Assessment Method Of Cerebral Autoregulation Based On Unsupervised Deep Learning

In recent years,artificial intelligence technology has experienced rapid development and is widely used in various fields of people’s lives.Among them,the medical and health innovation field has become one of the most promising and meaningful applications.Machine learning and deep learning related technologies have been widely used in many medical researches,providing large-scale qualitative and quantitative analysis methods,and giving birth to new types of diagnosis and prediction methods,further guiding doctors’ decision-making.However,the combination of the medical field and artificial intelligence still faces some technical difficulties,especially the lack of sufficient label data sets and applicable network models.This work focuses on the very important field of cerebral blood flow autoregulation in medicine.It pays attention to the contradiction that requires a ‘gold standard’ to mark autoregulation status and no method is recognized as the ‘gold standard’ to establish the labeled dataset,but the deep learning method often requires a large number of labeled datasets for training and learning.Therefore,this paper will look at the plight of the lack of labeled dataset from a new perspective.We mainly apply deep learning methods to the assessment and analysis of cerebral autoregulation,and respectively propose classification model based on semi-supervised manner and clustering model based on un-supervised manner,using unlabeled fragment data segments,simulation data,surrogate data,and real recorded data with label and without label for training and testing,the experiments have achieved conducive results.Therefore,based on the assessment of cerebral autoregulation,combined with machine learning and deep learning methods,this paper conducts the following main research contents:1.A semi-supervised classification model based on the variational autoencoder is proposed.The variational autoencoder and the gaussian mixture model are used to make the hidden variables from different categories approach different normal distribution probability densities,thereby establishing various hidden variables in the feature space.Thus,the density area of various hidden variables helps to improve the classification effect,and can generate different types of simulation data at the same time.The experimental result found that the distribution of the two types of simulation data generated by the variational autoencoder is significantly different,and the classifier has a significant effect on the learning of simulation data,reaching the accuracy of 99%.2.Apply a state-of-the-art unsupervised clustering methods on the surrogate data to carry out a proof-of-concept study.First we generate two types of surrogate data by different autoregulation indices from Tiecks’ model,which are regarded as normal and impaired autoregulation status under ideal conditions.And apply the joint training of reconstruction loss and clustering loss of deep embedded clustering model,in a completely un-supervised way,making the center layer learn more features that are conducive to clustering,thereby improving the effectiveness of clustering.Compared with the one-stage and two-stage clustering methods,the deep embedded clustering model with the joint optimization has significant separation performance,reaching a accuracy of 98.9% on the surrogate data,which proves the effectiveness in discriminating the normal and impaired autoregulation status of the deep embedded clustering model.3.Further optimize the un-supervised deep embedded clustering model and apply it to the real collected unlabeled data.In order to further enhance the credibility of the learning of dataset,we collect the unlabeled recorded data from a larger number of subjects;then apply the deep embedded clustering model for joint optimization,test and analyze the labeled dataset and the surrogate dataset.It is found that the joint optimization model can still achieve the best accuracy compared with other methods,and compared with training with surrogate data,the accuracy trained with real data can be improved by 4% on testset.This paper aims at the assessment of cerebral autoregulation,applying semi-supervised classification network and unsupervised deep clustering model to explore the performance on segment data through window sliding,simulation data generated by generation model,surrogate data through autoregulation indices and real recorded data from subjects respectively.In this way,a step by step transfer to the clinical reality can still achieve a better separation accuracy of normal and impaired cerebral autoregulation without the need for labeled datasets.The experimental results obtained in this paper are helpful to promote the development of the assessment classifier of cerebral autoregulation,which is enlightening and driving force for clinical analysis and diagnosis.