Bayesian Network Structure Learning Algorithm Research And Its Application In Epileptic Prediction

Epilepsy is a chronic brain disorder,characterized by recurrent sudden seizures,which not only pose a significant threat to the physical health of patients,but also disrupt the normal daily life of both patients and their families.Seizure episodes are dynamic processes,and if it is possible to issue a warning before a seizure occurs,timely intervention can prevent some accidental injuries.Therefore,real-time and accurate prediction of epileptic seizures holds great significance for patients with epilepsy.In this thesis,we conducted research on seizure prediction methods based on Bayesian networks,and the main work is as follows.(1)Aiming to address the problems of a large search space and a tendency to fall into local optima in Bayesian network structure learning process,a novel Bayesian network structure learning algorithm based on the hybrid of differential evolution and seagull optimization(HDE-SO)is proposed.This algorithm first employs the maximum spanning tree and simplified hill climbing method to construct a well-initialized population,effectively reducing the search space of Bayesian network structures.Then,an adaptive cosine adjustment factor is introduced for population partitioning,balancing the global search and local search capabilities of the algorithm.Finally,point crossover and column crossover operations are performed on each sub-population,and differential evolution and seagull optimization strategies are used for mutation operations,leading to the updating of subpopulations.The simulation experiments on the Asia and Car standard Bayesian networks have shown that the proposed algorithm has higher learning accuracy.(2)A seizure prediction method is proposed based on dynamic Bayesian network.Firstly,preprocessing,feature extraction and feature selection are performed on electroencephalogram(EEG)to determine the network nodes.Then,a static Bayesian network structure is constructed using the HDE-SO algorithm,and a transition network is established based on expert knowledge.The static Bayesian network is then unfolded along the time series to obtain a dynamic Bayesian network model.Finally,post-processing operations are performed on the inference values output by the dynamic Bayesian network to realize epileptic seizure prediction.On CHB-MIT dataset,the proposed method achieved an average sensitivity of 94.72%,an average false alarm rate of 0.1/h,and an average prediction time of 38.29 min。...