Prognostic Studies On Sepsis-associated Acute Kidney Injury Based On Machine Learning

BackgroundSepsis is life-threatening organ dysfunction caused by a dysregulated host response to infection,which is one of the major causes of health loss worldwide.Acute kidney injury(AKI)is one of the common complications in patients with sepsis,characterized by an abrupt increase in serum creatinine or decline in urine output.Sepsis-associated acute kidney injury(SA-AKI)is a clinical syndrome characterized by the simultaneous presence of consensus criteria for both sepsis and AKI.SA-AKI is closely associated with poor prognosis.However,the management of SA-AKI is limited to the control of risk factors for renal injury and symptomatic supportive treatment.The consensus report of the Acute Disease Quality Initiative(ADQI)Workgroup on acute kidney disease(AKD)and renal recovery proposed new definitions of transient and persistent AKI as well as AKD.Transient AKI is defined as the complete and sustained reversal of AKI within 48 h of AKI onset,while persistent AKI is defined as the continuance of AKI beyond 48 h from AKI onset.In addition,AKD is defined as acute or subacute damage and/or loss of kidney function for a duration of between7 and 90 d after exposure to an AKI initiating event.However,these definitions were mainly based on the results of previous studies and expert opinions.Further studies are needed to evaluate their applicability in different populations.Early discrimination between transient and persistent AKI in patients with sepsis is essential to evaluate patient prognosis,develop clinical management strategies,and determine the target population for clinical trials.In addition,given that SA-AKI is associated with rapid clinical evolution and significantly increased mortality,real-time mortality prediction models are needed to help identify high-risk patients early,dynamically evaluate treatment response,and provide timely clinical interventions.Recently,the rapid development in big data and advances in machine learning techniques,along with the data-rich environment in intensive care unit(ICU)settings,have provided unprecedented opportunities to establish prognostic models for patients with SA-AKI.Objectives1.The study aimed to investigate the association between renal recovery status after AKI and outcomes in a cohort of septic patients from China based on the ADQI consensus,so as to provide a basis for the prognostic assessment,management,and follow-up of patients.2.The study sought to combine big data and machine learning to distinguish early between transient and persistent AKI in patients with sepsis.3.The study aimed to develop and validate real-time prediction models for in-hospital death in patients with SA-AKI,based on the machine learning e Xtreme Gradient Boosting(XGBoost)algorithm combined with highly dimensional and granular big data of ICU patients.Methods1.The study retrospectively analyzed clinical data of septic patients who were admitted to the ICUs at the Second Xiangya Hospital of Central South University from April 2016 to December 2021.Patients who developed AKI were classified according to the ADQI consensus as transient or persistent AKI,and non-AKD or AKD.The primary outcomes included major adverse kidney events within 30 d(MAKE30)and death within 30 d,90 d,and 1 year.Logistic regression models were used to analyze the association of renal recovery status after AKI with MAKE30,and the odds ratio(OR)and 95% confidence interval(CI)were calculated.Cox proportional hazard regression models were used to analyze the association of renal recovery status after AKI with death within 30 d,90 d,and 1 year,and the hazard ratio(HR)and 95% CI were calculated.In addition,sensitivity analyses were performed to verify the robustness of the results in different patient subgroups.2.Eligible patients with SA-AKI were identified from the Medical Information Mart for Intensive Care(MIMIC)-Ⅲ database and classified as transient or persistent AKI according to the ADQI consensus.Patients were randomly assigned to the training and test sets in a ratio of 7:3.In the training set,five machine learning algorithms were used to establish prediction models for transient and persistent AKI,including Logistic regression,random forest,support vector machine,artificial neural network,and XGBoost.In the test set,the performance of the models was evaluated.On this basis,a simplified prediction model was also derived based on Logistic regression and features selected by the XGBoost algorithm and Least Absolute Shrinkage and Selection Operator regression.3.Eligible patients with SA-AKI in the MIMIC-Ⅳ database were identified and randomly allocated to the training,validation,and internal test sets in a ratio of 5:3:2.Eligible patients with SA-AKI in the e ICU Collaborative Research Database(e ICU-CRD)were screened as an external test set.The primary outcome was in-hospital death within 28 d after ICU admission.Each patient’s ICU stay within 28 d after ICU admission was separated into 12-hour windows.Clinical variables within each time window were used to predict the risk of in-hospital death in the following 48 h,72 h,and 120 h and within 28 d after ICU admission.In the training set,the XGBoost algorithm was used to establish prediction models.In the validation set,the hyperparameters of the models were optimized.In the internal and external test sets,the performance of the models was assessed.Additionally,sensitivity analysis was further performed to assess the predictive value of clinical data gathered in the early period after ICU admission for in-hospital death within 28 d.Results1.A total of 1,513 patients with sepsis were enrolled,of whom 950(62.8%)developed AKI.Among patients with SA-AKI,77.8% developed persistent AKI and 68.9% developed AKD.Compared with no AKI,transient AKI(OR 2.43,95% CI 1.48-4.00,P < 0.001)and persistent AKI(OR 35.73,95% CI 24.45-52.21,P < 0.001)were both independent predictors of MAKE30.Compared with no AKI,persistent AKI was an independent predictor of death within 30 d(HR 1.57,95% CI 1.23-2.02,P < 0.001),90 d(HR 1.47,95% CI 1.18-1.82,P = 0.001),and 1 year(HR1.47,95% CI 1.20-1.80,P < 0.001),but transient AKI was not.Similarly,compared with no AKI,AKI with no AKD(OR 2.85,95% CI 1.82-4.46,P < 0.001)and AKD(OR 50.96,95% CI 34.21-75.90,P < 0.001)were both independently associated with MAKE30.Compared with no AKI,AKD was an independent predictor of death within 30 d(HR 1.75,95%CI 1.36-2.25,P < 0.001),90 d(HR 1.64,95% CI 1.32-2.04,P = 0.001),and 1 year(HR 1.60,95% CI 1.31-1.97,P < 0.001),but AKI with no AKD was not.Sensitivity analyses showed that renal recovery status after AKI was closely associated with outcomes in multiple patient subgroups.2.A total of 5,984 patients with SA-AKI were included,of whom3,805(63.6%)developed persistent AKI.Among the five machine learning models,the artificial neural network model and the Logistic regression model exhibited the best predictive performance,with the area under the receiver operating characteristic curves(AUCs)being 0.76(95%CI 0.74-0.78)in the test set.The 14?variable simplified model showed good discrimination and calibration,with an AUC of 0.76(95% CI0.73-0.78)and Brier score of 0.187 in the test set.At the optimal cutoff of0.63,the simplified model achieved a sensitivity of 62.6% and a specificity of 76.2% in the test set.A nomogram and a risk calculator(http://xydsbakiteam.xyeyy.com)were further developed to facilitate the application of the simplified model.3.A total of 15,603 patients with SA-AKI were included,including12,132 from the MIMIC-Ⅳ database and 3,471 from the e ICU-CRD database.The XGBoost models showed good predictive performance for in-hospital death over different time periods,and were superior to the clinical risk scores.The AUCs of the XGBoost models ranged from 0.85(95% CI 0.84-0.86)to 0.81(95% CI 0.80-0.81)in the internal test set and from 0.82(95% CI 0.81-0.83)to 0.75(95% CI 0.74-0.76)in the external test set.In addition,the interpretability of the models was improved by the feature importance based on XGBoost and the SHapley Additive ex Planations method,which identified important predictors of in-hospital death in patients with SA-AKI.Sensitivity analysis showed that both the XGBoost model and the clinical risk scores showed poor predictive performance for in-hospital death within 28 d when only clinical data gathered within 12 h after ICU admission were used.Conclusions1.Both persistent AKI and AKD are associated with poor prognosis in patients with sepsis.2.Machine learning algorithms can distinguish early between transient and persistent AKI in patients with sepsis and identify important predictors.The simplified model is a practical tool for risk stratification and management of patients with SA-AKI.3.This study developed and validated real-time prediction models for in-hospital death in patients with SA-AKI based on machine learning,providing powerful tools for prognostic assessment and clinical decision-making.