A Research On Prediction Of Acute Kidney Injury Of Hospitalized Patientsfrom A Multi-center Clinical Data

Acute kidney injury(AKI)is one of the most common critical diseases in inpatients,which has high morbidity and mortality.Constructing the early risk prediction model of AKI and analyzing the key factors that affect its occurrence are very beneficial for AKI prevention,timely intervention in hospitalized patients with AKI risk,and implementing effective treatment.A total of 580,461 hospitalized patients(46395AKI,534066 non-AKI)from China collaborative study on AKI(CCS-AKI)as a multicenter including 16 hospitals were enrolled.Basic information,hematological indicators,medications,hospital department,Charlson comorbidity index,length of stay and other clinical information were collected.The admission departments,drug types,and hematological indicators were standardized,and the statistical values of hematological indicators(such as maximum value,minimum value,standard deviation,mean value,etc.),the days of drug used,the drug types and Charlson total score of Charlson comorbidity index were created,and a total of 114 features were obtained.A machine learning model called LightGBM was exploited for early prediction of AKI occurrence at three time point of 24?48?72 hours,and the important features were acquired,then were analyzed using SHAP(shapley additive explanations).In addition,the effect of serum creatinine(SCr)level on the performance of prediction model were elaborately compared and analyzed.All subjects were divided into training sets and independent validation sets with the ratio of 4:1 to train,validate and test the LightGBM model.The experimental results showed that the comprehensive evaluation F1 scores of the three time-period prediction models were 0.612 ? 0.705,the AUC values were 0.919 ? 0.946,and the accuracies were 0.946 ? 0.956.When the SCr levels were removed from the feature set,the LightGBM still exhibited prominent performance.Furthermore,important features mainly included length of stay,hospital department,age,the features related to SCr,the days of use of anti-infective drugs,chemotherapy drugs and diuretic drugs,the number of potential nephrotoxic drugs used,the last value of sodium(Na),urinalysis(UA),Chlorine(Cl),and potassium(K),and Charlsontotal score.This study suggested that LightGBM could effectively predict the risk of AKI in hospitalized patients and identify the important features related to AKI.Consequently,it has important guiding significance and clinical application value for AKI prevention,timely interventionin hospitalized patientswith AKI risk,and clinical treatment decision.