Detection Of ICU Sepsis And Prediction Of Mortality Based On Machine Learning

With the advent of the era of artificial intelligence and big data,intelligent medicine and precision medicine are gradually emerging.Judging the severity of the disease and predicting the risk of death are the focus and difficulty of research in the ICU.Sepsis is a systemic inflammatory response syndrome caused by infection,and it is one of the common high-risk complications and the leading cause of death in patients with ICU.Sepsis is a dangerous and rapidly developing disease,which can endanger life if it is not treated on time.Early prediction of sepsis and effective treatment are the keys to preventing the disease from worsening,improving the patient’s prognosis,and reducing the risk of death.Aiming at the current difficulty and low accuracy of clinical diagnosis of sepsis,based on machine learning methods,different ICU clinical data are used to carry out data mining research on ICU patients,and prediction models about early diagnosis and death risk of sepsis are established respectively.The main researches in this article are:1.According to the ICU clinical data set provided by the Physionet website,an integrated algorithm model based on weighted fusion is proposed to provide early warning of sepsis.Based on40-dimensional patient physiological variables per hour recorded in advance to predict the occurrence of sepsis.Firstly,preliminary statistics and comparative analysis are made on the distribution,variable dimensions,and sample size of the provided data set.Then,according to the time series and imbalance characteristics of the ICU data set,the data is pre-processed,including the use of forwarding filling of missing values and outlier processing.From a medical perspective,based on the understanding of variables,variable screening,and feature engineering are performed.According to the characteristics of time series,feature variables of different dimensions are constructed for model training.Finally,an integrated algorithm framework for model fusion is proposed.Model fusion is performed on random forest,XGBoost,and Lightgbm.The experimental results show that the weighted fusion model performs better than a single integrated algorithm.2.According to the current situation of sepsis prone to sepsis in ICU trauma patients,the early diagnosis of sepsis in trauma patients was carried out based on the MIMIC-III database.First,determine the research objectives,extract relevant research cohorts and predictive variables from MIMIC data,and mainly extract EHR data within 24 hours after patients enter the ICU,including variables such as patient demographic variables,vital signs,and laboratory tests.Then,the deep forest(gc Forest)algorithm is studied.Based on this,an improved deep forest algorithm(Improvedgc Forest)model is proposed for early diagnosis of sepsis.Through the comparison of experimental results,the improved deep forest algorithm model is better than other commonly used machine learning algorithm models,and it also obtains early warning factors that have an important impact on the diagnosis of sepsis.3.To further study the prognostic outcome of patients with traumatic sepsis in ICU-the risk of death,to select patients with sepsis from the previously extracted ICU traumatic patients,and to predict the risk of death.Aiming at the problems that the extracted data is more missing,five different missing value filling methods are used for comparison,and the best filling strategy is selected.Compare the performance of different machine learning algorithms(LR / SVM / XGBoost,etc.)and compare them with traditional medical scoring systems.The results show that the use of machine learning to predict the risk of death in patients with sepsis is better,with AUROC results increasing by 21%.