Research On Dynamic Bayes Network's Learning Algorithm And Application In Survive Prediction Of Intensive Care Unit

Background:Patients in intensive care unit?ICU?always have severe illness,along with quick change of vital signs.Malignant events occur frequently among these patients leading to lower survival probability.There are various scale methods or models for severity assessment or mortality prediction in ICU patients;however,we can only have a vague understanding of survival probability or health degree depending on the information at a specific time point or admission time.Early warning on their specific adverse vital signs changing trends or underlying organic damage cannot be offered by above traditional methods.Dynamic Bayes networks?DBNs?is a branch of directed acyclic graphs?DAGs?in the field of time series,as a complex version of hidden Markov model.DBNs can reflect causality by visual nodes and directed arcs.Referring to time slices conception,DBNs can construct a whole network by extracting relevant knowledge from expert knowledge system or directly studying relationship between the variables.nodes between adjacent time slices are connected to express their causal relationship.When all causal relationships were learned in the network,it can be said a complete topological structure has been learned.Then we can give these arcs certain parameters or learn parameter space from a data set to make final causal inference.DBN is like a multivariable time series,predicting or causal inferring variation trends of variables.But compare to linear equation and time series model,DBNs canbetter explain causal relationships or interaction effects between different variables.It provides a possible solution to predicting prognosis of ICU patients based on longitudinal data,however,how to apply DBNs to predict indices and mortality is a newly developing field,especially with mixed data type,remains to further investigation.There are still many challenges to construct a global optimal network.Objective:applying different DBN learn algorithm in constructing dynamic variation model of indices of ICU patients.Observe and assess the performance and prediction ability of each algorithm.Compare DBNs prediction methods with other ICU patient mortality prediction model and evaluate the application value of DBNs in ICU patients.Methods:1.DBN learning algorithm research based on data with different characteristicsThis research realizes na?ve DBNs and DBN with Gaussian mixture model?GMM?in R software based on the existing DBN learning algorithm.Na?ve DBNs is a type of network composed of discrete nodes?variables?but no continuous ones which can be learned by three main classes of algorithm?score-based algorithm,constraint-based algorithm and hybrid algorithm?.Score-based algorithm applied in this study includes hill-climbing searching method and tabu searching methods;Bayes information criterion?BIC?is chosed as score criterion in this class.Growth-shrinking algorithm,incremental association Markov blanket?IAMB?algorithm,fast IAMB algorithm,interleaved IAMB algorithm,IAMB algorithm with false discovery ratio?FDR?control,PC algorithm are realized in this study which are included in constraint-based algorithm.Hybrid algorithm,which includes max-min hill-climbing algorithm?MMPC?,restricted maximum algorithm?RSMAX?and ARCANE algorithm,combines former two algorithms,aiming to shorten learning time.Max-min parents-children algorithm and Si Hiton parents-children algorithm are two other local constraint algorithms.The above 13algorithms are implemented when data is completed;however,when data is incompleted,expectation-maximization?EM?algorithm is needed to iterate two score-based algorithms to find an optimal solution.All abovementioned 15 algorithms are belonging to na?ve DBN learning methods.In addition,in order to explore the performance of DBNs in mixture data set?containing continuous and discrete variables?,we constructed DBN with GMM.This model does not restrict distributions of continuous variables and can avoid information loss in discretization.So,it can improve prediction accuracy rate theoretically.Maximum likelihood estimation is used to learn parameters of DBNs.2.DBN Constructions Based on Longitudinal ICU Monitoring DataSepsis patients from MIMIC-?database is extracted as real example data source in this study.Patients'demographic information,vital signs and labotory examination results are acquired through Structured query language?SQL?.Completed dataset is built with cross link and polynomial interpolation.Then we use Chi Merge and percentage classification to discretize continuous variables to built a discrete data set.Finally,abovementioned 16 DBN learning algorithms were implemented to build DBNs in ICU.3.Assessment of DBNs Based on Longitudinal ICU Monitoring DataTen-folds cross validation was implemented in evaluation of DBNs,dividing dataset into training set and test set.Firstly,na?ve DBNs were intercompared by average prediction accuracy loss function and likelihood loss function of all nodes in second time slice.Then,we take area under receiver operating curve?AUROC?of 24 hours survive prediction as primary indicator,sensitivity,specificity and learning time as secondary indicators to assess the prediction model.Furthermore,performances and application value of DBNs are compared to logistic regression,support vector machine and scoring scale.Results:This paper illustrates DBN algorithms elaborately and realizes it in softwares.Na?ve DBN structure learning methods are universally classified into four major classes:score-based algorithm,constraint-based algorithm,hybrid algorithm and others.Score-based algorithm searches optimal structure through space searching algorithm?hill-climbing algorithm,etc.?and scoring candidaters by score criteria.Constraint-based algorithm identifies conditional correlations between nodes by Conditional independence test;then integrates the information into a DAG.Hybrid algorithm is a combined use of former two algorithms,using constraint-based algorithm to get a set of DAG candidates,then scoring them with score criteria to find the optimal one.Besides,mutual information and local discovery algorithm are used to learn DBN structure.DBN with GMM is an algorithm that applies GMM on the measures of continuous variables.Hidden node is introduced to indicate the different submodel of GMM.This algorithm can efficiently learn DBN structure under the condition that data is discrete-continuous mixed and not high dimention.This algorithm is comprehensive and easy implementing with GMM which is universally used in science.Furethermore,the model does not require strict demends in distribution of continuous variables.By screening appropriate cases,1994 sepsis patients in MIMIC-?database were extracted as real example dataset.Patients'sex,age,insurance,ethnity,type of admission,and source of admission,etc.were acquired as consistent information.Time of first record in ICU is taken as first time slice;four-hours is taken as time interval to extract patients'vital signs and laboratory examination results.Average values of above indices among the four hours were calculated as actual numerical value in each slice.Vital signs include heart rate,respiratory rate,oxyhemoglobin saturation,systolic pressure,diastolic pressure and body temperature;laboratory examinations include white blood cell count,hemoglobin,blood platelet count,blood glucose,creatinine,blood urea nitrogen,alanine aminotransaminase,and aspartate aminotransferase.By interpolation,discretesization and cross-link of T₀-T₁datasets,analysis dataset is built with 105399 records of two time-slices combination.16 DBN structure learning algorithms and maximum likelihood estimate are used to learn DBNs in ICU.Finally,we get thirteen na?ve DBN learned from comleted dataset,two DBNs from original dataset with EM algorithm and score-based searching methods,and one DBN with GMM among ICU sepsis patients.Ten-folds cross validation of prediction accuracy loss function and log-likelihood loss function among na?ve DBNs shows difference of prediction capability of these models.Score-based algorithms have lowest false prediction rate?19.02%?and lowest predction log-likelihood loss?36.14?.EM algorithm takes longest time when combined with hill-climbing algorithm?112.43 minutes?.Hill-climbing algorithm applied in complete dataset is best performed comprehensively and only need 1.35 seconds to learn DBN structure.PC algorithm takes longest time?71.47 seconds?except EM algorithms.Instant survival judgement model is built with logistic regression having values extracted from the last time slice of all patients,taking survive/death as response variable and nodes in DBN as random variables.Prediction is done with node values at 24 hours before last time slice as assumed current condition.We use DBNs to iterate 6 times of four-hours interval probability calculation from assumed current condition to get the predicted value at the last time slice of each patients.Then 24-hours survival prediction of DBNs is done by the abovementioned instant survival judgement model with these predicted values.There is no significant difference among the prediction AUROC of Na?ve DBNs which are around 0.866.Compared with DBN with GMM and other24-hours survival prediction model,AUROC of DBN with GMM?0.850?is just below the na?ve DBNs,but no statistical difference;however,predction ability of logistic regression?0.836?and support vector machine?SVM??0.828?algorithm are significantly below DBNs.ConclusionBy constructing DBN learning algorithms and implementing them in sepsis patients from MIMIC-?,this study compares different DBN learning algorithms.10-folds cross validation is implemented to compare AUROC,sensitivity and specifity of different algorithms.DBN algorithm learns and infers the causal relationship between nodes and then predict the node values at next time slice.It can provide more informations for doctors and nursed comparing with other prediction model.Cross validation in case study shows that PC algorithm,RSMAX,and Si hiton parents and children algorithm having the highest AUROC;PC algorithm has the highest sensitivity,DBN with GMM has the highest specifity.From a comprehensive perspective,PC algorithm,hill-climbing algorithm and MMPC are the best one in three major algorithm classes respectively.DBN with GMM can make more precise prediction,but has no difference in survival forcast when compared with na?ve DBNs.Furthermore,DBNs have stronger prediction ability than both logistic model and score scale.