness of the variables. For the comparison of continuous variables, normally distributed data were analyzed using the independent samples t-test, while non-normally distributed data were assessed with the Mann–Whitney U test. Cox models were developed to calculate the hazard ratio (HR) with a 95% confidence interval (CI) for the main outcome (i.e. (28-DACM)). Model I comprised age + race + sex + BMI, + Charlson Comorbidity Index. Model II: Model I + mean heart rate + MAP + mean respiratory rate + + mean temperature + + mean SpO2 + mean platelets + white blood cells + lactate + pH + creatinine + sodium + potassium, and INR. Model III Model II + SOFA score + OASIS score + APSIII. Model IV, based on Model II, was further adjusted for the use of vasoactive medications, mechanical ventilation, and isolated microorganisms. The cumulative incidence of 28-DACM was determined using Kaplan–Meier estimates, and differences were assessed using the log-rank test. To accurately assess the relationship between SI levels and the hazard ratio (HR) for 28-DACM, we utilized a restricted cubic spline (RCS) regression model, employing five knots as per Stone’s recommendation23 suitable for our sample size exceeding 500. The application of five knots is strategic for attaining a nuanced model fit that ensures the curve’s smoothness, thereby preventing overfitting and maintaining the precision of our estimates. This approach is particularly beneficial for capturing the subtleties of potential non-linear trends within our extensive dataset24. We conducted Pearson correlation analysis to assess the linear relationships between SI levels and other continuous variables included in Model IV. This step is a common practice in regression analysis to ensure that the assumptions of the linear model are not violated, particularly multicollinearity, which can inflate the variance of the regression coefficients and affect the stability of the model. This preliminary analysis ensures that our model is robust and the results are reliable. Statistical analyses were performed using R software version 4.3.1 (R Foundation for Statistical Computing, Vienna, Austria), and a 2-tailed P-value < 0.05 was considered statistically significant.
Sensitivity analysis
Subgroup analyses were conducted to evaluate the robustness of the primary results according to prespecified patient characteristics that could impact outcomes. To explore potential interaction effects, we examined the interactions between SI and key subgroup-defining variables by introducing product terms into our models. The subgroups included: age (< 60 vs ≥ 60 years), sex (female vs male), Charlson Comorbidity Index (< 6 vs 6 or higher)25, use of mechanical ventilation (yes vs no), and use of vasopressors (yes vs no). Hazard ratios (HRs) with 95% confidence intervals (CIs) were estimated from Cox proportional hazard models for each subgroup. Tests for interaction were performed to determine whether the association between SI levels and 28-DACM differed significantly across the subgroups.
