Research On Bayesian Network And Its Applications In The Prognosis Prediction For ICU Patients

The Bayesian network is a new mechanism for uncertain knowledge representation based on probability theory and graph theory. It is a graphical model that combines the directed acyclic graph with the conditional probability and indicates the dependencies between variables and causal relationship. Since Bayesian network provides a unique form to express uncertainty and to easily integrate priori knowledge with intuitive reasoning results and the like, it has been widely applied into fault detecting, medical diagnosing, traffic management, financial investment and market analysis, etc.In this paper, we thoroughly conducted research on structure learning and inference algorithms of Bayesian network. More importantly, Bayesian network has been successfully applied to the intensive care unit (ICU) for prognosis prediction in the Elderly. The main contributions of this thesis are listed as following.(1) After the characteristics of the Minimum Description Length (MDL) measure and K2algorithm are analyzed, Bayesian network structure learning algorithm based on MDL and K2(KMBN) is proposed in this paper. Experimental results show that KMBN algorithm is superior to both a conventional Bayesian Network structure learning algorithm, i.e. K2, and a structure learning algorithm based on simulated annealing and K2in respect of reliability and time complexity.(2) Bayesian inference is a key to employ uncertain knowledge embedded into Bayesian network. However, exact inference and approximate inference have been proved to be NP-hard problem. The junction tree algorithm is one of the most popular and general inference algorithms to realize exact inference. An Improved Adaptive Genetic Algorithm (IAGA) is proposed to optimize nodes’eliminating sequence during the Bayesian network triangulation through adjusting crossover and mutation operators in Genetic algorithm. Experimental results show that the convergence and prediction performance are improved.(3) In critically ill patients, the elderly account for a large proportion and take up many ICU resources. However, the treatment effect and prognosis are still unclear. Therefore, prognosis research for elderly critically ill patients is important. In this paper, an evaluation model of the prognosis for elderly critical patients based on Bayesian networks is constructed. Firstly, a Bayesian approach based on Minimum Description Length (MDL) and K2algorithm is utilized to obtain the optimal network structure, and then the maximum likelihood method is used for parameter learning. At last, Bayesian inference is employed to achieve the prediction results. Four-fold cross sampling experiment results show that the prediction accuracy of the model presented in this paper is superior to both conventional BP Neural Networks and K2algorithm, and the prediction accuracy has been improved by6.87%and27.20%, respectively. It is helpful for the doctors to estimate how much elderly critical patients benefit from ICU treatment.