| As a brain disease,stroke is one of the most common diseases in clinical medicine,with high mortality and disability rate.There are many risk factors leading to stroke.It is very important and challenging to detect the risk of individual stroke before onset.With the in-depth research of artificial intelligence in the field of intelligent medicine,the risk prediction of stroke based on machine learning method has become a research hotspot in recent years.The collection and annotation of stroke data often requires a lot of resources,and many of the collected datasets are imbalanced.The common machine learning classification algorithms usually incline to the features of the majority samples when building models,which makes the prediction accuracy of the minority samples low.As a machine learning method different from supervised and unsupervised learning,reinforcement learning obtains the cumulative expectation value in the process of continuous interaction between agent and environment to learn the optimal strategy.At present,scholars at home and abroad directly use reinforcement learning and environment interaction to build non-equilibrium data classification prediction model,which is a hot research topic in recent years.In this paper,a new deep reinforcement learning classification prediction model is proposed.Combined with the non-equilibrium characteristics of stroke data,from the data level,feature level and model level,the construction of stroke classification prediction model is studied.The main work of this paper is as follows:Firstly,through the cooperation with Shanxi Provincial People’s hospital to obtain stroke data samples,consult the neurology clinical experts,and preprocess the collected sample data.In this paper,a new deep reinforcement learning(nDRL)classification prediction model is proposed.The model regards the classification problem as a Markov decision process,and uses deep Q network algorithm to learn the corresponding classification strategy.The model can select the best feature subset,and then use the constructed deep Q network to make classification prediction for the input data.SVM,ELM and BA-ELM are used as the comparison methods.Experiments are carried out on UCI and TCD datasets.The results show that nDRL has better performance in stroke classification and prediction.Secondly,in the process of mapping the classification problem to Markov decision-making,because of the redundancy and irrelevance of the input samples,the calculation of model optimization is too large;in addition,the agent is easy to fall into the dilemma of "exploration" and "utilization" in the process of interacting with the environment,which reduces the performance of the model.Aiming at these defects,this paper proposes a new feature selection and new deep reinforcement learning(FS-nDRL)classification and prediction model based on feature selection.Firstly,the chi-square test feature selection algorithm is used to eliminate redundant features from the initial dataset,reduce the feature dimension to reduce the search range of action space in the sequence decision-making process,and reduce the computational complexity of classification prediction model construction;and the linear decay ε-greedy strategy is used to optimize the process of agent "exploration" and "utilization" to further improve the performance of the model.The experimental results show that the FS-nDRL classification and prediction model can obtain better classification and prediction accuracy and reduce the time complexity for UCI and TCD datasets.Thirdly,a new deep reinforcement learning imbalanced classification(nDRLI)model is proposed to deal with the imbalance of stroke data distribution,the proposed model improves the interaction rules between the agent and the environment to guide the agent to improve the performance of classification and prediction of unbalanced data,and makes the agent pay more attention to the minority samples.Firstly,the classification prediction process is regarded as an independent event.In each step,the agent gets a real-time reward and punishment.When the classification of a few classes is correct,the agent is given a higher reward than when the classification of the majority classes is correct;when the classification of a few classes is wrong,the agent is given a greater punishment than the classification of many classes.Taking some common datasets on KEEL and stroke screening datasets as the experimental objects,the performance of the model is verified.The experimental results show that,compared with the existing models CCR-ELM,MN-DRL and S-DRL,the proposed method nDRLI is superior to the existing models in terms of classification prediction accuracy,Gm and Se. |
