Interpretable Machine Learning for Early Prediction of Acute Kidney Disease (AKD) in Sepsis-Associated Acute Kidney Injury (SA-AKI): A Multicenter Cohort Study with External Validation 

Interpretable Machine Learning for Early Prediction of Acute Kidney Disease (AKD) in Sepsis-Associated Acute Kidney Injury (SA-AKI): A Multicenter Cohort Study with External Validation | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Interpretable Machine Learning for Early Prediction of Acute Kidney Disease (AKD) in Sepsis-Associated Acute Kidney Injury (SA-AKI): A Multicenter Cohort Study with External Validation Shuang Chen, Guang Li, Qingzhan Zeng, Xiancheng Xu, Chanlin Li, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7313497/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 6 You are reading this latest preprint version Abstract Sepsis-associated acute kidney injury (SA-AKI) represents a critical challenge in the management of critically ill patients, significantly contributing to morbidity and mortality in intensive care units (ICUs). This study aims to enhance the understanding and prediction of SA-AKI progression to acute kidney disease (AKD) by utilizing a comprehensive machine learning approach. Data from the MIMIC-IV and eICU-CRD databases were analyzed, incorporating 14 key clinical features identified through rigorous feature selection methods, including Boruta and LASSO regression. Eleven machine learning models were developed, with Gradient Boosting demonstrating the highest accuracy (78.94%) and optimal calibration characteristics. The external validation cohort revealed a decrease in model performance, emphasizing the risk of overfitting in complex models. Notably, the use of ACE inhibitors/ARBs was associated with a reduced risk of AKD progression, while nephrotoxic agents significantly increased this risk. Prognostic scoring systems, including SOFA and LODS, were found to correlate significantly with AKD outcomes, facilitating better risk stratification. Furthermore, a web-based risk calculator was developed to provide clinicians with an accessible tool for predicting the risk of SA-AKI progression to AKD based on individual patient data. In conclusion, this study underscores the importance of timely interventions and tailored treatment strategies in the management of SA-AKI, while also paving the way for future research to refine predictive models and improve clinical outcomes in critically ill patients. Sepsis-associated acute kidney injury (SA-AKI) Acute kidney disease (AKD) Interpretable machine learning Early prediction External validation Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Introduction Sepsis-associated acute kidney injury (SA-AKI) is a critical condition that significantly impacts the morbidity and mortality rates of patients admitted to intensive care units (ICUs) [1-3] . As the leading cause of acute kidney injury (AKI) in critically ill patients, sepsis is associated with severe complications, prolonged hospital stays, and increased healthcare costs. Despite advancements in understanding the pathophysiology of SA-AKI, effective diagnostic and therapeutic strategies remain limited, often resulting in delayed interventions and poor patient outcomes [4-6] . This gap in timely diagnosis and treatment underscores the urgent need for further research into SA-AKI to enhance early recognition and management in clinical settings. Current literature highlights the complex interplay of various risk factors contributing to the development of SA-AKI. Factors such as age, pre-existing comorbidities, and the severity of sepsis are increasingly recognized as critical determinants of patient outcomes [7-9] . For instance, older patients and those with underlying health conditions like hypertension and diabetes are at a heightened risk for developing AKI in the context of sepsis. Furthermore, the timing and appropriateness of therapeutic interventions, including the administration of antibiotics and the use of mechanical ventilation, play pivotal roles in influencing patient trajectories [10-12] . However, comprehensive insights into the multifactorial aspects of SA-AKI remain limited, necessitating a deeper exploration of clinical characteristics and treatment protocols. To address these research gaps, this study utilizes advanced machine learning methodologies applied to large-scale datasets from the MIMIC-IV and eICU-CRD databases, which provide rich clinical information on critically ill patients. The application of machine learning techniques offers a powerful framework for analyzing complex datasets and uncovering non-linear relationships among various clinical parameters [13,14] . These methods can facilitate the development of predictive models that identify patients at risk of progression from SA-AKI to acute kidney disease (AKD), thereby enabling timely interventions and improved patient care. The primary objective of this research is to delineate key clinical features and intervention strategies associated with SA-AKI while leveraging machine learning to create robust predictive models (Figure 1). By focusing on specific variables such as the use of vasopressors, mechanical ventilation, and the timing of antibiotic administration, this study aims to enhance the understanding of SA-AKI’s dynamics and its progression towards more severe renal impairment [15] . Ultimately, the goal is to develop a clinical decision-support tool that can assist healthcare providers in the early identification and management of SA-AKI, thus improving patient outcomes and reducing the associated burden on healthcare systems [16-18] . In summary, this research seeks to bridge the existing knowledge gaps surrounding SA-AKI by employing innovative machine learning techniques to analyze large clinical datasets. Through this approach, the study aims to contribute significantly to the understanding of SA-AKI, informing better clinical practices and enhancing patient management strategies in critical care settings. By elucidating the multifaceted risk factors and treatment protocols associated with SA-AKI, the findings are anticipated to foster improvements in the prevention and management of this serious condition in critically ill patients [19-21] . Materials and Methods Data Sources The MIMIC-IV (Medical Information Mart for Intensive Care IV) 3.1 and eICU-CRD (eICU Collaborative Research Database) are publicly accessible clinical databases extensively utilized in critical care research [22-26] . MIMIC-IV 3.1, jointly maintained by the Massachusetts Institute of Technology (MIT) and Beth Israel Deaconess Medical Center (BIDMC), integrates de-identified data from over 300,000 patients (including ~50,000 ICU admissions) across U.S. hospitals from 2008 to 2022. It encompasses multidimensional data such as demographics, vital signs, laboratory results, medication records, and diagnostic codes (ICD-9/10) . Similarly, the eICU-CRD, developed through a collaboration between Philips and MIT, includes data from 208 U.S. hospitals (2014–2015), covering >200,000 ICU patients with structured clinical data (e.g., vital signs, medications) and unstructured clinical notes. Both databases are compliant with ethical standards (CITI certification: 69327991), and their use aligns with sepsis-associated acute kidney injury (SA-AKI) research priorities outlined in recent guidelines [27] . Study Subjects Inclusion criteria: (1)Admission to ICU with confirmed diagnosis of sepsis or septic shock; (2)Age ≥18 years; (3)Sepsis-associated acute kidney injury; (4) First ICU admission episode. Exclusion criteria: (1)ICU length of stay <48 hours; (2) Survival time <7 days; (3) Incomplete data or missing critical variables. Data Extraction Variables extracted included:(1)Demographics: Age, sex, ICU admission time, antibiotic initiation time, AKI onset time, and comorbidities (hypertension, diabetes, heart failure, CKD, etc.). (2)Vital signs: Temperature, heart rate, blood pressure, respiratory rate, and weight (recorded from 1 day pre-AKI to 7 days post-AKI). (3)Laboratory tests: Hemoglobin, leukocyte count, lactate, creatinine, blood urea nitrogen (BUN), electrolytes (Na + /K + /Cl − ), coagulation markers (INR, PTT), and arterial blood gas parameters. (4)Interventions: Mechanical ventilation, vasopressors (norepinephrine, vasopressin), and continuous renal replacement therapy (CRRT). (5)Severity scores: Sequential Organ Failure Assessment (SOFA), Acute Physiology Score III (APS III), and Logistic Organ Dysfunction Score (LODS). Outcomes and Definitions The primary outcome was acute kidney disease (AKD), defined as a sustained decline in kidney function (creatinine-based criteria) persisting 7-90 days post-AKI onset, per the Acute Disease Quality Initiative (ADQI) consensus [28] . SA-AKI was classified as early (≤48 hours post-sepsis diagnosis) or late (48 hours–7 days). Secondary outcomes included mortality, recurrent AKI, and progression to CKD (>90 days) [29,30] . Data Processing and Statistical Analysis Data Cleaning: Outliers were removed based on clinical plausibility (e.g., creatinine >10 mg/dL). Variables with >10% missingness (e.g., urine output) were excluded; others (e.g., lactate) were imputed using mode substitution. The MIMIC-IV dataset was split into training (70%) and testing (30%) sets, while the eICU-CRD served as an external validation cohort. Synthetic Minority Over-sampling Technique (SMOTE) addressed class imbalance in the training set. Feature Selection: Firstly, multicollinearity is treated by calculating the feature-wise Pearson correlation coefficient matrix and setting the threshold at |r| > 0.9 to eliminate multicollinear features, with high-dimensional correlations visualized via a Rank2D heatmap. Secondly, primary feature screening is done using the Boruta algorithm,a wrapper - based method built on a random forest base classifier.Twenty independent experiments are conducted,and features stably selected across all iterations (Rank = 1) are retained based on the median distribution of feature importance rankings. Thirdly, secondary feature compression is carried out using LASSO regression,an embedded feature selection method.The optimal regularization coefficient λ is determined via 10-fold cross - validation,and features with nonzero coefficients are screened by finding the optimal sparse solution using the 1-SE rule. Finally,the feature set is determined by getting the common feature subset chosen by both Boruta and LASSO.This method integrates nonlinear feature importance evaluation and linear sparse constraints.It enhances stability through multiple Boruta iterations and the 1 - SE rule,and is verified through heatmaps,regularization path plots and network topology plots.It effectively improves the feature subset’s biological interpretability and model compatibility. Machine Learning Model Construction We establishes a multi-algorithm integrated machine learning forecasting framework, comprising 11 algorithms in three categories: conventional statistical learning models (Logistic Regression), single learner models (SVM, KNN, Decision Tree), and ensemble learning models (Random Forest, AdaBoost, Gradient Boosting, XGBoost, LightGBM, Bagging, Voting), forming a diverse model ensemble.For hyperparameter optimization, the RandomizedSearchCV strategy is used with ROC_AUC as the optimization target and 5 - fold stratified validation set. Each algorithm has its own parameter space: base learner parameters include decision tree depth (3 - 8 layers) and minimum samples in leaf nodes (2 - 5); ensemble algorithms set the number of base classifiers (10 - 50, step size 10) and learning rate (0.01 - 1, linear distribution); regularization parameters cover L1/L2 norms and elastic net. After parallel search (n_jobs = -1), the best parameter configurations for each algorithm are retained. To enhance model generalization, a two - level ensemble system is built. The primary ensemble uses Bagging, with the number of base classifiers set in a discrete search space of [31] , and diverse predictors are built through resampling with replacement. The advanced ensemble adopts a soft Voting mechanism, integrating the probability outputs of Logistic Regression, Random Forest and Gaussian Naive Bayes, with weight parameters optimized via grid search.A multi-stage model performance evaluation system is established. Initially, models are screened on the training set based on accuracy, precision, recall, F1 - score and ROC - AUC, with AUC as the core criterion for selecting candidate models. Calibration curves are used to assess each model’s probability prediction ability. Subsequently, secondary model selection is performed using an independent external validation set, with the model having the highest AUC value chosen as the optimal solution. For the preferred model, a systematic analysis is carried out. The confusion matrix is used to calculate specificity (Specificity = TN/(TN + FP)) and negative predictive value (NPV = TN/(TN + FN)). A precision - recall curve is plotted to evaluate performance under class imbalance. Finally, the SHAP framework is employed to interpret the decision - making mechanism, including global feature contribution ranking, individual sample decision attribution and feature interaction visualization. All experiments are set with a random seed (random_state = 0) to ensure reproducibility of results. Explainability analysis We use SHAP values to quantitatively break down the decision logic of “black-box” models. Its advantages are as follows: First, it creates an axiomatic explanation system based on game theory. Second, it offers sample-level interpretability. Third, it allows for visual verification of feature effect directions [32-34] . These advantages make the findings useful for clinical decision-making. They provide a transparent basis for clinical decisions, making predictive models more credible and clinically applicable. By offering a clear understanding of model predictions, this method helps clinicians trust and use these predictions in their work. It also helps find possible model biases or errors, which is important for patient safety. The sample-level interpretability from SHAP values further aids personalized patient care. Using SHAP values bridges the gap between advanced predictive analytics and practical clinical insights. They enhance model interpretability and create a transparent, trustworthy framework for predictions, supporting better-informed clinical decisions. Results This study selected 9,778 patients from the MIMIC-IV database to form the training cohort and 3,721 patients from the eICU-CRD database to constitute the external validation cohort (Figure 2). Initially, 63 features were incorporated into the training cohort. Subsequently, the Pearson correlation coefficient matrix among these features was calculated, leading to the removal of four highly correlated features (Pearson correlation coefficient > 0.9), specifically International Normalized Ratio (INR), Hematocrit, Red Blood Cells (RBC), and Renal disease (Figure 3). The remaining features underwent data dimensionality reduction using the Boruta feature selection method, which identified 36 features. Further features selection was performed using LASSO regression, narrowing down the list to 14 features. To refine the feature set and ensure the most relevant features were retained, we took the intersection of the features identified by both methods. This process ultimately determined 14 key features : angiotensin-converting enzyme inhibitor/angiotensin II receptor blocker (ACEI/ARB), Acute Physiology Score III (APS III), continuous renal replacement therapy (CRRT), Cerebrovascular Disease, Logistic Organ Dysfunction System (LODS), interval from infection detection to first antibiotic use (Los_inf._AB), mean blood pressure (MBP), Mechanical Ventilation, Paraplegia, respiratory rate, baseline serum creatinine (Baseline Scr), blood oxygen saturation (SpO 2 ), vasoactive Agent, and body weight (Figure 4). Using the training dataset, we constructed 11 machine learning models based on these 14 key features. The performance metrics of these models are illustrated in. In terms of accuracy, Gradient Boosting demonstrated the highest performance at 78.94%, followed closely by LightGBM with 78.29% and Bagging with 78.08%. For the area under the ROC curve (AUC), which is a critical metric for evaluating class discrimination ability, Logistic Regression, Adaboost, and Voting classifier all achieved the highest value of 0.84. This indicates that these models had the best ability to distinguish between different classes. In the recall rate (sensitivity) dimension, which measures the models' ability to identify positive samples, Logistic Regression (0.81) and KNN (0.79) showed superior performance. Notably, Gradient Boosting achieved a balance in both precision (0.57) and F1 score (0.57), making it a well-rounded model in these aspects. Meanwhile, XGBoost (0.51) and LightGBM (0.57) exhibited advantages in precision. Based on a comprehensive evaluation using the F1 score, which balances precision and recall, Logistic Regression (0.61), Adaboost (0.60), and Voting classifier (0.60) demonstrated the best predictive stability (Figure 5). In the external validation cohort, we selected the 14 relevant features identified in the training phase for model validation. The models generally exhibited performance decay when applied to the external cohort, with AUC values dropping to the range of 0.55–0.59, compared to 0.78–0.84 in the training set. Among these models, Adaboost experienced the most significant decay (ΔAUC = 0.28), while KNN showed relatively robust predictive characteristics with a recall rate of 0.47 and an F1 score of 0.52. It is particularly worth noting that Logistic Regression (Recall: 0.81→0.40) and Voting classifier (Recall: 0.68→0.40), which performed exceptionally well in the training set, experienced a decline of over 50% in recall rate during external validation. Additionally, the accuracy of Gradient Boosting decreased substantially from 78.94% to 54%, highlighting the risk of overfitting in complex models when they are applied to external data (Figure 6). The probability calibration ability of each model was further assessed using the Brier score (Figure 7). Regarding probability prediction accuracy, Gradient Boosting and LightGBM performed the best with a Brier score of 0.143, followed by Voting (0.146) and Bagging (0.148). These results suggest that ensemble learning methods have significant advantages in probability calibration. However, it is important to note that although Adaboost achieved an ROC AUC of 0.84, it had the highest Brier score of 0.241. This discrepancy indicates that the model may have systematic bias in probability prediction. Among traditional algorithms, Logistic Regression (0.167) and Random Forest (0.156) demonstrated better calibration performance compared to Decision Tree (0.175) and KNN (0.185). When combined with previous metric analyses, it is evident that Gradient Boosting not only maintained high accuracy (78.94%) but also achieved the best calibration characteristics. Although LightGBM had comparable calibration performance, its recall rate (0.48) was significantly lower than that of Logistic Regression (0.81). The evaluation of the predictive performance of the Logistic Regression model is detailed in Figure 8. Based on 2,934 test samples, the average precision of the Precision-Recall curve was calculated to be 0.63. This curve exhibited a typical downward trend, where precision gradually decreased as recall increased. The confusion matrix revealed that the model correctly classified 1,597 negative samples, corresponding to a specificity of 72.1%, and 585 positive samples, corresponding to a sensitivity of 81.3%. However, it also generated 617 false positives, resulting in a Type I error rate of 27.9%, and 135 false negatives, leading to a Type II error rate of 18.8%. These results indicate that while the model maintains moderate average precision, it demonstrates a relative advantage in discriminating negative classes. SHAP interpretability analysis provided a quantitative description of the feature impact intensity and direction within the Logistic Regression model (Figure 9). Mechanical Ventilation emerged as the feature with the highest contribution (SHAP = 0.79) to the prediction. An increase in its feature value was associated with a significant rise in the predicted risk probability. Vasoactive Agent (0.33) and CRRT (0.33) were identified as secondary driving factors, both exhibiting a positive correlation with the outcome and indicating a positive effect on the model's predictions. Notably, the use of ACEI/ARB was found to have a negative association with baseline creatinine. This suggests that an increase in the value of ACEI/ARB may lead to a reduction in the risk assessment. To facilitate the practical application of the model and make it accessible to clinicians, this study developed a web-based risk calculator ( https://sakiakd-fpohf6y9mmfki7gr3o67pv.streamlit.app/ ). This tool is designed to automatically calculate the risk of SA-AKI patients progressing to AKD by inputting the values of various variables. It aims to provide clinicians with a simple yet practical tool for early prediction of AKD in SA-AKI patients, thereby supporting clinical decision-making and potentially improving patient outcomes. Discussion SA-AKI is a critical condition observed in patients with severe sepsis, contributing significantly to morbidity and mortality within intensive care units (ICUs) [1] . This syndrome arises from a complex interplay of factors, including hemodynamic instability, inflammatory responses, and direct cellular injury, ultimately leading to renal dysfunction [9,35,36] . The incidence of SA-AKI has been reported to be alarmingly high, with estimates suggesting that it accounts for up to 70% of acute kidney injury cases in critically ill patients. The pathophysiology of SA-AKI remains incompletely understood, but recent studies indicate that mechanisms such as apoptosis and necroptosis may play pivotal roles in its development [30,37] . Moreover, the management of SA-AKI is further complicated by the lack of specific therapeutic interventions, underscoring the necessity for improved diagnostic and treatment strategies. In light of these challenges, our research aims to elucidate the clinical characteristics and treatment protocols associated with SA-AKI by leveraging extensive data from the MIMIC-IV and eICU-CRD databases. This study employs machine learning methodologies to identify critical risk factors and develop predictive models for the progression of SA-AKI to acute kidney disease (AKD) [38] . By examining various clinical parameters, including the timing of antibiotic administration and the use of mechanical ventilation, we aim to enhance understanding of SA-AKI’s complexities and improve patient outcomes [39] . The findings from this research will provide insights into optimal management strategies, potentially leading to better prognostic assessments and intervention opportunities for individuals suffering from SA-AKI. The innovation of this study lies in its application of advanced machine learning techniques to a large-scale dataset to identify critical risk factors associated with SA-AKI. By utilizing data from the MIMIC-IV and eICU-CRD databases, this research provides a comprehensive analysis that not only corroborates existing knowledge but also uncovers novel insights regarding the complex interplay of clinical parameters influencing AKD progression in septic patients. For instance, the identification of ACEI/ARBs as a protective factor against AKD progression aligns with previous findings in animal models but is one of the first to be confirmed in a human clinical context, thus bridging a significant gap in the literature regarding therapeutic strategies in SA-AKI management [40] . Moreover, the study heightens the understanding of how demographic factors, specifically age and pre-existing conditions like hypertension and diabetes, contribute to the risk of developing AKI, reinforcing the clinical implications of targeted management strategies [41,42] . The implications of these findings for clinical practice are profound. The established risk factors and treatment protocols suggest that timely interventions, such as early administration of ACEI/ARBs and careful monitoring of patients receiving nephrotoxic agents, can significantly improve patient outcomes [43-45] . This research advocates for the integration of machine learning-derived predictive models into clinical decision-making processes, which may help clinicians to identify high-risk patients and initiate early intervention strategies, thereby potentially reducing the morbidity and mortality associated with SA-AKI [39] . Furthermore, the development of a web-based risk calculator as a practical tool for clinicians exemplifies the application value of this research, facilitating early prediction of AKD in SA-AKI patients and enhancing the overall quality of patient care in intensive care settings [46] . Nevertheless, this study is not without limitations. The observational nature of the data used raises concerns about potential confounding variables that may not have been accounted for, thereby impacting the robustness of the findings. Additionally, the reliance on electronic health records means that the accuracy of the data is contingent upon the quality of documentation in clinical settings, which may vary [47] . Future research should aim to conduct prospective validation studies and explore the application of these predictive models in diverse clinical environments to assess their generalizability. Furthermore, integrating more granular patient-level data, including genetic and molecular markers, may provide deeper insights into the pathophysiology of SA-AKI and enhance predictive accuracy [48] . The limitations of this study must be acknowledged, particularly the absence of wet lab experiments, which restricts the validation of findings through biological mechanisms. Additionally, the reliance on multiple datasets may introduce inter-batch variability, potentially affecting the generalizability of our results. The lack of clinical validation analysis raises concerns about the applicability of the predictive models in real-world scenarios, emphasizing the need for further studies to confirm findings in diverse patient populations. Furthermore, the performance decay observed in external validation highlights the potential overfitting of complex models, necessitating cautious interpretation of model efficacy in clinical settings. In conclusion, this study provides significant insights into the factors associated with the progression of sepsis-associated acute kidney injury to acute kidney disease. By leveraging large-scale clinical data and machine learning techniques, we identified key predictors and established a practical risk calculator to aid clinicians in early intervention strategies. These findings underscore the critical need for timely and tailored management approaches in critically ill patients, ultimately enhancing patient outcomes and resource utilization in intensive care settings. Abbreviations SA-AKI: Sepsis-associated acute kidney injury AKD: Acute kidney disease AKI:Acute kidney injury ICUs: Intensive care units MIMIC-IV: Medical information mart for intensive care IV eICU-CRD:eICU collaborative research database CRRT: Continuous renal replacement therapy BUN: Blood urea nitrogen CKD: Chronic kidney disease SOFA: Sequential organ failure assessment APS III:Acute physiology score III LODS: Logistic organ dysfunction score. LASSO: Least absolute shrinkage and selection operator ACEI: Angiotensin converting enzyme inhibitors ARBs: Angiotensin II receptor blockers BIDMC: Beth israel deaconess medical center INR: International normalized ratio PTT: Partial thromboplastin time ADQI: Acute disease quality initiative SMOTE: Synthetic minority over-sampling technique SVM: Support vector machine KNN: K-nearest neighbors AdaBoost: Adaptive boosting XGBoost: Extreme gradient boosting LightGBM: Light gradient boosting machine RandomizedSearchCV: Randomized search cross-validation TN: True negative FP: False positive NPV: Negative predictive value FN: False negative SHAP: SHapley additive exPlanations RBC: Red blood cells MBP: Mean blood pressure Scr: Serum creatinine SpO2: Blood oxygen saturation Los_inf._AB: Interval from infection detection to first antibiotic use MSE: Mean squared error L1: Least absolute shrinkage and selection operator regression L2: Ridge regression 1-SE: One Standard Error Declarations ● Ethics approval and consent to participate Not applicable. ● Consent for publication All authors reviewed the manuscript and consented for publication. ● Availability of data and materials The datasets generated and/or analyzed during the current study are available in the MIMIC-IV v3.1 and eICU-CRD v2.0 repository/database at https://www.physionet.org/content/mimiciv/3.1/ and https://www.physionet.org/content/eicu-crd/2.0/ ● Competing Interests The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. ● Funding This study is partially supported by the Grant from Key Laboratory of Coronary Intraluminal Imaging and Functional Analysis of Dongguan City and Guizhou Provincial Health Commission Science and Technology Fund Project (gzwkj2025-005). ● Authors' contributions Shuang Chen (S.C.) and Guang Li (G.L.) contributed equally to this work. Conceptualized the study, developed machine learning algorithms, performed model training and internal validation. Qingzhan Zeng (Q.Z.Z.) curated multicenter clinical data, designed data collection protocols. Xiancheng Xu (X.C.X.) conducted statistical analyses and interpreted model performance metrics. Chanlin Li (C.L.L.) processed external validation datasets and performed feature engineering. Xiaoyue Li (X.Y.L.) assisted in clinical data annotation and quality control. Shaohong Li (S.H.L.) provided critical revisions for clinical relevance, methodology and finalized the manuscript. Heng Li (H.L.) (Corresponding Author) supervised the entire study, acquired funding, and finalized the manuscript. ● Acknowledgements Not applicable. References Chen Y, Jing H, Tang S, et al. Non-coding RNAs in Sepsis-Associated Acute Kidney Injury. Front Physiol. 2022;13:830924. Takeuchi T, Flannery AH, Liu LJ, et al. Epidemiology of sepsis-associated acute kidney injury in the ICU with contemporary consensus definitions. Crit Care. 2025;29(1):128. Mweene MD, Richards GA, Paget G, et al. Risk factors and outcomes of sepsis-associated acute kidney injury in intensive care units in Johannesburg, South Africa. S Afr Med J. 2022;112(12):919-923. Odum JD, Wong HR, Stanski NL. A Precision Medicine Approach to Biomarker Utilization in Pediatric Sepsis-Associated Acute Kidney Injury. Front Pediatr. 2021;9:632248. Peerapornratana S, Manrique-Caballero CL, Gómez H, et al. Acute kidney injury from sepsis: current concepts, epidemiology, pathophysiology, prevention and treatment. Kidney Int. 2019;96(5):1083-1099. Peng W, Li G. Research progress of microRNAs in sepsis-associated acute kidney injury. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2022;34(5):556-560. Saverymuthu A, Teo R, Zain JM, et al. Acute Kidney Injury following Rhabdomyolysis in Critically Ill Patients. J Crit Care Med. 2021;7(4):267-271 Louzada CF, Ferreira AR. Evaluation of the prevalence and factors associated with acute kidney injury in a pediatric intensive care unit. J Pediatr (Rio J). 2021;97(4):426-432. Magboul SM, Osman B, Elnour AA. The incidence, risk factors, and outcomes of acute kidney injury in the intensive care unit in Sudan. Int J Clin Pharm. 2020;42(6):1447-1455. Eswarappa M, Gireesh MS, Ravi V, et al. Spectrum of acute kidney injury in critically ill patients: A single center study from South India. Indian J Nephrol. 2014;24(5):280-285. Mahesh E, Nallamuthu P, Kumar M, et al. Clinical profile of geriatric acute kidney injury in a tertiary care center from south India. Saudi J Kidney Dis Transpl. 2017;28(4):886-890. Chen K, Lei Y, He Y, et al. Clinical outcomes of hospitalized COVID-19 patients with renal injury: a multi-hospital observational study from Wuhan. Sci Rep. 2021;11(1):15205. Baciu C, Xu C, Alim M, et al. Artificial intelligence applied to omics data in liver diseases: Enhancing clinical predictions. Front Artif Intell. 2022;5:1050439. Tseng AS, Noseworthy PA. Prediction of Atrial Fibrillation Using Machine Learning: A Review. Front Physiol. 2021;12:752317. Tan HK, Kaushik M, Tan CW, et al. Augmented Adsorptive Blood Purification during Continuous Veno-Venous Haemodiafiltration in a Severe Septic, Acute Kidney Injury Patient: Use of oXiris®: A Single Centre Case Report. Blood Purif. 2019;47 Suppl 3:1-6. Xu L, Sun P. Identification and management of sepsis associated-acute kidney injury. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2023;35(2):221-224. Zarbock A, Nadim MK, Pickkers P, et al. Sepsis-associated acute kidney injury: consensus report of the 28th Acute Disease Quality Initiative workgroup. Nat Rev Nephrol. 2023;19(6):401-417. Torres JSS, Tamayo-Giraldo FJ, Bejarano-Zuleta A, et al. Sepsis and post-sepsis syndrome: a multisystem challenge requiring comprehensive care and management—a review. Front Med (Lausanne). 2025;12:1560737. De Rosa S, Marengo M, Fiorentino M, et al. Extracorporeal blood purification therapies for sepsis-associated acute kidney injury in critically ill patients: expert opinion from the SIAARTI-SIN joint commission. J Nephrol. 2023;36(7):1731-1742. Ricci Z, Polito A, Polito A, et al. The implications and management of septic acute kidney injury. Nat Rev Nephrol. 2011;7(4):218-225. Parmar A, Langenberg C, Wan L, et al. Epidemiology of septic acute kidney injury. Curr Drug Targets. 2009;10(12):1169-1178. Yuan ZN, Xue YJ, Wang HJ, et al. A nomogram for predicting hospital mortality of critically ill patients with sepsis and cancer: a retrospective cohort study based on MIMIC-IV and eICU-CRD. BMJ Open. 2023;13(9):e072112. Sheng S, Li A, Zhang C, et al. Association between hemoglobin and in-hospital mortality in critically ill patients with sepsis: evidence from two large databases. BMC Infect Dis. 2024;24(1):1450. Schmidt L, Pigat L, Sheikhalishahi S, et al. Evaluating the SWIFT algorithm's efficacy in predicting hypoxemia across multiple critical care datasets. J Crit Care. 2025;89:155123. Qi Z, Dong L, Lin J, Duan M. Development and validation of a nomogram prediction model for early diagnosis of bloodstream infections in the intensive care unit. Front Cell Infect Microbiol. 2024;14:1348896. Peng X, Cai Y, Huang H, et al. A Predictive Model for Acute Kidney Injury Based on Leukocyte-Related Indicators in Hepatocellular Carcinoma Patients Admitted to the Intensive Care Unit. Mediators Inflamm. 2025;2025:7110012. Ramya K, Mukhopadhyay K, Kumar J. Predictive factors and risk scoring system for acute kidney injury (AKI) in sick neonates—a prospective cohort study. Eur J Pediatr. 2024;183(12):5419-5424. Patel M, Hornik C, Diamantidis C, et al. Patient-level factors increase risk of acute kidney disease in hospitalized children with acute kidney injury. Pediatr Nephrol. 2023;38(10):3465-3474. Shi B, Ye J, Chen W, et al. Prognosis of critically ill patients with early and late sepsis-associated acute kidney injury: an observational study based on the MIMIC-IV. Ren Fail. 2025;47(1):2441393. Kounatidis D, Vallianou NG, Psallida S, et al. Sepsis-Associated Acute Kidney Injury: Where Are We Now? Medicina (Kaunas). 2024;60(3):434. Tahir S, Ganie BA, Beigh TY, et al. Clinico-Etiological Spectrum and Outcome in Patients With Septic Acute Kidney Injury and Its Comparison With Non-septic Acute Kidney Injury: A Hospital-Based Prospective Study Conducted in a Tertiary Care Hospital in North India. Cureus. 2023;15(4):e37857. Yi F, Yang H, Chen D, et al. XGBoost-SHAP-based interpretable diagnostic framework for Alzheimer’s disease. BMC Med Inform Decis Mak. 2023;23(1):137. Roder J, Maguire L, Georgantas R, et al. Explaining multivariate molecular diagnostic tests via Shapley values. BMC Med Inform Decis Mak. 2021;21(1):211. Riasi A, Delrobaei M, Salari M. Personalized medication recommendations for Parkinson's disease patients using gated recurrent units and SHAP interpretability. Sci Rep. 2025;15(1):19074. Holland EM, Moss TJ. Acute Noncardiovascular Illness in the Cardiac Intensive Care Unit. J Am Coll Cardiol. 2017;69(16):1999-2007. . Glodowski SD, Wagener G. New insights into the mechanisms of acute kidney injury in the intensive care unit. J Clin Anesth. 2015;27(2):175-180. Wu Z, Deng J, Zhou H, et al. Programmed Cell Death in Sepsis Associated Acute Kidney Injury. Front Med (Lausanne). 2022;9:883028. Cheungpasitporn W, Thongprayoon C, Kashani KB. Artificial intelligence and machine learning's role in sepsis-associated acute kidney injury. Kidney Res Clin Pract. 2024;43(4):417-432. Huo L, Liu C, Yuan Y, et al. Pharmacological inhibition of ferroptosis as a therapeutic target for sepsis-associated organ damage. Eur J Med Chem. 2023;257:115438. Hasson DC, Watanabe-Chailland M, Romick-Rosendale L, et al. Choline supplementation attenuates experimental sepsis-associated acute kidney injury. Am J Physiol Renal Physiol. 2022;323(3):F255-F271. Jakubov K, Petr V, Zahradka I, et al. Acute Kidney Injury in Deceased Organ Donors: Risk Factors And Impacts on Transplantation Outcomes. Transplant Direct. 2024;10(12):e1730. Cheng Y, Nie S, Zhao X, et al. Incidence, risk factors and outcome of postoperative acute kidney injury in China. Nephrol Dial Transplant. 2024;39(6):967-977. Wang PT, Huang YB, Lin MY, et al. Prescriptions for angiotensin-converting enzyme inhibitors/angiotensin receptor blockers and monitoring of serum creatinine and potassium in patients with chronic kidney disease. Kaohsiung J Med Sci. 2012;28(9):477-483. Fang G, Annis IE, Farley JF, et al. Incidence of and Risk Factors for Severe Adverse Events in Elderly Patients Taking Angiotensin-Converting Enzyme Inhibitors or Angiotensin II Receptor Blockers after an Acute Myocardial Infarction. Pharmacotherapy. 2018;38(1):29-41. Chen JY, Tsai IJ, Pan HC, et al. The Impact of Angiotensin-Converting Enzyme Inhibitors or Angiotensin II Receptor Blockers on Clinical Outcomes of Acute Kidney Disease Patients: A Systematic Review and Meta-Analysis. Front Pharmacol. 2021;12:665250. Gao L, Wang GD, Yang XY, et al. Development of a risk prediction model for sepsis-related delirium based on multiple machine learning approaches and an online calculator. PLoS One. 2025;20(7):e0323831. Thoroddsen A, Sigurjónsdóttir G, Ehnfors M, et al. Accuracy, completeness and comprehensiveness of information on pressure ulcers recorded in the patient record. Scand J Caring Sci. 2013;27(1):84-91. Zhang Q, Yang B, Li X, et al. Biomarkers of cell cycle arrest, microcirculation dysfunction, and inflammation in the prediction of SA-AKI. Sci Rep. 2025;15(1):8023. Additional Declarations No competing interests reported. Supplementary Files eICUsakicleaning.csv eICUsakiraw.csv MIMICsakiraw.csv MIMICsakicleaning.csv eICUsaki.pdf MIMICsaki.pdf sakiakdshap.pdf Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 23 Sep, 2025 Reviewers invited by journal 11 Sep, 2025 Editor invited by journal 20 Aug, 2025 Editor assigned by journal 16 Aug, 2025 Submission checks completed at journal 12 Aug, 2025 First submitted to journal 12 Aug, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7313497","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":516606488,"identity":"f7621fe5-c175-42d1-8f95-c4c5b23c53b0","order_by":0,"name":"Shuang Chen","email":"","orcid":"","institution":"Songshan Lake TungWah Hospital","correspondingAuthor":false,"prefix":"","firstName":"Shuang","middleName":"","lastName":"Chen","suffix":""},{"id":516606489,"identity":"f8cd81e4-3f32-4128-9982-f79132b10a55","order_by":1,"name":"Guang Li","email":"","orcid":"","institution":"Guangdong Medical University","correspondingAuthor":false,"prefix":"","firstName":"Guang","middleName":"","lastName":"Li","suffix":""},{"id":516606490,"identity":"ff87b62f-a76d-4e0e-8f61-7406c47680c5","order_by":2,"name":"Qingzhan Zeng","email":"","orcid":"","institution":"Songshan Lake TungWah Hospital","correspondingAuthor":false,"prefix":"","firstName":"Qingzhan","middleName":"","lastName":"Zeng","suffix":""},{"id":516606492,"identity":"a75497d8-5377-405b-96a6-c2357332387f","order_by":3,"name":"Xiancheng Xu","email":"","orcid":"","institution":"Songshan Lake TungWah Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xiancheng","middleName":"","lastName":"Xu","suffix":""},{"id":516606494,"identity":"2855e530-70c9-435e-a1e5-a9bf27854beb","order_by":4,"name":"Chanlin Li","email":"","orcid":"","institution":"Dongguan TungWah Hospital","correspondingAuthor":false,"prefix":"","firstName":"Chanlin","middleName":"","lastName":"Li","suffix":""},{"id":516606495,"identity":"61c83bd0-c3de-4212-9fd4-70d94df281f9","order_by":5,"name":"Xiaoyue Li","email":"","orcid":"","institution":"the Fifth Affiliated Hospital of Zunyi Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xiaoyue","middleName":"","lastName":"Li","suffix":""},{"id":516606497,"identity":"3ebb567b-fa1d-45a6-94e6-9e3a9454741a","order_by":6,"name":"Shaohong Li","email":"","orcid":"","institution":"Songshan Lake TungWah Hospital","correspondingAuthor":false,"prefix":"","firstName":"Shaohong","middleName":"","lastName":"Li","suffix":""},{"id":516606498,"identity":"3244c315-1c77-40e8-89ea-3c063bfab927","order_by":7,"name":"Heng Li","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAs0lEQVRIiWNgGAWjYBACPgnGxgcfDGzk2NjbDxCnhU2CudlwRkGaMR/PmQRitbC3SfN8OJw4T8LBgEgt0o0NkjMM0tLbJBgSGH5UbCNCi8zBBgOgX3LbpBsPMPacuU2MwxIbEoG25LbJHEhgZmwjUsthHoPD6WwSCQZEa2lsBmpJIElLMyPQYYZtwEA+SJRf+CXSn//48MdGXr69/eCDHxVEaEEBB0hUPwpGwSgYBaMAFwAAfEc7eLs2MtEAAAAASUVORK5CYII=","orcid":"","institution":"Guangdong Medical University","correspondingAuthor":true,"prefix":"","firstName":"Heng","middleName":"","lastName":"Li","suffix":""}],"badges":[],"createdAt":"2025-08-07 00:38:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7313497/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7313497/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":91834630,"identity":"fed0f219-a09b-4a92-979c-648f5dd31db2","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1671858,"visible":true,"origin":"","legend":"","description":"","filename":"20250812AKIAKDe7.docx","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/3c08cc1ab8ec9436dcd9495d.docx"},{"id":91834631,"identity":"04c9244f-f387-4efd-8246-75184a05a5db","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9597,"visible":true,"origin":"","legend":"","description":"","filename":"dcb68fe23a784b20a1e6d59eb0d768ed.json","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/f1d6695a9a1c08e1bd690daa.json"},{"id":91836021,"identity":"aa7e5510-e6cd-4927-b0af-ea8f2256b36b","added_by":"auto","created_at":"2025-09-22 09:34:42","extension":"csv","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2403585,"visible":true,"origin":"","legend":"","description":"","filename":"MIMICsakicleaning.csv","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/bf7707c88a9c415e7abeabf9.csv"},{"id":91836024,"identity":"9c4b9d4c-e8d8-44de-91a0-b925d59cb10f","added_by":"auto","created_at":"2025-09-22 09:34:42","extension":"pdf","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4269768,"visible":true,"origin":"","legend":"","description":"","filename":"MIMICsaki.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/a180cf3d1e523bed89782f13.pdf"},{"id":91836026,"identity":"faeb3e3a-fa4d-4e4d-a436-b7ad4dce0d7c","added_by":"auto","created_at":"2025-09-22 09:34:42","extension":"csv","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3003131,"visible":true,"origin":"","legend":"","description":"","filename":"MIMICsakiraw.csv","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/c8d37028e8e1adf707760f3a.csv"},{"id":91834640,"identity":"74e345e6-999b-4406-ac0a-1333338c3495","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"csv","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":275608,"visible":true,"origin":"","legend":"","description":"","filename":"eICUsakicleaning.csv","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/3cd6c80e1bbf5f3636906334.csv"},{"id":91834645,"identity":"b6f908db-556c-40a7-9f19-2884d24a2f7d","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"pdf","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2736962,"visible":true,"origin":"","legend":"","description":"","filename":"eICUsaki.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/76277a77caa32e72c1d89bec.pdf"},{"id":91842155,"identity":"3eac1387-8e3f-40ea-90d5-e68cc84373c3","added_by":"auto","created_at":"2025-09-22 09:58:42","extension":"csv","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3684634,"visible":true,"origin":"","legend":"","description":"","filename":"eICUsakiraw.csv","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/fd6b037d88973f8d3d59d76d.csv"},{"id":91836041,"identity":"21e3b6f5-92f7-4d64-8185-0716fc1d87d0","added_by":"auto","created_at":"2025-09-22 09:34:43","extension":"pdf","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7137531,"visible":true,"origin":"","legend":"","description":"","filename":"sakiakdshap.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/795b9f91b795557a6d101566.pdf"},{"id":91834643,"identity":"90d61635-0a8c-4bb2-bedb-fdc0d6cb72a7","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"xml","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":111349,"visible":true,"origin":"","legend":"","description":"","filename":"dcb68fe23a784b20a1e6d59eb0d768ed1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/c60411f36b09f702f5d68fde.xml"},{"id":91838186,"identity":"71cd19af-ac3a-4a00-9e49-930e4c74d9ca","added_by":"auto","created_at":"2025-09-22 09:42:42","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":215316,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/5fc977294a091bb09b71479b.png"},{"id":91834647,"identity":"63253a1c-6bb3-4707-8a44-16e5e2fa4142","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":186452,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/a92df9d7b2f368b3895b3cd5.png"},{"id":91834658,"identity":"4e91313a-ad4b-4966-9e47-79eb2e7cb4ab","added_by":"auto","created_at":"2025-09-22 09:26:43","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":203349,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/51bf6e992812b50f47a306e2.png"},{"id":91836030,"identity":"328201ae-44e9-418a-847a-7a5f5da40f74","added_by":"auto","created_at":"2025-09-22 09:34:42","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":201763,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/ed48d0eccb02888b1dac1328.png"},{"id":91834661,"identity":"2d51fe9b-6a04-4a3b-8c0a-8df3b1c0d3da","added_by":"auto","created_at":"2025-09-22 09:26:43","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":203149,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/9b5f65f1421ec1f6a71610ff.png"},{"id":91836040,"identity":"c72b145c-2599-4ea2-9f4d-5806e975a3dc","added_by":"auto","created_at":"2025-09-22 09:34:43","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":196905,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/587309612f2df07ff2b35ded.png"},{"id":91836035,"identity":"fb8ca8c9-2abb-4c3f-8a42-644ae40b684b","added_by":"auto","created_at":"2025-09-22 09:34:43","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":243957,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/9701c00482f5060b9751fd8a.png"},{"id":91834664,"identity":"b680504b-ac76-4509-89f6-0b4565180c6b","added_by":"auto","created_at":"2025-09-22 09:26:43","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":61653,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/376a5a0632ccb95ba3fd2dd8.png"},{"id":91838190,"identity":"bd8f092e-9914-4629-b386-cf578946a970","added_by":"auto","created_at":"2025-09-22 09:42:43","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":106191,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/8098c765b0ab7d4b8aad5f56.png"},{"id":91834657,"identity":"8ccfd33c-fdc5-427e-86c8-4605c016788c","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":42648,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/5b27a7675761fc49fe9bc699.png"},{"id":91834649,"identity":"af2a8247-c808-4c11-bf7e-143f161995c1","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":43336,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/d937dad2307aaed024e22872.png"},{"id":91836039,"identity":"8ad978bd-1c87-4a3a-95ae-761b9cbb7d2a","added_by":"auto","created_at":"2025-09-22 09:34:43","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":49175,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/43e2377347f636fa76565b7c.png"},{"id":91834654,"identity":"81c9b07c-811b-4f35-9621-8c23d4d7476d","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":39836,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/583aef814b49d6a5806534c4.png"},{"id":91834669,"identity":"b92e26f8-0efd-41de-a495-b382271a133c","added_by":"auto","created_at":"2025-09-22 09:26:43","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":45302,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/4db0dafc690bbbe1c5d5eb38.png"},{"id":91839119,"identity":"c8700905-10d6-446a-b74a-52025e3bf766","added_by":"auto","created_at":"2025-09-22 09:50:43","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":43374,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/6e88538e7cb92c7ebb84eda3.png"},{"id":91834665,"identity":"b15b66dd-2a25-47fd-9823-e6e5e166aad3","added_by":"auto","created_at":"2025-09-22 09:26:43","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":41467,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/3fd3d89030d6d64d0517504e.png"},{"id":91836033,"identity":"3c5cf7aa-065c-4859-bf93-0a949d3edd61","added_by":"auto","created_at":"2025-09-22 09:34:43","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":17188,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/88fadedf56039e6e5c111b80.png"},{"id":91836037,"identity":"1f3c6e63-2f09-408e-a7ae-636f5466873c","added_by":"auto","created_at":"2025-09-22 09:34:43","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":19253,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/99ec14fb8daa246560ca38ae.png"},{"id":91834671,"identity":"9acfa46c-1c92-4e25-ad79-7a5ffdf5721f","added_by":"auto","created_at":"2025-09-22 09:26:43","extension":"xml","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":108815,"visible":true,"origin":"","legend":"","description":"","filename":"dcb68fe23a784b20a1e6d59eb0d768ed1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/bb1e681f9fc2de106f06630f.xml"},{"id":91834672,"identity":"d9c30132-2a55-4398-8515-c36b21518aa7","added_by":"auto","created_at":"2025-09-22 09:26:43","extension":"html","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":120578,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/8912bd60c7badaedbe59aaee.html"},{"id":91834628,"identity":"d50f8332-9d5f-4aa1-9f0e-96720fac770e","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":103231,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow of the study.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/a4730af8e3ed7bf2198923bb.png"},{"id":91834629,"identity":"2c137add-dd2c-4ffa-8151-0e895376c409","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":81935,"visible":true,"origin":"","legend":"\u003cp\u003eInclusion and exclusion of SAKI population in MIMIC-IV and eICU-CRD database.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/95d00273d509af1c2476c7d8.png"},{"id":91836019,"identity":"deeb2253-5bed-4edc-b26f-e43f19348580","added_by":"auto","created_at":"2025-09-22 09:34:42","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":125382,"visible":true,"origin":"","legend":"\u003cp\u003ePearson correlation coefficient matrix.Features with Pearson correlation coefficients \u0026gt; 0.9: (PT, INR), (Hemoglobin, Hematocrit), (RBC, Hematocrit), (RBC, Hemoglobin), (CKD, Renal Disease).\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/f8ce5bb30469ed3df54cdafc.png"},{"id":91834634,"identity":"8a26393c-ee1e-4c12-b509-75fe6b5edbba","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":95173,"visible":true,"origin":"","legend":"\u003cp\u003eFeature screening process.\u003c/p\u003e\n\u003cp\u003eA. Boruta-ranked feature importance distribution. B. Relationship between Lasso regression Mean Squared Error (MSE) and regularization parameter α, C. Lasso regression coefficient paths.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/cea51b1a6747a760ffafaf9b.png"},{"id":91834635,"identity":"37ef7b43-242f-4531-b0c4-10ff8108bcd4","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":99776,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of Internal Data Classifier Performance Metrics\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/d7f946f8a026d93c98abfccd.png"},{"id":91836022,"identity":"236d8256-28c7-4db4-ada2-7c5888366a4e","added_by":"auto","created_at":"2025-09-22 09:34:42","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":97674,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of External Data Classifier Performance Metrics.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/15e45c1e6a05694f4509753c.png"},{"id":91834636,"identity":"03273c98-9dc6-4609-9d73-6ac46bb860b1","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":100299,"visible":true,"origin":"","legend":"\u003cp\u003eCalibration curves in 11 working modles.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/21a1a0e44c34ded921111d1c.png"},{"id":91836025,"identity":"9031b175-1f5c-4efa-ba04-fbb065e2550b","added_by":"auto","created_at":"2025-09-22 09:34:42","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":29455,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion Matrix in Logistic Regression model.\u003c/p\u003e\n\u003cp\u003eA. Precision–recall curve for the logistic regression model. B. Confusion matrix for the logistic regression model\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/a5ee188729a398cd738e3b16.png"},{"id":91834638,"identity":"9d07fa23-9a71-42e7-88d4-9d5fee4296e0","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":45921,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP-based feature importance analysis of the logistic regression model.\u003c/p\u003e\n\u003cp\u003eA. Average impact of each feature on model predictions. B. Detailed impact of each feature on individual predictions.\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/872ad28816e8b0dc8606ba78.png"},{"id":91842156,"identity":"f60f4a3f-9875-4848-8a97-c67681b63828","added_by":"auto","created_at":"2025-09-22 09:58:48","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1088893,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/20749705-b201-44af-94de-a111ace828db.pdf"},{"id":91834632,"identity":"aac34efe-baa5-4237-b27d-f6c748117465","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"csv","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":275608,"visible":true,"origin":"","legend":"","description":"","filename":"eICUsakicleaning.csv","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/9a1743c508a481d18d82d9b9.csv"},{"id":91834641,"identity":"9defce74-1bae-4a5b-a409-84d00d5124dd","added_by":"auto","created_at":"2025-09-22 09:26:42","extension":"csv","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":3684634,"visible":true,"origin":"","legend":"","description":"","filename":"eICUsakiraw.csv","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/a8142f1a2cf885db71996dd4.csv"},{"id":91836023,"identity":"ed22069b-df8d-4116-ae6d-0079f0e4e226","added_by":"auto","created_at":"2025-09-22 09:34:42","extension":"csv","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":3003131,"visible":true,"origin":"","legend":"","description":"","filename":"MIMICsakiraw.csv","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/8ba287f8b92189bfdc655e24.csv"},{"id":91838191,"identity":"279f8c01-ee0f-4538-b7d2-f652b8147239","added_by":"auto","created_at":"2025-09-22 09:42:43","extension":"csv","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":2403585,"visible":true,"origin":"","legend":"","description":"","filename":"MIMICsakicleaning.csv","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/90d2efa32659c40b1dcb1684.csv"},{"id":91839116,"identity":"0935faa9-66b5-4b12-a633-1758635c7502","added_by":"auto","created_at":"2025-09-22 09:50:42","extension":"pdf","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":2736962,"visible":true,"origin":"","legend":"","description":"","filename":"eICUsaki.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/0ac501de1bcc0a55824a6650.pdf"},{"id":91838188,"identity":"ef412a43-b6e5-432d-8786-03a94d2a30df","added_by":"auto","created_at":"2025-09-22 09:42:43","extension":"pdf","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":4269768,"visible":true,"origin":"","legend":"","description":"","filename":"MIMICsaki.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/b9f5bd6567bd668d564033c1.pdf"},{"id":91839118,"identity":"768fa879-64bd-4218-aacc-97929fdc9b8d","added_by":"auto","created_at":"2025-09-22 09:50:42","extension":"pdf","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":7137531,"visible":true,"origin":"","legend":"","description":"","filename":"sakiakdshap.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7313497/v1/52c0a94b9b742c6213ef1f4a.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Interpretable Machine Learning for Early Prediction of Acute Kidney Disease (AKD) in Sepsis-Associated Acute Kidney Injury (SA-AKI): A Multicenter Cohort Study with External Validation ","fulltext":[{"header":"Introduction","content":"\u003cp\u003eSepsis-associated acute kidney injury (SA-AKI) is a critical condition that significantly impacts the morbidity and mortality rates of patients admitted to intensive care units (ICUs)\u003csup\u003e[1-3]\u003c/sup\u003e. As the leading cause of acute kidney injury (AKI) in critically ill patients, sepsis is associated with severe complications, prolonged hospital stays, and increased healthcare costs. Despite advancements in understanding the pathophysiology of SA-AKI, effective diagnostic and therapeutic strategies remain limited, often resulting in delayed interventions and poor patient outcomes\u003csup\u003e[4-6]\u003c/sup\u003e. This gap in timely diagnosis and treatment underscores the urgent need for further research into SA-AKI to enhance early recognition and management in clinical settings.\u003c/p\u003e\n\u003cp\u003eCurrent literature highlights the complex interplay of various risk factors contributing to the development of SA-AKI. Factors such as age, pre-existing comorbidities, and the severity of sepsis are increasingly recognized as critical determinants of patient outcomes\u003csup\u003e[7-9]\u003c/sup\u003e. For instance, older patients and those with underlying health conditions like hypertension and diabetes are at a heightened risk for developing AKI in the context of sepsis. Furthermore, the timing and appropriateness of therapeutic interventions, including the administration of antibiotics and the use of mechanical ventilation, play pivotal roles in influencing patient trajectories\u003csup\u003e[10-12]\u003c/sup\u003e. However, comprehensive insights into the multifactorial aspects of SA-AKI remain limited, necessitating a deeper exploration of clinical characteristics and treatment protocols.\u003c/p\u003e\n\u003cp\u003eTo address these research gaps, this study utilizes advanced machine learning methodologies applied to large-scale datasets from the MIMIC-IV and eICU-CRD databases, which provide rich clinical information on critically ill patients. The application of machine learning techniques offers a powerful framework for analyzing complex datasets and uncovering non-linear relationships among various clinical parameters\u003csup\u003e[13,14]\u003c/sup\u003e. These methods can facilitate the development of predictive models that identify patients at risk of progression from SA-AKI to acute kidney disease (AKD), thereby enabling timely interventions and improved patient care.\u003c/p\u003e\n\u003cp\u003eThe primary objective of this research is to delineate key clinical features and intervention strategies associated with SA-AKI while leveraging machine learning to create robust predictive models (Figure 1). By focusing on specific variables such as the use of vasopressors, mechanical ventilation, and the timing of antibiotic administration, this study aims to enhance the understanding of SA-AKI’s dynamics and its progression towards more severe renal impairment\u003csup\u003e[15]\u003c/sup\u003e. Ultimately, the goal is to develop a clinical decision-support tool that can assist healthcare providers in the early identification and management of SA-AKI, thus improving patient outcomes and reducing the associated burden on healthcare systems\u003csup\u003e[16-18]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eIn summary, this research seeks to bridge the existing knowledge gaps surrounding SA-AKI by employing innovative machine learning techniques to analyze large clinical datasets. Through this approach, the study aims to contribute significantly to the understanding of SA-AKI, informing better clinical practices and enhancing patient management strategies in critical care settings. By elucidating the multifaceted risk factors and treatment protocols associated with SA-AKI, the findings are anticipated to foster improvements in the prevention and management of this serious condition in critically ill patients\u003csup\u003e[19-21]\u003c/sup\u003e.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003eData Sources\u003c/p\u003e\n\u003cp\u003eThe MIMIC-IV (Medical Information Mart for Intensive Care IV) 3.1 and eICU-CRD (eICU Collaborative Research Database) are publicly accessible clinical databases extensively utilized in critical care research\u003csup\u003e[22-26]\u003c/sup\u003e. MIMIC-IV 3.1, jointly maintained by the Massachusetts Institute of Technology (MIT) and Beth Israel Deaconess Medical Center (BIDMC), integrates de-identified data from over 300,000 patients (including ~50,000 ICU admissions) across U.S. hospitals from 2008 to 2022. It encompasses multidimensional data such as demographics, vital signs, laboratory results, medication records, and diagnostic codes (ICD-9/10) . Similarly, the eICU-CRD, developed through a collaboration between Philips and MIT, includes data from 208 U.S. hospitals (2014\u0026ndash;2015), covering \u0026gt;200,000 ICU patients with structured clinical data (e.g., vital signs, medications) and unstructured clinical notes. Both databases are compliant with ethical standards (CITI certification: 69327991), and their use aligns with sepsis-associated acute kidney injury (SA-AKI) research priorities outlined in recent guidelines\u003csup\u003e[27]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eStudy Subjects\u003c/p\u003e\n\u003cp\u003eInclusion criteria: (1)Admission to ICU with confirmed diagnosis of sepsis or septic shock; (2)Age \u0026ge;18 years; (3)Sepsis-associated acute kidney injury; (4) First ICU admission episode. Exclusion criteria: (1)ICU length of stay \u0026lt;48 hours; (2) Survival time \u0026lt;7 days; (3) Incomplete data or missing critical variables.\u003c/p\u003e\n\u003cp\u003eData Extraction\u003c/p\u003e\n\u003cp\u003eVariables extracted included:(1)Demographics: Age, sex, ICU admission time, antibiotic initiation time, AKI onset time, and comorbidities (hypertension, diabetes, heart failure, CKD, etc.). (2)Vital signs: Temperature, heart rate, blood pressure, respiratory rate, and weight (recorded from 1 day pre-AKI to 7 days post-AKI). (3)Laboratory tests: Hemoglobin, leukocyte count, lactate, creatinine, blood urea nitrogen (BUN), electrolytes (Na\u003csup\u003e+\u003c/sup\u003e/K\u003csup\u003e+\u003c/sup\u003e/Cl\u003csup\u003e\u0026minus;\u003c/sup\u003e), coagulation markers (INR, PTT), and arterial blood gas parameters. (4)Interventions: Mechanical ventilation, vasopressors (norepinephrine, vasopressin), and continuous renal replacement therapy (CRRT). (5)Severity scores: Sequential Organ Failure Assessment (SOFA), Acute Physiology Score III (APS III), and Logistic Organ Dysfunction Score (LODS).\u003c/p\u003e\n\u003cp\u003eOutcomes and Definitions\u003c/p\u003e\n\u003cp\u003eThe primary outcome was acute kidney disease (AKD), defined as a sustained decline in kidney function (creatinine-based criteria) persisting 7-90 days post-AKI onset, per the Acute Disease Quality Initiative (ADQI) consensus\u003csup\u003e[28]\u003c/sup\u003e. SA-AKI was classified as early (\u0026le;48 hours post-sepsis diagnosis) or late (48 hours\u0026ndash;7 days). Secondary outcomes included mortality, recurrent AKI, and progression to CKD (\u0026gt;90 days)\u003csup\u003e[29,30]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eData Processing and Statistical Analysis\u003c/p\u003e\n\u003cp\u003eData Cleaning: Outliers were removed based on clinical plausibility (e.g., creatinine \u0026gt;10 mg/dL). Variables with \u0026gt;10% missingness (e.g., urine output) were excluded; others (e.g., lactate) were imputed using mode substitution. The MIMIC-IV dataset was split into training (70%) and testing (30%) sets, while the eICU-CRD served as an external validation cohort. Synthetic Minority Over-sampling Technique (SMOTE) addressed class imbalance in the training set.\u003c/p\u003e\n\u003cp\u003eFeature Selection:\u003c/p\u003e\n\u003cp\u003eFirstly, multicollinearity is treated by calculating the feature-wise Pearson correlation coefficient matrix and setting the threshold at |r| \u0026gt; 0.9 to eliminate multicollinear features, with high-dimensional correlations visualized via a Rank2D heatmap. Secondly, primary feature screening is done using the Boruta algorithm,a wrapper - based method built on a random forest base classifier.Twenty independent experiments are conducted,and features stably selected across all iterations (Rank = 1) are retained based on the median distribution of feature importance rankings. Thirdly, secondary feature compression is carried out using LASSO regression,an embedded feature selection method.The optimal regularization coefficient \u0026lambda; is determined via 10-fold cross - validation,and features with nonzero coefficients are screened by finding the optimal sparse solution using the 1-SE rule. Finally,the feature set is determined by getting the common feature subset chosen by both Boruta and LASSO.This method integrates nonlinear feature importance evaluation and linear sparse constraints.It enhances stability through multiple Boruta iterations and the 1 - SE rule,and is verified through heatmaps,regularization path plots and network topology plots.It effectively improves the feature subset\u0026rsquo;s biological interpretability and model compatibility.\u003c/p\u003e\n\u003cp\u003eMachine Learning Model Construction\u003c/p\u003e\n\u003cp\u003eWe establishes a multi-algorithm integrated machine learning forecasting framework, comprising 11 algorithms in three categories: conventional statistical learning models (Logistic Regression), single learner models (SVM, KNN, Decision Tree), and ensemble learning models (Random Forest, AdaBoost, Gradient Boosting, XGBoost, LightGBM, Bagging, Voting), forming a diverse model ensemble.For hyperparameter optimization, the RandomizedSearchCV strategy is used with ROC_AUC as the optimization target and 5 - fold stratified validation set. Each algorithm has its own parameter space: base learner parameters include decision tree depth (3 - 8 layers) and minimum samples in leaf nodes (2 - 5); ensemble algorithms set the number of base classifiers (10 - 50, step size 10) and learning rate (0.01 - 1, linear distribution); regularization parameters cover L1/L2 norms and elastic net. After parallel search (n_jobs = -1), the best parameter configurations for each algorithm are retained.\u003c/p\u003e\n\u003cp\u003eTo enhance model generalization, a two - level ensemble system is built. The primary ensemble uses Bagging, with the number of base classifiers set in a discrete search space of \u003csup\u003e[31]\u003c/sup\u003e, and diverse predictors are built through resampling with replacement. The advanced ensemble adopts a soft Voting mechanism, integrating the probability outputs of Logistic Regression, Random Forest and Gaussian Naive Bayes, with weight parameters optimized via grid search.A multi-stage model performance evaluation system is established. Initially, models are screened on the training set based on accuracy, precision, recall, F1 - score and ROC - AUC, with AUC as the core criterion for selecting candidate models. Calibration curves are used to assess each model\u0026rsquo;s probability prediction ability. Subsequently, secondary model selection is performed using an independent external validation set, with the model having the highest AUC value chosen as the optimal solution.\u003c/p\u003e\n\u003cp\u003eFor the preferred model, a systematic analysis is carried out. The confusion matrix is used to calculate specificity (Specificity = TN/(TN + FP)) and negative predictive value (NPV = TN/(TN + FN)). A precision - recall curve is plotted to evaluate performance under class imbalance. Finally, the SHAP framework is employed to interpret the decision - making mechanism, including global feature contribution ranking, individual sample decision attribution and feature interaction visualization. All experiments are set with a random seed (random_state = 0) to ensure reproducibility of results.\u003c/p\u003e\n\u003cp\u003eExplainability analysis\u003c/p\u003e\n\u003cp\u003eWe use SHAP values to quantitatively break down the decision logic of \u0026ldquo;black-box\u0026rdquo; models. Its advantages are as follows: First, it creates an axiomatic explanation system based on game theory. Second, it offers sample-level interpretability. Third, it allows for visual verification of feature effect directions\u003csup\u003e[32-34]\u003c/sup\u003e. These advantages make the findings useful for clinical decision-making. They provide a transparent basis for clinical decisions, making predictive models more credible and clinically applicable. By offering a clear understanding of model predictions, this method helps clinicians trust and use these predictions in their work. It also helps find possible model biases or errors, which is important for patient safety. The sample-level interpretability from SHAP values further aids personalized patient care. Using SHAP values bridges the gap between advanced predictive analytics and practical clinical insights. They enhance model interpretability and create a transparent, trustworthy framework for predictions, supporting better-informed clinical decisions.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eThis study selected 9,778 patients from the MIMIC-IV database to form the training cohort and 3,721 patients from the eICU-CRD database to constitute the external validation cohort (Figure 2). Initially, 63 features were incorporated into the training cohort. Subsequently, the Pearson correlation coefficient matrix among these features was calculated, leading to the removal of four highly correlated features (Pearson correlation coefficient \u0026gt; 0.9), specifically International Normalized Ratio (INR), Hematocrit, Red Blood Cells (RBC), and Renal disease (Figure 3). The remaining features underwent data dimensionality reduction using the Boruta feature selection method, which identified 36 features.\u003c/p\u003e\n\u003cp\u003eFurther features selection was performed using LASSO regression, narrowing down the list to 14 features. To refine the feature set and ensure the most relevant features were retained, we took the intersection of the features identified by both methods. This process ultimately determined 14 key features\u003cstrong\u003e\u003cem\u003e\u003cu\u003e\u0026nbsp;\u003c/u\u003e\u003c/em\u003e\u003c/strong\u003e: angiotensin-converting enzyme inhibitor/angiotensin II receptor blocker (ACEI/ARB), Acute Physiology Score III (APS III), continuous renal replacement therapy (CRRT), Cerebrovascular Disease, Logistic Organ Dysfunction System (LODS), interval from infection detection to first antibiotic use (Los_inf._AB), mean blood pressure (MBP), Mechanical Ventilation, Paraplegia, respiratory rate, baseline serum creatinine (Baseline Scr), blood oxygen saturation (SpO\u003csub\u003e2\u003c/sub\u003e), vasoactive Agent, and body weight\u003cstrong\u003e\u003cem\u003e\u0026nbsp;\u003c/em\u003e\u003c/strong\u003e(Figure 4).\u003c/p\u003e\n\u003cp\u003eUsing the training dataset, we constructed 11 machine learning models based on these 14 key features. The performance metrics of these models are illustrated in. In terms of accuracy, Gradient Boosting demonstrated the highest performance at 78.94%, followed closely by LightGBM with 78.29% and Bagging with 78.08%. For the area under the ROC curve (AUC), which is a critical metric for evaluating class discrimination ability, Logistic Regression, Adaboost, and Voting classifier all achieved the highest value of 0.84. This indicates that these models had the best ability to distinguish between different classes. In the recall rate (sensitivity) dimension, which measures the models\u0026apos; ability to identify positive samples, Logistic Regression (0.81) and KNN (0.79) showed superior performance. Notably, Gradient Boosting achieved a balance in both precision (0.57) and F1 score (0.57), making it a well-rounded model in these aspects. Meanwhile, XGBoost (0.51) and LightGBM (0.57) exhibited advantages in precision. Based on a comprehensive evaluation using the F1 score, which balances precision and recall, Logistic Regression (0.61), Adaboost (0.60), and Voting classifier (0.60) demonstrated the best predictive stability (Figure 5).\u003c/p\u003e\n\u003cp\u003eIn the external validation cohort, we selected the 14 relevant features identified in the training phase for model validation. The models generally exhibited performance decay when applied to the external cohort, with AUC values dropping to the range of 0.55\u0026ndash;0.59, compared to 0.78\u0026ndash;0.84 in the training set. Among these models, Adaboost experienced the most significant decay (\u0026Delta;AUC = 0.28), while KNN showed relatively robust predictive characteristics with a recall rate of 0.47 and an F1 score of 0.52. It is particularly worth noting that Logistic Regression (Recall: 0.81\u0026rarr;0.40) and Voting classifier (Recall: 0.68\u0026rarr;0.40), which performed exceptionally well in the training set, experienced a decline of over 50% in recall rate during external validation. Additionally, the accuracy of Gradient Boosting decreased substantially from 78.94% to 54%, highlighting the risk of overfitting in complex models when they are applied to external data (Figure 6).\u003c/p\u003e\n\u003cp\u003eThe probability calibration ability of each model was further assessed using the Brier score (Figure 7). Regarding probability prediction accuracy, Gradient Boosting and LightGBM performed the best with a Brier score of 0.143, followed by Voting (0.146) and Bagging (0.148). These results suggest that ensemble learning methods have significant advantages in probability calibration. However, it is important to note that although Adaboost achieved an ROC AUC of 0.84, it had the highest Brier score of 0.241. This discrepancy indicates that the model may have systematic bias in probability prediction. Among traditional algorithms, Logistic Regression (0.167) and Random Forest (0.156) demonstrated better calibration performance compared to Decision Tree (0.175) and KNN (0.185). When combined with previous metric analyses, it is evident that Gradient Boosting not only maintained high accuracy (78.94%) but also achieved the best calibration characteristics. Although LightGBM had comparable calibration performance, its recall rate (0.48) was significantly lower than that of Logistic Regression (0.81).\u003c/p\u003e\n\u003cp\u003eThe evaluation of the predictive performance of the Logistic Regression model is detailed in Figure 8. Based on 2,934 test samples, the average precision of the Precision-Recall curve was calculated to be 0.63. This curve exhibited a typical downward trend, where precision gradually decreased as recall increased. The confusion matrix revealed that the model correctly classified 1,597 negative samples, corresponding to a specificity of 72.1%, and 585 positive samples, corresponding to a sensitivity of 81.3%. However, it also generated 617 false positives, resulting in a Type I error rate of 27.9%, and 135 false negatives, leading to a Type II error rate of 18.8%. These results indicate that while the model maintains moderate average precision, it demonstrates a relative advantage in discriminating negative classes.\u003c/p\u003e\n\u003cp\u003eSHAP interpretability analysis provided a quantitative description of the feature impact intensity and direction within the Logistic Regression model (Figure 9). Mechanical Ventilation emerged as the feature with the highest contribution (SHAP = 0.79) to the prediction. An increase in its feature value was associated with a significant rise in the predicted risk probability. Vasoactive Agent (0.33) and CRRT (0.33) were identified as secondary driving factors, both exhibiting a positive correlation with the outcome and indicating a positive effect on the model\u0026apos;s predictions. Notably, the use of ACEI/ARB was found to have a negative association with baseline creatinine. This suggests that an increase in the value of ACEI/ARB may lead to a reduction in the risk assessment.\u003c/p\u003e\n\u003cp\u003eTo facilitate the practical application of the model and make it accessible to clinicians, this study developed a web-based risk calculator (\u003cem\u003e\u003cu\u003ehttps://sakiakd-fpohf6y9mmfki7gr3o67pv.streamlit.app/\u003c/u\u003e\u003c/em\u003e). This tool is designed to automatically calculate the risk of SA-AKI patients progressing to AKD by inputting the values of various variables. It aims to provide clinicians with a simple yet practical tool for early prediction of AKD in SA-AKI patients, thereby supporting clinical decision-making and potentially improving patient outcomes.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eSA-AKI is a critical condition observed in patients with severe sepsis, contributing significantly to morbidity and mortality within intensive care units (ICUs)\u003csup\u003e[1]\u003c/sup\u003e. This syndrome arises from a complex interplay of factors, including hemodynamic instability, inflammatory responses, and direct cellular injury, ultimately leading to renal dysfunction\u003csup\u003e[9,35,36]\u003c/sup\u003e. The incidence of SA-AKI has been reported to be alarmingly high, with estimates suggesting that it accounts for up to 70% of acute kidney injury cases in critically ill patients. The pathophysiology of SA-AKI remains incompletely understood, but recent studies indicate that mechanisms such as apoptosis and necroptosis may play pivotal roles in its development\u003csup\u003e[30,37]\u003c/sup\u003e. Moreover, the management of SA-AKI is further complicated by the lack of specific therapeutic interventions, underscoring the necessity for improved diagnostic and treatment strategies.\u003c/p\u003e\n\u003cp\u003eIn light of these challenges, our research aims to elucidate the clinical characteristics and treatment protocols associated with SA-AKI by leveraging extensive data from the MIMIC-IV and eICU-CRD databases. This study employs machine learning methodologies to identify critical risk factors and develop predictive models for the progression of SA-AKI to acute kidney disease (AKD)\u003csup\u003e[38]\u003c/sup\u003e. By examining various clinical parameters, including the timing of antibiotic administration and the use of mechanical ventilation, we aim to enhance understanding of SA-AKI’s complexities and improve patient outcomes\u003csup\u003e[39]\u003c/sup\u003e. The findings from this research will provide insights into optimal management strategies, potentially leading to better prognostic assessments and intervention opportunities for individuals suffering from SA-AKI.\u003c/p\u003e\n\u003cp\u003eThe innovation of this study lies in its application of advanced machine learning techniques to a large-scale dataset to identify critical risk factors associated with SA-AKI. By utilizing data from the MIMIC-IV and eICU-CRD databases, this research provides a comprehensive analysis that not only corroborates existing knowledge but also uncovers novel insights regarding the complex interplay of clinical parameters influencing AKD progression in septic patients. For instance, the identification of ACEI/ARBs as a protective factor against AKD progression aligns with previous findings in animal models but is one of the first to be confirmed in a human clinical context, thus bridging a significant gap in the literature regarding therapeutic strategies in SA-AKI management\u003csup\u003e[40]\u003c/sup\u003e. Moreover, the study heightens the understanding of how demographic factors, specifically age and pre-existing conditions like hypertension and diabetes, contribute to the risk of developing AKI, reinforcing the clinical implications of targeted management strategies\u003csup\u003e[41,42]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eThe implications of these findings for clinical practice are profound. The established risk factors and treatment protocols suggest that timely interventions, such as early administration of ACEI/ARBs and careful monitoring of patients receiving nephrotoxic agents, can significantly improve patient outcomes\u003csup\u003e[43-45]\u003c/sup\u003e. This research advocates for the integration of machine learning-derived predictive models into clinical decision-making processes, which may help clinicians to identify high-risk patients and initiate early intervention strategies, thereby potentially reducing the morbidity and mortality associated with SA-AKI\u003csup\u003e[39]\u003c/sup\u003e. Furthermore, the development of a web-based risk calculator as a practical tool for clinicians exemplifies the application value of this research, facilitating early prediction of AKD in SA-AKI patients and enhancing the overall quality of patient care in intensive care settings\u003csup\u003e[46]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eNevertheless, this study is not without limitations. The observational nature of the data used raises concerns about potential confounding variables that may not have been accounted for, thereby impacting the robustness of the findings. Additionally, the reliance on electronic health records means that the accuracy of the data is contingent upon the quality of documentation in clinical settings, which may vary\u003csup\u003e[47]\u003c/sup\u003e. Future research should aim to conduct prospective validation studies and explore the application of these predictive models in diverse clinical environments to assess their generalizability. Furthermore, integrating more granular patient-level data, including genetic and molecular markers, may provide deeper insights into the pathophysiology of SA-AKI and enhance predictive accuracy\u003csup\u003e[48]\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eThe limitations of this study must be acknowledged, particularly the absence of wet lab experiments, which restricts the validation of findings through biological mechanisms. Additionally, the reliance on multiple datasets may introduce inter-batch variability, potentially affecting the generalizability of our results. The lack of clinical validation analysis raises concerns about the applicability of the predictive models in real-world scenarios, emphasizing the need for further studies to confirm findings in diverse patient populations. Furthermore, the performance decay observed in external validation highlights the potential overfitting of complex models, necessitating cautious interpretation of model efficacy in clinical settings.\u003c/p\u003e\n\u003cp\u003eIn conclusion, this study provides significant insights into the factors associated with the progression of sepsis-associated acute kidney injury to acute kidney disease. By leveraging large-scale clinical data and machine learning techniques, we identified key predictors and established a practical risk calculator to aid clinicians in early intervention strategies. These findings underscore the critical need for timely and tailored management approaches in critically ill patients, ultimately enhancing patient outcomes and resource utilization in intensive care settings.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eSA-AKI: Sepsis-associated acute kidney injury\u003c/p\u003e\n\u003cp\u003eAKD: Acute kidney disease\u003c/p\u003e\n\u003cp\u003eAKI:Acute kidney injury\u003c/p\u003e\n\u003cp\u003eICUs: Intensive care units\u003c/p\u003e\n\u003cp\u003eMIMIC-IV: Medical information mart for intensive care IV\u003c/p\u003e\n\u003cp\u003eeICU-CRD:eICU collaborative research database\u003c/p\u003e\n\u003cp\u003eCRRT: Continuous renal replacement therapy\u003c/p\u003e\n\u003cp\u003eBUN: Blood urea nitrogen\u003c/p\u003e\n\u003cp\u003eCKD: Chronic kidney disease\u003c/p\u003e\n\u003cp\u003eSOFA: Sequential organ failure assessment\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAPS III:Acute physiology score III\u003c/p\u003e\n\u003cp\u003eLODS: Logistic organ dysfunction score.\u003c/p\u003e\n\u003cp\u003eLASSO: Least absolute shrinkage and selection operator\u003c/p\u003e\n\u003cp\u003eACEI: Angiotensin converting enzyme inhibitors\u003c/p\u003e\n\u003cp\u003eARBs: Angiotensin II receptor blockers\u003c/p\u003e\n\u003cp\u003eBIDMC: Beth israel deaconess medical center\u003c/p\u003e\n\u003cp\u003eINR: International normalized ratio\u003c/p\u003e\n\u003cp\u003ePTT: Partial thromboplastin time\u003c/p\u003e\n\u003cp\u003eADQI: Acute disease quality initiative\u003c/p\u003e\n\u003cp\u003eSMOTE: Synthetic minority over-sampling technique\u003c/p\u003e\n\u003cp\u003eSVM: Support vector machine\u003c/p\u003e\n\u003cp\u003eKNN: K-nearest neighbors\u003c/p\u003e\n\u003cp\u003eAdaBoost: Adaptive boosting\u003c/p\u003e\n\u003cp\u003eXGBoost: Extreme gradient boosting\u003c/p\u003e\n\u003cp\u003eLightGBM: Light gradient boosting machine\u003c/p\u003e\n\u003cp\u003eRandomizedSearchCV: Randomized search cross-validation\u003c/p\u003e\n\u003cp\u003eTN: True negative\u003c/p\u003e\n\u003cp\u003eFP: False positive\u003c/p\u003e\n\u003cp\u003eNPV: Negative predictive value\u003c/p\u003e\n\u003cp\u003eFN: False negative\u003c/p\u003e\n\u003cp\u003eSHAP: SHapley additive exPlanations\u003c/p\u003e\n\u003cp\u003eRBC: Red blood cells\u003c/p\u003e\n\u003cp\u003eMBP: Mean blood pressure\u003c/p\u003e\n\u003cp\u003eScr: Serum creatinine\u003c/p\u003e\n\u003cp\u003eSpO2: Blood oxygen saturation\u003c/p\u003e\n\u003cp\u003eLos_inf._AB: Interval from infection detection to first antibiotic use\u003c/p\u003e\n\u003cp\u003eMSE: Mean squared error\u003c/p\u003e\n\u003cp\u003eL1: Least absolute shrinkage and selection operator regression\u003c/p\u003e\n\u003cp\u003eL2: Ridge regression\u003c/p\u003e\n\u003cp\u003e1-SE: One Standard Error\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e● Ethics approval and consent to participate\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e● Consent for publication\u003c/p\u003e\n\u003cp\u003eAll authors reviewed the manuscript and consented for publication.\u003c/p\u003e\n\u003cp\u003e● Availability of data and materials\u003c/p\u003e\n\u003cp\u003eThe datasets generated and/or analyzed during the current study are available in the MIMIC-IV v3.1 and eICU-CRD v2.0 repository/database at https://www.physionet.org/content/mimiciv/3.1/ and https://www.physionet.org/content/eicu-crd/2.0/\u003c/p\u003e\n\u003cp\u003e● Competing Interests\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e\n\u003cp\u003e● Funding\u003c/p\u003e\n\u003cp\u003eThis study is partially supported by the Grant from Key Laboratory of Coronary Intraluminal Imaging and Functional Analysis of Dongguan City and Guizhou Provincial Health Commission Science and Technology Fund Project (gzwkj2025-005).\u003c/p\u003e\n\u003cp\u003e● Authors\u0026apos; contributions\u003c/p\u003e\n\u003cp\u003eShuang Chen (S.C.) and Guang Li (G.L.) contributed equally to this work. Conceptualized the study, developed machine learning algorithms, performed model training and internal validation. Qingzhan Zeng (Q.Z.Z.) curated multicenter clinical data, designed data collection protocols. Xiancheng Xu (X.C.X.) conducted statistical analyses and interpreted model performance metrics. Chanlin Li (C.L.L.) processed external validation datasets and performed feature engineering. Xiaoyue Li (X.Y.L.) assisted in clinical data annotation and quality control. Shaohong Li (S.H.L.) provided critical revisions for clinical relevance, methodology and finalized the manuscript. Heng Li (H.L.) (Corresponding Author) supervised the entire study, acquired funding, and finalized the manuscript.\u003c/p\u003e\n\u003cp\u003e● Acknowledgements\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eChen Y, Jing H, Tang S, et al. Non-coding RNAs in Sepsis-Associated Acute Kidney Injury. Front Physiol. 2022;13:830924. \u003c/li\u003e\n\u003cli\u003eTakeuchi T, Flannery AH, Liu LJ, et al. Epidemiology of sepsis-associated acute kidney injury in the ICU with contemporary consensus definitions. Crit Care. 2025;29(1):128. \u003c/li\u003e\n\u003cli\u003eMweene MD, Richards GA, Paget G, et al. Risk factors and outcomes of sepsis-associated acute kidney injury in intensive care units in Johannesburg, South Africa. S Afr Med J. 2022;112(12):919-923. \u003c/li\u003e\n\u003cli\u003eOdum JD, Wong HR, Stanski NL. A Precision Medicine Approach to Biomarker Utilization in Pediatric Sepsis-Associated Acute Kidney Injury. Front Pediatr. 2021;9:632248. \u003c/li\u003e\n\u003cli\u003ePeerapornratana S, Manrique-Caballero CL, G\u0026oacute;mez H, et al. Acute kidney injury from sepsis: current concepts, epidemiology, pathophysiology, prevention and treatment. Kidney Int. 2019;96(5):1083-1099. \u003c/li\u003e\n\u003cli\u003ePeng W, Li G. Research progress of microRNAs in sepsis-associated acute kidney injury. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2022;34(5):556-560. \u003c/li\u003e\n\u003cli\u003eSaverymuthu A, Teo R, Zain JM, et al. Acute Kidney Injury following Rhabdomyolysis in Critically Ill Patients. J Crit Care Med. 2021;7(4):267-271 \u003c/li\u003e\n\u003cli\u003eLouzada CF, Ferreira AR. Evaluation of the prevalence and factors associated with acute kidney injury in a pediatric intensive care unit. J Pediatr (Rio J). 2021;97(4):426-432. \u003c/li\u003e\n\u003cli\u003eMagboul SM, Osman B, Elnour AA. The incidence, risk factors, and outcomes of acute kidney injury in the intensive care unit in Sudan. Int J Clin Pharm. 2020;42(6):1447-1455. \u003c/li\u003e\n\u003cli\u003eEswarappa M, Gireesh MS, Ravi V, et al. Spectrum of acute kidney injury in critically ill patients: A single center study from South India. Indian J Nephrol. 2014;24(5):280-285. \u003c/li\u003e\n\u003cli\u003eMahesh E, Nallamuthu P, Kumar M, et al. Clinical profile of geriatric acute kidney injury in a tertiary care center from south India. Saudi J Kidney Dis Transpl. 2017;28(4):886-890. \u003c/li\u003e\n\u003cli\u003eChen K, Lei Y, He Y, et al. Clinical outcomes of hospitalized COVID-19 patients with renal injury: a multi-hospital observational study from Wuhan. Sci Rep. 2021;11(1):15205. \u003c/li\u003e\n\u003cli\u003eBaciu C, Xu C, Alim M, et al. Artificial intelligence applied to omics data in liver diseases: Enhancing clinical predictions. Front Artif Intell. 2022;5:1050439. \u003c/li\u003e\n\u003cli\u003eTseng AS, Noseworthy PA. Prediction of Atrial Fibrillation Using Machine Learning: A Review. Front Physiol. 2021;12:752317. \u003c/li\u003e\n\u003cli\u003eTan HK, Kaushik M, Tan CW, et al. Augmented Adsorptive Blood Purification during Continuous Veno-Venous Haemodiafiltration in a Severe Septic, Acute Kidney Injury Patient: Use of oXiris\u0026reg;: A Single Centre Case Report. Blood Purif. 2019;47 Suppl 3:1-6. \u003c/li\u003e\n\u003cli\u003eXu L, Sun P. Identification and management of sepsis associated-acute kidney injury. Zhonghua Wei Zhong Bing Ji Jiu Yi Xue. 2023;35(2):221-224. \u003c/li\u003e\n\u003cli\u003eZarbock A, Nadim MK, Pickkers P, et al. Sepsis-associated acute kidney injury: consensus report of the 28th Acute Disease Quality Initiative workgroup. Nat Rev Nephrol. 2023;19(6):401-417. \u003c/li\u003e\n\u003cli\u003eTorres JSS, Tamayo-Giraldo FJ, Bejarano-Zuleta A, et al. Sepsis and post-sepsis syndrome: a multisystem challenge requiring comprehensive care and management\u0026mdash;a review. Front Med (Lausanne). 2025;12:1560737.\u003c/li\u003e\n\u003cli\u003eDe Rosa S, Marengo M, Fiorentino M, et al. Extracorporeal blood purification therapies for sepsis-associated acute kidney injury in critically ill patients: expert opinion from the SIAARTI-SIN joint commission. J Nephrol. 2023;36(7):1731-1742. \u003c/li\u003e\n\u003cli\u003eRicci Z, Polito A, Polito A, et al. The implications and management of septic acute kidney injury. Nat Rev Nephrol. 2011;7(4):218-225. \u003c/li\u003e\n\u003cli\u003eParmar A, Langenberg C, Wan L, et al. Epidemiology of septic acute kidney injury. Curr Drug Targets. 2009;10(12):1169-1178. \u003c/li\u003e\n\u003cli\u003eYuan ZN, Xue YJ, Wang HJ, et al. A nomogram for predicting hospital mortality of critically ill patients with sepsis and cancer: a retrospective cohort study based on MIMIC-IV and eICU-CRD. BMJ Open. 2023;13(9):e072112. \u003c/li\u003e\n\u003cli\u003eSheng S, Li A, Zhang C, et al. Association between hemoglobin and in-hospital mortality in critically ill patients with sepsis: evidence from two large databases. BMC Infect Dis. 2024;24(1):1450. \u003c/li\u003e\n\u003cli\u003eSchmidt L, Pigat L, Sheikhalishahi S, et al. Evaluating the SWIFT algorithm\u0026apos;s efficacy in predicting hypoxemia across multiple critical care datasets. J Crit Care. 2025;89:155123. \u003c/li\u003e\n\u003cli\u003eQi Z, Dong L, Lin J, Duan M. Development and validation of a nomogram prediction model for early diagnosis of bloodstream infections in the intensive care unit. Front Cell Infect Microbiol. 2024;14:1348896. \u003c/li\u003e\n\u003cli\u003ePeng X, Cai Y, Huang H, et al. A Predictive Model for Acute Kidney Injury Based on Leukocyte-Related Indicators in Hepatocellular Carcinoma Patients Admitted to the Intensive Care Unit. Mediators Inflamm. 2025;2025:7110012.\u003c/li\u003e\n\u003cli\u003eRamya K, Mukhopadhyay K, Kumar J. Predictive factors and risk scoring system for acute kidney injury (AKI) in sick neonates\u0026mdash;a prospective cohort study. Eur J Pediatr. 2024;183(12):5419-5424. \u003c/li\u003e\n\u003cli\u003ePatel M, Hornik C, Diamantidis C, et al. Patient-level factors increase risk of acute kidney disease in hospitalized children with acute kidney injury. Pediatr Nephrol. 2023;38(10):3465-3474. \u003c/li\u003e\n\u003cli\u003eShi B, Ye J, Chen W, et al. Prognosis of critically ill patients with early and late sepsis-associated acute kidney injury: an observational study based on the MIMIC-IV. Ren Fail. 2025;47(1):2441393. \u003c/li\u003e\n\u003cli\u003eKounatidis D, Vallianou NG, Psallida S, et al. Sepsis-Associated Acute Kidney Injury: Where Are We Now? Medicina (Kaunas). 2024;60(3):434. \u003c/li\u003e\n\u003cli\u003eTahir S, Ganie BA, Beigh TY, et al. Clinico-Etiological Spectrum and Outcome in Patients With Septic Acute Kidney Injury and Its Comparison With Non-septic Acute Kidney Injury: A Hospital-Based Prospective Study Conducted in a Tertiary Care Hospital in North India. Cureus. 2023;15(4):e37857. \u003c/li\u003e\n\u003cli\u003eYi F, Yang H, Chen D, et al. XGBoost-SHAP-based interpretable diagnostic framework for Alzheimer\u0026rsquo;s disease. BMC Med Inform Decis Mak. 2023;23(1):137. \u003c/li\u003e\n\u003cli\u003eRoder J, Maguire L, Georgantas R, et al. Explaining multivariate molecular diagnostic tests via Shapley values. BMC Med Inform Decis Mak. 2021;21(1):211. \u003c/li\u003e\n\u003cli\u003eRiasi A, Delrobaei M, Salari M. Personalized medication recommendations for Parkinson\u0026apos;s disease patients using gated recurrent units and SHAP interpretability. Sci Rep. 2025;15(1):19074. \u003c/li\u003e\n\u003cli\u003eHolland EM, Moss TJ. Acute Noncardiovascular Illness in the Cardiac Intensive Care Unit. J Am Coll Cardiol. 2017;69(16):1999-2007.\u003c/li\u003e\n\u003cli\u003e. Glodowski SD, Wagener G. New insights into the mechanisms of acute kidney injury in the intensive care unit. J Clin Anesth. 2015;27(2):175-180. \u003c/li\u003e\n\u003cli\u003eWu Z, Deng J, Zhou H, et al. Programmed Cell Death in Sepsis Associated Acute Kidney Injury. Front Med (Lausanne). 2022;9:883028. \u003c/li\u003e\n\u003cli\u003eCheungpasitporn W, Thongprayoon C, Kashani KB. Artificial intelligence and machine learning\u0026apos;s role in sepsis-associated acute kidney injury. Kidney Res Clin Pract. 2024;43(4):417-432. \u003c/li\u003e\n\u003cli\u003eHuo L, Liu C, Yuan Y, et al. Pharmacological inhibition of ferroptosis as a therapeutic target for sepsis-associated organ damage. Eur J Med Chem. 2023;257:115438. \u003c/li\u003e\n\u003cli\u003eHasson DC, Watanabe-Chailland M, Romick-Rosendale L, et al. Choline supplementation attenuates experimental sepsis-associated acute kidney injury. Am J Physiol Renal Physiol. 2022;323(3):F255-F271. \u003c/li\u003e\n\u003cli\u003eJakubov K, Petr V, Zahradka I, et al. Acute Kidney Injury in Deceased Organ Donors: Risk Factors And Impacts on Transplantation Outcomes. Transplant Direct. 2024;10(12):e1730. \u003c/li\u003e\n\u003cli\u003eCheng Y, Nie S, Zhao X, et al. Incidence, risk factors and outcome of postoperative acute kidney injury in China. Nephrol Dial Transplant. 2024;39(6):967-977. \u003c/li\u003e\n\u003cli\u003eWang PT, Huang YB, Lin MY, et al. Prescriptions for angiotensin-converting enzyme inhibitors/angiotensin receptor blockers and monitoring of serum creatinine and potassium in patients with chronic kidney disease. Kaohsiung J Med Sci. 2012;28(9):477-483. \u003c/li\u003e\n\u003cli\u003eFang G, Annis IE, Farley JF, et al. Incidence of and Risk Factors for Severe Adverse Events in Elderly Patients Taking Angiotensin-Converting Enzyme Inhibitors or Angiotensin II Receptor Blockers after an Acute Myocardial Infarction. Pharmacotherapy. 2018;38(1):29-41. \u003c/li\u003e\n\u003cli\u003eChen JY, Tsai IJ, Pan HC, et al. The Impact of Angiotensin-Converting Enzyme Inhibitors or Angiotensin II Receptor Blockers on Clinical Outcomes of Acute Kidney Disease Patients: A Systematic Review and Meta-Analysis. Front Pharmacol. 2021;12:665250. \u003c/li\u003e\n\u003cli\u003eGao L, Wang GD, Yang XY, et al. Development of a risk prediction model for sepsis-related delirium based on multiple machine learning approaches and an online calculator. PLoS One. 2025;20(7):e0323831.\u003c/li\u003e\n\u003cli\u003eThoroddsen A, Sigurj\u0026oacute;nsd\u0026oacute;ttir G, Ehnfors M, et al. Accuracy, completeness and comprehensiveness of information on pressure ulcers recorded in the patient record. Scand J Caring Sci. 2013;27(1):84-91. \u003c/li\u003e\n\u003cli\u003eZhang Q, Yang B, Li X, et al. Biomarkers of cell cycle arrest, microcirculation dysfunction, and inflammation in the prediction of SA-AKI. Sci Rep. 2025;15(1):8023.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Sepsis-associated acute kidney injury (SA-AKI), Acute kidney disease (AKD), Interpretable machine learning, Early prediction, External validation","lastPublishedDoi":"10.21203/rs.3.rs-7313497/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7313497/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eSepsis-associated acute kidney injury (SA-AKI) represents a critical challenge in the management of critically ill patients, significantly contributing to morbidity and mortality in intensive care units (ICUs). This study aims to enhance the understanding and prediction of SA-AKI progression to acute kidney disease (AKD) by utilizing a comprehensive machine learning approach. Data from the MIMIC-IV and eICU-CRD databases were analyzed, incorporating 14 key clinical features identified through rigorous feature selection methods, including Boruta and LASSO regression. Eleven machine learning models were developed, with Gradient Boosting demonstrating the highest accuracy (78.94%) and optimal calibration characteristics. The external validation cohort revealed a decrease in model performance, emphasizing the risk of overfitting in complex models. Notably, the use of ACE inhibitors/ARBs was associated with a reduced risk of AKD progression, while nephrotoxic agents significantly increased this risk. Prognostic scoring systems, including SOFA and LODS, were found to correlate significantly with AKD outcomes, facilitating better risk stratification. Furthermore, a web-based risk calculator was developed to provide clinicians with an accessible tool for predicting the risk of SA-AKI progression to AKD based on individual patient data. In conclusion, this study underscores the importance of timely interventions and tailored treatment strategies in the management of SA-AKI, while also paving the way for future research to refine predictive models and improve clinical outcomes in critically ill patients.\u003c/p\u003e","manuscriptTitle":"Interpretable Machine Learning for Early Prediction of Acute Kidney Disease (AKD) in Sepsis-Associated Acute Kidney Injury (SA-AKI): A Multicenter Cohort Study with External Validation ","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-22 09:26:37","doi":"10.21203/rs.3.rs-7313497/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"135122241483304056619534973716251095452","date":"2025-09-23T17:25:27+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-09-11T06:04:47+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-08-20T10:12:21+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-08-16T18:08:48+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-08-13T03:39:24+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Informatics and Decision Making","date":"2025-08-13T03:36:18+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f5367188-56ad-448f-9587-1e171a3c0bc0","owner":[],"postedDate":"September 22nd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2025-09-22T09:26:37+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-22 09:26:37","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7313497","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7313497","identity":"rs-7313497","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}