Study design and setting
The present study utilized the TTE method based on observational data to assess the impact of early albumin use on the development of SA-AKI in sepsis patients. This study employed the clone-censor-weight (CCW) method to reduce immortal time bias, along with a new-user design to address current user bias [14, 15]. Specifically, to mitigate current user bias, only records of patients receiving albumin for the first time were selected. A detailed explanation of the CCW method is provided in Additional file 1: Methods.
Patients admitted to the ICU between 2008 and 2022 were included in the study, with sepsis diagnosed according to Sepsis-3 criteria, which define sepsis as suspected infection combined with an acute increase in Sequential Organ Failure Assessment (SOFA) score ≥ 2 points and acute kidney injury (AKI) identified using the Kidney Disease Improving Global Outcomes (KDIGO) guidelines [16, 17]. The definition of SA-AKI in this study follows the criteria established by the 28th ADQI workgroup, which defines SA-AKI as AKI occurring within 7 d of sepsis onset [7]. The patient’s survival time was calculated from the baseline to the time of death. The study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines [18].
Data source
All data for this study were sourced from the Medical Information Mart for Intensive Care IV (MIMIC-IV), a publicly accessible database that is widely utilized in medical research [19,20,21]. MIMIC-IV v3.1 comprises hospitalization records from approximately 364,627 patients who received emergency or intensive care treatment at Beth Israel Deaconess Medical Center in Boston from 2008 to 2022 [22]. The database ensures patient anonymity, eliminating the need for informed consent. Access to the data was granted to the authors upon completion of the necessary training and certification [20].
Eligibility criteria
Patients over the age of 18 who had been diagnosed with sepsis, and who were admitted to the ICU for the first time and had an ICU stay longer than 24 h, participated in this study. The exclusion criteria were as follows: 1) patients who developed AKI before the onset of sepsis, 2) patients who received albumin before sepsis onset, and 3) patients with incomplete or incorrect time records. The exclusion criteria did not include post-baseline information [23].
TTE
TTE is an approach that constructs a framework, like RCT, from observational data, simulating a design for an ideal trial that could be implemented in practice. By explicitly delineating the temporal configuration of the research question, the inclusion criteria, the intervention strategies, the outcome definitions, and the analytical protocols, TTE seeks to minimize biases such as immortal time bias and prevalent user bias, thereby enhancing the interpretability and the causal validity of the observational findings [13, 14]. Additional file 1: Table S1 provides a comparison of TTE with RCT and observational studies. In this study, following the principles of TTE [13, 14, 23, 24], the study initiation point, exposure assessment window, intervention and control allocation strategies, and outcome monitoring during follow-up were defined with precision. Using the TTE method, we established a causal inference-compliant framework (Additional file 1: Table S2) to better estimate the potential causal impact of early albumin administration on sepsis-related outcomes.
Treatment strategy and assignment
In this study, the time of sepsis onset was defined as the baseline time. The population was classified into the “Albumin group” (treatment group, n = 27,088) and the “No albumin group” (control group, n = 27,088) based on whether patients received albumin within 24 h after the onset of sepsis [25]. It is not equivalent to a placebo arm.
Follow-up and outcomes
Patients were observed until the following outcomes occurred: the onset of SA-AKI, death, discharge, or the expiration of the 7-day follow-up period.
Covariates
Baseline, time-dependent, and post-baseline covariates were selected for adjustments [24]. The type and codes of each variable are shown in Additional file 1: Table S3. The selection of covariates was primarily informed by literature review [7, 26] and expert clinical knowledge, to identify variables that may plausibly confound the relationship between exposure and outcome. For instance, the SOFA score has been demonstrated to reflect illness severity, which is closely related to liquid management strategies [27]. Comorbidities such as chronic kidney disease (CKD) and diabetes have been demonstrated to modify the outcome risk [7, 26]. A detailed explanation regarding the selection of covariates for inverse probability-censored weighting (IPCW) is provided in Additional file 1: Methods. The final list of adjusted covariates includes age, sex, race, year of admission, type of ICU, weight, length of stay before ICU admission, SOFA score, Acute Physiology Score III (APSIII) score, Charlson Comorbidity Index (CCI), crystalloids dosage, time to antibiotic use, hypertension, diabetes, CKD, artificial colloid use, nephrotoxic drug exposure, vasoactive agent use, mechanical ventilation, and renal replacement therapy (RRT).
Sensitivity and subgroup analyses
To ensure the robustness of the results, 3 sensitivity analyses were conducted. In the first 2 sensitivity analyses, the follow-up grace periods were adjusted to 12 h and 36 h to assess whether the effect of early albumin use remained consistent across different grace periods. In another sensitivity analysis, patients who had received RRT were excluded, and all subsequent sensitivity analyses were repeated using CCW and the previously described analytical procedures. Finally, subgroup analyses were performed based on age, sex, race, CKD, and septic shock to evaluate the potential influence of these variables on outcomes. All 95% confidence intervals (CIs) for sensitivity and subgroup analyses were obtained using a nonparametric bootstrap method with 1000 repetitions.
Statistical analysis
This study utilized a TTE framework. In observational data, patient treatment intention is unknown. Therefore, the causal contrasts in the TTE are the per-protocol effect [28]. Continuous variables that did not meet the assumption of a normal distribution were reported as medians with interquartile ranges (IQRs) and compared using the Wilcoxon rank-sum test. Categorical variables were presented as frequencies and percentages, with group differences assessed using the chi-square test.
Initially, each patient in the original dataset was replicated, thus creating 2 identical individuals with identical baseline characteristics. These individuals were then assigned to the treatment and control groups. Following the generation of the cloned dataset, human censoring was applied, meaning that patients who deviated from the planned protocol were censored. Specifically, clones in the treatment group who did not receive albumin within the follow-up grace period and clones in the control group who did receive albumin were censored (Additional file 1: Fig. S1). It is noteworthy that most patients do not receive albumin treatment immediately following the onset of sepsis. Consequently, a follow-up grace period was introduced, and the length of the follow-up grace period was set at 24 h [23, 29]. Finally, IPCW was applied to mitigate selection bias that had been introduced by censoring. The weights were truncated at the 1st and 99th percentiles [30]. Additionally, the mean values of covariates over time were visualized before and after weighting (Additional file 1: Figs. S2, S3).
Subsequent analyses were conducted using CCW datasets. The risk difference was calculated as the relative difference in the incidence of the outcome between the treatment and control groups during the specified follow-up period [28]. The development of weighted cumulative risk curves and weighted Kaplan-Meier survival curves was plotted to illustrate the differences in SA-AKI risk between the treatment and control groups, as well as the differences in 7-day survival. In the SA-AKI analysis, death was considered a competing event. To calculate 95% CIs for the differences in SA-AKI risk, 7-day survival, restricted mean time lost (RMTL) [31], and restricted mean survival time (RMST) [32], a nonparametric bootstrap method with 1000 repetitions was used. However, it should be noted that the use of hazard ratios is not recommended for causal inference [28, 33].
Only variables with a missing rate of 20% or less were included in the final analysis. A comprehensive summary of missingness across all candidate variables is provided in Additional file 1: Fig. S4. For variables with a missing rate below this threshold, random forest (RF) imputation was performed using the “mice” package in R [34]. To ensure robustness, sensitivity analyses were performed using predictive mean matching (PMM) and weighted predictive mean matching (midastouch) (Additional file 1: Fig. S5).
Statistical analyses were performed using R v4.4.2 software (https://www.r-project.org/). Data from the MIMIC-IV databases were extracted using Structured Query Language (SQL) [35], leveraging SQL scripts from the official MIT (Massachusetts Institute of Technology) Laboratory for Computational Physiology GitHub repository (https://github.com/MIT-LCP/mimic-code). P < 0.05 (two-tailed) was considered statistically significant.