In this retrospective cohort study, the data were sourced from eICU-CRD, an online international database in the United States (13). The data were accessed and extracted in compliance with the data usage protocol of the PhysioNet review committee, following examination and certification (record ID: 40859994). The database, jointly established by the eICU Institute and Massachusetts Institute of Technology’s Laboratory for Computational Physiology, is a vast public data resource encompassing more than 200,000 ICU patient records from 208 medical institutions in the United States between 2014 and 2015.

To ensure the protection of patient information and eliminate the need for additional informed consent from patients, all patient information involved in the study was de-identified. This study was conducted in accordance with the principles of the Declaration of Helsinki (14).

The study population comprised all patients hospitalized for sepsis. Sepsis was diagnosed based on the presence of a suspected or confirmed infection in conjunction with a rapid increase of more than 2 points in the Sequential Organ Failure Assessment (SOFA) score, derived from the Acute Physiology and Chronic Health Evaluation (APACHE) IV scale (4, 15). Infection identification was based on the International Classification of Diseases, Ninth Edition (ICD-9) codes recorded in the eICU-CRD. The following exclusion criteria were applied during the patient selection process: (1) patients not entering the ICU for the first time, (2) patients with ICU stays of less than 24 h, (3) patients younger than 18 years of age, (4) patients lacking ICU treatment outcome information, and (5) patients lacking sACR data or system recording errors.

The eICU database encompasses a wide array of patient information, including basic demographic data; physiological monitoring parameters from bedside monitors; diseases diagnosed according to the ICD-9, Clinical Modification (ICD-9-CM) codes; and other laboratory results collected during routine medical care. The research team concentrated their efforts on gathering data from the initial 24-h period following the patient’s admission. Vital sign data, including body temperature, respiratory rate, heart rate, and mean arterial pressure (MAP), were extracted from the Apache Aps Var table. Baseline patient characteristics, including age, sex, race, and body mass index (BMI), were collected from the patient and patient results tables. Comorbidities, including diabetes, liver failure, cirrhosis, and metastatic cancer, were identified using the APACHE IV scoring system. Relevant information was extracted from diagnostic tables. Laboratory indicators, including serum albumin, serum creatinine, alanine aminotransferase (ALT), aspartate aminotransferase (AST), and total cholesterol, were extracted from the laboratory table using SQL codes. We used the time of measurement (lab result offset) to select the first recorded values within 24 h of ICU admission as baseline indicators for analysis. The severity of the patient’s condition at admission was assessed using a range of scoring systems, including the SOFA, APACHE IV, Glasgow Coma Scale (GCS), and Acute Physiology Score III.

The study outcome was the mortality rate within 28 days of ICU admission.

The patients were divided into four distinct cohorts, with each cohort corresponding to a quartile of sACR. Subsequently, a comprehensive description of the characteristics of each cohort was provided. Continuous variables that exhibited a normal distribution are expressed as mean ± standard deviation. By contrast, continuous variables that did not adhere to a normal distribution are expressed as median with interquartile range. One-way analysis of variance was used to compare normally distributed continuous variables across groups, whereas the non-parametric Kruskal–Wallis test was used for continuous variables that were not normally distributed. Numerical representations of categorical data are presented as count (N) and relative frequency (percentage). Statistical analyses of group disparities were conducted using the chi-square test.

Univariate and multivariate logistic regression models were used to estimate the correlation between sACR and 28-day mortality. The results are expressed as odds ratio (OR) with a 95% confidence interval (95% CI). The principle of adjusting covariates is determined based on clinical significance and change in the matched OR by at least 10% (16). The following covariates were adjusted for age, sex, BMI, ethnicity, temperature, respiratory rate, heart rate, MAP, blood urea nitrogen level, calcium level, AST level, ALT level, platelet count, hemoglobin level, white blood cell count, GCS score, SOFA score, chronic obstructive pulmonary disease, congestive heart failure, acute myocardial infarction, diabetes, pneumonia, hepatic failure, metastatic cancer, intubation status, mechanical ventilation use, and dialysis.

A generalized additive model (GAM) was used to investigate the dose–response relationship between sACR and 28-day mortality. Subsequently, a two-piecewise linear regression model was employed to investigate the threshold effect of sACR on mortality. When a clear relationship between sACR and 28-day mortality was observed on the smooth curve, software automatically calculated the inflection point using a recursive method after adjusting for potential confounders. Furthermore, a log-likelihood ratio test was conducted to compare univariate and two-piece linear regression models. Considering the heterogeneity of the population, we employed stratified linear regression models for subgroup analyses across different baseline indicators and presented the results as part of our sensitivity analysis. Statistical analyses were conducted using R software (version 4.2.0; R Foundation) and EmpowerStats software¹ (X&Y Solutions, Inc., Boston, MA), and the level of statistical significance was set at a p-value of <0.05.

A flowchart illustrating the study’s inclusion and exclusion criteria is shown in Figure 1. Table 1 presents a comprehensive account of the baseline characteristics of the 9,690 participants, including demographic data, vital signs, laboratory results, infection sites, comorbidities, and disease severity. The participants were divided into quartiles based on their sACR levels: Q1 (0.08–0.96), Q2 (0.96–1.74), Q3 (1.74–2.84), and Q4 (2.84–7.71). The group with lower sACR values exhibited a higher proportion of female patients (Q1: 1,355, 55.97%; Q2: 1,277, 52.73%; Q3: 1,306, 53.94%; Q4: 1,036, 42.72%) and patients with higher BMI (29.98 ± 9.26 kg/m²), along with significant increases in indicators of liver and kidney function impairment, including serum creatinine (Q1: 3.83, 2.93–5.41), ALT (Q1: 29.00, 17.00–70.00), and AST levels (Q1: 45.00, 23.00–119.25). Furthermore, these patients demonstrated markedly elevated Acute Physiology Score III, SOFA score, and APACHE IV score. The 28-day ICU mortality rates differed significantly among the groups (p < 0.001), with the number of deaths in each group as follows: Q1, 391 (16.15%); Q2, 283 (11.68%); Q3, 190 (7.85%); and Q4, 104 (4.29%).

Table 2 presents the various factors associated with sACR and 28-day mortality using a univariate logistic model. The analysis indicated that sACR levels were negatively correlated with the risk of 28-day mortality (OR = 0.65, 95% CI: 0.61–0.70, p < 0.0001). Regarding disease severity scores, an increase in the Acute Physiology Score III and SOFA score was significantly associated with an elevated risk of mortality. It is noteworthy that the group with a SOFA score of 6–21 points exhibited a significantly elevated risk of mortality compared with the lowest-scoring group (0–2 points) (OR = 5.17, 95% CI: 4.16–6.42, p < 0.0001). Furthermore, an increase in the APACHE IV score was significantly correlated with a marked increase in the risk of death. The highest-scoring group (82–193 points) exhibited a markedly elevated risk of mortality compared with the lowest-scoring group (13–60 points). The OR was 9.99 (95% CI: 7.81–12.80, p < 0.0001).

Table 3 presents the findings of the multivariate logistic regression model for sACR and 28-day mortality. In Model I, after adjusting for age, sex, BMI, and race, the inverse correlation between sACR and 28-day mortality remained significant (OR = 0.65, 95% CI: 0.61–0.69, p < 0.0001). Furthermore, in Model II, when controlling for various potential confounding factors, including vital signs, laboratory data, disease severity scores, comorbidities, and treatment modalities, the correlation between sACR and 28-day mortality was somewhat attenuated but remained statistically significant (OR = 0.78, 95% CI: 0.71–0.87, p < 0.0001). Moreover, quartile analysis of sACR demonstrated that, in comparison with the lowest quartile (Q1), the risk of death was significantly reduced in the other three quartiles (Q2, Q3, and Q4) in the unadjusted model. After adjustment for Models I and II, the mortality risk increased in Q2, Q3, and Q4 but remained significantly lower than that in Q1. The trend test yielded a statistically significant result in all models (p < 0.0001), indicating a significant dose–response relationship between sACR and 28-day mortality Table 4.

Considering the heterogeneity of the population, we conducted stratified regression analyses for all baseline indicators; the results are presented in Supplementary Table S1. The results indicated that the direction of the effect size was consistent across various subgroups, including sex, age, presence of intubation, mechanical ventilation, diabetes, and renal failure.

We investigated the potential non-linear correlation between sACR and 28-day mortality in the entire cohort using GAM and smooth curve fitting while accounting for variables such as age, sex, BMI, race, vital signs, and disease severity scores (Figure 2). As illustrated, there was a decreasing trend in 28-day mortality in the ICU with an increase in sACR, indicating a non-linear correlation.

The threshold effect of sACR on 28-day mortality was analyzed using a two-piecewise linear regression model. Model I demonstrates that for each unit increase in sACR, the OR for 28-day mortality was 0.78 (95% CI: 0.71–0.87), with a p-value <0.0001, indicating a significant association between sACR and the reduced risk of death. Model II demonstrated a threshold effect when the sACR index was less than 4.79, wherein a significant association was observed between sACR and a reduced risk of death (OR = 0.75, 95% CI: 0.67–0.84, p < 0.0001). Conversely, above this threshold, the observed association was not statistically significant (OR = 1.19, 95% CI: 0.72–1.96, p = 0.49). Several covariates, including age, sex, and BMI, were adjusted for in the analysis. The results demonstrated a non-linear relationship between sACR and 28-day mortality rate, with a particularly pronounced association below a specific threshold.