Early identification of acute kidney injury progression in critically ill patients with sepsis: interpretable machine learning approach Authors: Xiaodong You, Yanlong Chen, Jiahui Chen, Xinyi Mao, Yifei Wang, Zhongman Zhang, Zheng Zhou, Wei Li, Yong Mei, Yi Zhu, Peipei Huang, Xufeng Chen Published: 2026 ABSTRACT ABSTRACTBackgroundAcute kidney injury (AKI) is a common and severe complication of sepsis and is often associated with a poor prognosis. However, there is still a lack of an effective prediction model for early identification of AKI progression in critical septic patients, defined as AKI stage 1 or 2 to stage 3 within 7 days after diagnosis of sepsis-associated AKI (SA-AKI).MethodsWe extracted the clinical data of patients with SA-AKI from the Medical Information Mart for Intensive Care (MIMIC) datasets, eICU Collaborative Research Database (eICU-CRD) and Salzburg Intensive Care database (SICdb), with the MIMIC-IV (version 3.1) database used for training and internal validation, the MIMIC-III Clinical Database CareVue subset used as temporal validation, and the eICU-CRD and SICdb used as external validation. Lasso regression and recursive feature elimination were used for feature selection. Six machine learning (ML) algorithms, including k-nearest neighbors, logistic regression, naïve Bayes, random forest (RF), support vector machine and decision tree, were utilized to establish the prediction model. Model performance was assessed using receiver operating characteristic curves, calibration curves and decision curve analysis. SHapley Additive exPlanations (SHAP) method was used for the interpretation of the models.ResultsThe MIMIC-IV, MIMIC-III subset, eICU-CRD and SICdb included 9193, 2178, 10 332 and 1701 patients with SA-AKI. Twelve variables were selected for model construction, including weight, liver disease, mechanical ventilation, systolic blood pressure, hemoglobin, glucose, blood urea nitrogen, creatinine, chloride, anion gap and urine output. An RF model achieved the best performance in both internal, temporal and external validation (area under the curve is 0.779, 0.758 and 0.713, respectively). A user-friendly platform was built to early predict SA-AKI progression for clinician use.ConclusionML could be a useful tool for predicting AKI progression in septic patients. We developed an RF model to predict the risk of SA-AKI progression, which may provide a reference for early identification and prompt intervention of high-risk group. What was known: Previous studies have often primarily concentrated on early diagnosis and prognostic assessment of sepsis-associated acute kidney injury (SA-AKI). Previous studies have often primarily concentrated on early diagnosis and prognostic assessment of sepsis-associated acute kidney injury (SA-AKI). A variety of diagnostic and prognostic models for SA-AKI based on traditional regression methods or machine learning (ML) methods have been reported recently. A variety of diagnostic and prognostic models for SA-AKI based on traditional regression methods or machine learning (ML) methods have been reported recently. However, there is still a lack of research on early identification of SA-AKI progression in the intensive care unit (ICU), and its prediction models remain undeveloped. However, there is still a lack of research on early identification of SA-AKI progression in the intensive care unit (ICU), and its prediction models remain undeveloped. This study adds: This study introduced a useful and reliable prediction model for SA-AKI progression in the ICU using ML algorithms. This study introduced a useful and reliable prediction model for SA-AKI progression in the ICU using ML algorithms. We also constructed a user-friendly online platform to facilitate its clinical application. We also constructed a user-friendly online platform to facilitate its clinical application. Potential impact: This prediction model could help clinicians understand the underlying mechanism of SA-AKI progression more comprehensively, and assist them in early identification of AKI progression in critically ill patients with sepsis. This prediction model could help clinicians understand the underlying mechanism of SA-AKI progression more comprehensively, and assist them in early identification of AKI progression in critically ill patients with sepsis. The user-friendly online platform we built in this study further facilitates the clinical application and could help clinicians achieve early intervention and management of the high-risk group. The user-friendly online platform we built in this study further facilitates the clinical application and could help clinicians achieve early intervention and management of the high-risk group. INTRODUCTION Acute kidney injury (AKI) is a common and severe complication resulting from multiple factors, which affects more than 50% of patients in the intensive care unit (ICU) [1]. Severe infection and sepsis are the most frequent causes of AKI, which is well-known as sepsis-associated AKI (SA-AKI) [2]. It is reported that the global incidence of SA-AKI might be approximately 240.4 per 100 000 population [3, 4]. Due to the complex pathophysiology of SA-AKI, including systemic and renal inflammation, altered microcirculatory and endothelial function, and complement activation, septic patients who developed AKI often have a poor prognosis, with an in-hospital mortality of approximately 25% [5]. There remain great challenges in the prevention and management of SA-AKI. According to the Kidney Disease: Improving Global Outcomes (KDIGO) criteria, AKI was classified into three stages based on severity, encompassing stages 1, 2 and 3 [6]. Compared with septic patients with AKI stage 1 or 2, those with AKI stage 3 tend to have a higher risk of mortality and are more likely to develop chronic kidney disease and kidney failure [3, 7]. Therefore, early identification of AKI progression is clinically critical for improving the overall prognosis of patients with SA-AKI admitted to the ICU. Previous studies have often primarily concentrated on early diagnosis and prognostic assessment of SA-AKI [8–11]. There is still a lack of research on early identification of SA-AKI progression in the ICU. Due to severe tubular necrosis, consistent microcirculatory dysfunction, and greater systemic and renal inflammation of SA-AKI stage 3, precise prediction of SA-AKI progression is crucial for reducing complications of these patients and enhancing their quality of life [5, 7]. Although KDIGO criteria clearly defined AKI staging and diagnosis based on urine output and creatinine, their clinical utility in early alert of SA-AKI progression is limited. This is mainly because serum creatinine levels increase late after kidney impairment and are easily influenced by individual baseline characteristics [12, 13]. Additionally, several new biomarkers such as tissue inhibitors of metalloproteinase 2 and insulin-like growth factor binding protein 7 have been proven to be early and accurate indicators for AKI progression [14]. However, due to their high cost and insufficient clinical evidence, these biomarkers have not been widely applied in clinical practice. Machine learning (ML) is an important branch of artificial intelligence and has achieved rapid development and wide application in the field of medicine [15, 16]. In recent years, several ML models have been constructed for early prediction of SA-AKI and their accuracy has also been validated [8, 9]. Besides, prognostic models based on ML methods have also been designed for patients with SA-AKI in the ICU, and they exhibited great performance in predicting the risk of short-term mortality [10, 11]. However, there is no study examining the prediction of SA-AKI progression. In this study, we aimed to build a model for the early identification of AKI progression in critically ill patients with sepsis using ML methods, and to construct a user-friendly online platform to help clinicians achieve early identification and intervention of high-risk populations. MATERIALS AND METHODS Data source The data used in this study came from the Medical Information Mart for Intensive Care (MIMIC), eICU Collaborative Research Database (eICU-CRD) and Salzburg Intensive Care database (SICdb), with the MIMIC-IV (version 3.1) database used for training and internal validation, the MIMIC-III Clinical Database CareVue subset used as temporal validation, and the eICU-CRD and SICdb used as external validation [17–19]. The MIMIC-III Clinical Database CareVue subset is a subset of the MIMIC-III database and contains clinical information on 23 692 patients between 2001 and 2008, so there was no overlap with the MIMIC-IV database which includes 546 028 admissions from 2008 to 2022. As the MIMIC database, eICU-CRD and SICdb were de-identified, informed consent from patients was not required. Participates Patients diagnosed with sepsis within 24 h after ICU admission were included. Then, those patients who developed AKI (stage 1 or 2) within 48 h of their initial sepsis diagnosis were further selected. Septic patients were determined by Sepsis-3 criteria, and AKI was defined by KDIGO criteria. Patients meeting the following criteria were excluded: (i) <18 years old; (ii) with pre-existing chronic kidney disease; (iii) ICU stay <48 h; and (iv) receiving renal replacement treatment before diagnosis of SA-AKI. The process of patient screening is shown in Fig. 1. Flowchart of the study population selection. Definition and outcome Sepsis was defined as a life-threatening organ dysfunction caused by a dysregulated host response to infection according to Sepsis-3 criteria [20]. Organ dysfunction was assessed using the Sequential Organ Failure Assessment (SOFA) score, with an increase of 2 or more points indicating sepsis. AKI was defined in consistency with KDIGO criteria as follows: an increase in serum creatinine of at least 0.3 mg/dL within 48 h; or an increase in serum creatinine to at least 1.5 times the baseline within the previous 7 days; or a urine output of <0.5 mL/kg/h for 6 h or more. AKI is further classified into three stages of severity based on serum creatinine or urine output according to KDIGO criteria. If creatinine stage and urine output stage were inconsistent, the higher AKI stage derived from either criterion was assigned for staging. The primary outcome of this study was SA-AKI progression, defined as AKI stage 1/2 to AKI stage 3 within 7 days after diagnosis of SA-AKI. SA-AKI was defined as AKI stage 1/2 development within 48 h after their initial sepsis diagnosis. The secondary outcome was 28-day mortality. Data extraction and pre-processing Data extraction was conducted using Navicat Premium software (version 16.3.8) via structured query language. We collected the following variables: (i) demographic characteristics: age, gender, height and weight; (ii) comorbidities: hypertension, diabetes, acute myocardial infarction, congestive heart failure, chronic pulmonary disease and liver disease; (iii) interventions: vasoactive agents and mechanical ventilation (MV); (iv) vital signs: heart rate, systolic blood pressure, diastolic blood pressure, mean blood pressure, respiratory rate, temperature, saturation of peripheral oxygen (SpO2); (v) measurements: white blood cell, platelet, hemoglobin, blood glucose, blood urea nitrogen (BUN), serum creatinine, sodium, potassium, chlorine, anion gap, and urine output; and (vi) survival information: length of ICU stay and 28-day mortality. The maximum and minimum values of the first day after ICU admission were taken for multiple measurements, except for SpO2. The percentage of missing values was calculated for all continuous variables to minimize bias (Supplementary data, Fig. S1). Variables with missing values exceeding 20% were excluded, and for variables with missing values <20%, the K-Nearest Neighbors (KNN) imputer method was employed. Outliers were identified using the interquartile range (IQR) method and handled utilizing the winsorization technique, with 5% and 95% regarded as cut-off points [21]. Statistical analysis The normality of continuous variables was determined using the Shapiro–Wilk test. Continuous variables with normal distribution are expressed as mean ± standard deviation, while others are expressed as median (IQR). Categorical variables are presented as counts (%). We used the Kruskal‒Wallis test to compare continuous variables and the chi-square test to compare categorical variables. Lasso regression and recursive feature elimination (RFE) were utilized to select features. The iterative analysis was performed using a 10-fold cross-validation method. After feature selection, six ML algorithms, including KNN, logistic regression (LR), naïve Bayes (NB), random forest (RF), support vector machine (SVM) and decision tree (DT), were used to construct the prediction model. The features selected by lasso regression and RFE were included in the model. A 10-fold cross-validation was employed to prevent overfitting and ensure the stability of the model. The MIMIC-IV data served as an internal validation set, the MIMIC-III subset was used as the temporal validation set, and the eICU-CRD and SICdb were used for external validation. The models were validated in the internal validation set, temporal validation set and external validation sets after construction, respectively. The performance of the prediction models was assessed by receiver operating characteristic (ROC) curves, calibration curves and decision curve analysis (DCA). The optimal model was chosen according to calibration capability, DCA and area under the curve (AUC). Specifically, the best model was chosen based on AUC if its calibration capability (slope >0.6) and clinical utility were satisfied. The SHapley Additive exPlanations (SHAP) method was used to explain the best model. SHAP explains the prediction model by assigning each feature a contribution value [22]. It provides both global and local interpretability, making complex models more transparent and understandable. Feature importance was assessed by the global interpretability of SHAP, and individual prediction was achieved by the local interpretability of SHAP. To achieve better clinical promotion and facilitate the ML model clinician application, we built a user-friendly online platform using the R package “Shiny.” After inputting the corresponding data of the patients with SA-AKI, the prediction can be made automatically and obtained quickly. A two-tailed P-value <.05 was considered statistically significant. All statistical analyses were conducted using R software (version 4.3.3). RESULTS Baseline characteristics A total of 11 371 patients with SA-AKI from the MIMIC database were enrolled in this study, with 9193 from the MIMIC-IV database and 2178 from the MIMIC-III subset, of whom 2147 (23.4%) and 664 (30.5%) developed AKI progression, respectively. In the eICU database, there were 10 332 patients with SA-AKI admitted to the ICU, of whom 2570 (24.9%) developed AKI progression. In the SICdb, a total of 1701 patients with SA-AKI were included, of whom 353 (20.8%) developed AKI progression. The median time (days) from SA-AKI onset to hospital discharge of MIMIC cohort, eICU cohort and SICdb cohort was 7.27 (4.41, 13.30), 6.97 (4.04, 12.20) and 7.00 (4.00, 13.00), respectively. Table 1 shows the baseline characteristics of patients in the MIMIC-IV cohort. Male patients with SA-AKI were more likely to develop AKI progression compared with female patients during their ICU stay (52.5% vs 47.5%, P < .001). Patients who developed AKI progression had higher weight and exhibited a higher incidence of diabetes, congestive heart failure, chronic pulmonary disease and liver disease. More details can be seen in Table 1. The baseline characteristics of patients in the MIMIC-III subset cohort, the eICU-CRD cohort and the SICdb cohort are listed in Supplementary data, Tables S1, S2 and S3, respectively. Table 1:Baseline characteristics of septic patients with and without AKI progression.CharacteristicsOverallNon-progressionProgression P-value N = 9193 N = 7046 N = 2147Age, years, median (IQR)67.7 (21.3)67.7 (21.0)67.4 (22.2).926Gender<.001 Male, n (%)5353 (58.2)4225 (60.0)1128 (52.5) Female, n (%)3840 (41.8)2821 (40.0)1019 (47.5)Weight, kg, median (IQR)82.4 (28.4)81.6 (27.6)85.0 (33.0)<.001Hypertension, n (%)5685 (61.8)4365 (62.0)1320 (61.5).926Diabetes, n (%)2530 (27.5)1887 (26.8)643 (29.9).016Acute myocardial infarction, n (%)1471 (16.0)1117 (15.9)354 (16.5).781Congestive heart failure, n (%)2308 (25.1)1657 (23.5)651 (30.3)<.001Chronic pulmonary disease, n (%)2309 (25.1)1709 (24.3)600 (27.9).003Liver disease, n (%)1400 (15.2)841 (11.9)559 (26.0)<.001Vasoactive agents, n (%)4316 (46.9)3266 (46.4)1050 (48.9).116Mechanical ventilation, n (%)5851 (63.6)4464 (63.4)1387 (64.6).575HR max, median (IQR)103 (27)102 (26)107 (29)<.001HR min, median (IQR)70 (19)70 (19)72 (23)<.001SBP max, mmHg, median (IQR)144 (27)144 (26)145 (29).252SBP min, mmHg, median (IQR)87 (16)88 (15)85 (17)<.001DBP max, mmHg, median (IQR)83 (23)82 (22)86 (25)<.001DBP min, mmHg, median (IQR)45 (12)45 (11)44 (13)<.001MBP max, mmHg, median (IQR)100 (23)99 (22)102 (25)<.001MBP min, mmHg, median (IQR)58 (12)58 (12)57 (14)<.001RR max, median (IQR)27 (8)27 (7)28 (9)<.001RR min, median (IQR)12 (5)12 (4)13 (6)<.001Temperature max, ℃, median (IQR)37.4 (0.9)37.4 (0.9)37.3 (0.9)<.001Temperature min, ℃, median (IQR)36.4 (0.7)36.4 (0.7)36.4 (0.7).801SpO2 min, median (IQR)93 (5)93 (4)92 (6)<.001WBC max, K/uL, median (IQR)14.3 (8.6)14.2 (8.2)14.5 (9.5).070WBC min, K/uL, median (IQR)10.0 (6.4)9.9 (6.1)10.1 (7.4).404Platelet max, K/uL, median (IQR)197 (120)198 (115)196 (146).629Platelet min, K/uL, median (IQR)153 (108)154 (102)150 (128).011Hemoglobin max, g/dL, median (IQR)11.6 (2.9)11.7 (2.8)11.4 (3.3)<.001Hemoglobin min, g/dL, median (IQR)9.9 (3.0)10.0 (2.9)9.7 (3.4)<.001Glucose max, mg/dL, median (IQR)143 (69)139 (63)159 (85)<.001Glucose min, mg/dL, median (IQR)114 (38)114 (36)114 (43).404BUN max, mg/dL, median (IQR)20.0 (15.0)19.0 (13.0)27.0 (24.0)<.001BUN min, mg/dL, median (IQR)17.0 (13.0)16.0 (10.0)21.0 (19.5)<.001Creatinine max, mg/dL, median (IQR)1.0 (0.6)1.0 (0.5)1.3 (1.2)<.001Creatinine min, mg/dL, median (IQR)0.9 (0.4)0.8 (0.3)1.0 (0.8)<.001Sodium max, mmol/L, median (IQR)140.0 (5.0)140.0 (4.0)140.0 (6.0).081Sodium min, mmol/L, median (IQR)137.0 (5.0)137.0 (5.0)137.0 (7.0)<.001Potassium max, mmol/L, median (IQR)4.5 (0.8)4.4 (0.7)4.5 (1.0)<.001Potassium min, mmol/L, median (IQR)3.9 (0.7)3.9 (0.6)3.8 (0.7).008Chloride max, mmol/L, median (IQR)107.0 (8.0)108.0 (7.0)106.0 (9.0)<.001Chloride min, mmol/L, median (IQR)103.0 (8.0)104.0 (7.0)102.0 (9.0)<.001Anion gap max, mmol/L, median (IQR)15 (5)15 (5)17 (6)<.001Anion gap min, mmol/L, median (IQR)12 (4)12 (4)13 (4)<.001Urine output, mL, median (IQR)1360 (1060)1480 (1020)910 (925)<.001ICU stay, days, median (IQR)7.8 (8.6)7.1 (7.2)11.2 (13.3)<.00128-day mortality, n (%)1413 (15.4)763 (10.8)650 (30.3)<.001HR, heart rate; SBP, systolic blood pressure; DBP, diastolic blood pressure; MBP, mean blood pressure; RR, respiratory rate; WBC, white blood cell. Baseline characteristics of septic patients with and without AKI progression. HR, heart rate; SBP, systolic blood pressure; DBP, diastolic blood pressure; MBP, mean blood pressure; RR, respiratory rate; WBC, white blood cell. Feature selection Lasso regression was conducted and screened 22 variables. Twenty variables were determined by RFE. Finally, 12 variables which were important features selected by both Lasso and RFE algorithm were chosen for model construction, including weight, liver disease, MV, systolic blood pressure min, hemoglobin max, glucose max, BUN max, creatinine max, chloride max, chloride min, AG max and urine output. The process of feature selection can be seen in Fig. 2. Screening of important variables based on lasso regression and recursive feature elimination. (A) Variation characteristics of the coefficient of variables. (B) The selection process of the optimal value of the parameter λ in the lasso regression model by the cross-validation method. (C) The feature selection process of recursive feature elimination. (D) Venn diagram for the variables intersection of lasso regression and recursive feature elimination. Model construction and validation We constructed six ML models to identify AKI progression in critically ill patients with sepsis using the features screened by lasso regression. Six models, including KNN, LR, NB, RF, SVM and DT, all exhibited considerable prediction performance for SA-AKI progression. In the internal validation set, the RF model exhibited the best performance among the six models. Figure 3A reveals that RF had a better discriminative ability compared with other models [AUC = 0.779, 95% confidence interval (CI) 0.767–0.789]. The remaining five models (KNN, LR, NB, SVM and DT) also demonstrated good predictive power: SVM (AUC = 0.772, 95% CI 0.760–0.783), LR (AUC = 0.763, 95% CI 0.751–0.775), NB (AUC = 0.762, 95% CI 0.750–0.773), KNN (AUC = 0.681, 95% CI 0.668–0.694) and DT (AUC = 0.745, 95% CI 0.732–0.757). The calibration plot demonstrated that they had good accuracy between the actual probability and predictive probability (Fig. 3B). The Brier score and calibration slope are also presented in Fig. 3B, and the RF model achieved the best fit. The DCA curves of six models are shown in Fig. 3E. Results of the F1 score, accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) for the six models are presented in Supplementary data, Table S4. RF model achieved an F1 score of 0.874, accuracy of 0.800, sensitivity of 0.907, specificity of 0.449, PPV of 0.844 and NPV of 0.595. Discriminative, calibration and clinical capability of the models. (A) ROC curve in the internal validation set. (B) Calibration curve in the internal validation set. (C) ROC curve in the temporal validation set. (D) Calibration curve in the temporal validation set. (E) DCA analysis in the internal validation set. (F) DCA analysis in the temporal validation set. Similar results were observed in the temporal validation set. Six ML models constructed above demonstrated robust generalizability. The ROC curves in Fig. 3C reveal that the RF model still had a good discriminative ability (AUC = 0.758, 95% CI 0.736–0.781). The calibration capability and decision curves of six ML models are displayed in Fig. 3D and F. In the external validation datasets, the RF model also achieved the best discriminative capability among the six ML models (eICU-CRD, AUC = 0.713, 95% CI 0.701–0.725; SICdb, AUC = 0.743, 95% CI 0.729–0.758). Additionally, the calibration curves and DCA curves can be seen in Supplementary data, Figs S1 and S2. Interpretability analysis Figure 4A presented a summary swarm plot which provided a global interpretability of the RF model. It depicted the entire distribution of each feature’s impact on the model output. Each dot corresponds to a specific instance, with its position denoting the SHAP value and its color representing the feature value. The features are ordered by their mean contribution to the model output, the top five of which included urine output, creatinine max, BUN max, MV and AG max, illustrating that these factors were key predictors for SA-AKI progression in the ICU. Model interpretation based on the SHAP method. (A) SHAP summary plot of the 12 features of the RF model. Each dot corresponds to a specific instance, with its position denoting the SHAP value and its color representing the feature value. The importance of variables is ranked based on mean contribution to model output from top to bottom. (B, C) Waterfall plots visualized the process of individual prediction. The yellow arrow represents positive contributions to the prediction, while the purple arrow denotes negative influences. The f(x) value signifies the actual SHAP value, and the corresponding prediction is estimated probability of SA-AKI progression. The waterfall plot displayed the process of individual prediction. The yellow arrow represents positive contributions to the prediction, while the purple arrow denotes negative influences. The f(x) value signified the actual SHAP value, and then the corresponding risk of SA-AKI progression can be obtained. Figure 4B displayed a prediction process of a patient with a lower risk of SA-AKI progression, and Fig. 4C showed the RF model predicted values for an individual with a higher risk of SA-AKI progression. Model application A user-friendly online platform was built based on the RF model, which can be accessed at the following website: https://njmuyxd.shinyapps.io/AKI_new/. This online platform enabled clinicians to input the information and early metrics of patients with SA-AKI and promptly assess their risk of AKI progression. DISCUSSION In this study, we constructed a prediction tool for early identification of SA-AKI progression in the ICU. Six ML algorithms were utilized to establish the prediction model using 12 variables screened by lasso regression and RFE. Our results demonstrated that the RF model exhibited the best performance with a better discriminative, calibration and clinical capability compared with other ML models. The user-friendly online platform we built in this study facilitates the clinical application and could assist clinicians in early identification of AKI progression in critically ill patients with sepsis, thereby achieving the intervention and management of the high-risk group. Although the pathophysiology of SA-AKI is not well understood, a growing body of research has been conducted to explore its potential mechanisms [5, 23, 24]. Currently, the prevailing viewpoints hold that the pathogenesis of SA-AKI involves multiple mechanisms, including renal hypotension and associated ischemia, inflammation and apoptosis, macrovascular and microvascular dysfunction, and immunologic and autonomic dysregulation [5, 25]. The complex pathophysiologic processes not only bring great challenges to the treatment and management of septic patients with SA-AKI but also serve as the primary causes of their adverse outcomes. Therefore, early identification and prevention play a critical role in improving outcomes and survivorship in patients with SA-AKI. Many attempts have been made recently to build predictive models for the early identification and prognostic assessment of SA-AKI in the whole population or a particular one with a certain comorbidity. Yue et al. constructed an XGBoost model with an AUC of 0.817 for early prediction of AKI in septic patients [9]. Also, other researchers have built prognostic models for predicting short-term mortality in patients with AKI [10]. Unfortunately, few studies have concentrated on early identification of SA-AKI progression. Although Wei et al. developed a prediction model for AKI progression, the extension of the prediction model to other populations was limited greatly due to the lack of external validation [7]. Additionally, the existence of sepsis could influence significantly the progression of AKI in septic patients, so the prediction model built by Wei et al. may not apply to patients with SA-AKI admitted to the ICU. In this study, we concentrated on patients with SA-AKI and aimed to create a prediction model for risk of AKI progression instead of early diagnosis and prognosis. For septic patients, their kidney may be influenced by systemic inflammation and temporary ischemia; this kind of kidney injury status tend to be transient. Once their renal function begins continuous deterioration, they may face a poor prognosis. Therefore, our findings may assist early identification of these patients and achieve prompt intervention. Once a higher risk of AKI progression was determined by the prediction model, immediate protection of renal function should be considered, such as discontinuing nephrotoxic drugs and avoiding excessive fluid intake. Additionally, more frequent renal function metrics monitoring is recommended. RF is a popular ML algorithm for handling large and noisy datasets and has a higher accuracy compared with decision trees [26]. RF could handle unbalanced sample distributions well and build the final model by using different subsets of the original data, and the model averages the output from the individual trees, which reduces overfitting effectively. In recent years, RF has been widely applied in the medical field, exhibiting outstanding performance in various areas such as cardiovascular disease, neurology and oncology [27–32]. In this study, we developed an RF model for predicting AKI progression in critically ill patients with sepsis, which exhibited the best discriminative ability among six ML models. However, when ML models were generalized to external cohorts, their calibration performance were not so good. One main potential cause was that the heterogeneity between the development and external validation cohorts may affect the model calibration performance significantly. Additionally, some clinical valuable predictors such as vasoactive agents use or SOFA score were not included in our model. Given their established associations with the severity of AKI, their omission may contribute to model miscalibration [33, 34]. Even so, RF still exhibited the best calibration, and considering that RF algorithm could handle unbalanced sample distributions and reduce overfitting effectively, the final model was constructed based on RF algorithm. SHAP is a mature method for model interpretation which could provide us with a more comprehensive understanding of feature contribution and the prediction process [35, 36]. The SHAP summary plot could help us identify the key factors for SA-AKI progression, and warn us to early handle these risk factors to improve the overall prognosis of patients. The individual waterfall plot revealed the prediction process of specific patients, which can guide us in effective decision-making on these patients. After SHAP interpretation, the rank of variable importance of 12 predictors was determined. Consistent with the findings of Ahn et al. that obesity is associated with early SA-AKI in patients in the ICU due to complex mechanisms, weight served as a risk factor for SA-AKI progression, which may make us consider whether to initiate renal protection early to improve the prognosis of these patients [37]. Besides weight, blood glucose was also determined as a predictor for risk of AKI progression. This may be attributed to a reduction in renal perfusion caused by endothelial function impairment induced by hyperglycemia [38]. Interestingly, we observed that chloride was a key predictor for AKI progression in critically ill patients with sepsis. Consistent with our findings, a retrospective study demonstrated that hyperchloremia is common in severe sepsis and septic shock and is independently associated with AKI [39]. Hyperchloremia has detrimental effects on renal function through multiple mechanisms, such as decrease in renal blood flow and glomerular filtration rate (GFR) due to thromboxane release and the activation of renin–angiotensin–aldosterone system induced by chloride. Furthermore, chloride can initiate the tubuloglomerular feedback mechanism, resulting in afferent arteriolar vasoconstriction, mesangial contraction and further decrease in GFR. To facilitate clinical application, we designed a user-friendly online platform that allows clinicians to predict the risk of SA-AKI progression. Once a high risk of AKI progression is alerted, clinicians should dynamically monitor the above key indicators, especially urine output, serum creatinine, BUN and ventilator parameters, promptly correct existing abnormalities, such as optimizing volume status or ventilator parameters, and prevent irreversible renal function deterioration. Notably, AKI staging based on both KDIGO creatinine and urine output criteria tends to be inconsistent due to their different time sensitivity, so the higher stage was recognized as the actual AKI staging when creatinine and urine output staging exhibited discrepancy. Although this may lead to earlier or higher AKI staging, early identification of AKI in clinical practice was crucial for improving their prognosis. Urine output is a very sensitive indicator of kidney injury, but it can be affected by many factors, especially fluid infusion. While creatinine is a relatively specific indicator, its change was slow so early capture of kidney injury cannot be achieved. Therefore, the worse criterion was prioritized for staging in our study, which may be more favorable for early identification and prompt intervention on AKI. Several limitations to this study should be considered. Firstly, not all potential variables were included in this study due to missing data, such as height, lactate and cardiac biomarkers. Additionally, some clinical valuable predictors such as vasoactive agents use or SOFA score were not included in our model. This may introduce potential bias. Secondly, due to the different periods of the two databases, some biases may not be avoided, affecting the accuracy of the results. Thirdly, competing events bias may exist due to early discharge or death. Finally, the model was built based on baseline data, while some of them may be captured after SA-AKI diagnosis, which may introduce some bias. Further studies are required to validate the accuracy of our model and its generalization. CONCLUSION ML could be a useful and reliable tool for predicting AKI progression in septic patients. We developed an RF model to predict the risk of SA-AKI progression, which may provide a reference for early identification and prompt intervention in the high-risk group. Supplementary Material