| Stroke is one of the main diseases that endanger the life of middle-aged and old people And nowadays,there is no effective treatment for stroke disease.In order to reduce the occurrence of stroke disease,clinical doctors think the best method is that they intervene in the behavior of stroke patients’ early stages.In recent years,deep learning has been widely used in different areas,including the fields of image,language,voice and so on.Furthermore,in the existing stroke prediction research,compared with traditional methods,the performance of deep learning is outstanding.However,to build a good stroke prediction model,deep learning methods demand a large amount of labeled data.However,the actual stroke data are distributed in the form of small data subsets in hospitals in different regions.Intuitively,the approach to deal with the problem of the amount of data required for deep learning is to directly integrate all the relevant data and train it to build a good prediction model.However,in the real world,because of the strict privacy protection policy in the medical system,the medical data between different hospitals and regions can not be shared.In reality,there are two imbalances in stroke data.On the one hand,intuitively,large-scale hospitals have more stroke data and there are few stroke data in community hospitals.On the other hand,the other imbalance is that the number of patients with stroke is much less than the number of patients without stroke.In order to solve the above problems,we will discuss our proposed method as follows:1.In this paper,a hybrid deep transfer learning approach is used to construct stroke prediction model.The specific data were collected from three years before the occurrence of the stroke disease and demographic data such as sex and age.The prediction model predicted whether stroke would occur after three years.The proposed model based on hybrid deep transfer learning approach is composed of three parts.First,we use the method of Generative Instance Transfer(GIT)method to transfer the synthetic data from different regions for the distributed and non-shareable stroke data.Second,we used the Active Instance Transfer(AIT)method to filter out the most informative transferred samples from the generated samples Thirdly,to exploit the relevant data within the hospital,we achieved Network Weight Transfer(NWT)method for hypertension,diabetes,and other chronic diseases that are different but related to stroke.Therefore,the issues of small data,data imbalance,distributed data were solved by the above methods.In the experimental part,in order to verify the proposed approach in different scenarios,we conduct experiment in the simulation environment,we applied this method to stroke and chronic disease data from three hospitals with geographical distribution,and compared it with existing stroke prediction methods2.The network weight transfer learning approach can make use of the data of multiple source domains.The models of different source domains are trained and carried out in a certain order.Finally,the network weight is transferred to the target domain.Although the performance is better than using a single source domain,it is time-consuming to find the optimal transfer structure model,especially when the network structure is very complex Therefore,in this paper,we propose a network structure optimization method based on Bayesian Optimization,in which the function distribution is estimated by using Gaussian Process and then the next possible optimal point is evaluated by using a specific sampling function.Finally,we validate the proposed method by experiment.And the future work of this research is discussed in the last section. |
PDF Full Text Request