Research And Implementation Of Death Risk Prediction Model For ICU Septic Patients Based On Machine Learning

The death risk assessment of patients in Intensive Care Unit(ICU)is of great help to determine the treatment plan and improve the survival probability of patients.Most of the existing death risk prediction models based on machine learning algorithm need to take the patient’s blood routine examination results as the input features of the model,which has many features and takes a long time to collect.To solve the above problems,this thesis proposes a prediction model of death risk for ICU septic patients based on machine learning.Only clinical real-time features are used to predict the death risk of patients in the next five hours,and the number of input features is reduced by 71%while maintaining the prediction accuracy.The main work of this thesis includes the following three aspects:(1)Because the original data samples contain many features,an improved wrapped feature selection method is proposed.Aiming at the problems existing in the traditional wrapped feature selection method,random permutation and full combination are introduced to optimize.From hundreds of input features,17 features with high correlation with patients’death risk are selected,which lay the foundation for the follow-up prediction of death risk.(2)After feature selection,the amount of data in the data set is still huge.Most of the data are continuous time series data,which is relatively complex,and the time series data can’t be directly used for eXtreme Gradient Boosting(XGBoost)prediction.Therefore,this thesis first constructs Long Short-Term Memory(LSTM)network,and realizes the data dimensionality reduction of preserving the time series information by obtaining the hidden layer state.Then build the XGBoost death risk prediction model.Since the mutation of patient time series data may have more information than the data itself,the maximum,minimum and variance of data and the dimensionality reduction results of data are taken as the input of XGBoost model.Finally,a comparative experiment is carried out on the public data set.The experiment shows that the prediction model proposed in this thesis reduces the number of input features by 71%on the premise of maintaining the prediction accuracy.(3)Based on the proposed prediction model and the data set used for model training,a death risk prediction system for ICU septic patients based on machine learning is realized,including data receiving layer,data analysis layer and application display layer.The system can provide users with functions such as real-time monitoring of patients’ body and prediction of death risk.