| Objectives1.To describe the clinical features related to death and the occurrence of organ dysfunction and its severity in patients with sepsis;2.To develop prediction models of death at different time points based on clinical data and sequence characteristic data,as well as the best prediction model of organ dysfunction(shock,respiratory function,coagulation function,liver function and renal dysfunction)and its severity of sepsis patients at different time points;3.To compare and optimize prediction models,and to validate them externally in real-world data based on the effect of death and organ dysfunction prediction in sepsis patients.Methods1.A retrospective cohort study was conducted in data from the MIMIC-III database and the Xiangya Hospital ICU inpatient sepsis database from 2016-2021.Using general characteristics of sepsis patients and clinical characteristics from day 1 to day 5 after ICU admission as the predictive variables,death and organ dysfunction(respiratory dysfunction,coagulation dysfunction,liver dysfunction,renal dysfunction,shock,)as the outcome variables,the prediction models of death and organ dysfunction and its severity of sepsis patients were developed with the sequential data.2.After data preprocessing,Sepsis patients in MIMIC-III database were randomly divided into training set and internal validation set.The general characteristics and clinical information during the hospitalization on day 1,3,and 5 among patients were defined as the predictor variables in the training set.Six classifier algorithms(Random Forest Classifier,Logistic Regression,Ridge Classifier,MLP Classifier,support vector machine,Gaussian naive Bayes),7 feature selection algorithms(linear support vector Classifier,Elastic Net,Ridge Regression,Logistic Regression,Random gradient descent classifier,Random Forest Classifier,Lasso),3 data imbalance processing algorithms(the synthesis of a few oversampling techniques,the combination of oversampling and undersampling,and undersampling methods)and 2 data preprocessing methods(standard deviation standardization,robust standardization)were combined to develop the 6 * 7 * 3 * 2=252 kinds of the combination algorithms.The early warning and severity prediction types of sepsis death and multiple organ dysfunction were respectively established and selected42 prediction models,each model corresponds to a group of optimal combination algorithms.3.The prediction accuracy,sensitivity,specificity,F1 score and the Receiver Operating Characteristic AUC(Area Under the curve,AUC)of each model were calculated,respectively,and the prediction performance of each model was comprehensively evaluated by AUC to screen out the optimal prediction model.The models in training set and the internal validation set were performed with 5-fold cross validation to assess the models,meanwhile,the optimal prediction model was externally verified with the retrospective cohort data from Xiangya Hospital sepsis database.4.Based on the basic data and the clinical predictor variables during hospitalization from day 1 to day 5 among sepsis patients in the MIMICIII database,LSTM neural network was used to develop the temporal and non-temporal prediction model of death and organ dysfunction(respiratory dysfunction,coagulation dysfunction,liver dysfunction,renal dysfunction,shock)in sepsis patients,respectively.Results1.Basic characteristics: A total of 5784 patients with sepsis were included in the MIMIC–III database,and 1652 septic patients in the Xiangya Hospital database,respectively.During the first 5 days before the hospital stay,the incidence rate of patients in the MIMIC-III with respiratory dysfunction,coagulation dysfunction,liver dysfunction,renal dysfunction,septic shock,and deaths within 30 days,180 days and 365 days were 457(81.3%),286(25.5%),63(5.5%),138(43.5%),17(1.2%),1095(18.9%),1604(27.7%)and 1830(31.6%),respectively.And the incidence rate of patients in the Xiangya Hospital database were232(41.7%),120(11.5%),205(75.7%),184(23.3%),70(8.4%),331(20.0%),357(21.6%)and 367(22.2%),respectively.2.In the development of the prediction model of sepsis death,three death prediction models with a moderate predictive value were developed in this study.The AUC of these models were 0.870(95%CI:0.867-0.872),0.811(95%CI:0.807-0.815),0.801(95%CI:0.795-0.807),respectively.The optimal model was the random forest model for predicting death events within 30 days of sepsis with the predictors screened by LASSO based on the clinical information for day 1,3 and 5.4627 cases in training set and an 1157 cases in internal validation set,and with selected 127 predictors were included in the development of the models.The external validation results of the data of sepsis patients in Xiangya Hospital showed that the AUC of each model maintained above 0.690,with an accuracy of more than 74%.3.Predictive model of sepsis-related organ dysfunction: 39 predictive models of sepsis-related organ dysfunction were developed in this study,and 23 models had good predictive value of AUC greater than 0.7.The optimize models were as follows:(1)In the prediction model of sepsis-related respiratory dysfunction,the five models predicting the severity of sepsis-related respiratory dysfunction with high predictive value of AUC greater than 0.8 were developed in this study.The AUC of the models were 0.854(95%CI:0.847-0.862),0.850(95%CI:0.845-0.855),0.840(95%CI:0.829-0.851),0.839(95%CI:0.831-0.847),0.825(95%CI:0.819-0.832),respectively.The optimal model was the multilayer perception prediction models of severity of sepsis-related respiratory dysfunction with the predictors screened by ridge regression based on the clinical information for day 1 clinical information to predict day 3compared to day 1.1979 cases in training set and an 495 cases in internal validation set,and with selected 22 predictors were included in the development of the models.The maximum AUC for external verification was 0.848.(2)In the prediction model of sepsis-related coagulation dysfunction,Three early warning models of sepsis-related coagulation dysfunction and six predictive models for its severity were developed in this study and the AUC was between 0.712 and 0.776.The optimal model was the random forest model for predicting of sepsis-related coagulation dysfunction with the predictors screened by the stochastic gradient-descent classifier based on the clinical information for day 1 and day 3 to predicting day 5compared to day 1.The AUC of the model was 0.776(95% CI:0.759-0.792).1196 cases in training set and an 300 cases in internal validation set,and with selected 31 predictors were included in the development of the models.(3)In the predictive models of sepsis-related liver injury,the three models predicting the severity of sepsis-related liver injury with a good predictive value of AUC were developed in this study.The AUC of the models were 0.826(95% CI:0.809-0.844)、0.795(95% CI:0.769-0.822)、0.789(95% CI:0.759-0.820),respectively.The optimal model was the random forest model for predicting of sepsis-related liver injury with the predictors screened by the LASSO based on the clinical information for day 1 and day 3 to predicting day 5.880 cases in training set and an 221 cases in internal validation set,and with selected 76 predictors were included in the development of the models.(4)In the prediction models of sepsis-related kidney injury,one early warning model of sepsis-related renal dysfunction and five severity prediction models all with AUC greater than 0.7 were developed,and the highest AUC of the model was 0.962(95%CI:0.956-968).The optimal model was the random forest model for predicting the severity of sepsisrelated renal dysfunction with the predictors screened by the linear support-vector classifier based on the clinical information for day 1 and day 3 to predicting day 5 compared to day 1.986 cases in training set and an 247 cases in internal validation set,and with selected 9 predictors were included in the development of the models.The external validation performance was acceptable.4.In this study,36 temporal models of adverse prognosis of sepsis were established based on LSTM.The optimal models were as follows:(1)In the death temporal model of sepsis patients,four wellperforming models were developed in this study.Based on the use of temporal and non-temporal variables,high predictive value were found in the model of LSTM-Day5-Static,LSTM-Day-5-Static-Linear,LSTMDay X-Static and LSTM-Day-Static-Static-Linear,which The highest AUC were from 0.927 to 0.953.The LSTM-Day X-Static-Linear model established by inputting temporal variables and non-temporal variables into LSTM and Linear layer,can dynamically predict death in sepsis patients with the first day of clinical information.The AUC of patient death is 0.939.This model needs external validation and clinical transformation.(2)In the temporal model of sepsis-related respiratory dysfunction,two excellent performance models were obtained in this study.LSTMDay5-Static and LSTM-Day X-Static models based on temporal variables and non-temporal variables and both input into LSTM had high predictive value,with AUC of 0.736 and 0.793,respectively.(3)In the temporal model of sepsis-related coagulation dysfunction,a model with excellent performance was obtained in this study.The AUC of LSTM-Day5-Static-Linear model based on the clinical information for day4 to predicting day 5 is 0.722.(4)In the temporal model of sepsis-related liver dysfunction,two excellent performance models were developed in this study.The LSTMDay5 and LSTM-Day X models based on the temporal variables had high predictive value,with an AUC of 0.805 and 0.815,respectively.Conclusions1.Based on the clinical information of days 1,3,and 5,the predictive models for death prediction developed in this study have good performance.The models are the the random forest model for predicting patient death at30 days with the screened predictors by LASSO regression,and the models for predicting patient death at 180 days and 365 days with the screened predictors by elastic network regression and random forest algorithm,respectively.2.Based on the clinical information of day 1 of sepsis patients,the early warning model of liver dysfunction and the prediction model for the severity of respiratory dysfunction developed in this study have excellent performance.The models are a support vector machine model for early warning of sepsis-related liver dysfunction on day 3 of admission with the screened the predictors by elastic network regression,and a multilayer perceptual prediction model of predicting respiratory function severity on day 3 compared to day 1 with the screened the predictors by ridge regression,respectively.The findings implicated that those models have the important clinical significance for the early identification of sepsis patients with liver dysfunction and the deterioration of respiratory dysfunction.3.Based on the clinical sequence information on days 1 and 3 for sepsis patients,with the screened the predictors by stochastic gradient descent and a linear support vector classifier,the optimal random forest model to predict the severity of coagulation dysfunction and renal dysfunction on day 5 compared to day 1 in sepsis patients have been developed in this study.The random forest model for predicting the occurrence of liver dysfunction on day 5 with screened the predictors by LASSO have a good prediction effect.4.Both the temporal variable and non-temporal variable prediction models based on LSTM can effectively predict death in sepsis patients,and the temporal variables and non-temporal variables play a vital role in the modeling process.The best performance on dynamically predicting patients death on day X+1 are the model of LSTM-Day X-Static and LSTM-Day X-Static-Linear based on the inputted temporal variables and non-temporal variables,respectively.Both of the models have excellent future prospects of clinical transformation.5.The LSTM temporal models effectively introduced by time series information have better prediction ability of sepsis organ dysfunction(Respiratory dysfunction,coagulation dysfunction,liver dysfunction,and renal dysfunction).The LSTM-Day5 and LSTM-Day X models developed only by the temporal variables had the best performance in predicting liver dysfunction,while the LSTM-Day5-Static and LSTM-Day X-Static models based on temporal and non-temporal variables into LSTM have excellent performance in predicting respiratory dysfunction,respectively.LSTMDay5-Static-Linear model has a excellent predictive effect on sepsisrelated coagulation dysfunction.Figures: 46,Tables: 79,References: 135... |
PDF Full Text Request