This study is reported following the REporting of studies Conducted using Observational Routinely collected health Data (RECORD) statement (16). Medical ethical approval and the informed consent were exempted due to the retrospective study design and anonymous information collection.

We extracted patient data from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database (https://mimic.physionet.org/about/mimic/). The description of MIMIC-IV database is available elsewhere (17). In brief, the MIMIC-IV database is a large and publicly available database comprising more than 70,000 patients in the ICU of the Beth Israel Deaconess Medical Center in Boston, MA, USA, between 2008 and 2019. All data were collected before the coronavirus disease 2019 outbreak. To apply for access to the database, we have passed the National Institutes of Health Web-based training course and Protecting Human Research Participants examination (No. 9555299). Data extraction was performed in the Structured Query Language with Navicat Premium (version 15).

In the present study, patients older than 18 years of age admitted to the ICU with an ICU length of stay longer than 72 hours were screened for possible inclusion. Since one patient may be admitted to the ICU multiple times, we only counted the first ICU admission due to non-independence of the outcome among subsequent ICU admissions. Patients were included if they met the Sepsis-3.0 definition upon ICU admission. The details of sepsis-3.0 definition include the presence of an infection with signs of organ dysfunction, which are represented by an increase in the Sequential [Sepsis-related] Organ Failure Assessment (SOFA) score of two points or more (18). Infection was identified from the international classification of diseases (ICD)-9 or ICD-10 code in the MIMIC-IV database. We excluded patients if they developed AKI before ICU admission or after 7 days of ICU admission.

We defined the occurrence of AKI within the first 7 days after ICU admission as the primary outcome of interest. Included patients for whom AKI occurred within 7 days after ICU admission were classified as the AKI group, and the rest of the patients comprised the non-AKI group. The AKI was diagnosed according to the Kidney Disease: Improving Global Outcomes (KDIGO) criteria. KDIGO criteria are as follows (19): increase in serum creatinine (SCr) to ≥ 1.5 times baseline must have occurred within the prior 7 days, or an increase of SCR ≥ 0.3 mg/dl within 48 hours, or urine volume < 0.5 ml/kg/hour for 6 h or more. We used the admission SCr as a baseline value, in accordance with previous studies (20, 21). Patients were categorized into the AKI group (KDIGO stage 3) if they received continuous renal replacement therapy (CRRT) within the first 7 days after ICU admission. Severe AKI was defined as KDIGO stage 2 or higher, otherwise it was defined as mild AKI (22, 23). Secondary outcomes included the ICU and hospital mortality, which were defined as the occurrence of death during the ICU and hospital stay, respectively.

We did several sensitivity and subgroup analyses for the primary outcomes. First, given the considerable impact of age on sex hormones (7), we conducted a stratified analysis according to age in decades (≤40, (40–50], (50–60], (60–70], (70–80], (80–90] and>90 years). Second, to determine whether the illness severity of AKI would influence the results, we further compared patients without AKI to those with mild AKI or with severe AKI. Lastly, we also performed analyses adjusted for the presence of comorbidities.

Baseline characteristics within the first 24 h after ICU admission were collected, including age, sex, body mass index (BMI), ethnicity and admission type. Comorbidities including hypertension, coronary atherosclerosis, heart failure, diabetes mellitus, chronic obstructive pulmonary disease (COPD), cerebral infarction, chronic liver disease, chronic kidney disease, and tumors were also collected for analysis based on the recorded ICD-9 or ICD-10 codes in the MIMIC-IV database. The severity as measured by the Sequential Organ Failure Assessment (SOFA) score, the Acute Physiology Score III (APS III) and the Oxford Acute Severity of Illness Score (OASIS) were calculated upon ICU admission. The use of vasopressor, mechanical ventilation, diuretic, aminoglycoside, statin and angiotensin-converting enzyme inhibitors/angiotensin receptor blockers were also recorded. For the AKI group, an intervention was considered positive if conducted before the onset of AKI and negative otherwise; for the non-AKI group, it was based on the records within 7 days of ICU admission. Vital signs included the heart rate, mean arterial pressure, respiratory rate, temperature and peripheral blood oxygen saturation (SpO₂). Laboratory findings including white blood cell, hemoglobin, platelet, pondus hydrogenii, bicarbonate, blood urea nitrogen, creatinine, potassium, sodium, chloride, glucose, prothrombin time, activated partial thromboplastin time and lactate were measured during the first 24 h in the ICU. If a variable was recorded more than once, we only used the first value for analysis. Detailed information on missing data is available in Supplementary Table S1 . The random Forest-based imputation method was used to impute missing values for these variables (24). The random forest algorithm was implemented using the R package “missForest” (version 1.4; https://CRAN.R-project.org/package=missForest).

We performed the statistical analysis and created pictures by using R statistical software (version 3.6.1, https://cran.r-project.org/), GraphPad Prizm (version 8.0, San Diego, CA, https://www.graphpad.com/scientific-software/prism/) or SPSS software (version 26.0, IBM, USA, https://www.ibm.com/analytics/spss-statistics-software). For continuous variables, we first investigated the normality using the Kolmogorov-Smirnov test. Normally distributed variables are expressed as mean ± standard deviation (SD) and were compared using Student’s t-test. Non-normally distributed variables are presented as the median and interquartile range (IQR) and were analyzed using the non-parametric Mann–Whitney U test. Categorical variables are expressed as count with percentage (%) and were compared with the chi-square test or Fisher’s exact test as appropriate. All tests were two-tailed. A p-value <0.05 was considered significant.

We then used the multivariable logistic regression analyses to assess the independent association between sex and the incidence of SA-AKI. Five different models were built. The first one only included sex, and the second model included sex, age, BMI, ethnicity and admission type. In the third model, we further included comorbidities as covariates. The fourth model additionally included other exposure variables with p-value <0.1 in univariate analyses for multivariable analysis, due to their potential influence on the patient’s primary outcome. In the final model, all baseline variables were included in order to provide a comprehensive assessment. In all models, the effect of sex (women=referent) on primary outcome was presented as an odds ratio (OR) with a 95% confidence interval (95%CI).

Further, we applied three machine learning models to screen for the risk factors: decision tree model, random forest model and Extreme Gradient Boosting (XGBoost) model. A decision tree algorithm is a basic classification method that constructs a model based on the feature of data using a tree structure (25). Feature selection, tree generation, and pruning are the basic steps of building decision trees, and an object-relational mapping relationship was eventually generated (26). In this study, we used the Classification and Regression Tree method with the rpart package (version 4.1.16, https://CRAN.R-project.org/package=rpart) for constructing the decision tree in the R language. Random forest is an ensemble algorithm that combines multiple decision trees, and there is no correlation between each decision tree (27). The voting method is used to discriminate and classify data, and the maximum number of votes is taken as the final classification result (28). In this study, we used the RandomForest package (version 4.7-1, https://CRAN.R-project.org/package=randomForest) in the R language for analysis. XGBoost is an ensemble algorithm composed of multiple decision trees and a gradient boost machine (29). The main advantage of XGBoost is to combine multithreading, data compression, and fragmentation methods to improve the efficiency of the algorithm as much as possible (30). In this study, XGBoost was implemented using the xgboost package (version 1.5.2.1, https://CRAN.R-project.org/package=xgboost) in the R language. Machine learning models have exhibited several advantages over conventional statistical methods, and are gradually used in recent studies (31, 32).

Lastly, we estimated the effect of sex on ICU and hospital mortality for patients with SA-AKI. We used propensity score matching (PSM) to balance baseline variables between men and women. Specifically, matching was performed using R package ‘matching’ (version 4.9-11, https://CRAN.R-project.org/package=Matching), with a ratio of 1:1 and a caliper width of 0.1 without replacement. We considered a standardized mean difference (SMD) of less than 0.1 as acceptable. After propensity matching, differences in ICU and hospital mortality were compared between the two groups. This method has been widely used to compare the mortality between different groups, due to its excellent ability to control measured confounding in observational studies (33, 34).

The process of patient selection is shown in Figure 1 . Finally, a total of 6463 patients were included in our study. Of them, 3673 (56.8%) were men and 2790 (43.2%) were women, with mean age of 66.42 ± 16.58 years. The most common comorbidity was hypertension, followed by heart failure, diabetes mellitus and chronic kidney disease. Men were younger, less likely to have hypertension, heart failure, COPD and cerebral infarction than women; however, men were more likely to have coronary atherosclerosis and chronic kidney disease. Men had lower OASIS and higher SOFA scores than women, but the APS III score was similar between them. In terms of the interventions, more men had mechanical ventilation and statin, while fewer men had diuretic use. The overall incidence of SA-AKI was 83.1% (5368/6463), and men had a slightly higher but not significant incidence than women (83.8% vs. 82.1%, p = 0.068). The detailed information about baseline characteristics of the total cohort, men and women are shown in Table 1 .

In order to search for factors that might influence the incidence of SA-AKI, we compared the baseline characteristics between patients with and without SA-AKI. As shown in Table 2 , we found that the AKI group was older, had a higher BMI, and was more likely to be admitted from the emergency department. The proportion of patients with heart failure, diabetes mellitus, COPD, chronic liver disease and chronic kidney disease were higher in the AKI group, while more patients had tumors in the non-AKI group. In addition, patients with SA-AKI had higher severity scores on admission when compared to those without SA-AKI, including APS III, OASIS and SOFA scores. Patients with SA-AKI were more likely to have vasopressor use and mechanical ventilation when compared to those without SA-AKI; however, a lower proportion of patients in the AKI group had diuretic, aminoglycoside, statin and ACEI/ARBs. Patients with SA-AKI had significantly higher ICU and hospital mortality rates, compared with the patients without SA-AKI (ICU mortality, 19.84% vs. 3.65, p<0.001; hospital mortality, 25.09% vs. 7.76%, p<0.001).

We examined the association between sex and SA-AKI in univariate and extended logistic regression models. We found that men were associated with a higher likelihood of having SA-AKI, but this association became insignificant in the full adjusted model (OR, 1.137; 95% CI, 0.949-1.361; p = 0.163; Table 3 ). Similarly, multiple logistic regression revealed that no significant differences in SA-AKI incidence between sexes when comparing patients without SA-AKI to those with mild SA-AKI or with severe SA-AKI (non-AKI vs. mild AKI, OR 1.161, 95% CI 0.962-1.402, p = 0.120, Table S2 ; non-AKI vs. severe AKI, OR 1.175; 95% CI 0.912-1.513; p = 0.211, Table S3 ). Then, we examined the effect of age on sex-related outcomes. When age was stratified per decade, there was no association between sexes and SA-AKI incidence except for higher rates for men in those aged 50 to 60 years (men vs. women, 19.23% vs. 20.66%, p = 0.196; Figure 2 ). However, in logistic regression adjusted for all baseline variables, the SA-AKI incidence in men and women was approximately the same for all ages ( Figure 3 ). Additionally, we conducted a series of subgroup analyses to determine the effect of comorbidity on the incidence of SA-AKI. The results remained consistent except for patients with cerebral infarction (men vs. women, OR 2.062, 95%CI 1.088-3.905, p=0.026, Table S4 ).

We used three machine learning methods (decision tree model, random forest model and XGBoost model) to evaluate risk factors of SA-AKI in patients with sepsis. Of all patients, BMI, OASIS, diuretic, APS III and age had the five highest values in importance of the models ( Figures 4A –C ). Similar results were also found in men ( Figures 5A –C ) and women ( Figures 6A –C ).

We finally assessed the effect of sex on in-hospital and ICU mortality in patients with SA-AKI. Before propensity-score matching, age, the proportion of patients with coronary atherosclerosis, OASIS, SOFA, the proportion of mechanical ventilation and diuretic usage, hemoglobin, platelet, BUN, creatinine and potassium were different between the men and women groups ( Table 4 ). We found that men had lower but not statistically significant ICU and hospital mortality than women (ICU mortality, men vs. women, 19.23% vs. 20.66%, p = 0.196, Figure 7A ; hospital mortality, men vs. women, 24.20% vs. 26.29%, p = 0.082, Figure 7C ). With the use of propensity-score matching (1:1 matching ratio), 1866 pairs of matched SA-AKI patients were created. After matching, we found no significant imbalance between the two groups, with the standardized mean difference being<0.1 for all variables ( Figure S1 ; Table 4 ). No differences in ICU and hospital mortality between men and women were found (ICU mortality, men vs. women, 19.72% vs. 20.42%, p = 0.595, Figure 7B ; hospital mortality, men vs. women, 25.62% vs. 25.77%, p = 0.991, Figure 7D ).