Research On Multi-label Classification Algorithm Of Medical Data Stream Based On Deep Sequence Mode

Medical data streams contain rich human pathological information.Accurate classification of medical data stream based on AI technology is currently a hot and challenging research topic.A segment of medical data stream usually contains multiple disease anomalies,so a sample usually corresponds to multiple class labels.Multi-label classification of medical data stream can identify multiple concurrent diseases simultaneously,and has high medical diagnostic value.Currently,the multi-label classification of medical data stream still faces the following two challenges:(1)how to fully utilize the relevance between labels to improve the overall classification accuracy;(2)how to improve the classification accuracy of the model for lowfrequency classes while avoiding the loss of classification accuracy for high-frequency and medium-frequency classes.To solve these key issues of multi label classification for medical data streams mentioned above,this paper conducts the following research:(1)For the problem of mining label relevance between multi-label medical data stream,this paper proposes a Class-Driven Graph Attention network learning framework for Multilabel classification(C-DGAM).Specifically,C-DGAM consists of three parts: a temporal context attention module,a class prompt graph attention module,and a class active dynamic graph attention module.Firstly,the temporal context attention module performs graph embedding learning on the medical data stream sample to obtain vectorized class node,i.e.,class feature vectors.Each class feature vector represents the attributes of a specific class,thereby forming a corresponding relationship with the label.Secondly,the class prompt graph attention module learns the first-order label relevance and updates the class feature vector,where the first-order label relevance describes the relationship between different label pairs.Then,the class active dynamic graph attention module learns the second-order label relevance and updates the class feature vector again for classification output,where the second-order label relevance describes the bidirectional relationship within a pair of labels.By learning the two orders label relevance,the overall classification accuracy of multi-label medical data stream can be improved.(2)For the problem of class imbalance in multi-label medical data stream,this paper proposes a Dynamic Graph Attention Autoencoder based Multi-Tasks learning framework for multi-label classification(DGAAE-MT).It consists of three parts: a temporal class semantic activation module,a dynamic graph attention autoencoder module,and a multi-tasks cooperative training algorithm framework.Specifically,the multi-tasks cooperative training algorithm performs class-balanced sampling and uniform sampling without sample replay on the multi-label dataset to construct two subsets.Then,the two subsets are subject to graph embedding learning using the temporal class semantic activation module to obtain two sets of vectorized class nodes,i.e.,class feature vectors.Next,the two sets of class feature vectors are input into the dual-branch structure,where two dynamic graph attention autoencoders with different initialization learn the two orders label relevance.The multi-tasks cooperative training algorithm improves the classification accuracy of low-frequency categories through the classbalanced branch and compensates for the loss of classification accuracy for high-frequency and medium-frequency categories through the uniform sampling branch.Then,it improves the overall classification accuracy of multi-label medical data stream with class imbalance.