Research On ICU Patient Death Risk Prediction Combined With Medical Treatment Process Mining

The risk assessment of the patients in icu is of great significance to the determination of the follow-up treatment and the reduction of the cost of the rescue.In the past,the work was mainly based on the severity of the disease and some machine learning algorithms to predict the risk of death.Both methods use only the life and physiological variables of the patient to describe the severity of the disease.In fact,the patient’s medical treatment process also implies that the valuable information of the severity of the patient’s condition can provide the decision information for the patient’s death risk.In order to use this information,this paper probes into the drug and drug treatment patterns of the patients,and the life and physiological variables of the patient,and the risk of death of the patient.This paper proposes a method to predict the risk of death by combining the mining of patients’ medical process with their physiological feature.The main research contents are as follows:(1)the LDA drug efficacy topic clustering:assuming the same diseases of patients in the treatment of produced in the process of the treatment of drugs according to its efficacy is divided into multiple topics,and the drug efficacy and polynomial distributed subject distribution,at the same time patients daily treatment drugs and is subject to the drug efficacy theme,according to a polynomial distribution which can use the LDA theme model training from the history of the patient’s medication history logs,patients with each drug efficacy of diagnosis and treatment,subject distribution and the drug efficacy under the theme of drug distribution,this can lay a foundation for the subsequent medical process;(2)Prediction model of death risk based on probability suffix tree and random forests:after complete drug clustering,through calculating the similarity of patients daily drug efficacy subject distribution,drug efficacy of different treatment,clustering scheme combination,and the efficacy of diagnosis and treatment,distribution within the same cluster use unified combination drug efficacy label to mark,the patients with drug treatment processes can be converted into a combination drug efficacy sequence tag;It is assumed that the drug efficacy combination of patients’ daily medication is subject to the variable-order markov model,so the drug treatment process of patients can be modeled with the help of probability suffix tree.Then,patients were divided into four categories according to the similarity of the drugs they took and the final treatment results,with each category corresponding to a probability suffix tree model.To calculate the similarity between the drug treatment sequence of the patients to be predicted and the above four probabilistic suffix tree models.Using the similarity and the data of the patient’s physical signs as input,the random forest classification model was used to train the prediction model of the patient’s risk of death;(3)Prediction model of death risk based on LSTM neural network and random forests:Patients during hospitalization of drugs is a kind of time series data,and both short-term and long-term memory network,can learn to depend on the information for a long time,is a good way to solve the problem of long dependence,through daily drug distribution subject,theme LDA model obtained by theme distribution as its characteristics,using network model,both short-term and long-term memory of time-series data classification prediction,to predict probability data as input,and the patient’s various signs,using random forest classification model,training the patient’s death risk prediction model;The feasibility and effectiveness of the proposed regimen were evaluated in combination with the data of prescription medication and physiological signs of patients with MIMIC-III clinical data of sepsis and pneumonia.From the experimental results,based on probability suffix tree and random forest combined with double model method is more suitable for septicemia the shorter length of hospital stay,the diseases were relatively fixed treatment mode,and for complex treatment,treatment mode,length of hospital stay longer disease using LSTM both short-term and long-term memory network and with the method of random forest effect is better.The precision rate,recall rate,F1 value and other indicators obtained by these two methods are better than the traditional feature-based methods.